motivation and organizational principles for anatomical

Journal of the American Medical Informatics Association Volume 5 Number 1 Jan / Feb 1998 17

Research Paper n

Motivation and OrganizationalPrinciples for AnatomicalKnowledge Representation:The Digital Anatomist SymbolicKnowledge Base

CORNELIUS ROSSE, MD, DSC, JOSE L. MEJINO, MD, BHARATH R. MODAYUR, PHD,REX JAKOBOVITS, KEVIN P. HINSHAW, JAMES F. BRINKLEY, MD, PHD

A b s t r a c t Objective: Conceptualization of the physical objects and spaces that constitutethe human body at the macroscopic level of organization, specified as a machine-parseableontology that, in its human-readable form, is comprehensible to both expert and novice users ofanatomical information.

Design: Conceived as an anatomical enhancement of the UMLS Semantic Network andMetathesaurus, the anatomical ontology was formulated by specifying defining attributes anddifferentia for classes and subclasses of physical anatomical entities based on their partitive andspatial relationships. The validity of the classification was assessed by instantiating the ontologyfor the thorax. Several transitive relationships were used for symbolically modeling aspects of thephysical organization of the thorax.

Results: By declaring Organ as the macroscopic organizational unit of the body, and defining theentities that constitute organs and higher level entities constituted by organs, all anatomicalentities could be assigned to one of three top level classes (Anatomical structure, Anatomical spatialentity and Body substance). The ontology accommodates both the systemic and regional(topographical) views of anatomy, as well as diverse clinical naming conventions of anatomicalentities.

Conclusions: The ontology formulated for the thorax is extendible to microscopic and cellularlevels, as well as to other body parts, in that its classes subsume essentially all anatomicalentities that constitute the body. Explicit definitions of these entities and their relationshipsprovide the first requirement for standards in anatomical concept representation. Conceived froman anatomical viewpoint, the ontology can be generalized and mapped to other biomedicaldomains and problem solving tasks that require anatomical knowledge.

n JAMIA. 1998;5:17–40.

Affiliations of the authors: The Digital Anatomist Program, De-partments of Biological Structure and Computer Science, Uni-versity of Washington, Seattle, WA.

Grant support by National Library of Medicine contractLM13506 and grants LM04925 and LM05982.

Correspondence and reprints: Cornelius Rosse, Department ofBiological Structure, Box 357420, University of Washington, Se-attle, WA 98195. E-mail: [email protected]

Received for publication: 8/6/97; accepted for publication:9/15/97.

Human anatomy is a fundamental science that un-derlies all fields of medicine. Indeed, it is difficult tomake a statement in any medical context without ex-plicitly or implicitly invoking anatomical concepts.The physical examination, medical imaging and otherprocedures, as well as the basic elements of the med-ical history, all generate clinical data that pertain toanatomical entities in the human body. The interpre-tation of these data relies on an implicit understand-ing of anatomy. The inferences entailed in such rea-soning call upon cognitive or computationalprocessing of abstractions about physical entities of

Dow

nloaded from https://academ

ic.oup.com/jam

ia/article/5/1/17/768860 by guest on 01 September 2022

18 ROSSE ET AL., Digital Anatomist Symbolic Knowledge Base

the body, making use of relationships that existamong anatomical concepts.

Continuing advances in imaging, coupled with thegrowing use of imaging in diverse fields of patientcare, have highlighted anatomy as an information do-main critical for clinical medicine. Although the num-ber of anatomical terms in current use is large, andtheir occurrence in hard copy sources and controlledmedical vocabularies is extensive, there is little con-sistency in their representations. Each of the currentmajor medical vocabulary projects arranges anatomi-cal terms according to its own conceptualization.1 – 4

Such disparity in approach is perhaps not surprising,since the organizational principles which could pro-vide the foundation for a reusable anatomical infor-mation resource have not been articulated.

Since many types of clinical data may be regarded asattributes manifested by various classes of anatomicalentities, we argue that standards in anatomical con-cept representation could facilitate the establishmentof standards in clinical concept representation. Theneed for a formalized, canonical model of anatomyhas, in fact, been advocated as a prerequisite for clin-ical concept representation.5,6 This need can best bemet by a coordinated research effort focused on de-veloping symbolic representations of anatomicalknowledge; an effort that should parallel the intenseinterest and research activity currently directed to-ward visualizing human anatomy.

We have been developing a symbolic knowledge basein anatomy as one of the information resources in theDigital Anatomist distributed framework for anatom-ical information.7 – 9 Our objective is a system in whichconceptualizations of anatomical entities, and the re-lationships among them, are specified as an ontologythat models the physical organization of the body. Weargue that, in order for such a knowledge base to bereusable in different fields of patient care, biomedicaleducation and research, it should be developed froman anatomical viewpoint.

Structure of this Paper

We begin by defining the term anatomy. We next con-sider the types of physical objects and spaces that con-stitute the body, an essential step in formulating asymbolic model of the body’s physical organization.This part of the paper ends with an assessment ofcurrently available representations of anatomical con-cepts. Then, we state the requirements of a knowledgebase of anatomy and propose defining attributes ac-cording to which anatomical entities may be classi-fied. We support the proposed classification by type

definitions and examples. A description of the gran-ularity of concept representation, term assignments,and relationships is followed by the implementationof the ontology. Finally, we consider the emergentproperties and the evaluation of the knowledge base,and conclude the section with the evolutionary en-hancements we currently envision for the knowledgebase. The final section summarizes the features of theDigital Anatomist Knowledge Base that distinguish itfrom other symbolic representations of anatomy, andconsiders its contribution to the setting of standardsin anatomical concept representation. In conclusion,we describe ways in which the availability of theknowledge base may facilitate the development ofproblem-targeted applications that require anatomicalknowledge.

Anatomy and Anatomical Entities

Definitions of Anatomy

A contraction of two Greek roots, ana- (up), and tem-nein (to cut), hence anatemnein (to cut up or dissect),the term anatomy was used historically to denote ei-ther dissection of a cadaver (e.g., performing or at-tending an anatomy, such as that depicted by Rem-brandt in 163210), or a treatise on the same subject(e.g., Vesalius’ Anatomy, 154311). With the advance-ment of methods for analyzing the human body, suchas the introduction of the microscope, the term beganto acquire broader meanings.

For the purposes of developing the knowledge base,we define anatomy in terms of two distinct concepts:

n The ordered aggregate of physical objects andspaces that are assembled according to predeter-mined spatial relationships or patterns that are in-fluenced by the coordinated expression of groupsof genes, and thereby constitute an organism andits physical subdivisions; in other words, the orga-nizational structure that, in the material sense, con-stitutes a living organism and its parts. To distin-guish this concept from that of the seconddefinition, we designate it as anatomy (structure),when so doing serves the purpose of clarity.

n The science concerned with the discovery, analysisand representation of anatomy (structure), and theprocesses responsible for its establishment, alongwith the conceptual entities required for under-standing, explaining and drawing conclusions fromanatomical observations. To distinguish this con-cept from anatomy (structure), we designate it anat-omy (science).

Dow


ic.oup.com/jam



The second definition implies that a primary objectiveof anatomy (science) is the conceptualization of phys-ical entities that constitute anatomy (structure). It is apeculiarity of anatomy (science) that most of the con-cepts it deals with are physical entities (e.g., thorax,heart, gland, lymphocyte), whereas the chief conceptdomain in other biomedical sciences involves non-physical entities (inflammation, nausea, diabetes mel-litus, tachycardia, schizophrenia, pregnancy). None-theless, the underlying physical entities with whichthese other biomedical sciences are concerned are, infact, anatomical entities. Thus, conceptualizations ofanatomy (structure) are critical for modeling biomed-ical concept domains in fields other than anatomy, aswell as in anatomy itself.

Physical Anatomical Entities

The cell theory, formulated in 1838 and 1839 by Schlei-den and Schwann, defined the cell as the fundamentalorganizational unit of plants and animals.12,13 The hu-man body is formed from a single cell, the zygote, inwhich the nuclei of a male and a female gamete fuseand mingle their genetic material. The progeny of thezygote, generated by successive cell divisions, areconstituted of organelles and macromolecules synthe-sized by the parent cells. Cells also produce moleculesthat act as signals for inducing and regulating the as-sembly of newly formed cells into aggregates accord-ing to spatial patterns that are genetically conserved.Governed by these genetic and molecular processes,the cell aggregates become folded and rearranged intomicroscopic and macroscopic physical objects. Spacesare generated within and between them, and becomefilled with body substances. The physical objects,spaces and substances formed by these cellular andmorphogenetic processes range in size from moleculesto major body parts and the entire body (organism)itself; together, they constitute anatomy (structure).

All these anatomical entities manifest both physicaland functional attributes, according to which theymay be characterized and classified. The great diver-sity of anatomical entities that exists in the body ac-counts for its structural complexity. An understandingof this complexity requires conceptualization of a hi-erarchy of physical units or parts, and of the relation-ships that hold among them. Such conceptualizationis the domain of anatomy (science), and more partic-ularly, the purpose of a knowledge base.

Organizational units and relationships at the molec-ular level can be defined with quantitative methodsof physics, and physical chemistry (e.g., x-ray crystal-lography and NMR spectroscopy), as well as with as-says of molecular and cellular functions. In contrast,

the criteria by which multicellular units of anatomicalorganization can be defined and characterized aremore subjective. Consequently, there is a good deal ofambiguity in the current definitions of microscopicand macroscopic anatomical entities and in the namesby which groups or classes of them are identified. Theemergence of computer-based methods for knowl-edge representation and information processing callsfor an evaluation and possible modification of theclassification schemes which, for more than a century,have served to represent and communicate anatomi-cal knowledge.

Owing to the large body and scope of information,the subject domain of the knowledge base must ini-tially be restricted, for practical reasons, to one orga-nizational level of anatomy (structure). Since the an-atomical data that are generated in clinical practicepertain predominantly to anatomical entities visible tothe unaided eye, we have restricted the subject do-main of the knowledge base to the macroscopic anat-omy of the human body. We use the term macroscopicanatomy to include anatomical entities visible with theunaided eye in both the living and the deceased hu-man body; it is distinct from gross anatomy, which im-plies cadaver anatomy alone. The distinction betweenthese terms is relevant to the further subdivision ofmacroscopic anatomy into canonical and instantiatedanatomy, discussed below.

Representations of Macroscopic Anatomy

We have previously drawn distinctions between spa-tial (or graphical) and symbolic representations ofphysical anatomical entities and have proposed thatsuch reusable representations be developed in paral-lel, with methods best suited for each, together withmethods that link the two representations.7,14 This ap-proach has been adopted by others15,16 and is also be-ing pursued by the Visible Human project.17 Both spa-tial and symbolic representations may be concernedwith either canonical or instantiated anatomy.

Canonical and Instantiated Anatomy

Knowledge of macroscopic anatomical entities isbased on the historical exploration of the body and itsparts by qualitative methods (e.g., cadaver and sur-gical dissection, the physical examination, physicaland optical sectioning). These methods, however, canbe applied only to individual instances of anatomicalstructure. The anatomical data that have resulted havenot been systematically recorded; rather, generaliza-tions deduced from such qualitative observationshave been propagated by generations of anatomists.A synthesis of such generalizations, sanctioned im-plicitly by accepted usage, may be regarded as canon-

Dow


ic.oup.com/jam



ical anatomy. Because of the highly conserved natureof morphogenetic processes, the canonical model hassufficed for learning anatomy and for applying thisknowledge by humans to clinical problems. The cur-rent canonical model of anatomy, therefore, is a usefulconceptualization of physical anatomical entities.Such a model is implicit in the graphics and narrativetexts that constitute the recorded legacy of anatomy(science), and it also exists as an abstraction in theminds of humans who understand anatomy.

In contrast to canonical anatomy, instantiated anatomycomprises the anatomical data generated by invasiveand non-invasive methods in clinical practice. Thesedata are systematically documented in the clinical rec-ord, and amount to a new field which deals with themacroscopic anatomy (structure) of individual, living,human subjects. When problem-targeted applicationsneed to rely on anatomical knowledge for sorting dataof instantiated anatomy, and making inferences basedon these data, the ranges of variation in size, shape,spatial relations and other anatomical attributes,which seem trivial in terms of the canonical model,assume practical importance. Such sorting and infer-ence, however, must rely on knowledge of canonicalanatomy and therefore the canonical model mustserve as a foundation for a knowledge base in anat-omy.

It is one of our objectives to formulate a canonicalsymbolic model which also accommodates instan-tiated anatomy.

Symbolic Representations

A large number of anatomical terms are embedded incontrolled medical vocabularies.1 – 3,18 However, the de-velopment of these resources is motivated largely bythe need to computerize the clinical record. Thus theywere not developed from a strictly anatomical view-point, and lack concepts and semantic relationshipsthat are required for modeling the physical organi-zation of the body. A review of some of the availablerepresentations is warranted in order to justify the cre-ation of yet another anatomical knowledge source.

Nomina Anatomica,19 written in Latin and availableonly in hard copy, was published originally in 1895.Now in its sixth edition, it is the only comprehensivecompendium of anatomical terms that has been de-veloped from an anatomical viewpoint and sanc-tioned by the International Nomenclature Committee.Nonetheless, the specificity and granularity of itsterms are inadequate for meeting the needs of schol-arly treatises in anatomy or many fields of clinicalpractice. Terms in Nomina Anatomica (NA) are organ-ized according to the study of so-called systems of the

body (e.g., osteologia, the study of bones; splanch-nologia, the study of internal organs). However, itfails to define the systems adequately and does notattempt a classification of the concepts denoted by itsterms, shortcomings that may account for some in-consistencies evident in its organization. Splanchno-logia, for example, includes the liver but not the heart,which is included in Angiologia, the study of bloodvessels.

To a varying extent, the same shortcomings are to befound in the large body of published anatomical in-formation. As a result, the specificity in availablesources is insufficient for defining classes of anatom-ical entities and their relationships in ways that wouldmeet the requirements for logic-based knowledge rep-resentation. More important, the principles for for-mulating such definitions in narrative text, as well asthe definitions themselves, have only been impliedand not explicitly articulated. A few examples will un-derscore the relevance of this point.

Which class of entity is the heart? The answer is likelyto be: an organ, rather than a blood vessel, as specifiedby NA. If it is Organ, presumably so are the spleenand the uterus; but according to what criteria? Are thesciatic nerve, the abdomen, the biceps and the kneejoint also members of the class Organ? Is the rightatrium an organ? If it is not Organ but a part of anorgan (the heart), is the myocardium (which forms themuscular wall of the heart, as well as that of the rightatrium) also a member of the same class? Should thisclass be designated as Organ part? Should the cavityof the right atrium also be classified as an Organ part?

The anatomical terms in these examples denote con-cepts that manifest sets of morphological and func-tional properties—in other words, anatomical attrib-utes. On the basis of these attributes, anatomistsintuitively conceptualize differences and similaritiesbetween the concepts, even though these attributesare not explicitly defined.

Because of this state of affairs, ad hoc attempts at an-atomical knowledge representation are forced to re-sort to other arbitrary classification schemes (e.g.classes such as simple structure, paired structure, sym-metrical structure,20 conventional cavity, true cavity, actualcavity,21 thoracic structure, cardiovascular structure4), orthey simply omit assigning terms to classes. Segmentsof anatomical information have been incorporated asillustrative examples into knowledge representationmethods (see, for instance, Lucas22), or as indispen-sable components of clinical applications.23 – 26 Of thelarge controlled vocabularies, SNOMED2 includes anextensive and detailed compendium of terms in axes

Dow


ic.oup.com/jam



that relate them to anatomy. However, type defini-tions do not assign these terms to classes, nor are therelationships among these terms specified, and thefixed, code-dependent hierarchy is too limiting to ac-commodate a consistent conceptualization of anat-omy.

The GALEN3 and the Read Codes18 projects adopt amore flexible approach. Each, however, develops anindependent and different representation scheme foranatomical concepts.4,21 Moreover, the anatomicalprinciples on which the schemes are based are notdefined, leading to inconsistencies in the classification.For instance, in the anatomical symbolic knowledgebase of the Read Codes,4 the list of subordinate con-cepts of the Chambers of the heart includes the fibrouspericardium and the pericardial cavity.27 It is hard tosee how a membrane and a space, both of which areoutside the heart itself, can be subordinated to a setof entities that are part of the heart proper (the cardiacchambers).

The other project, GALEN,3,21 regards anatomical en-tities primarily as the location of disease processes;consequently its anatomy module lacks sufficientspecificity and taxonomic structure for a coherentsymbolic representation of anatomy. The Representa-tion and Integration Language of GALEN,3 however,seems well suited for generating expressions aboutanatomical concepts by the compositional approach,as well as associating diverse attributes with theseconcepts.

The UMLS Metathesaurus1,28 provides text definitionsfor a modest number of anatomical concepts andspecifies their semantic type and their relations. Theseclasses, however, are too broad for the purposes ofmodeling a knowledge domain as complex as anat-omy. Nevertheless, the nodes and relationships in theUMLS Semantic Network have served us well as use-ful starting points for structuring anatomical knowl-edge.

As this review indicates, there is currently a substan-tial duplication of effort in the compilation of anatom-ical terms within various ongoing projects concernedwith clinical concept representation. We contend thatthe problem lies in the failure of traditional knowl-edge sources to define concepts of anatomical entitieswith sufficient specificity and consistency. A standard-ized representation of anatomy (structure) in the clin-ical concept representation projects is impeded by theprimary orientation of these projects: they are directedtoward the clinical user rather than toward anatomy(science). Fundamental concepts, according to whichanatomical entities may be organized, cannot beculled from radiological or clinical discharge reports,

or from publications in diverse fields of clinical med-icine. Writers of such communications assume a cer-tain level of anatomical knowledge and understand-ing in the reader, but the basic concepts that supportthat understanding remain unarticulated and aretherefore lacking from the respective controlled vo-cabulary projects.

The first requirement for standardizing anatomicalknowledge representation seems to be an ontologygenerated from an anatomical viewpoint. Domain on-tologies have been widely advocated as the first re-quirement for developing logic-based formalisms thatcan model defined information domains of medi-cine.29 – 34 In addition to comprehensive concept rep-resentation, an anatomical oncology should alsomodel the structure of anatomical knowledge itself. Inorder to assure its soundness and reusability, such aresource must be based on explicit definitions of an-atomical entities and the relationships that holdamong them.

The Evolving Digital Anatomist SymbolicKnowledge Base

Our work on symbolic representation of anatomicalknowledge began as an enhancement of the UMLSSemantic Network and Metathesaurus.1,28 Rather thanadopting one of the more complex knowledge repre-sentation systems and languages, we retained and ex-tended the semantic net of the UMLS and enhancedit with well defined semantics. This strategy allowedus to define the requirements for representing ana-tomical knowledge. Currently, the Digital Anatomistsymbolic knowledge base consists of hierarchies, con-structed on the basis of defined relationships, and def-initions. Our experimental approach to developingthe knowledge base has focused on anatomical enti-ties that comprise the thorax. The foundation of theknowledge base is the -is a- hierarchy, in which over8700 concepts have been assigned to more than 200classes and subclasses; we refer to this hierarchy asthe anatomical ontology.

Requirements of the Anatomical Ontology

At the outset, we imposed a number of requirementsthat the ontology had to satisfy. These requirementshave been met and are listed below.

n The subject of the anatomical ontology had to berestricted to anatomy (science). Anatomy (science)is defined above, and ontology has been defined, inthe broadest sense, as an explicit specification of aconceptualization.31 For the time being, the scope of

Dow


ic.oup.com/jam



the ontology has been further restricted to entitiespertaining to macroscopic anatomy.

n The anatomical ontology had to specify, first of all,conceptualizations of the physical entities that con-stitute the body. This requirement must be met be-fore such concepts as states, processes, and activi-ties can be associated with physical anatomicalentities. For the time being, we have limited the on-tology to the static state of physical anatomical en-tities.

n The ontology had to model canonical anatomy andprovide ways for the knowledge base to accom-modate and organize data pertaining to instantiatedanatomy.

n The ontology had to define the physical units thatconstitute anatomy (structure), in order to modelthe physical organization of the body. Although theontology is currently limited to the macroscopiclevel of organization, physical units that constitutemacroscopic entities had to be defined at the micro-scopic level, as well.

n Generic, partitive, and certain spatial relationshipsthat exist, both horizontally and vertically, amongthe macroscopic and microscopic units of physicalorganization had to be defined.

n Although developed for macroscopic anatomy, theontology had to be scalable and extendible to otherfields of anatomy (embryology and developmentalbiology, microscopic anatomy, cellular and molec-ular biology, and neuroanatomy).

n The representations of anatomical entities had to beparseable by machine and comprehensible, in theirhuman-readable form, to both expert and noviceusers of anatomical information.

n The ontology had to include all macroscopicallydiscernible anatomical entities arranged in a typehierarchy. Therefore, the defining attributes ofclasses or subclasses in this hierarchy had to bestated in terms of anatomical attributes which areinherited by members of a class or subclass; like-wise, the differentia between members of a class orsubclass had to be defined in terms of anatomicalattributes.

n A symbolic model had to rearrange entities of theanatomical ontology according to various defined

relationships that describe the physical organiza-tion of the entire organism (in the current context,the human body).

n To ensure consistency of the classification and theevaluation of the ontology by teachers and studentsof anatomy, the constraints imposed on the mean-ing of terms had to be formulated as definitions inhuman-readable text, and subsequently transcribedin logic-based notation.

An ontology that meets these requirements is distinctfrom knowledge itself and also from its existing rep-resentations in narrative text, in that it puts explicitconstraints on the terms and the structure of theknowledge.32 These constraints emerge largely fromthe conceptualization of the units of anatomical or-ganization and their relationships.

Defining Anatomical Attributes

In medical dictionaries35,36 and textbooks of anatomy(see, for example, Hollinshead’s37 and Gray’s38) defini-tions of anatomical entities tend to be formulated interms of functional attributes.* These attributes, how-ever, fail to specify the constraints by which anatom-ical entities may be grouped together or distinguishedfrom one another. In order to model the physical or-ganization of the body, defining attributes of anatom-ical entities need to be stated in terms of their con-stituent parts, and in terms of the entities which they,in turn, constitute. A physical unit of anatomical or-ganization must, therefore, be defined in terms of theparts of which it is composed and the higher orderunits of which it, in turn, is a part.

We have, therefore, formulated definitions of physicalanatomical entities by specifying constraints in termsof the units that make up the body, and the relation-ships that hold among these units. In doing so, wehave found it necessary to restrict and specify themeaning of several terms that denote partitive rela-tionships (subdivision, part, component). Differentiaare specified in the definitions primarily by the rela-tionships -consists of-, and its inverse, -constitutes-. Weuse these relationships with a minor modification oftheir UMLS definitions.†

*E.g., lung, an organ of respiration that effects the aeration ofblood; heart, a viscus of cardiac muscle that maintains the cir-culation of blood.35

†Consists of: structurally made up, in whole or in part, of somephysical units, material or matter. This includes composed of,made of, formed of (UMLS definition1 modified by adding‘‘physical units’’; RIN: constitutes).

Dow


ic.oup.com/jam



F i g u r e 1 The classification of physicalanatomical entities based on the -is a-relationship. The first-generation off-spring of Anatomical structure and Organpart are also shown.

Classification and Definitions

Since the objective of the anatomical ontology is tomodel the physical organization of the body, we de-clared Physical anatomical entity as the top level con-cept of the ontology. We classify all physical anatom-ical entities into one of three classes: Anatomicalstructure, Anatomical spatial entity and Body substance(Fig. 1). These classes are also referred to as root con-cepts because we assign, based on the -is a- relation-ship, all physical entities in the body, as concepts, tothese three classes. These assignments result in threeparallel type hierarchies.

In each type hierarchy, several generations of sub-classes link the root concept (class) to leaf concepts orinstances. We distinguish between leaf concepts of ca-nonical anatomy and instantiated anatomy (definedabove). When desirable, a canonical leaf concept mayitself serve as a parent for leaf concepts of instantiatedanatomy: the canonical instance becomes a subclass.For example, ‘‘Left tibia’’ is a canonical instance andincludes among its parents such subclasses as Tibia,Long bone, Bone(organ) and Organ, as well as the classAnatomical structure. The ‘‘Left tibia’’ of an individualsubject (e.g., that of John Doe), however, is a leaf con-cept of instantiated anatomy; when designated withan appropriate extension (e.g., ‘‘Left tibia[JD]’’) it be-comes a child of the subclass Left tibia. Classes andsubclasses in the anatomical ontology correspond tosemantic types of the UMLS Semantic Network,which are higher level nodes in the net.

The remainder of this section provides the definitionsof the three classes (root concepts) of the anatomical

ontology and those of some of their offspring. Foreach class and subclass, the definition states the genusand the defining anatomical attributes that distin-guish its members (the entities included within theclass or subclass) from those of the parent class orsubclass. We have evaluated and revised the sub-classes and their definitions during the process ofmaking concept assignments in the ontology, pursu-ing in parallel top-down and bottom-up approaches.

Although our approach so far has been largely limitedto the thorax, this body part actually includes almost allsubclasses of anatomical entities that are found in thebody. Thus the three root concepts and their immediateoffspring, illustrated in Figure 1, subsume virtually allmacroscopically visible entities in the body. It should,therefore, be possible to assign to these classes and sub-classes all physical entities in the living or deceasedbody, regardless of whether or not they have beennamed previously. The same holds true for representa-tions of macroscopic anatomical entities in medical im-ages (e.g., groups of pixels and voxels).

Although concentrating predominantly on the thorax,we have found it necessary to include examples fromother body parts to clarify and validate the classifi-cation and definitions. Some of these examples arecited in the following sections.

Classes of Physical Anatomical Entity

The definition of Physical anatomical entity is impliedby that of anatomy (structure) above. Of the three rootconcepts or classes of the ontology (Fig. 1), Anatomicalstructure plays a dominant role; the other two, Ana-

Dow


ic.oup.com/jam



tomical spatial entity and Body substance, are defined inrelation to Anatomical structure.‡

Anatomical structure is a physical anatomical en-tity and a physical object which is generated by pro-cesses that are affected by the coordinated expressionof groups of genes; it consists of parts that are them-selves anatomical structures. The parts are spatiallyrelated to one another according to patterns whichare also affected by the coordinated expression ofgroups of genes.

The largest anatomical structure is the whole organ-ism; the smallest (for the current purpose of devel-oping the knowledge base), is a cell, which is thefundamental organizational unit of plants and ani-mals. (Examples: heart, right ventricle, mitral valve,myocardium, endothelium, lymphocyte, fibroblast,thorax, cardiovascular system.)

Anatomical spatial entity is a physical anatomicalentity and a spatial entity of three or fewer dimen-sions, which is associated with the exterior or interiorof anatomical structures. (Examples: thoracic cavity,pericardial cavity, epigastrium, femoral triangle, di-aphragmatic surface of heart, transpyloric plane,midclavicular line, midinguinal point.)

Body substance is a physical anatomical entity anda substance§ in gaseous, liquid, semisolid or solidstate, with or without the admixture of cells, whichis produced by anatomical structures or derived frominhaled and ingested substances that become modi-fied by anatomical structures as they pass into orthrough the body. (Examples: intercellular matrix,saliva, semen, cerebrospinal fluid, inhaled air, urine,feces, blood, lymph.)

Arbitrarily limiting the definition of Anatomical struc-ture to objects that are constituted by cells is justifiablein the present circumstances because the immediatepurpose of the knowledge base is to model macro-scopic anatomy. It will be necessary, however, to ex-tend the symbolic model, with the collaboration of ap-propriate domain experts, to cell components(organelles such as mitochondrion, ribosome), largeand small molecules, (e.g., myoglobin, T-cell receptor,cyclic AMP), and embryonic structures (developmen-tal stages of fully-formed structures as well as tran-sient structures that are transformed or eliminated by

‡In the definitions, concepts that are classes and subclasses inthe ontology are shown in bold face, relationships defined byUMLS as RL or RIN (or their synonyms included in the textualUMLS definitions)1 are shown in italics, and relationships thatare children of defined UMLS relationships (designated asUWDA relationships) are in underlined italics.

§UMLS definition of substance: A material with definite or fairlydefinite chemical composition.1

the morphogenetic process), since these entities alsosatisfy the definition of Anatomical structure.

The definition of Anatomical structure that we proposeis more restrictive than that of UMLS, whose defini-tion of the same term includes ‘‘pathological part ofanatomy’’ (e.g. tumors, abscess).1 Since such patholog-ical entities are not generated by the processes re-sponsible for establishing normal anatomy, we haveexcluded them from the anatomical ontology. Themost specific semantic type for macroscopic anatomyin UMLS is Body Part, Organ or Organ Component, de-fined as a collection of cells and tissues which are lo-calized to a specific area or combine and carry out oneor more specialized functions of an organism.1 (Ex-amples: thorax, vagus nerve, heart, left ventricle,myocardium, anterior leaflet of mitral valve.) This se-mantic type is subsumed by Anatomical structure in theanatomical ontology that we are developing. Most ofour work has entailed giving depth and specificity tomembers of the class Body Part, Organ, Organ Compo-nent of the UMLS semantic network.

Subclasses of Anatomical Structure

Whereas the cell is usually considered the fundamen-tal organizational unit of the body, assigning this roleto the concept Cell does not serve a useful purpose forformulating abstractions of macroscopic anatomy.Rather, we have hypothesized that of the subclassesof Anatomical structure shown in Figure 1, Organ beregarded as the basic organizational unit of macro-scopic anatomy. We have proposed definitions ofother subclasses of Anatomical structure in terms of therelationships they hold to Organ. We have tested thishypothesis by populating the ontology with anatom-ical structures located in the thorax. Although makingthese assignments called for a cyclic revision of theproposed definitions and the nodes of the ontology,the validity of Organ as the basic organizational unitof macroscopic anatomy was, in the final analysis, up-held.

We have assigned macroscopic and microscopic ana-tomical structures that constitute organs to the sub-class Organ part, and we have defined Body part andOrgan system as those subclasses of Anatomical struc-ture that are constituted by organs. We have formu-lated the differentia that distinguish all these conceptsfrom one another in terms of the anatomical structuresthat constitute each of them and the structures thatthey, in turn, constitute. All these subclasses of Ana-tomical structure also inherit the second defining at-tribute of their parent; namely, the parts that consti-tute members of each of the subclasses manifestspatial patterns of organization that are characteristicof each.

Dow


ic.oup.com/jam



The following paragraphs state the constraints for Or-gan and Organ part in narrative text, supported by rel-evant examples; the subsequent section defines Bodypart and Organ system.

Organ and Its Parts. These terms are defined as fol-lows:

Organ is an anatomical structure that consists of themaximal set of organ parts so connected to one an-other that together they constitute a self-containedunit of macroscopic anatomy, distinct both morpho-logically and functionally from other such units. To-gether with other organs, an organ constitutes an or-gan system or a body part. An organ is divisible intoorgan parts but not organs. (Examples: femur, bi-ceps, liver, heart, aorta, sciatic nerve, ovary.)

In the current iteration of the anatomical ontology,three members of the subclass Organ are canonical in-stances (brain, spinal cord, skin) and the others aresubclasses (Fig. 2). The latter subsume one or moregenerations of additional subclasses; only those of Vis-cus are displayed in Figure 2. There are no canonical

instances of Organ part; its subclasses include Tissue,Organ component, and Organ subdivision (Fig. 1).

Organ part is an anatomical structure that consistsof one or more types of cells and intercellular matrix(which is a body substance); organ parts are con-nected to one another to constitute anatomical struc-tures of increasing size and structural complexity,which together account for the emergent morpholog-ical and functional attributes of an organ. An organpart is divisible into other organ parts, the smallestof which is a tissue. (Examples: endothelium, os-teon, cortical bone, neck of femur, bronchopulmon-ary segment, middle lobe of right lung).

Tissue is an organ part that consists of similarly spe-cialized cells and intercellular matrix, aggregatedaccording to specific spatial relationships; togetherwith other tissues, it constitutes an organ compo-nent. [Examples: epithelium, muscle (tissue), con-nective tissue, neural tissue, lymphoid tissue.]

In addition to tissues, similarly specialized cells mayalso aggregate in body substances (e.g., cell aggre-gates and sediments in blood and urine). Spatial re-lationships in these aggregates, however, are not spec-ified; therefore, these aggregates do not satisfy thedefinition of Tissue.

Tissues do not exist as discrete anatomical structuresin the body. Accumulation of a particular tissue leadsto spatial associations with one or more additional tis-sue types, resulting in the formation of anatomicalstructures of a higher order, which we have desig-nated as Organ component.

Organ component is an organ part that consists of aprincipal and one or more subsidiary tissues; con-nected to other organ components, it constitutes anorgan subdivision. (Examples: osteon, acinus, sub-mucosa, capillary, papillary muscle, anterior leafletof mitral valve, capsule of kidney, cortical bone,muscle fasciculus, anterior rootlet of spinal nerve.)

bF i g u r e 2 The subclasses of Organ and its subclasses,displayed in a screen capture from the hierarchy editorof Knowledge Base Manager, the authoring program forthe symbolic knowledge base. The symbol >> indicatesthat the node in the hierarchy (formed in this instanceby the -is a- relationship) has at least one generation ofchildren that is not shown. Double clicking on a termdisplays its immediate offspring by one tab indentation.Some of the terms denoting organ subclasses are ex-tended by the qualifier ‘‘(organ)’’ because these terms arealso used to designate members of other subclasses ofAnatomical structure, such as Bone (tissue), Muscle (tis-sue). The tree is opened up for Viscus, the first Organsubclass, in which examples of canonical instances aredisplayed to illustrate the classification.

Dow


ic.oup.com/jam



F i g u r e 3 Seman-tic network con-structed with -is a-(solid lines) and-part of- (interruptedlines) relationshipsto model aspects ofknowledge pertain-ing to the rightatrium. Terms de-noting parts of theheart are shown inthe plane of theshaded quadrangle;the subclasses towhich these entitiesare assigned are dis-played above thisplane.

The associations of a principal tissue with subsidiarytissues delineate organ components which manifestgreat morphological diversity in terms of size, shapeand patterns of spatial organization. One or moretypes of organ components connected together forman organ part of a higher order, which we have des-ignated as Organ subdivision.

Organ subdivision is an organ part that consists oftwo or more organ components; connected to otherorgan subdivisions, it constitutes an organ. An organsubdivision is demarcated from other subdivisionsof the same organ by one or more anatomical fea-tures (defined below under Subclasses of Anatomi-cal Spatial Entity). (Examples: right atrium, mitralvalve, left lobe of liver, neck of femur, short head ofbiceps, arch of aorta).

Spatial relationships among organ components andorgan subdivisions account for the shape, internal ar-chitecture and some of the physiological properties ofan organ, as a whole. Different subclasses of Organhave different kinds of subdivisions (e.g., shaft, lobe,chamber), and different kinds of organ components.

The definitions of Organ and Organ part resolve theconundrum cited above, under Symbolic Representa-tions, regarding the classification of the heart and itsparts. Figure 3 illustrates these definitions and pro-vides examples for the relationships that exist be-tween an organ and its various organ parts.

The heart is a Hollow viscus,\ which is a Viscus, whichin turn is an Organ (Fig. 2). The right atrium is not anorgan, because it is not a maximal set of organ partsthat constitute a self-contained, distinct unit of mac-roscopic anatomy (constraints that are satisfied by theheart itself). For instance, the myocardium (or heartmuscle, forming the middle layer of the trilaminarwall of the heart), is a constituent not only of the rightatrium but of other parts of the heart as well. Theright atrium is an instance of Cardiac chamber, a sub-class of Organ subdivision. On the exterior of the heart,the right atrium is demarcated from other cardiacchambers by the coronary and interatrial sulci, whichare anatomical features. Moreover, the right atrium isnot an Organ component, because it can be further sub-divided into anatomical structures that are more com-plex than a tissue. As noted earlier, one such Organcomponent is the myocardium, which consists of

\Hollow viscus is a viscus, some organ parts of which constitutea wall; the viscus wall surrounds the organ (viscus) cavity. (Ex-amples: esophagus, heart, stomach, uterus.) Viscus is an organthat consists of organ parts that are embryologically derived fromendoderm, splanchnic mesoderm or intermediate mesoderm; itis located within the body cavity (or in an extension of the bodycavity, the scrotum); together with other organs, the viscus con-stitutes the respiratory, gastrointestinal, urinary, reproductive orimmune systems, or is the central organ of the cardiovascularsystem. (Examples: heart, lung, esophagus, kidney, ovary,spleen.)

Dow


ic.oup.com/jam



myocardial fiber bundles or fasciculi of varying size.These fasciculi, in turn, consist of cardiac muscle (theprincipal tissue), and connective tissue. The latterserves as a subsidiary tissue, which ensheathes aggre-gations of cardiac muscle and defines the myocardialfasciculi and their spatial pattern of disposition. Thusmyocardial fasciculi of different sizes and heart mus-cle itself are also organ components.

The spatial organization of the fasciculi—manifest inpart as the muscular ridges on the interior surface ofthe wall of the heart—is characteristic of a given car-diac chamber (Organ subdivision). For instance, as aconsequence of abnormal developmental processes,the right ventricle may be located on the left ratherthan the right side of the heart. In such a case, thischamber can still be identified as the right ventricle,based on its myocardial fiber architecture, and is in-deed so designated (‘‘anatomical right ventricle’’ lo-cated on the left).

The cavity of the heart, and that of the right atrium,do not satisfy the constraints of Organ subdivision,even though they are parts of the heart and may berelated to it by the -part of- relationship. The cavity ofthe heart is classified as an Organ cavity, and the cavityof the right atrium as an Organ cavity subdivision, bothof which are subclasses of Anatomical spatial entity(Fig. 3), discussed below.

Similar introspective analysis has led us to establishsubclasses of Organ, shown in Figure 2, and to assignother anatomical structures as subclasses of Organpart. As a result we have classified as Organ severalanatomical structures that are not usually thought ofas such, whereas other structures conventionally re-garded as organs are classified as Organ subdivision.

For instance, we classify skin as an organ rather thanan organ system, because no part of the skin is a self-contained anatomical unit, its subdivisions (skin ofthe chest, skin of the palm) are demarcated by ana-tomical features, and its other constituent parts bestmeet the definition of Organ component. Similar rea-soning justifies the classification of Fascia as a subclassof Organ. On the other hand, the rectum, colon andcecum are each classified as Organ subdivision, ratherthan Organ. In no case do the organ parts of thesestructures constitute a self-contained, distinct unit ofmacroscopic anatomy; the structures are demarcatedfrom one another by anatomical features on the largeintestine, and the spatial organization of their organcomponents manifests differences that are character-istic of each of these anatomical structures. It is thelarge intestine that satisfies the constraints for the def-

inition of Organ. A similar argument may be made forclassifying the trachea and bronchi as Organ subdivi-sion, and only the full tracheobronchial tree as Organ.

The classification of Joint has presented a particulardifficulty.} Its constituent parts include subdivisionsof two or more bones, each of which is a distinct or-gan. However, all constituent parts of a joint satisfythe constraints for Organ subdivision and Organ com-ponent; therefore, a joint is classified as Organ.

In addition to their defining attributes, organs and or-gan parts exhibit a number of other attributes thatcharacterize them and assist in making class assign-ments and distinctions among the members of thesetwo subclasses. Embryological derivation, location,anatomical feature, and physiologic function are suchattributes. Organs are located in body parts, and maybe represented as -part of- a body part, whereas organparts are located only in organs, and may be repre-sented as -part ofan organ. Organs are not locatedwithin other organs and do not form parts of otherorgans. Exceptions to this assertion are bone (organ)which, as already noted, contributes to the formationof joints, and blood vessels, lymphatic vessels andnerves (each a subclass of Organ), which arborizewithin other organ types. Another exception is a sub-division of a serous sac (the visceral layer of the se-rous membrane that forms the sac), which may be soadherent to the adjacent viscus that it is regarded aspart of the viscus wall, as well as a subdivision of thesac itself (e.g., epicardium is the visceral layer of theserous pericardium; serosa of the small intestine is thevisceral layer of the peritoneum). In each case, it is asubdivision of these organs (bone, blood vessel, nerve,serous sac) that is incorporated in another organ.

The differentia for distinguishing between various or-gan subclasses can be formulated by filling in valuesfor different organ parts and their spatial relations.The result will be the assignment of some organs tomore than one subclass. For instance, the ovary is clas-sified as both a parenchymatous viscus and a genitalorgan. Likewise, the thymus is classified as a lym-phoid organ and a parenchymatous viscus. Similar ex-amples can be cited for subclasses of nerve. Such mul-tiple assignments establish the basis for multipleinheritance.

}Joint is an organ that constitutes an anatomical junction; itconsists of two or more adjacent bones, parts of which are in-terconnected by organ parts that consist of various types of con-nective tissue. Together with other joints and bones, a joint con-stitutes the skeletal system. (Examples: pubic symphysis, kneejoint, temporomandibular joint.)

Dow


ic.oup.com/jam



Body Part and Organ System. The two higher level or-ganizational units of the body are constituted of or-gans, as specified by their definitions.

Body part is an anatomical structure that consists ofmembers of diverse subclasses of organ, one ofwhich is a set of bones, and another is skin, a sub-division of which completely or partially surroundsthe body part; together with all other body parts, abody part constitutes the body. (Examples: head,trunk, thorax, upper limb, forearm, finger, bodywall.)

Organ system is an anatomical structure that con-sists of all members of one or more organ subclass;these members are interconnected by anatomicalstructures or body substances. (Examples: skeletalsystem, cardiovascular system, gastrointestinal sys-tem, immune system.)

Both terms, Body part and Organ system, are used byanatomical sources and controlled vocabularies, andboth carry various connotations. Nomina Anatomica,together with most other anatomical and clinicalsources, uses Body part synonymously with Body re-gion or region. For instance, the thorax is regarded asboth a body part and a region, as are the upper limband the hand. As already cited above, UMLS providesone inclusive definition for a semantic type desig-nated ‘‘Body Part, Organ or Organ Component.’’ An-atomical sources do not define organ system explicitlybut use the term in the same sense as NA. (See Sym-bolic Representations, above.) UMLS defines BodySystem as: ‘‘A complex of anatomical structures thatperforms a common function.’’1 Both Organ system(e.g., cardiovascular system) and Body part (e.g., upperlimb, or hand) satisfy the constraints of this definition.The definitions we propose constrain the meanings ofthe terms Organ system and Body part. Furthermore,both terms are distinguished from Body region, whichis dealt with in the next section.

That defining attributes of Body part are inherited bymembers of the subclass can be appreciated if oneconsiders that each is supported by a specific skeletalframe, each contains various types of organs (e.g.,muscles, nerves, blood vessels, viscera), and most ofthe surfaces of each are covered by skin. Differentiaof the members of the subclass will be given by fillingin the values for the defining attributes, particularlythose for the set of bones.

The differentia for members of the subclass Organ sys-tem can be specified in terms of the organ subclasses

that constitute the system, and in terms of the partic-ular anatomical structures or substances that intercon-nect them. For instance, the skeletal system consistsof all members of the Bone (organ) subclass, which areinterconnected by all members of the Joint subclass.In this instance, the interconnecting entity is an inte-gral component of this particular organ system itself.This is the case in the majority of organ systems (e.g.respiratory, cardiovascular, gastrointestinal). In the he-matopoietic and immune systems, however, the con-stituent organs are spatially separated and the inter-connection between them is provided by blood andlymph, body substances contained in the cardiovas-cular and lymphatic systems. In all cases, the inter-connections are as important for the manifestation ofphysiologic function as are the organs that constitutethe system.

A number of body systems do not satisfy the defini-tion of Organ system that we propose. For instance, weclassify the conduction system of the heart as an Or-gan component (see Fig. 5), and the portal system as anOrgan system subdivision. The defining attribute of Or-gan system subdivision is that it consists of a specificset, rather than all members, of an organ subclass. Forinstance, the rib cage is an organ system subdivision,because its constituent organs are a set of bones con-sisting of thoracic vertebrae, ribs and the sternum,rather than all bones. The portal system includes thatset of veins that are spatially associated with the gas-trointestinal system; it is a subdivision of the cardio-vascular system. The definition also specifies that thecentral, peripheral, and autonomic nervous systemseach be classified as Organ system subdivision, ratherthan as Organ system.

The classification of anatomical structures that wepropose deviates in some respects from generally heldviews. Traditionally, anatomy (structure) has been de-scribed along one or the other of two axes. On the onehand, systemic anatomy (science) organizes anatomi-cal structures into systems on the basis of the physi-ologic functions they share, and is sometimes calledfunctional anatomy; on the other, regional or topo-graphical anatomy (science) organizes anatomicalstructures on the basis of their location in ‘‘regions,’’that is, body parts, and is sometimes also called mor-phological anatomy. Some textbooks and treatises de-scribe anatomy (structure) along the systemic axis38,39;others along the regional axis.37,40 Such different viewsor axes have dominated the conceptualization of sev-eral symbolic models of anatomy (e.g., Read Codes,4

Voxelman16).

The structural ontology we propose, and the defini-tions that support it unify these two axes of anatomy

Dow


ic.oup.com/jam



F i g u r e 4 Members of the class Anatomical spatial entity,displayed by KB Manager. The first generation sub-classes of Anatomical spatial entity are comprehensive;trees are partially opened up for Body space and Anatom-ical junction to show their subclasses and some canonicalinstances of Body space.

(science). The unification results from regarding Organas the basic organizational unit of macroscopic anat-omy, and from specifying relationships among mem-bers of the same or different organ subclasses in waysthat constitute the higher order anatomical structuresof Organ system and Body part. This conceptualizationallows an association of attributes, such as physiologicand pathologic function, not only with Organ system,but also with Body part, Organ, and Organ part as well.

Subclasses of Anatomical Spatial Entity

Anatomical spatial entity is defined above underClasses of Physical Anatomical Entity. Many anatom-ical terms relate to spatial concepts about the body,but neither specific definitions nor an ontology ofthese concepts have been proposed. Some symbolicmodels do not distinguish them from structures.UMLS defines two semantic types for spatial entities:Body Space or Junction, ‘‘an area enclosed or sur-rounded by body parts or organs or the place wheretwo anatomical structures meet or connect’’; and BodyLocation or Region, ‘‘an area, subdivision or region ofthe body demarcated for the purpose of topographicaldescription.’’1 These definitions include such diverseentities as the midinguinal point, midclavicular line,sagittal plane, epigastrium, diaphragmatic surface ofheart, pleural cavity, sinus venarum of right atrium,orifice of right coronary artery, and mediastinum. Theexamples illustrate the heterogeneity of one-, two-, orthree-dimensional spatial concepts. Our purpose is tosort these entities into classes according to the attrib-utes they manifest in relation to anatomical structures.

The subclasses of Anatomical spatial entity are shownin Figure 4. The first generation of subclasses includeBody space, Body region, Anatomical landmark, Anatomi-cal junction, and Anatomical feature. All have second orthird generation subclasses as parents of canonical in-stances. Here we provide definitions for those sub-classes of Anatomical spatial entity that require expla-nation.

Body Space. This term is defined as follows:

Body space is a three-dimensional anatomical spa-tial entity, which is generated by morphogenetic orother physiologic processes that generate anatomicalstructures; it is enclosed by anatomical structures andcontains one or more anatomical structures or bodysubstances. (Examples: celom, thoracic cavity, lessersac of peritoneum, cavity of right atrium, lumen ofblood vessel, mediastinum, anterior compartment offorearm, intervertebral foramen.)

The constraints of the definition exclude spaces gen-erated by pathological processes, such as the cavities

of abscesses and cysts, but include pathological en-largement of anatomical body spaces. Members of thesubclass are shown in Figure 4. The following exam-ples illustrate the differentia on the basis of whichmembers of the subclass may be distinguished interms of embryological derivation, location, bounda-ries and contents.

Body cavity is a body space that is embryologicallyderived from the intraembryonic celom, is located inthe trunk, is enclosed by the body wall, and containsserous sacs, viscera and other organs. There is onlyone body cavity; it is a canonical instance of bodyspace.

Body cavity subdivision is a body space that is apart of the body cavity; it is enclosed by a body wallsubdivision and is demarcated from another bodycavity subdivision by an anatomical structure or a

Dow


ic.oup.com/jam



conduit#; it contains one or more serous sacs, vis-cera and other organs; together with the other bodycavity subdivisions, it constitutes the body cavity.Canonical instances: thoracic cavity, abdominopel-vic cavity, abdominal cavity, pelvic cavity.

The specificity of these definitions may be judged bycomparing them with those of sibling subclasses ofBody space (Fig. 4). Anatomically and clinically, a dis-tinction needs to be made between a Body cavity sub-division, such as the thoracic cavity, and the cavity ofthe serous sac that lines that body cavity subdivision:in this case the pleural cavity. The pleural cavity is asubclass of Organ cavity, since it is the cavity of a se-rous sac, which is an organ (Fig. 2). The definitionsillustrate the use of differentia for constraining themeaning of anatomical spatial entities in relation toanatomical structures.

Organ cavity is a body space that is enclosed by or-gan parts and contains one or more body sub-stances; in the case of bone (organ), it contains bonemarrow. (Examples: lumen of blood vessel or tra-cheobronchial tree, pericardial cavity, cavity ofstomach or heart, uterine cavity, medullary cavity offemur.)

Serous cavity is an organ cavity that is enclosed by aserous membrane and contains serous fluid. (Ex-amples: pleural cavity, peritoneal cavity, subdeltoidbursa, synovial cavity of hip joint.)

Serous membrane is an Organ part of two organ types:Serous sac (e.g., pleural sac, peritoneal sac, bursa) andof synovial joint, a subclass of Joint. Therefore onlyserous sacs and synovial joints have serous cavities.Accordingly, a joint cavity is a serous cavity, which isdistinct from joint space, a term used in radiology. Thejoint space is filled by articular cartilage (an organcomponent, which is translucent to x-rays), ratherthan by serous fluid. Therefore, joint space needs to berepresented as a radiological attribute of both articu-lar cartilage and synovial joint.

Body Region and Anatomical Landmark. In contrast tobody spaces, body regions and landmarks are ratherarbitrarily defined spatial concepts. Their extensiveuse in both anatomical and clinical descriptions, how-ever, requires that the terms denoting these conceptsbe specified. Body region, in particular, calls for clari-

#Conduit is a body space that connects two or more body spaceswith one another and is surrounded by subdivisions of two ormore organs; it contains subdivisions of organs, or one or morebody substances. (Examples: intervertebral foramen [canal],pharyngotympanic tube, inguinal canal, carpal canal [tunnel],superior thoracic aperture, pelvic inlet.)

fication because, as noted above, the term has beenused synonymously with Body part.

Body region is a two-dimensional anatomical spa-tial entity, that is demarcated by anatomical featuresor anatomical landmarks on the external or internalsurfaces of anatomical structures. It serves the pur-pose of topographical description, and contains an-atomical features and the surface projections of an-atomical structures and spatial entities that arelocated subjacent to the area. [Examples: epigas-trium, precordium, palm of hand (region), axilla (re-gion), triangle of Koch, right iliac fossa.]

The examples will suggest that the names of membersof this subclass are frequently used to imply 3-Drather than 2-D spatial entities. For instance, ‘‘theepigastrium contains the stomach and the liver,’’ or‘‘the palm of the hand contains the lumbrical mus-cles.’’ The first example is an incorrect assertion (theepigastrium contains the surface projections of visceralocated within the abdominal cavity); the second ex-ample implies a compartment rather than a region.Therefore, the term ‘‘palm of hand’’ has two mean-ings, a region and a compartment, both of which canbe specified by the values of the defining attributes ofthe two subclasses of Anatomical spatial entity. Whennecessary, an extension associated with the termshould indicate the relevant concept (palm of hand[compartment], palm of hand [region]).

Anatomical landmarks include visible and palpableanatomical entities as well as arbitrary lines, planesand points. For example, the umbilicus, nipple andcardiac impulse are visible landmarks, and the sternalangle and apex beat are palpable landmarks. A num-ber of arbitrary planes, lines and points, defined inrelation to a variety of anatomical structures, havebeen sanctioned by long term usage (e.g., coronalplane, subcostal plane, vertebral level, midaxillaryline, Nelaton’s line, McBurney’s point).

Anatomical Junction. An anatomical junction impliesphysical continuity rather than adjacency, exemplifiedby the inosculation of blood vessels and hollow vis-cera with one another, or the intermingling of organcomponents of muscles, aponeuroses, nerves andnerve fiber tracts in such junctions as raphes, plexusesand decussations.

Anatomical junction is a two-, or three-dimensionalanatomical spatial entity where two or more ana-tomical structures or body spaces meet and establishphysical continuity with one another or with the ex-terior, or intermingle their organ components. (Ex-amples: brachial plexus, optic chiasm, anococcygealraphe, linea alba, orifice [ostium] of left coronary

Dow


ic.oup.com/jam



artery, anus, gastroesophageal junction, pylorus,knee joint.)

The differentia for the subclasses shown in Figure 4can be stated as those anatomical entities that inter-mingle or establish continuity at the various junctions.The definition is best satisfied by the junction of bodyspaces (Orifice and Anastomosis), and the branchingpoints of nerves, blood vessels and ducts. However,it is not entirely satisfactory to limit Anatomical junc-tion to body spaces, because distinct anatomical struc-tures (objects) are also formed by the junction of an-atomical structures: plexuses of nerves and vessels, aswell as raphes and decussations, are objects, whichconsist of the commingling of anatomical structures.For this reason, in addition to assigning such anatom-ical entities to the spatial concept Anatomical junction,we have also assigned them to the subclasses of An-atomical structure according to the defining attributesthey satisfy. For instance, Nerve plexus and Vascularplexus satisfy the definitions of both Organ system sub-division and Anatomical junction. As discussed above,Joint satisfies constraints in the definitions of Organand Anatomical junction.

Anatomical Feature. Many named anatomical entitiesare rather difficult to classify either as Anatomicalstructure or Anatomical spatial entity, and these are bestdesignated as Anatomical feature. These concepts arewidely used for describing anatomical structures. Be-cause the anatomical attributes that are shared bythese concepts can be stated most satisfactorily interms of spatial concepts, we classify anatomical fea-tures as a subclass of Anatomical spatial entity.

Anatomical feature is an anatomical spatial entitywhich is a modulation of the external or internal sur-face, or of the internal organizational pattern, ofbody parts, organs and organ parts. (Examples:facet, surface, margin, border, apex, pole, tubercle,spine, gyrus, sulcus, metameric segmentation, mul-tipennate fascicular architecture, acinar architec-ture.)

Most anatomical features are related to body parts,organs and organ parts by the -part of- relationship(e.g., the apex of the heart is a part of the heart). Theyserve descriptive purposes and as such, do not exhibitphysiologic functions.

We have subdivided this subclass into External feature,Internal feature and Organizational pattern, in order toenhance the specificity of representations of anatomy(structure). Descendants of External feature are suchmodulations of an external surface as border or mar-gin, apex or pole, base, prominence (e.g., spine, crest,

tubercle, gyrus) and depression (e.g., fossa, fissure,sulcus or groove). Internal feature includes similarmodulations of the internal surface. Distinctions be-tween internal and external features are required if theanatomical ontology is to model the physical organi-zation of the body. For instance, the knowledge cap-tured by assigning the fossa ovalis as an internal fea-ture to the right atrium (Fig. 3), and the coronary sulcusas an external feature to the heart, is more useful, bothclinically and anatomically, then if both concepts wereentered in the ontology as canonical instances of An-atomical feature of the heart.

Spatial concepts such as lobulation, segmentation, me-tamerism, trabeculation, acinar architecture and fas-cicular architecture capture the spatial relationshipsaccording to which anatomical structures are organ-ized into higher level units. We include these conceptsin the subclass Organizational pattern.

Body Substance

Body substance is defined above under Classes ofPhysical Anatomical Entity. The definition is based onthat of the same semantic type in the UMLS SemanticNetwork.1 The Digital Anatomist definition specifiesbody substances in relation to anatomical structures,making it possible to state the differentia for membersof the subclass in terms of the anatomical structuresthat produce or contain them. Most of the examplescited to illustrate the definition satisfy the constraintsof the definition without ambiguity. Our assignmentof blood and lymph to Body substance, however, callsfor justification.

Blood and lymph have traditionally been regarded byanatomical sources as tissues. Other fluids of the body,however, in which cells are suspended (e.g., cerebro-spinal fluid, semen, synovial fluid) have not been clas-sified as tissues. In the anatomical ontology, we assignblood, lymph and all other body fluids in which cellsare suspended to Body substance, because all satisfy theconstraints of the definition of Body substance. We donot assign these body fluids to Tissue, because noneof them satisfies one of the defining attributes we pro-posed for Tissue; namely, specific spatial organizationof its constituent parts.

Concept Representation

We have entered in the anatomical ontology, asunique concepts, all physical entities that are macro-scopically discernible in the thorax. A granularity ofgreater resolution calls for microscopic methods to an-alyze anatomy (structure). Associating every conceptwith a discrete physical entity in the body allowed usto safeguard against representing one physical entityby more than one concept, even if a concept is known

Dow


ic.oup.com/jam



by several names. Once this constraint was satisfied,we assigned each unique concept to one or moreclasses or subclasses in the ontology for which it sat-isfied the type definition.

Our commitment to represent each visible entity ex-plicitly, as a unique concept is one of the notable fea-tures that distinguish the Digital Anatomist ontologyfrom available hard copy and machine-parsed sourcesof anatomical information. An example may help toillustrate what we believe is the advantage of our ap-proach. Genetically determined organizational pat-terns conserved in the vertebrate body dictate thatcertain subclasses of anatomical structures occur notonly bilaterally but also in metameric sets or othersegmental patterns that are based on acinar architec-tures. In contrast to ours, several structured vocabu-laries take a predominantly compositional ap-proach2 – 4 and provide procedures for joining a term(e.g., rib) to numerical and laterality modifiers in or-der to represent members of a set. We chose an enu-merative approach and have entered a unique conceptin the ontology for each discrete, visible entity, andhave associated it with a unique concept identifierand a preferred term. For instance, the Right third ribis a unique concept in the ontology, as is the Superiorarticular facet of the head of the right third rib.

Although macroscopic anatomical entities are numer-ous, their number is finite and can be readily man-aged by commercial database programs. Thus, for thepurposes of an ontology that models the physical or-ganization of the body, the comprehensiveness andveracity of a symbolic model obtainable by an enu-merative approach outweigh the procedural disad-vantage which, after all, has to be overcome onlyonce. A compositional approach is suited for makingstatements about anatomy (structure) by those whohave a knowledge of anatomy, but it fails to satisfythe objective of our ontology: a comprehensive andcoherent, symbolic model of anatomy (structure)which, in its human-readable form, is comprehensibleto both novice and expert users. Meeting this objectiveis a practical requirement for evaluating the ontologyby teachers and students of anatomy.

Entering concepts at the level of granularity that wehave chosen, leads to a heightened awareness of dif-ferences and variations in recurring anatomical enti-ties and their parts. An example involves the 11 pairsof intercostal nerves and their distribution pattern: al-though each pair resembles other members of the set,each has branches that innervate different subdivi-sions of the skin and serous membranes associatedwith the body wall (pleura, pericardium, peritoneum).Unless each nerve is represented independently, theparticular nerves that send branches to the breast, for

example, or to the diaphragm, or to the peritoneumapposed to the diaphragm’s abdominal surface, maybe readily overlooked. In other words, unless themembers of a set are represented as distinct concepts, itwill be procedurally intractable to associate branchesof different members of the set with different anatom-ical structures.

Such detailed representations can provide the knowl-edge necessary to make inferences about sources ofreferred pain, for instance, without having to providea problem-targeted application concerned with thedifferential diagnosis with all possible manifestationsof referred pain. Likewise, the specific morphologicaldifferences between the 12 thoracic vertebrae may nothave an obvious clinical relevance but the spatial re-lations that change from vertebra to vertebra certainlydo, and these relationships cannot be represented un-less each vertebra exists as a distinct concept. Knowl-edge of these different relations is necessary in orderto deduce which of the adjacent anatomical structureswill be affected by lesions of a particular vertebra.Knowledge-based applications that target problemsrequiring such detail in anatomy (structure) will callfor the explicit representation of these concepts. En-tering such information in the knowledge base whenthe need for it arises is economically inefficient and islikely to result in inconsistencies.

Notwithstanding these justifications for the approachwe have taken, we recognize the value and advantagesof the compositional or generative approach. In expres-sive representation systems such as Ontolingua29 andGALEN,3,21 common relationships can be described byaxioms that hold for enumerated sets of anatomical con-cepts without having to repeat those axioms for eachmember of the set. It is indeed desirable to merge enu-merative and generative representations once we mi-grate the simple frame system we have implemented toa more expressive system, such as Ontolingua.

In order to ensure that the canonical symbolic modelswe generate can also accommodate instantiated anat-omy, we have assigned variants of normal anatomy(structure) to a Variant subclass of particular anatom-ical concepts. For instance, Third coronary artery is as-signed to Variant artery, a child of Artery in the ontol-ogy. At this stage, however, anatomical variants havenot been systematically instantiated. Our experienceargues for comprehensive representation of anatomi-cal variants in the anatomical ontology.

Anatomical variants result from modulation of theprocesses that establish the canonical anatomicalpattern of organization, without adversely affectingphysiologic function. In some cases (e.g., coronary ar-teries, bronchopulmonary segments), the incidence of

Dow


ic.oup.com/jam



variants may be higher than that of the canonical pat-tern. When the modulation or disruption of morpho-genetic processes leads to the persistence of embry-onic structures that are normally eliminated duringmorphogenesis, or an abnormal structure developsthat interferes with physiologic function, the structureshould be classified as a Congenital abnormality. Anatrial septal defect, or an ileal (Meckel’s) diverticulum,for instance, is classified as Congenital abnormality ofthe interatrial septum or ileum, respectively. Such asemantic type already exists in UMLS.

Term Assignments

We have assigned 14,916 terms to the 8,763 conceptswe have entered to date in the ontology. Preferredterms were based on widely used American37 andEuropean38 textbooks of anatomy. We disallowedhomonyms. Where two concepts have been tradition-ally denoted by the same term, we associated, in pa-renthesis, a different modifier with each term to as-sure the uniqueness of preferred terms [e.g., muscle(organ), muscle (tissue)]. We also entered as syno-nyms all other terms that we could identify as havingbeen associated with the concept (Fig. 5); in some in-stances there are as many as six synonyms, but usu-ally not more than two. Segmentally recurring con-cepts and their parts (e.g., branches of arteries andnerves), as a rule, have no synonyms.

We have cross-referenced the terms that we have en-tered with three sources: Nomina Anatomica,19 SNO-MED,2 and the UMLS Metathesaurus1 (Fig. 5). Of the15,000 terms, 1,850 occur in SNOMED, and less than700 in NA. Since NA tends to record the name of setsof entities as a Latin plural string, we have used theLatin singular string as a synonym for a member ofthe set, when doing so seemed appropriate (e.g., Ar-teria intercostalis posterior dextra for Right posterior in-tercostal artery, when the NA entry is limited to Aa.intercostales posteriores).

Relationships

The generic -is a- relationship has served for fully in-stantiating an ontology with canonical concepts for amajor body part, the thorax. The ontology establishestype classes for anatomical physical entities. However,in order to model the physical organization of anat-omy (structure), a number of other relationships mustbe represented explicitly. The definitions we have for-mulated imply that chief among these are partitiverelationships, as defined by the UMLS.1 Others are re-lationships of spatial adjacency and those that de-scribe the branching and union of vessels and nerves.

We have formulated hierarchies using the transitive-part of- relationship. Figure 5 illustrates the high gran-

ularity of anatomical knowledge that can be rep-resented with this relationship. For symbolically mod-eling relationships among instances of nerves,arteries, veins, and lymphatic vessels, we have de-fined two anatomical relationships:

-branch of-, a smaller, peripheral anatomical struc-ture, given off by a larger, central one or into whicha larger structure divides. Pertains in particular tothe arborization of arteries, nerves, ganglia andbronchi. Inverse relationship: -has branch-.

-tributary of-, a smaller peripheral anatomical struc-ture, particularly a vessel, that combines with an-other to form a larger, more central one. Pertains inparticular to the confluence of veins, lymphatic ves-sels and ducts. Inverse relationship: -has tributary-.

Using these two relationships, we have linked over3,500 concepts of thoracic anatomy to 80 root con-cepts. Some of the trees extend to 7–8 generations ofnodes in the -branch of- and -tributary of- hierarchies.

We are currently in the process of defining and im-plementing transitive anatomical relationships ofspatial adjacency. In a fully segmented geometric an-atomical dataset or model, such adjacencies can berepresented quantitatively by sets of coordinates. Co-ordinate-free methods have also been proposed fordescribing in qualitative terms spatial relationships in2-D medical images.41 However, these descriptions, aswell as coordinates, must also be stated in terms in-tuitive to humans, using the established naming con-ventions for spatial relationships. These are the termsthat appear in standard anatomical and clinical publi-cations, and in medical records.

Our current purpose is to represent canonical rela-tionships of anatomical adjacency that have unambig-uous semantics and a long history of established us-age. The canonical spatial adjacencies are:

-anterior-, and its inverse, -posterior--superior-, and its inverse, -inferior--lateral-, and its inverse, -medial-

Any of these attributes may be joined to any other inthe set, except to its own inverse, in order to describewith considerable precision binary spatial adjacenciesbetween anatomical structures and anatomical spatialentities. Conjunction of an attribute with its inversespecifies direction, which pertains to orientation,viewing, or to passage of x-rays, projectiles, instru-ments, or of body substances and cells (including thespread of exudates, pus or cancer cells).

Our objective is to specify the conceptualization of ca-nonical, as well as other, anatomical spatial relation-

Dow


ic.oup.com/jam



F i g u r e 5 Screen capture from the authoring program KB Manager, showing a segment of the -part of- hierarchy forthe heart to illustrate concept granularity, term assignments and cross references with other vocabularies. The symbol>> indicates that the node in the hierarchy has at least one generation of children that is not shown. Immediate offspringof a node are shown by one tab indentation. All components of the Conducting system of the heart could, for instance,be displayed by double clicking on the term and the successive generations of its children. Note that those componentsof the conducting system that are associated with the right atrium are displayed as parts of the Myocardium of rightatrium, providing a symbolic representation of useful spatial information. The preferred name of a concept is highlightedin the Hierarchy Editor panel and also appears in the top bar of the Concept Inspector panel. Of the five synonymsassociated with the concept, one is highlighted and appears in the Term Inspector panel, which provides informationabout the term: its role (synonym), concept identifier (UWDA ID) and its SNOMED identifier. Selecting Nodus sinu-atrialis among the synonyms would identify the authority for the terms as Nomina Anatomica.

ships. The spatial adjacency ontology will representthe symbolic equivalents of a geometric constraintnetwork,42 which relates, and also predicts, the rela-tive position of one anatomical entity to others. Rep-resentation of such spatial knowledge will be of par-ticular value in the interpretation of medical images,

in which the location of invisible anatomical entitiesmust be inferred from those that are revealed by theimaging procedure.

We believe that a similar approach must be pursuedfor specifying other relationships that hold among an-

Dow


ic.oup.com/jam



atomical entities and non-anatomical concepts ofother biomedical domains. There is a need for devel-oping domain ontologies that specify not only phys-iologic function associated with different anatomicalentities, but also representations of anatomical entitieswith various medical imaging modalities, so thatthese ontologies can be mapped to the anatomical on-tology.

Process of Generating the Ontology

The conceptual framework of the ontology resultedlargely from discussions that took place over an aca-demic year in the context of a course at the Universityof Washington: Anatomical Knowledge Representa-tion (CS 590BR); all the authors participated in thecourse. Principles for formulating symbolic represen-tations of physical anatomical entities were proposedby one of the authors (CR) and were approved follow-ing rounds of discussions supplemented by presen-tations of publications from the relevant literature.The same process was followed for sanctioning nar-rative text definitions of classes, subclasses and rela-tionships. The Knowledge Manager tool (designed byJFB) was used by JLM and CR for data entry. In gen-eral, this proceeded according to subclasses of physi-cal anatomical entities (e.g., viscera, nerves, bones, se-rous cavities); in each case switching back and forthbetween the -is a- and -part of- hierarchies. Conceptassignments to subclasses were strictly guided by thedefining attributes specified in the definitions. Fullypopulating a specific subclass with instances presentin the thorax tested the validity of the definitions notonly for the subclass itself, but for entities of which itis constituted, and also for entities that the subclassin turn constitutes. Problematic instances were pre-sented at weekly meetings of class CS 590BR, result-ing at times in the modification of definitions, andreassignments of concepts. For instance, the initial as-signment of Joint was to Anatomical junction in accordwith the UMLS semantic type definition of this con-cept. As the definitions for various anatomical entitiesbecame clarified, Joint was reassigned to Organ systemsubdivision, and later to Organ. Once subclasses of An-atomical spatial entity were also defined, it was recog-nized that Joint also satisfied defining attributes of An-atomical junction. Currently it is assigned to bothOrgan and Anatomical junction, because it meets thedefinition of both these concepts.

Apart from those for segmentally recurring structures,all concept assignments were made one by one usingthe Knowledge Manager. When a subclass was com-pleted, comprehensiveness of the entered data waschecked by consulting the text book references,37,38

Nomina Anatomica and SNOMED.

Implementation

The symbolic knowledge base is integrated in the Dig-ital Anatomist distributed framework.8,9 Currently, theknowledge base is represented by tables in a relationaldatabase, and by associated text files which describedefinitions and other textual attributes. The terms ta-ble contains the preferred names and synonyms forall anatomical concepts, including the Nomina Ana-tomica Latin string. Each concept is assigned a uniqueID, which remains constant. The terms table also re-cords the associated SNOMED and UMLS ID, if theconcept exists in these sources.

Terms are related by means of the links table, whichrecords binary semantic relationships. The relationaldatabase is accessed by the Sybase commercial rela-tional database server, which is in turn accessed bythe Knowledge Base Manager authoring program, atool for entering knowledge in the knowledge base.Figure 5 is a screen capture from the Knowledge BaseManager program, which is written in NeXTStep forthe NeXT computer.

Applications and Evaluation

The application that has driven the development ofthe symbolic knowledge base is the Atlas client pro-gram in the Digital Anatomist distributed frame-work.9 Designed to support anatomy education, theprogram retrieves and integrates information from thespatial database and the symbolic knowledge basemodules of the framework (Fig. 6). Unlike clinical us-age, which targets instantiated anatomy and is usuallyrestricted to selected body parts and organ systems,education deals with the canonical anatomy of thewhole body. Therefore, during the formative phase ofthe knowledge base, its educational uses are ideal forevaluating comprehensiveness of concept representa-tion and the logical structuring of the information. Forinstance, in addition to retrieving names of anatomi-cal entities by clicking on the image, the Web clientversion of Atlas also generates a so called ‘‘pin dia-gram’’ on the fly, which automatically displays thenames of all entities contained in an image.9,43 Suchlabeling provides a check on the comprehensivenessof symbolic concept representation and also assuresthat not more than one preferred name is associatedwith each entity.

In addition to their use by health science students ina variety of courses at the University of Washington,the Digital Anatomist information sources are con-sulted world wide. For instance, during a recent oneand a half year period, the Web client Atlas was ac-cessed by over 33,000 sites from 94 countries.9 Feed-

Dow


ic.oup.com/jam



F i g u r e 6 Screen capture from the web atlas of Thoracic Viscera, illustrating the association of a term (retrieved fromthe Symbolic Knowledge Base by the Symbolic Knowledge Server), on the fly, with a structure present in a 3-Dreconstruction (retrieved from the Spatial Database by the Web Server and CGI Programs9,43). The term posterior leftventricular branch of left coronary artery was selected in the -branch of- hierarchy of the Symbolic Knowledge Base, accessedby the Web client; clicking on the term provides a list of images in which the concept is present. A lateral view of theheart, reconstructed by Skandha program8 from 1-mm cryosections of a cadaver,44,45 was chosen; a leader associates theselected term (shown above the image) with the corresponding anatomical structure in the 3-D reconstruction.

back received from such extensive usage has beenhelpful in identifying errors and inconsistencies interm assignments.

As described above, formative evaluation of the clas-sification was an integral part of formulating and in-

stantiating the ontology. Assisted by guidelines forstructuring ontologies,34,46,47 we have revised and val-idated candidate subclasses and their defining attrib-utes as we entered specific data about a major bodypart. Parallel approaches from the top down and fromthe bottom up were indispensable in this cyclic pro-

Dow


ic.oup.com/jam



cess. The anatomical ontology has been evaluated bythe National Library of Medicine and will be incor-porated into the 1998 edition of UMLS. Since the an-atomical information we have represented is stable,maintenance of the knowledge base does not pose aparticular problem. The anatomy ontology and ter-minology will be maintained as a component of theMetathesaurus.

Availability of the Digital Anatomist Symbolic Knowl-edge Base through UMLS will open up the possibilityfor its empirical evaluation by knowledge base de-velopers, independent of our group. Our motivationfor describing the underlying principles and rationalefor the anatomical ontology in this publication is toinvite and promote such evaluation. Although ourformative evaluation based on the classification ofseveral thousands of concepts suggests to us that wehave reached an asymptote in defining the ontology,a number of questions remain to be answered. Forinstance: How unambiguous is the ontology forknowledge base developers in different fields of med-ical informatics and how easy is it to adapt it to spe-cific applications? How well do classes and subclassesof the ontology subsume anatomical concepts associ-ated with body parts other than the thorax, and whatis the error rate in such modeling?

Emergent Properties

Although the current usage of the symbolic knowl-edge base is limited largely to the naming of anatom-ical entities, there are at least five properties of therepresentations we have generated that facilitate con-ceptualization of anatomy (structure):

n Clarity. The explicit definitions of classes and sub-classes in the context of the ontology have intro-duced a degree of clarity into the conceptualizationand description of anatomy (structure) that is lack-ing from both hard copy and other machine-parsedsources of anatomical information. The representa-tions seem promising for facilitating the learning ofanatomy, a hypothesis that will be tested throughthe use of the Digital Anatomist Atlas.9

n Portability. The structural ontology is portable,can be made available on line, and readily lendsitself for labeling and organizing any image datasetof human macroscopic anatomy, be it from a livingor a deceased subject.9 The ease of associating termsin the structural ontology with spatial representa-tions of anatomical entities further facilitates con-ceptualization of anatomy (structure).

n Display of information at different levels of gran-ularity. The sheer wealth of anatomical informa-

tion presents a difficult and longstanding conun-drum in anatomy: how to filter and accessanatomical detail at a level appropriate to eachuser’s expertise and specific needs. Historically, theproblem has been addressed by producing text-books and treatises that are specifically targeted fordifferent user populations with widely disparateneeds for detailed anatomical information (alliedhealth students, medical and dental students, andtrainees and practitioners in the clinical specialties).The possibility of opening up trees in the -part of- -branch of- and -tributary of- hierarchies, which dis-play information at different levels of abstraction,solves this problem. For instance, if desired, the -branch of- hierarchy can display the smallest visiblebranch of an artery in relation to its parent vesselby clicking on successive generations of nodes; al-ternatively, the display can be limited to majorbranches of a parent artery, or similarly the parentartery itself. Such nodal levels can specify the gran-ularity of anatomical spatial data that are requiredfor the problem at hand. Display of greater detailbecomes meaningful when differences in serially re-curring structures (e.g., branches of intercostalnerves, as cited before) assume practical importanceand are needed to support inference.

n Integration of systemic and regional anatomy. Byregarding Organ as the basic macroscopic organi-zational unit of the body, the knowledge base read-ily provides both a systemic and regional view ofanatomy. The concepts defined by the anatomicalontology can be displayed in the -part of- hierarchyto model both systemic and regional organizationof anatomical entities in a form that is intuitive tohumans.

n Support for inference. The high level of granu-larity in concept representation supports inferenceof a number of relationships which, without therequisite anatomical detail, would need explicit rep-resentation. For instance, as noted above, the nervesupply of the diaphragm can be inferred from the-branch of- hierarchy, obviating the need for explic-itly representing the -supplies- relationship or attrib-ute. It may readily be inferred that the right andleft phrenic nerves and a subset of the right and leftintercostal nerves supply the diaphragm, becausethese are the only nerves that are represented ashaving diaphragmatic branches.

Evolutionary Enhancements

It will be evident from the preceding sections that thesimple scheme of the -is a- ontology, supplemented byother hierarchies such as -part of-, can capture detailed

Dow


ic.oup.com/jam



anatomical knowledge. However, the lack of sufficientsemantic expressivity for the concepts and relation-ships in a semantic net motivates us to consider moreformal knowledge representation schemes.48 Havingrepresented a substantial body of anatomical knowl-edge enables us to evaluate available knowledge rep-resentation systems for extending and enhancing theDigital Anatomist Symbolic Knowledge Base.

The next step in the evolution of the symbolic knowl-edge base will be to commit to a representation lan-guage and scheme. Recognizing the advantages of anenvironment that enables reuse of domain knowledgeand problem-solving methods, we are looking at sys-tems that provide tools for promoting the reuse of theknowledge base we are building. A number of suchsystems have been advocated for biomedical knowl-edge representation and application development.3,29,30

Reusable environments, such as those provided byPROTEGE-II,33 promote the development of domainontologies, which are decoupled from the specificproblem-solving methods that use domain knowl-edge. The anatomical ontology could, therefore, bemapped onto the problem solving methods, which aredescribed in terms of domain independent, abstractproblem solving concepts.

The fundamental components of the Digital AnatomistSymbolic Knowledge Base we describe in this reportwere developed from an anatomical viewpoint basedon our experience and expertise in anatomy. The nextphase of our work will benefit from collaborations withinvestigators whose experience and expertise is inmethodologies of knowledge representation.

Conclusions and Discussion

Focusing on a single organizational level of the hu-man body, the macroscopic level, we have proposedprinciples according to which we have implementeda consistent symbolic representation of the physicalentities that constitute the human body. The anatom-ical ontology we have formulated is distinguishedfrom other symbolic representations of anatomy in thefollowing respects:

n It is formulated from an anatomical viewpoint.

n Its nodes are defined by sets of anatomical attri-butes, which are also explicitly defined.

n It defines Organ as the basic organizational unit ofmacroscopic anatomy.

n It models the physical organization of the body anddisplays this organization in a manner that is bothintuitive to humans and can be parsed by machine.

By fully instantiating the ontology for a major bodypart, the thorax, we have shown that all macroscopicentities can be assigned to one of three classes andtheir subclasses. The ontology is extendible, further-more, to the microscopic organizational level, in thateach cell in the body can be assigned to a canonicalinstance in one of the classes we have defined formacroscopic anatomical entities. In the ontology, de-fining attributes of a parent class or subclass are in-herited by its descendants; these are, in turn, distin-guished from their parent and their siblings bydifferentia, expressed as the defining anatomical at-tributes of each. We have used established anatomicalterminology for denoting canonical instances of ana-tomical entities, and have assigned specific meaningto the terms that we have used to designate classesand subclasses of these entities.

Many elements of the knowledge structure we pro-pose are implied in hard copy sources of canonicalanatomy. However, when we have judged it to beboth necessary and prudent, we have intentionally setsome precedents in organizing anatomical knowl-edge. As a result, the semantics, definitions, and theclassification we propose remove many of the currentambiguities in anatomical information and establishthe first requirements for logic-based notations ofanatomy. The ontology should support the develop-ment of applications for reasoning along the horizon-tal axis of anatomy, an information domain funda-mental to virtually all fields of medicine. Making useof the knowledge we have represented, such appli-cations should provide the means for empirical eval-uation of the knowledge base. The outcome of theseevaluations, and the ensuing implemented revisionsof the knowledge base, will establish standards foranatomical knowledge representation. Standardizedrepresentation of anatomical knowledge is an impor-tant objective for realizing standards in clinical datarepresentation, because macroscopic anatomy is rela-tively stable and forms the basis for many types ofclinical data, particularly those generated by the phys-ical examination and medical imaging.

In addition to being available as a knowledge sourcethat can be mapped to other domain ontologies, theDigital Anatomist Symbolic Knowledge Base has im-mediate practical applications in its current imple-mentation. Through the Digital Anatomist Atlas clientprogram,9 the knowledge base makes available overthe Internet and the Web well-defined terminology forannotating spatial datasets such as the Visible HumanProject49 and various medical images. The synonymsassociated with anatomical entities accommodate andunify different usages of terminology prevalent in dif-ferent medical and surgical specialties. The knowl-

Dow


ic.oup.com/jam



edge base provides a structure for the classification ofanatomical entities, which should assist in the storage,sorting and retrieval of medical and other anatomicalimage data. Its display in human-readable form canpromote the conceptualization of anatomy. Anatomy isthe first subject—and one of the most challenging andtime-consuming subjects—introduced in the trainingof all health care professionals. There is a need forlogic-based, machine-parsed representations of ana-tomical knowledge for the creation of intelligent ed-ucational programs in anatomy. The Digital Anato-mist ontology establishes a basic requirement for suchapplications.

We hope that both the immediately realizable and thepotential uses of the anatomical ontology we havegenerated will serve as persuasive arguments for in-vesting in the representation of deep knowledge alongthe horizontal axis of the basic and clinical sciences.An approach that is oriented entirely toward solvingproblems is in danger of keeping its operation, evenif narrowly targeted, on a superficial level. To para-phrase Blois,50 medical problems, including the learn-ing of anatomy, require vertical reasoning. Unlessknowledge sources are developed along the horizon-tal axes of the basic and clinical sciences to supportsuch reasoning, the ad hoc approaches to problemswill be both shallow and costly. As demonstrated byour work, the properties that emerge from efforts fo-cused on representing deep knowledge in a defineddomain, can yield immediate practical benefits. Theirmost important contribution, however, is that theyempower applications by the inferences that are madepossible through the reusable knowledge. It will beinteresting to see whether the availability of a re-source for symbolic anatomical knowledge will exerta motivating effect on the development of problem-targeted applications in biomedical education and pa-tient care to an extent similar to that exerted by theavailability of the Visible Human spatial dataset onapproaches to visualizing anatomy. At the least, thework we have accomplished to date should facilitatethe encoding of anatomical knowledge for the entirehuman body in a schema that makes this knowledgewidely accessible and usable. We regard the ontologywe have formulated as the first iteration and the foun-dation of a knowledge base in anatomy, and we invitecomments and feedback to assist in its revision andrefinement.

The authors thank the following individuals: Dr. Alexa T.McCray, National Library of Medicine, for her advice and sup-port throughout this project and for reviewing an early draft ofthe manuscript; Kraig Eno and Jeff Prothero, for implementingthe Knowledge Base Manager program; Drs. Keith E. Campbell,Mark A. Musen, John W. Prothero, Henk Roelink, and Mark S.

Tuttle for critically reading the manuscript; and Dr. PenelopeGaddum-Rosse, for editing the manuscript.

References n

1. National Library of Medicine. UMLS knowledge sources,7th ed. Unified Medical Language System. Bethesda, MD:National Library of Medicine, 1996.

2. Cote RA, Rothwell DJ, Palotay JL, Beckett RS, Brochu L(eds). SNOMED International: The systematized nomencla-ture of human and veterinary medicine. Northfield, IL: Col-lege of American Pathologists, 1993.

3. Rector AL, Bechhofer S, Goble CA, Horrocks I, Nowlan WA,Solomon WD. The GRAIL concept modeling language formedical terminology. Artif Intell Med. 1997;9:139–71.

4. Schulz EB, Price C, Brown PJB. Symbolic anatomic knowl-edge representation in the Read Codes Version 3: Structureand application. J Am Med Inform Assoc. 1997;4:38–48.

5. Campbell KE, Das AK, Musen MA. A logical foundation forrepresentation of clinical data. J Am Med Inform Assoc.1994;1:218–32.

6. Evans DA, Cimino JJ, Hersh WR, Huff SM, Bell DS. Towarda medical-concept representation language. J Am Med In-form Assoc. 1994;1:207–17.

7. Brinkley JF, Prothero JS, Prothero JW, Rosse C. A frameworkfor the design of knowledge-based systems in structural bi-ology. In: Kingsland LC (ed). Proc 13th Annu Symp ComputAppl Med Care (SCAMC 89). Los Alamitos, CA: IEEE Com-puter Society Press, 1989:61–5.

8. Brinkley JF, Rosse C. The Digital Anatomist distributedframework and its applications to knowledge-based medi-cal imaging. J Am Med Inform Assoc. 1997;4:165–83.

9. Brinkley JF, Bradley SW, Sundsten JW, Rosse C. The DigitalAnatomist information system and its use in the generationand delivery of Web-based anatomy atlases. Comput Bio-med Res. 1997; in press.

10. Heckscher WS. Rembrandt’s Anatomy of Dr. Nicolaas Tulp.New York: New York University Press, 1958.

11. Vesalius Andreas. De Humanis Corporis Fabrica. (On thefabric of the human body.) Basel: Oporinus, 1543.

12. Schleiden MJ. Beitrage zur Phytogenesis. Muller’s Archiv,1838. Translation in Sydenham Soc. 12, London, 1847.

13. Schwann T. Mikroscopische Untersuchungen uber dieUbereinstimmung in der Struktur und dem Wachstum derThiere und Pflanzen. Berlin, 1839. Translation in SydenhamSoc. 12, London, 1847.

14. Rosse C. The potential of computerized representations ofanatomy in the training of health care providers. Acad Med.1995;70:499–505.

15. Hohne KH, Pflesser B, Pommert A, et al. A new represen-tation of knowledge concerning human anatomy and func-tion. Nat Med. 1995;1:506–11.

16. Pommert A, Schubert R, Riemer M, Schiemann T, Tiede U,Hohne KH. Symbolic modeling of human anatomy forvisualization and simulation. In: Visualization in Biomedi-cal Computing. Bellingham, WA: Proceedings Series, SPIE.The International Society for Optical Engineering, 1994;2359:412–23.

17. National Library of Medicine. Labeling and structuring ofanatomical image data of the visible human project. Requestfor proposals (RFP) number: NLM 95-112/SLC, 1995.

18. O’Neil MJ, Payne C, Read JD. Read Codes Version 3: A userled terminology. Meth Inform Med. 1995;34:187–92.

19. International Anatomical Nomenclature Committee. Nom-

Dow


ic.oup.com/jam



ina Anatomica, 6th ed. Edinburgh: Churchill–Livingstone,1989.

20. Niggemann J. Representation of neuroanatomical knowl-edge: the description language ADL. In: Czap H, NedobityW (eds): Proceedings of the 2nd International Congress onTerminology and Knowledge Engineering: Applications(TKE 90). Frankfurt: INDEKS Verlag, 1990;200–9.

21. Rector AL, Gangemi A, Galeazzi E, Glowinski AJ, Rossi-Mori A. The GALEN CORE model schemata for anatomy:Towards a re-usable application-independent model of medi-cal concepts. In: Barahona P, Veloso M, Bryant J (ed): Pro-ceedings of the 12th International Congress of the EuropeanFederation for Medical Informatics (MIE 94), Lisbon. IOSPress, 1994;229–33.

22. Lucas P. The representation of medical reasoning models inresolution-based theorem provers. Artif Intell Med. 1993;5:395–414.

23. Yalcinalp LU, Sterling L. Diagnosing jaundice expert sys-tem. Comput Math Appl 1990;20:125–40.

24. Horn W. Utilizing detailed anatomical knowledge for hy-pothesis formation and hypothesis testing in rheumatolog-ical decision support. Artif Intell Med. 1991;3:21–39.

25. Cawley MG, Natarajan K. Development of a model for usein medical image interpretation. In: Proceedings of the 5thAlvey Vision Conference (AVC 89). Sheffield, UK: Univer-sity of Sheffield, 1989;305–8.

26. Menhardt W. Iconic fuzzy sets for MR image segmentation.In: Todd-Pokropek AE, Viergever MA (eds): Medical im-ages: formation, handling and evaluation. Berlin:Springer–Verlag, 1992:579–91.

27. Message MA, Anderson RH. Towards a new terminologyfor clinical anatomy, with special reference to the heart. ClinAnat. 1996;9:317–29.

28. Tuttle MS, Blois MS, Erlbaum MS, Nelson SJ, Sherertz DD.Toward a bio-medical thesaurus: building the foundation ofthe UMLS. In: Greenes RA (ed): Proc 12th Annu Symp Com-put Appl Med Care (SCAMC 88). New York: IEEE Com-puter Society Press, 1988;191–5.

29. Neches R, Fikes R, Finin T, et al. Enabling technology forknowledge sharing. AI Magazine. 1991;12:36–56.

30. Musen MA. Dimensions of knowledge sharing and reuse.Comput Biomed Res. 1992;25:435–67.

31. Gruber TR. A translation approach to portable ontologyspecifications. Knowledge Acquisition. 1993;5:199–220.

32. van Heijst G, Falasconi S, Abu-Hanna A, Schreiber G, Ste-fanelli M. A case study in ontology library construction. Ar-tif Intell Med. 1995;7:227–55.

33. Tu SW, Eriksson H, Gennari JH, Shahar Y, Musen MA. On-tology-based configuration of problem-solving methodsand generation of knowledge-acquisition tools: applicationof PROTEGE-II to protocol-based decision support. Artif In-tell Med. 1995;7:257–89.

34. Zweigenbaum P, Bachimont B, Bouaud J, Charlet J, Bois-

vieux J-F. Issues in the structuring and acquisition of anontology for medical language understanding. Meth InformMed. 1995;34:15–24.

35. Dorland’s Illustrated Medical Dictionary, 27th ed. Philadel-phia: Saunders, 1988.

36. Stedman’s Medical Dictionary, 25th ed. Baltimore: Williams& Wilkins, 1990.

37. Rosse C, Gaddum-Rosse P. Hollinshead’s Textbook of Anat-omy, 5th ed. Philadelphia: Lippincott–Raven, 1997.

38. Williams PL, Warwick R, Dyson M, Bannister LH. Gray’sAnatomy, 37th ed. New York: Churchill–Livingstone, 1989.

39. Ferner H, Staubesand J. Benninghoff/Goerttler’s Lehrbuchder Anatomie des Menschen. Munchen: Urban & Schwar-zenberg, 1980.

40. McMinn RMH. Last’s Anatomy, 8th ed. New York:Churchill–Livingstone, 1990.

41. Tagare HD, Vos F, Jaffe CC, Duncan JS. Arrangement: a spa-tial relation comparing part embeddings and its use in med-ical image comparisons. In: Proceedings of the 13th Inter-national Conference on Information Processing in MedicalImaging (IPMI 93). New York: IEEE Computer Society, 1993,pp. 132–48.

42. Brinkley JF. Hierarchical geometric constraint networks asa representation for spatial structural knowledge. In: FrisseME (ed): Proc 16th Annu Symp Comput Appl Med Care(SCAMC 92). New York: McGraw-Hill, 1992:140–4.

43. Bradley SW, Rosse C, Brinkley JF. Web-based access to anonline atlas of anatomy: the Digital Anatomist commongateway interface. In: Gardner RM (ed): Proc 19th AnnuSymp Comput Appl Med Care (SCAMC 95). Philadelphia:Hanley & Belfus, 1995:512–6.

44. Conley DM, Kastella KG, Sundsten JW, Rauschning W,Rosse C. Computer-generated three-dimensional reconstruc-tion of the mediastinum correlated with sectional and ra-diological anatomy. Clin Anat. 1992;5:185–202.

45. Conley D, Rosse C. The Digital Anatomist: interactive atlasof thoracic viscera (CD-ROM). Seattle: Health Sciences Cen-ter for Educational Resources, University of Washington,1996.

46. Cimino JJ, Hripcsak G, Johnson SB, Clayton PD. Designingan introspective, multipurpose, controlled medical vocab-ulary. In: Kingsland LC (ed): Proc 13th Annu Symp ComputAppl Med Care (SCAMC 89). Los Alamitos, CA: IEEE Com-puter Society Press, 1989;513–8.

47. Rada R, Ghaoui C, Russell J, Taylor M. Approaches to theconstruction of a medical informatics glossary and thesau-rus. Med Inf (Lond). 1993;18:69–78.

48. Rich E, Knight K. Artificial Intelligence, 2nd ed. New York:McGraw-Hill, 1991.

49. Spitzer V, Ackerman MJ, Scherzinger AL, Whitlock D. TheVisible Human Male: A technical report. J Am Med InformAssoc. 1996;3:118–30.

50. Blois MS. Medicine and the nature of vertical reasoning. NEngl J Med. 1988;318:847–51.

Dow


ic.oup.com/jam


motivation and organizational principles for anatomical

Documents