Improved detection of lung cancer arising in diffuse lung diseases on chest radiographs using temporal subtraction1
out of 8
Post on 04-Jan-2017
Embed Size (px)
<ul><li><p>Improved Detection of Lung Cancer Arisingin Diffuse Lung Diseases on Chest</p><p>Radiographs Using Temporal Subtraction1</p><p>Hiroko Okazaki, MD, Katsumi Nakamura, MD, Hideyuki Watanabe, MD, Yuichi Matsuki, MD, Toshimi Uozumi, MD,Shingo Kakeda, MD, Kouji Kamada, MD, Nobuhiro Oda, RT, Hajime Nakata, MD,</p><p>Shigehiko Katsuragawa, PhD, Kunio Doi, PhD</p><p>Rationale and Objectives. The purpose of this study was to evaluate the usefulness of temporal subtraction for the detec-tion of lung cancer arising in pneumoconiosis, idiopathic pulmonary fibrosis, and pulmonary emphysema.</p><p>Materials and Methods. Fifteen cases of lung cancer arising in diffuse lung diseases, including three cases of pneumoconio-sis, six of idiopathic pulmonary fibrosis, and six of pulmonary emphysema, were evaluated. Pathologic proof was obtained bysurgery or transbronchial lung biopsy. The average interval between previous and current radiographs was 356 days (range, 31947days). All chest radiographs were obtained with a computed radiography system, and temporal subtraction images were pro-duced by subtracting of a previous image from a current one with a nonlinear image-warping technique. The effect of the tem-poral subtraction image was evaluated by observer performance study with receiver operating characteristic analysis.</p><p>Results. The average observer performance with temporal subtraction was significantly improved (Az 0.935) comparedwith that without temporal subtraction (Az 0.857, P .0001).</p><p>Conclusion. The temporal subtraction technique is useful for the detection of lung cancer arising in pneumoconiosis, idio-pathic pulmonary fibrosis, and pulmonary emphysema.</p><p>Key Words. Temporal subtraction image; chest radiograph; diffuse lung diseases; lung cancer; receiver operating charac-teristic curve. AUR, 2004Temporal subtraction is a method in which a previouschest radiograph is subtracted from a current one for en-</p><p>Acad Radiol 2004; 11:498505</p><p>1 From the Department of Radiology, University of Occupational and Envi-ronmental Health, School of Medicine, Japan (H.O., K.N., H.W., Y.M., T.U.,S.K., K.K., N.O., H.N.); the Section of Health Administration, Iwakuni-Ohtake Works, Mitsui Chemicals, Inc, 1-2 Waki 6-chome Waki-cho Kuga-gun Yamaguchi 740-0061 Japan (H.O.); School of Health Sciences, Kum-amoto University, Kumamoto, Japan (S.K.); and Kurt Rossman Laboratoriesfor Radiologic Image Research, Department of Radiology, The Universityof Chicago, Chicago, IL (K.D.). Received September 22, 2003; revision re-quested November 17; revision received December 3; acceptedDecember 8. Address correspondence to H.O. e-mail:Hiroko1.Okazaki@mitsui-chem.co.jp</p><p> AUR, 2004</p><p>doi:10.1016/S1076-6332(03)00820-1</p><p>498hancement of interval changes (13). It has been reportedthat the temporal subtraction technique on chest radiogra-phy is useful for the detection of pulmonary nodules andinterval changes of diffuse infiltrative lung diseases (48). Kakeda et al (7) reported that the temporal subtractiontechnique is useful for improving detection accuracy forperipheral lung nodules on digital chest radiographs.However, in their study, most lung nodes were presentwithout pre-existing lung disease. It seemed necessary toperform an additional study to determine whether tempo-ral subtraction would be useful for detecting nodules onthe background of pre-existing diffuse lung changes. Theincidence of lung malignancy is increased in pneumoconi-</p><p>osis, idiopathic pulmonary fibrosis, and pulmonary em-</p></li><li><p>Academic Radiology, Vol 11, No 5, May 2004 DETECTION OF LUNG CANCER USING TSphysema (914). Early detection of lung cancer is diffi-cult in such cases because opacity caused by lung canceroverlaps pre-existing lung disease.</p><p>In this study, we evaluated the usefulness of temporalsubtraction specifically for the detection of lung cancerarising in patients with pneumoconiosis, idiopathic pul-monary fibrosis, and pulmonary emphysema.</p><p>MATERIAL AND METHODS</p><p>Subjects</p><p>Consecutive cases of lung cancer with pneumoconiosis,idiopathic pulmonary fibrosis, and pulmonary emphysemawere reviewed from a thoracic computed tomography(CT) file between January 1994 and October 1999 at ourinstitution by two board-certified radiologists. We selected15 cases (three pneumoconiosis, six idiopathic pulmonaryfibrosis, and six pulmonary emphysema) that had theirlung cancers proved and had current and previous chestradiographs available. The cases with emphysema hadassociated fibrotic linear or strand opacities involving atleast one third of the lung zone bilaterally. This selectionwas performed by two board-certified radiologists whofirst reviewed only previous and current images, and didnot participate in the subsequent observer study. Thepathologic diagnosis of lung cancer was confirmed bysurgical resection or transbronchial lung biopsy in allcases (10 adenocarcinomas, five squamous cell carcino-mas). The sizes of the lung cancers ranged from 2.06.2cm (average, 3.8 cm) in diameter. All patients were maleranging from 5984 years in age (average, 72.4 years).The time intervals between previous and current radio-graphs ranged from 31947 days (average, 356 days).</p><p>For cases without lung cancer, 15 cases with pre-exist-ing lung diseases similar to the cases with lung cancerwere selected by the same two board-certified radiolo-gists. The possibility of lung cancer was excluded by CTfindings. These cases consisted of 13 male and two fe-male individuals ranging from 5678 years in age (aver-age, 66.9 years). The time intervals between previous andcurrent radiographs ranged from 321,051 days (average,312 days).</p><p>Temporal Subtraction ImageAll patients had chest radiographs taken with a Fuji</p><p>Computed Radiographic (CR) system (Fuji Medical Sys-tems Co Ltd, Tokyo, Japan). All chest radiographs wereexposed at 100 kV with a 10:1 grid, and obtained with a</p><p>model KXO-1250 generator (Toshiba, Tokyo, Japan). Theimaging plate (model ST-V; Fuji Photo Film) size was35 43 cm (matrix size, 1,760 2,140). The pixel sizeand gray level of the CR images were 0.2 mm and a 10-bit gray scale, respectively. The image data of chest ra-diographs were transferred to an O2 workstation (SiliconGraphics, Mountain View, CA).</p><p>Temporal subtraction images were created by use of atemporal subtraction program developed at the Universityof Chicago (Chicago, IL) (1). Previous and current im-ages were subsampled to 586 586 pixels, and the pre-vious image was shifted and rotated to correct for varia-tions in patient positioning between the previous and cur-rent images. After global matching was performed withlow-resolution images, a number of template regions ofinterest (ROIs) (32 32 pixels) and the correspondingsearch area ROIs (64 64 pixels) were selected from theprevious and current images, respectively. Local matchingof a template ROI with a corresponding search area ROIproduced shift values for x and y directions for all pairsof selected ROIs by use of a cross-correlation technique.The previous image was nonlinearly warped according tolocal shift vectors for best matching. For further match-ing, image warping was repeated. Finally, a temporal sub-traction image was created by subtraction of the secondwarped previous image from a current image (1,3). Ittook approximately 15 seconds to produce a temporalsubtraction image from a pair of current and previous CRimages. When multiple pervious images were available,the same two board-certified radiologists determinedbest-match images that showed least misregistrationartifacts with satisfactory background subtraction. Thiswas a subjective judgment reached by consensus. But nocase was rejected because of poor matching.</p><p>Observer Performance StudyTwelve observers, six attending radiologists and six</p><p>radiology residents with 13 years experience, partici-pated in the study. They evaluated the presence or ab-sence of lung cancer on chest radiographs without andwith temporal subtraction images. A continuous ratingscale with a line-marking method (4) was used for record-ing each observers confidence level regarding the pres-ence or absence of lung cancer. The continuous ratingscale was an 8-cm line, the left end of which indicateddefinitely absent and the right end indicated definitelypresent. Current and previous radiographs were shownfirst for conventional interpretation, and the observermarked his or her confidence level by using a black pen-</p><p>cil on the 8-cm line. The temporal subtraction image was</p><p>499</p></li><li><p>OKAZAKI ET AL Academic Radiology, Vol 11, No 5, May 2004then displayed on a CRT monitor. Viewing this imagetogether with the current and previous radiographs, theobserver marked his or her confidence level by using ared pencil on the same line, if this level was differentfrom the first confidence level.</p><p>Before the test, observers were shown 10 trainingcases that were not included in the observer test so theycould learn the rating method as well as the characteris-tics and patterns of artifacts caused by mismatching ofnormal structures with temporal subtraction. This observerstudy was analyzed by receiver operating characteristicanalysis (LABROC5 program by C. Metz, University ofChicago) (15). The index Az, which represents the areaunder the best-fit binomial receiver operating characteris-tic curve when it is plotted in a unit square, was calcu-lated for each fitted curve. The statistical significance ofthe difference between Az values was evaluated with atwo-tailed t test for paired data.</p><p>For additional analysis, we evaluated the cases inwhich the confidence level on the continuous rating scalehad changed markedly. We assumed that temporal sub-</p><p>Figure 1. Receiver operating characteristion of lung cancer arising in diffuse lungtion images.traction had an apparent effect on an observers diagnosis,</p><p>500namely clinical relevance when there was a difference of20% or more in the scores between the first and secondratings. This selection of 20% as a dividing point wasarbitrarily used in the previous study (7). When the sec-ond rating score moved to the correct or wrong directionby more than 20%, we decided that temporal subtractionwas beneficial or detrimental, respectively. The change inclinical relevance was analyzed according to the presenceor absence of lung cancer and the type of diffuse lungdisease.</p><p>RESULTS</p><p>By use of temporal subtraction images, the overall av-erage Az values significantly improved (P .0001), asdid the diagnostic performance. The average Az valuesfor 12 radiologists without and with temporal subtractionimages were 0.857 and 0.935, respectively (Fig 1, Table1). The average Az values increased significantly for bothattending radiologists, from 0.879 to 0.935 (P .028),and radiology residents, from 0.836 to 0.934 (P .001).</p><p>rves including all radiologists for detec-ses with and without temporal subtrac-tic cudiseaIt should be noted that the average Az value for residents</p></li><li><p>Academic Radiology, Vol 11, No 5, May 2004 DETECTION OF LUNG CANCER USING TSwith temporal subtraction images was higher than that forattending radiologists with the CR images alone.</p><p>The numbers of beneficial and detrimental cases af-fected by temporal subtraction images are shown in Fig-ure 2. The number of beneficial cases was much largerthan that of detrimental ones. According to the presence</p><p>Table 1Az Values of ROC Curve for Detecting Lung Cancer Arising inDiffuse Lung Diseases With and Without TemporalSubtraction Images</p><p>ObserverWithout Temporal</p><p>Subtraction ImagesWith Temporal</p><p>Subtraction Images</p><p>Attending radiologistsA 0.912 0.947B 0.866 0.958C 0.776 0.896D 0.953 0.945E 0.923 0.960F 0.841 0.905Average 0.879 0.935*</p><p>Radiology residentsG 0.903 0.958H 0.813 0.914I 0.869 0.934J 0.853 0.953K 0.804 0.918L 0.775 0.926Average 0.836 0.934*</p><p>All radiologistsAverage 0.857 0.935*</p><p>*P .05 (two tailed paired t test).</p><p>Figure 2. Change in clinical relevance af</p><p>and without lung cancer.or absence of lung cancer in beneficial cases, most ob-servers found temporal subtraction to be beneficial incases with lung cancer. The total number of beneficialcases was 72 (20%) among 360 observations (Table 2).Forty-one (22.8%) had lung cancer (Fig 3) and 31(17.2%) were without lung cancer (Fig 4). In cases withlung cancer, they were more frequent in idiopathic pul-monary fibrosis or pulmonary emphysema than in pneu-moconiosis. However, the number of beneficial caseswithout lung cancer was largest for pneumoconiosis.There were only 11 detrimental cases (3.1%) among 360observations; four with lung cancer and seven withoutlung cancer. Among the cases with lung cancer, therewere none with pneumoconiosis, and only a small per-centage with other diseases. Among cases without lungcancer, not more than 5% were detrimental in each dis-ease.</p><p>DISCUSSION</p><p>It is well known that lung cancer increased in inci-dence in patients with idiopathic interstitial pneumonia,asbestosis, pulmonary silicosis, and pulmonary emphy-sema (914). Follow-up study by chest radiograph is nor-mally used and is very important for the detection of lungcancer in these patients. However, early detection of lungcancer is often difficult because pre-existing pulmonarylesions obscure subtle cancer and irregular shapes of can-cer growth on chest radiographs. This is particularly so inthe case of pneumoconiosis and idiopathic pulmonaryfibrosis because of overlying diffuse lung opacities.</p><p>d by temporal subtraction images withfecte501</p></li><li><p>OKAZAKI ET AL Academic Radiology, Vol 11, No 5, May 2004Temporal subtraction is a technique that helps in thedetection of new lesions on serial chest radiographs byenhancing interval changes (1). Improvement in the detec-tion of lung nodules on chest radiographs by temporalsubtraction has been reported (57). We undertook thecurrent study to determine the usefulness of temporal sub-traction for detecting lung cancer arising in the presenceof pre-existing diffuse lung diseases. This technique isexpected to be useful in detection of lung cancer arisingin pre-existing pulmonary lesions because the overlyingopacities, if unchanged, can be eliminated. In this study,the average Az value of both attending radiologists andresidents increased significantly with the use of temporalsubtraction images. The main reason for the increase inAz values was probably the successful removal of obscur-ing diffuse opacities by the temporal subtraction tech-nique. Before this study, we were uncertain about the per-formance of the temporal subtraction technique in diffuselung diseases. However, background diffuse lung changeswere satisfactorily removed in most cases, and the accu-racy of diagnosing developing lung cancer was improved.</p><p>Comparison of experienced with inexperienced radiolo-gists showed that the diagnostic accuracy of the residentsgroup with temporal subtraction images was higher thanthat of the attendings group without these images; there-fore, the temporal subtraction image was especially usefulfor the less-experienced readers.</p><p>As for beneficial and detrimental cases, 20% (72/360)were considered as beneficial. The percentage o...</p></li></ul>
View more >
Restrictive lung diseases Acute and Chronic. Restrictive lung diseases (Diffuse Interstitial Lung Disease) A heterogeneous group of disorders characterized.
Automatic calculation of total lung capacity from automatically traced lung boundaries in postero-anterior and lateral digital chest radiographs
Dr Mat Callister - Cancer Research UK early detection of lung cancer in Leeds Mat Callister Consultant Chest Physician Leeds Teaching Hospitals NHS Trust On behalf of Leeds Lung Cancer Early ...
Effect of an artificial neural network on radiologists' performance in the differential diagnosis of interstitial lung disease using chest radiographs.
Usefulness of Computerized Method for Lung Nodule Detection in Digital Chest Radiographs Using Temporal Subtraction Images
Classification of Lung Tumor Using SVM - of Lung Tumor Using SVM 1Ms. Swati P. Tidke, 2 ... Diagnosis System (CAD) for early detection of lung cancer nodules from the Chest Computer Tomography ...