texture and mathematical morphology for hot-spot detection ... · texture and mathematical...

Texture and Mathematical Morphologyfor Hot-Spot Detection in Whole Slide Images

of Meningiomas and Oligodendrogliomas

Zaneta Swiderska1(B), Tomasz Markiewicz1,2, Bartlomiej Grala2,and Wojciech Kozlowski2

1 Warsaw University of Technology, 1 Politechniki Sq., 00-661 Warsaw, [email protected], [email protected]

2 Military Institute of Medicine, 128 Szaserow Str., 04-141 Warsaw, Poland{bgrala,wkozlowski}@wim.mil.pl

Abstract. The paper presents a combined method for an automatichot-spot areas selection in the whole slide images to support thepathomorphological diagnostic procedure. The studied slides repre-sent the meningiomas and oligodendrogliomas tumour stained with theKi-67/MIB-1 immunohistochemical reaction. The presented methodbased on mathematical morphology and texture analysis helps to deter-mine the tumour proliferation index complementing medical informa-tion for prognosis and treatment. The major functions of the algorithminclude detection of immunopositive cells in the tumour area and theidentification and elimination of hemorrhages areas from the specimenmap. The results of the numerical experiments confirm high efficiency ofthe proposed solutions.

Keywords: Image processing · Mathematical morphology · Texture

1 Introduction

Histopathological examination of tissues using immunostaining tests is a basicmethod of identifying tumours. It often serves as a tool supporting the choice ofoptimal therapy and defines the prognostic indicators. Tumour proliferation incentral nervous system tumours can be characterised with the widely used Ki-67/MIB-1 marker. The immunopositive (proliferated) cell nuclei are marked withbrown whereas ther cell nuclei are marked with blue, and their raitio gives theproliferation index. In meningiomas and oligodendrogliomas (the most frequentprimary intracranial tumours) this index is used for classification of tumoursinto meningothelial (WHO I), atypical (WHO II), anaplastic (WHO III), andoligodendrogliomas (WHO II and III), and to it correlates with tumour recur-rences. Thus, the automatic quantitative evaluation of the specimen can offersa very useful tool for the pathologists.

The several solutions to this problem have been proposed in literature [1,4,9].Most approaches are based on the quantitative evaluation of the percentage ofc© Springer International Publishing Switzerland 2015G. Azzopardi and N. Petkov (Eds.): CAIP 2015, Part II, LNCS 9257, pp. 1–12, 2015.DOI: 10.1007/978-3-319-23117-4 1

2 Z. Swiderska et al.

immunopositive cells in the areas of greatest tumour proliferation. The areaswith a high immunopositive reactions are called hot-spots. In the paper [8] isproposed a method for the increasing of visibility of the positive nuclei at lowresolution image in manual examination. In [7] the author presented the auto-mated selection method of hot spots algorithm. This method is based on adap-tive step finding techniques for the increasing the computational efficiency andperformance of hot spot detection. A problem of the spatial distribution of theselected hot spot fields was not taken into account in this paper. To the best ofour knowledge, algorithm containing hotspot selection and problem of the hotspot spatial distribution, for whole slide image, not yet been presented in theliterature. For objective evaluation, some scattering of selected fields is also rec-ommended. For example, one of the schemes is the choice of fields of view of thehighest levels of Ki-67 and then choosing several adjacent regions with positivereaction in short distance from this field, followed by the search for other areasof specimen. The major problem in the tested formulations are hemorrhages,which are stained with brown creating false-positive results. This paper presentsa computational method for improved identification of tumour proliferation andareas of hemorrhages.

A lot of algorithms for medical image processing are based on mathematicalmorphology and texture analysis. In many cases these methods are successfullyused to analyze single microscopic images. Recently, texture analysis and clas-sification have been applied to processing and context analysis of large imagescovering the whole specimen [2,11]. The whole histological specimen acquired ona microscopic glass scanner is called a whole slide image (WSI). The size of thisslide is usually very large (for example 100000×80000 pixels). The contextualanalysis of such images requires matching the image resolution to the task andalso developing new methods of image analysis and classification.

The example part of WSI of tumour, blood and background stained withKi-67 is presented in Fig. 1. To perform the proper quantitative evaluation ofthe proliferative Ki-67 index, a hot-spot regions must be identified. One of themis outlined in Fig. 1 as a part of tumour with increased Ki-67 proliferation index,compared to such index for the neighbourhood tumour area. Normally WSI mayinclude a lot of hot-spot regions. However, in the specimen exist also the densebrown areas indicated by the arrows. There are the touching blood neutrophils.So, correct classification of tumour and hemorrhages areas allow to accurateidentification of hot-spot areas and may help improve sample analysis.

In spite of some progress in this field there is still a need for an extensivestudy aimed at supporting clinical decisions [6]. In this work we propose a newmethod for an identification of the hot-spot areas in brain tumour images. Thetextural and classification methods are applied to differentiate the tumour areasagainst the hemorrhage areas. At the same time we apply the mathematical mor-phology based analysis to estimate the number of immunopositive cells. Also, weintroduce a penalty criterion for the selection of the hot-spot areas. The quan-tification results of an automatic selection of tumour proliferation is comparedwith standard visual selection performed by the expert.

Texture and Mathematical Morphology for Hot-Spot Detection 3

Fig. 1. An example of a part of the whole slide image representing the brain tumourwith the outlined hot-spot area. The hemorrhages are marked by the arrows.

2 Material and Methods

Fifteen cases of meningiomas and oligodendrogliomas subject to Ki-67/MIB-1immunohistochemical staining were obtained from the archives of Departmentof Pathomorphology from the Military Institute of Medicine in Warsaw, Poland.Acquisition of WSI was performed on the Aperio ScanScope scanner. The imageswere acquired under magnification 400x with a resolution 0.279 μm per pixels.Due to a very large size of images in the contextual analysis of the specimen it wasnecessary to reduce the resolution to enable direct examination and visualisation.We have chosen eight-fold reduction of the resolution to enable the evaluationperformed both manually and by a computer.

The proposed scheme of the image processing is organized in the followingstages: a) defining a map of specimen, b) elimination of the areas containingblood cells (hemorrhages), c) selection of sequential fields of the hot-spots basedon stained cells segmentation. This stage requires defining the punishment func-tion preventing from identifying hot-spot areas located in close proximity.

The map of specimen is produced by using the thresholding procedure andmorphological filtering [12]. The differentiation of the tumour with hemorrhageareas is solved using the texture analysis [14] and classification. The key stepof the identification of the high immunopositive cell concentration (hot-spots) isimplemented in two different ways. One of them is based on the colour represen-tation, whilst the second on the mathematical morphology operations. The finalanalysis of Ki-67 index in manually and automatically selected hot-spot areas isperformed on the full resolution images. The proposed computational approachis validated by the quantitative comparison of both methods.

2.1 Detection of Specimen Map

The first stage of image processing is to define a map of specimen, and possi-ble view field localizations, encompassing the complete tumour. To start with,an image produced by the morphological operation of opening is subjected tobrightness equalization. It is performed using a structuring element shaped as adisc with a large beam (100 pixels). The operation expressed in the form [12]

f = f/ΘfSE (1)


is performed independently for every RGB colour components. Afterwards, differ-encing image B and R components is processed with Otsu thresholding method.Developed algorithm is based on advantage in the brown color component (neg-ative values in the differencing image are set to zero). Morphological operationssuch as: erosion, dilatation, fill holes are conducted on the resulting image. Smallareas are eliminated from final specimen map in order to remove the unnecessarypart of image. This reduces the number of operations and shorten the time of imageprocessing.

2.2 Texture Analysis and Classification for Hemorrhage Exclusion

The next step of image processing is detecting the areas with hemorrhage andthen excluding them from the quantitative analysis. This is an important taskbecause the proliferative Ki-67 index can be significantly overestimated if thehemorrhage areas are processed like hot-spot areas. Thus, the whole tissue struc-tures must be analysed using the description of a texture in the local fields.

Our approach to texture analysis is based on the normalized probabilityapplied to the pixel intensity of the image. In defining texture descriptors weapply the histograms of the sum (s) and difference (d) of images [13,14], whereNΩ is count of pixels in the analyzed neighborhood area for selected texturedescriptor, and μΩ is mean intensity of the neighborhood. These images areformed from the original image by applying the relative translation (it is thesum or difference of the original image and his translation of 3 pixels in thehorizontal axis). We use the modified formulas of Unser features [13] to a localdefinition, which can be expressed in the following form for e.g. homogeneity(sixth feature):

f6 =∑

x∈Ω1

1+d(x)2 /NΩ (2)

An important step is to determine the image resolution and radius resulting inthe best characterization of the local structures in specimens. Also, the resolutionof the objects should enable to represent nuclei of the tumour cells by at leasta few pixels. Too small radius causes a little ambient impact on the propertiesfor a given pixel and a strong heterogeneity. On the other side, too large radiuscauses excessive generalization. We decided to use Unser method due to theireasy implementation in the Matlab environment. In the future research we wanttry to use different methods of quantifying texture.

Another challenge is the choice of colour components for best differentia-tion between tumour and hemorrhage. An original colour representation of theanalyzed images is the RGB. However, as a result of the immunohistochem-ical reaction, one may observe the occurrence of the chromogen not only inimmunopositive cells, but also of changeable quantity in blood cells. This maylead to similar values in the RGB components. Also, we aim to establish thetexture description of the tumour areas uniformly, irrespective of the percentageof immunopositive cells. Using RGB components we observed a significant vari-ability in values of texture characteristics. So, it is desirable to propose such animage colour representation in which the used descriptors will be independent


Fig. 2. The box-and-wiskers plot for areas of tumour (left black boxes) and hemor-rhages (right grey boxes), for colour componnents and three features: mean, energy,and homogeneity.

on percentage of the immunopositive cells. At the same time they should allowfor differentiation of the blood from the tumour areas. We achieved an imagecolour representation, independent on percentage of the immunopositive cellsby introducing the additional colour representation in the form of a sum of u(from CIE Luv colour space) multiplied by 512 and C (CMYK colour space)components. The defined textures were determined for each of the componentsin the RGB, CMYK colour spaces, and for the combined u and C representation.The latter representation unifies the image description in the areas of tumour,and at the same time assumes lower values in the areas of blood. The box-and-wiskers plot presented in the Fig. 2 containes the comparison of feature rangesfor areas of tumour (left black boxes) with for hemorrhages (right grey boxes),for colour componnents and three features: mean, energy, and homogeneity. Itcan be observed that the proposed u and C representation has not worse orbetter ability to differentiate the studied areas of tissue that the other (single)colour components. In this way we have determined 64 features as descriptors ofthe defined patterns. To assess the suitability of individual features in the dif-ferentiation of tumour and hemorrhage, the Fishers linear discriminant for thefeature f assessment was applied [3]. It is defined for recognition of two classesas follows

F12(f) =∣∣∣μ1−μ2σ1+σ2

∣∣∣ (3)

where μ1,μ2 are the average values and σ1, σ2 are the standard deviations of thefeature f , for the first and second class, respectively. The features of high valuesof Fisher measure represent good input attributes for the classifier. To classifythe data into the tumour area and hemorrhage, the Support Vector Machine(SVM) with Gaussian kernel function was selected. The primary advantage ofthis classifier is its good generalization ability resulting from the maximizing theseparation margin between classes during learning.

2.3 Recognition and Counting Immunopositive Cells for Hot-SpotLocalization

The accurate detection of immunopositive cells within the tumour area is a keystep in the algorithm of hot-spot area localization. These fields correspond to


areas with dense immunopositive reactions, so it is important to distinguishbetween immunopositive and immunonegative cells. We assume the higher con-centration of immunopositive cells is directly related to the higher value of Ki-67index.

The first approach presented in this peper is based on thresholding of differ-ential image of B and G colour components. In order to obtain the most adequatemap of immunopositive cell nuclei, values of G component were multiplied by0.85 factor before subtracting both components. Through this operation, onlythe brownish areas of the image receive value greater than zero. The resultingimage is multiplied by the previously obtained specimen mask. The next step isto threshold this image using the threshold value obtained by the Otsu method[10]. The purpose is to remove the components other than immunopositive cells,for example areas representing the colouration of stromal.

The second method is based on the evaluation of the spatial relation of thestained brown objects to their neighboring environment. It is carried out by mor-phological operations performed on the component u of CIE Luv representationof colours after image transformation. This component is strictly associated withthe red colour and it is the best for differentiating the immunopositive cells fromthe remaining part of the image. In order to detect objects with pixel intensitystanding out significantly from the environmental components, the extendedregional minima transformation was applied. The regional minima connectingpixels with a constant intensity value and whose external boundary pixels havea higher value, are detected. The key parameter of the extended regional minimatransformation is the choice of h value, representing the criterion for the mini-mum difference between the intensity of the point in a local minimum and itsclose environment. This value was determined in an experimental way and wasset on the level of 45. It should be noted that this method is more independent ofcolour intensity of immunohistochemical staining. Finally, for the isolated areasrepresenting the tumour cells, a binary mask locating the centers of tumourcells/areas is created. The final image map contains the spatial distribution ofcenters representing the immunopositive tumour cells.

2.4 Localization of Hot-Spot Areas

The localization of areas representing hot-spots is based on finding the local max-ima with the highest density of immunopositive cell nuclei. The image of densitydistribution can be obtained, inter alia, by counting the number of objects ineach window (field of view) or by averaging the filtration performed on the binarymask. In practice, it may happen that only one area of considerable dominanceproliferative Ki-67 index occurs within the analyzed specimen. In such case anautomatic quantitative analysis can lead to the selection of the hot-spot areasonly in this particular region. However, one of the guidelines for quantitativeassessment of tumour scan is to identify the fields for the analysis within multi-ple areas of tumour proliferation. In order to force the selection of the hot-spotsin the entire scanned image, the penalty function is proposed and applied. Itassociates the distance between the designated areas and position of another


candidate for hot-spot. Now, the positions of the hot-spots depend on the valueof the local maxima and their localization in the scan. This operation is realizedby the multiplication of the received mask of specimen by the mask of distances.The penalty term for the candidate centre placed in (x,y) position takes thevalue defined by the following formula

penalty = 1 − ρ∑ 14√(x−xi)2+(y−yi)2 (4)

in which xi, yi represent the coordinates of the neighbouring (existing) centres.The value of ρ was chosen experimentally and equals 0.2. The increase of ρ valueresults in enforcing greater scattering of the designated areas. In the case whereone hot-spot area is dominating, the penalty term allows to determine the hot-spot positions in different localizations of the tumour. As a result, the algorithmdetermines a set of hot-spot localizations representing high immunopositive reac-tion in diverse locations. The final analysis of Ki-67 index in these selections isperformed on full resolution images with the help of the algorithm describedin[5].

In order to compare and verify the proposed procedure of an automatic selec-tion of hot-spot fields, we introduce the procedure marking the area of tumourproliferation. To identify the areas of tumour proliferation, the regional max-ima were extended by neighborhood areas only if these areas meet the followingcriterion:

η ≥ αηmax (5)where ηmax is the local maxima of Ki-67 index, and α is the scaling coefficientdetermined experimentally and equal 0.7 in our experiments. Reduction of αvalue results in an increase of the neighbourhood areas connected to the maxima.This analysis allows determining to what extent the manually and automaticallyselected hot-spot fields represent different regions of tumour proliferation in thespecimen.

3 Results

Fifteen cases of meningiomas and oligodendrogliomas were subject to a quantita-tive analysis. Using the presented algorithm we determined ten hot-spot fields ineach scan and made their diagnostic evaluation by the quantitative analysis. Theaim of this quantitative analysis was to determine the number of immunopos-itive and immunonegative cells, and then to determine the percentage of theimmunopositive cells for each of the areas (Ki-67 index).

In the first step of a WSI processing, a specimen map is determined. Theexemplary result is presented in Fig. 3 with three well outlined tissue scraps.Next, a texture description was calculated for each specimen pixel and its sur-roundings. The applied radius of the neighborhood region Ω was 12 pixels inthe learning process and 11 pixels for the testing purposes (these values gave thebest results in the classification step). The number of features forming the input


vector to the classifier was reduced to 25, each corresponding to Fisher discrim-inant value higher than 0.5. There were 4 features from RGB space, 15 fromCMYK and 6 from combined u and C space. The most of the chosen featuresrepresented the energy, homogeneity and contrast. In learning SVM classifierwe have chosen the regularization coefficient C=150 and sigma of Gaussian ker-nel function equal 4.5. Data used for the learning process were selected by theexpert from different images. It was 18 areas representing hemorrhages, and 32areas representing cells ( areas with tumour and areas with normal cells). Foreach of area, 20 locations of Ω were randomly selected . This gives a collectionof 1080 learning data. Data used to the test process originated from differentformulations.

a) b)

Fig. 3. Specimen area identification in the whole slide image for unprocessed sample(a) and as a binary mask (b).

The aim of the classification process is to determine if the pixel belongsto the normal tissue structure, or to the area with hemorrhages. In this waywe obtain a map of areas containing hemorrhages. An example is presented inFig. 4 with identified specimen map and hemorrhages (Fig. 4b and c). Imposingthe maps of hemorrhage and specimen results in the map of specimen area thatcan be quantitatively analysed.

In the following step, the tumour proliferation areas were identified in WSIby applying the thresholding (algorithm I) or extended regional minima (algo-rithm II) approaches. Both methods were associated with the region growingfor demonstrate which areas are focus of interest (Fig. 5). Significant differencesof the results there are visible. Nevertheless, all selected regions are located in

a) b) c)

Fig. 4. Hemorrhage area detection in original image (a) marked on dark grey on lightgrey specimen map (b) and focused on the detected hemorrhage(c).


a) b)

Fig. 5. Identification of the proliferation areas based on the algorithm I (a) and algo-rithm II (b).

the higher proliteration areas of the specimens. The detailed quantitative anal-ysis better explain the significance of these selections on the assessment of theanalyzed material.

Finally, the selection of 10 hot-spot fields of view was performed. In orderto assess the variability and reproducibility of the obtained results, we askedan independent expert, to perform twice the manual selection of the hot-spotareas for each of the preparations (Series I and Series II) within an interval of 6months. Figure 6 shows the virtual slide with 10 hot-spots marked by an expertin Series I (marked as ∗), 10 hot-spot areas in Series II (marked as �), 10 hot-spot areas identified by the algorithm I (marked as ◦) and 10 hot-spot areasdesignated by the algorithm II (marked as �).

Although there are some differences in locations in each pair of selections, mostof them are in the same areas of specimen. In the next step the values of Ki-67index were calculated in the areas designated by the program, and then comparedwith the average values of two results given by an expert. The recognition and cellcounting was carried out at the resolution of selected areas, compatible with thatused in the algorithm presented in the paper [5]. The quantitative results served

Fig. 6. Virtual slide with the hot-spot areas marked by an expert (Series I-∗, SeriesII-�) and the hot-spots designated by the algorithm I (◦) and algorithm II (�).


Fig. 7. The comparative results of analysis of 15 specimens for which areas of interestwere determined by the medical expert and by the developed algorithms.

to assess the effectiveness of the hot-spot localizations made by our automatic sys-tem. Quantitative analysis also showed the differences in the numerical results ofimage analysis between two implemented algorithms.

Figure 7 presents the comparison of the average values of the Ki-67 indexfor the fields of view selected by an expert (the mean values for Series I andSeries II) and by the developed algorithms I and II. The comparison was donefor fifteen WSI. In three specimens, it was necessary to manually remove somefields selected by developed algorithms due to artifacts (specimen damage). Thedetailed results are included in Table 1. We have observed that:

Table 1. The quantitative analysis of Ki-67 index in the designated fields of viewfor the hot-spot areas selected by the expert (average for series I and II), and by thedeveloped algorithms (algorithm I and algorithm II).

Expert Automatic

Case Series I Series II Average Algorithm I Algorithm II

1 18.39 17.69 18.04 20.45 23.642 56.11 54.54 55.32 57.96 55.033 13.61 13.43 13.52 13.58 14.074 28.47 30.30 29.39 29.56 29.675 27.28 26.91 27.10 27.23 27.996 29.41 34.31 31.86 33.13 34.027 54.71 59.35 57.03 55.03 45.398 60.12 50.48 55.30 43.60 58.189 21.11 19.40 20.26 23.21 23.6410 24.63 24.06 24.35 23.51 24.7111 9.82 9.43 9.62 12.64 14.8812 8.28 9.05 8.66 10.96 11.7313 1.72 1.75 1.74 2.73 3.8314 2.95 3.06 3.01 3.68 5.8515 1.61 1.39 1.50 1.63 2.19

Mean: 23.78 23.93 24.99


– in fourteen cases the values of the Ki-67 index for the fields determinedautomatically by algorithm I were comparable (the maximum difference wasless than 3%) with the expert results,

– in one case the automatic result of algorithm I was 11.7% lower than theexpert results,

– in ten cases the values of the Ki-67 index for the fields determined automat-ically by algorithm II were comparable (the maximum difference was lessthan 3%) with the expert results,

– in four cases the automatic results of algorithm II were 3-6% higher thanthe expert results,

– in one case the automatic result of algorithm II was 11.6% lower than theexpert results.

There are some differences between the average values of Ki-67 index for theareas selected by expert (two series) and by the developed algorithms I andII. The algorithm I appears to give results closer to expert opinion, comparedto algorithm II. However, the algorithm II can more accurately recognize thehot-spot areas with the higher Ki-67 index. The Wilcoxon matched pairs testconfirms that no significant differences exist between the mean of experts resultsand Algorithm I (Z=1.7, p=0.088), but there is a significant difference whencompared with Algorithm II (Z=2.44, p=0.014). Also, it should be noted thatin the case of the analysis of low reaction specimens, fields selected by developedalgorithms had similar or higher reaction than the reaction in fields selected bythe expert. It is clear that in the low reaction specimen, finding hot-spot areasmanually is much more difficult compared to the high reaction cases. Our resultsconfirm the advantage of automatic evaluation over the manual assessment.

4 Conclusions

We have presented the effective method for an automatic localization of thehot-spot areas in meningioma and oligodendrogliomas tumours. Two differentapproaches have been suggested. The algorithm II has shown the advantageover the algorithm I in accurate detection of hot-spot areas in the presence ofstaining artifacts. The use of maps of specimen and the elimination of hem-orrhage areas have reduced the size of an image under analysis and also thecomputational time. The presented methods have good reproducibility (char-acterized by the repeatability of results) which gives them an advantage overthe traditional, manual way of identification of hot-spot areas in meningiomaand oligodendrogliomas. Future research will be include: study different meth-ods of quantifying texture, determine dependence between penalty factor andlocalization of hot spot areas; detection and elimination artefacts on slide.

Acknowledgements. This work has been supported by the National Centre for

Research and Development (PBS2/A9/21/2013 grant), Poland.


References

1. Colman, H., Giannini, C., Huang, L., Gonzalez, J., Hess, K., Bruner, J., Fuller, G.,Langford, L., Pelloski, C., Aaron, J., Burger, P., Aldape, K.: Assessment and prog-nostic significance of mitotic index using the mitosis marker phospho-histone h3 inlow and intermediate-grade infiltrating astrocytomas. Am. J. Surg. Pathol. 30(5),657–664 (2006)

2. Cruz-Roa, A., Gonzalez, F., Galaro, J., Judkins, A., Ellison, D., Baccon, J., Mad-abhushi, A., Romero, E.: A visual latent semantic approach for automatic analysisand interpretation of anaplastic medulloblastoma virtual slides. Med. Image Com-put. Comput. Assist. Interv. 15(Pt 1), 157–164 (2012)

3. Duda, R., Hart, P., Stork, P.: Pattern classification and scene analysis. Wiley,New York (2003)

4. Gavrielides, M.A., Conway, C., OFlaherty, N., Gallas, B.D., Hewitt, S.M.: Observerperformance in the use of digital and optical microscopy for the interpretation oftissue-based biomarkers. Anal. Cell. Pathol. 2014, ID 157308 (2014)

5. Grala, B., Markiewicz, T., Kozlowski, W., Osowski, S., Slodkowska, J., Papierz, W.:New automated image analysis method for the assessment of ki-67 labeling index inmeningiomas. Folia Histo. Cyto. 47(4), 587–592 (2009)

6. Kothari, S., Phan, J., Stokes, T., Wang, M.: Pathology imaging informatics forquan-titative analysis of whole-slide images. J. Am. Med. Inform. Assoc. 20(6),1099–1108 (2013)

7. Lu, H., Papathomas, T.G., van Zessen, D., Palli, I., de Krijger, R.R.,van der Spek, P.J., Dinjens, W.N.M., Stubbs, A.P.: Automated selection of hotspots(ash): enhanced automated segmentation and adaptive step finding for ki67 hotspotdetection in adrenal cortical cancer. Diagn. Pathol. 9, 216 (2014)

8. Molin, J., Devan, K.S., Wardell, K., Lundstrom, C.: Feature-enhancing zoom tofacilitate ki-67 hot spot detection. In: Proc. SPIE 9041, Medical Imaging 2014:Digital Pathology, San Diego, USA, February 15, 2014 (2014)

9. Nakasu, S., Li, D.H., Okabe, H., Nakajima, M., Matsuda, M.: Significance of mib-1 staining indices in meningiomas. comparison of two counting methods. Am. J.Surg. Pathol. 25(4), 472–478 (2001)

10. Otsu, N.: A threshold selection method from gray-level histograms. IEEE Trans.Sys. Man. Cyber. 9(1), 62–66 (1979)

11. Roullier, V., Lezoray, O., Ta, V., Elmoataz, A.: A visual latent semantic approachfor automatic analysis and interpretation of anaplastic medulloblastoma virtualslides. Comput. Med. Imaging Graph. 35(7–8), 603–615 (2011)

12. Soille, P.: Morphological Image Analysis, Principles and Applications, 2nd edn.Springer, Berlin (2003)

13. Unser, M.: Sum and difference histograms for texture classification. IEEE Trans.Pattern Anal. Mach. Intell. 8(1), 118–125 (2013)

14. Wagner, P.: Texture analysis. In: Jahne, B., Haussecker, H., Geissler, P. (Eds.)Handbook of Computer Vision and Applications, vol. 2, chapter 12. AcademicPress, San Diego (1999)

Texture and Mathematical Morphology for Hot-Spot Detection in Whole Slide Images of Meningiomas and Oligodendrogliomas1 Introduction2 Material and Methods2.1 Detection of Specimen Map2.2 Texture Analysis and Classification for Hemorrhage Exclusion2.3 Recognition and Counting Immunopositive Cells for Hot-Spot Localization2.4 Localization of Hot-Spot Areas

3 Results4 ConclusionsReferences

texture and mathematical morphology for hot-spot detection ... · texture and mathematical...

Documents