1.  Computerized Detection of Pulmonary Nodules   Currently, radiologists can fail to detect lung nodules in up to 30% of actually positive cases.  If a computerized scheme could alert the radiologist to locations of suspected nodules, then potentially the number of missed nodules could be reduced.  We developed such a computerized scheme that involves a difference-image approach and various feature-extraction techniques.  Two processed images were obtained from one chest radiograph:  one in which the signal of the nodule was enhanced and the other in which it was suppressed.  Both linear and nonlinear filtering operations were used.  The difference between the two processed images yielded an image of the signal superimposed on a simplified background.  Feature-extraction techniques were then applied to the difference image to distinguish nodules from normal anatomic background patterns.  The computerized detection scheme was used in the evaluation of posteroanterior chest radiographs from 120 clinical cases that included nodules of various subtlety and sizes (5-30 mm).  The presence and location of the nodules were verified by means of computed tomography or follow-up radiography.  The computer program achieved a true-positive detection rate of approximately 75% with an average of approximately one false-positive detection per chest image.  Computer outputs indicating locations of potential lesions were marked by arrows on the chest images, which were displayed on the viewstation. (235, 274, 276, 350, 351, 361, 434, 473, 474, 475, 533, 583, 584, 585, 626, 650, 677, 692, 694, 737, 743, 744, 787, 886)
2.  CAD Scheme for Lung Nodule Detection by Use of a Multiple-Template Matching Technique   A key to the successful clinical application of a CAD scheme is to ensure that there are only a small number of false positives that are incorrectly reported as nodules by the scheme.  In order to significantly reduce the number of false positives in our CAD scheme, we developed a multiple-template matching technique, in which a test candidate can be identified as a false positive and thus eliminated, if its largest cross-correlation value with non-nodule templates was larger than that with nodule templates.  We determined cross-correlation values for test candidates with nodule templates and non-nodule templates, and created a large number of nodule templates and non-nodule templates in order to achieve a good performance.  A large number of false positives (44.3%) were removed with reduction of a very small number of true positives (2.3%) by use of the multiple-template matching technique.  This technique can be used to significantly improve the performance of CAD schemes for lung nodule detection in chest radiographs.  (769, 770)
3.  Effect of a CAD Scheme on Radiologists’ Performance in Detection of Lung Nodules   We evaluated the effect of a CAD scheme on radiologists’ performance in the detection of lung nodules, and also examined a new method of receiver operating characteristic (ROC) analysis.  One hundred and twenty radiographs (60 normal and 60 abnormal with lung nodules of varying subtlety) were used.  Sixteen radiologists (two thoracic, six general, and eight residents) participated in an observer study in which they read chest radiographs (with and without computer output).  The radiologists’ performance was evaluated with ROC analysis with two different methods (independent testing and sequential testing) and a continuous rating scale.  Az (area under the best fit binormal ROC curve) values obtained from ROC analysis with and without CAD output were 0.940 and 0.894, respectively, in the independent test and 0.948 and 0.906, respectively, in the sequential test.  Findings with both methods indicated that the CAD scheme statistically significantly improved diagnostic accuracy, particularly for radiologists with less experience (P<0.001).  Reading time was not increased when CAD was used.  Therefore, the CAD scheme can assist radiologists in the detection of lung nodules on chest radiographs.  (511, 674)
4.  Computerized Analysis of the Likelihood of Malignancy in Solitary Pulmonary Nodules   We developed a CAD scheme by using an artificial neural network (ANN) to assist radiologists in the distinction of benign and malignant pulmonary nodules.  Fifty-six chest radiographs of 34 primary lung cancers and 22 benign nodules were digitized with a 0.175-mm pixel size and a 10-bit grayscale.  Eight subjective image features were evaluated and recorded by radiologists in each case.  A computerized method was developed to extract objective features that could be correlated with the subjective features.  An ANN was used to distinguish benign from malignant nodules on the basis of subjective or objective features.  The performance of the ANN was compared with that of the radiologists by means of ROC analysis.  Performance of the ANN was considerably greater with objective features (Az = 0.854) than with subjective features (Az = 0.761).  Performance of the ANN was also greater than that of the radiologists (Az = 0.752).  The computerized scheme has the potential to improve the diagnostic accuracy of radiologists in the distinction of benign and malignant solitary pulmonary nodules.  (733, 791)
5.  Effect of CAD on Radiologists’ Performance for Distinction between Benign and Malignant Nodules   We evaluated the usefulness of a CAD scheme on radiologists’ performance in the distinction between benign and malignant pulmonary nodules in chest radiographs.  Fifty-three chest radiographs including 31 primary lung cancers and 22 benign nodules were used.  The likelihood measure of malignancy for each nodule was determined by use of an automated computerized scheme.  Sixteen radiologists (nine attendings and seven residents) participated in an observer study in which cases were interpreted first without and then with use of CAD.  The average Az values of radiologists without and with CAD were 0.743 and 0.817, respectively.  The performance of radiologists was improved significantly when CAD was used (P=0.002).  However, the performance (Az=0.889) of the computer alone exceeded these results by a substantial margin.  The average change in radiologists’ confidence level from without to with CAD was highly correlated (R=0.845) with the likelihood measure of malignancy, which was presented as computer output.  Therefore, this type of CAD scheme has the potential to improve the accuracy of radiologists in the classification of benign and malignant solitary pulmonary nodules, probably because the performance of the computer output was very high and particularly superior to that of radiologists.   (832, 887)
6.  Computerized Detection and Characterization of Interstitial Disease   Evaluation of diffuse interstitial disease in chest radiographs is one of the most difficult problems in diagnostic radiology.  This difficulty is due to (1) the numerous patterns and complex variations that are involved, (2) the lack of firmly established correlation between radiologic and pathologic findings, and (3) variations among radiologists in terms that they use to describe radiographic patterns, which are not defined objectively.  We developed an automated method for determining physical measures of lung textures in digital chest radiographs to detect and characterize interstitial lung disease.  Multiple regions of interest were selected automatically (using a gradient-distribution technique) and analyzed throughout the lung region.  With our method, the underlying background density variations caused by the gross lung and chest wall anatomy were corrected for which isolated the fluctuating patterns of the underlying lung texture for subsequent computer analysis.  The power spectrum of lung texture, obtained from the two-dimensional Fourier transform, was filtered by the visual system response of the human observer.  The magnitude and coarseness (or fineness) of the lung textures were then quantified by the root-mean-square (rms) variation and the first moment of the power spectrum, respectively.  Results indicated that the rms variations and/or the first moments of the texture of abnormal lungs with various interstitial diseases were clearly different from those of normal lungs.  Computer outputs were marked on the chest image using symbols that indicated the severity and pattern type of the infiltrates.  (241, 250, 257, 285, 286, 368, 462, 566, 844)
7.  Application of Artificial Neural Networks for Computerized Detection of Interstitial Lung Disease   We developed a CAD scheme by using ANNs on quantitative analysis of image data.  Three separate ANNs were applied for detection of interstitial disease on digitized chest images.  The first ANN was trained with horizontal profiles in regions of interest (ROIs) selected from normal and abnormal chest radiographs for distinguishing between normal and abnormal patterns.  For training and testing of the second ANN, the vertical output patterns obtained from the first ANN were used for each ROI.  The output value of the second ANN was used to distinguish between normal and abnormal ROIs with interstitial infiltrates.  If the ratio of the number of abnormal ROIs to the total number of all ROIs in a chest image was greater than a specified threshold level, the image was classified as abnormal.  In addition, the third ANN was applied to distinguish between normal and abnormal chest images.  The combination of the rule-base method and the third ANN also was applied to the classification between normal and abnormal chest images.  The performance of the ANNs was evaluated by means of ROC analysis.  The average Az value for distinguishing between normal and abnormal cases was 0.976 ± 0.012 for 100 chest radiographs that were not used in training of ANNs.  The results indicated that the ANN trained with image data can learn some statistical properties associated with interstitial infiltrates in chest radiographs.  (604)
8.  Effect of CAD Scheme on Radiologists’ Detection of Interstitial Infiltrates   We evaluated the impact of a CAD scheme on radiologists’ interpretation of chest radiographs with interstitial infiltrates by using ROC analysis.  Twenty normal and 20 abnormal chest radiographs were used.  Each radiograph was divided into 4 quadrants.  A total of 129 quadrants (80 normal and 49 abnormal quadrants) were used for testing by excluding 31 equivocal quadrants.  Sixteen observers (ten residents and six attending radiologists) participated in this study.  The radiologists’ performance with, and without the computer output, which indicates the normal and abnormal cases by various markers, was evaluated by ROC analysis.  Az values obtained with and without the CAD output were 0.970 and 0.949 (P=0.0002), respectively, for all observers; 0.969 and 0.943 (P=0.006), respectively, for the residents’ subgroup; 0.972 and 0.960 (P=0.162), respectively, for the attending radiologists’ subgroup.  Az value for the computerized scheme by itself was 0.943.  Thus, the CAD scheme can assist radiologists in the diagnosis or exclusion of interstitial diseases on chest radiographs.  (460, 613)
9.  Neural Network Approach for Differential Diagnosis of Interstitial Lung Diseases   A neural network approach was applied for the differential diagnosis of interstitial lung diseases.  The ANNs were designed to distinguish between 11 interstitial lung diseases on the basis of 10 clinical parameters and 16 radiologic findings extracted by chest radiologists.  Thus, the ANNs consisted of 26 input units and 11 output units.  One hundred fifty actual clinical cases, 110 cases from previously published articles, and 110 hypothetical cases were used for training and testing the ANNs by using a round-robin (or leave-one-out) technique.  ANN performance was evaluated with ROC analysis.  The Az obtained with actual clinical cases was 0.947, and both the sensitivity and specificity of the ANNs were approximately 90% in terms of indicating the correct diagnosis with the two largest output values among the 11 diseases.  (265, 266, 844, 897)
10.  Effect of ANN on Radiologists’ Performance in Differential Diagnosis of Interstitial Lung Disease   The aim of this study was to evaluate the effect of the output from the artificial neural network on radiologists’ diagnostic accuracy.  Thirty-three clinical cases (three cases for each lung disease) were selected.  Chest radiographs were viewed by eight radiologists (four attending physicians and four residents) with and without network output, which indicated the likelihood of each of the 11 possible diagnoses in each case.  The radiologists’ performance in distinguishing among the 11 interstitial lung diseases was evaluated by ROC analysis with a continuous rating scale.  When chest radiographs were viewed in conjunction with network output, a statistically significant improvement in diagnostic accuracy was achieved (P<0.001).  The Az value was .826 without network output and .911 with network output.  An artificial neural network can provide a useful “second opinion” to assist radiologists in the differential diagnosis of interstitial lung disease using chest radiographs.  (633, 634, 790, 844, 897)
11.  Computerized Automated Analysis of Heart Sizes   Cardiac size is an important and useful diagnostic parameter in chest radiographs.  We developed an automated method for determining a number of parameters related to the size and shape of the heart and of the lungs in chest radiographs.  To obtain standard patterns of the cardiac shadow as “gold standards,” four radiologists traced their best estimates of the entire contour of the heart, including the largely invisible inferior margin, on 11 radiographs.  These contours were analyzed by Fourier transform, and the results were used as a guide to obtain a shift-variant cosine function which was applied to predict the cardiac contour by fitting a limited number of detected heart boundary points.  These points were obtained from analysis of edge gradients in two orthogonal directions.  A simple observer study indicated that the contours of the heart shadows computed for 60 chest radiographs were generally acceptable to radiologists for estimation of the size and area of the projected heart.  We also detected the edges of the rib cage and the diaphragm, which enabled us to determine the projected thoracic area.  From these results, we calculated the cardiothoracic ratio and other parameters, such as the ratio of the projected heart area to the projected thoracic area.  Using this information, the probability of cardiomegaly was calculated. (293, 325)
12.  Computerized Detection of Pneumothorax   To aid radiologists in the diagnosis of pneumothorax from chest radiographs, an automated method for detection of subtle pneumothorax was developed.  The computerized method was based on the detection of a fine curved-line pattern, which is a unique feature of radiographic findings of pneumothorax.  Initially, regions of interest (ROIs) were determined in each upper lung area, where subtle pneumothoraces commonly appear.  The pneumothorax pattern was enhanced by the selection of edge gradients within a limited range of orientations.  Rib edges included in this edge-enhanced image were removed, based on the locations of posterior ribs, which were determined separately.  A subtle curved line due to pneumothorax was then detected by means of the Hough transform.  The detected pneumothorax pattern was marked on the chest image displayed on a CRT monitor.  With the present computer method applied to 50 chest images (28 normals and 22 abnormals with pneumothorax), we were able to detect 77% of pneumothoraces, with 0.44 false-positives per image.  (329, 357)
13.  Temporal Subtraction of Sequential Chest Radiographs   Detection of abnormalities in chest radiographs is limited by the complex and variable nature of the normal background anatomical structures.  In the case of patients who have had a previous chest radiograph, an opportunity exists to enhance areas of interval change selectively, including regions with new or altered pathology, by using the previous radiograph as a subtraction mask.  To perform temporal subtraction, conventional film-screen radiographs were digitized.  The current and previous images of a given patient were subjected to density and contrast normalization.  Using previously developed algorithms, the midline and ribcage outlines were identified.  A non-linear warping technique was applied to the previous image to correct for differences in projection, and patient positioning. The temporal subtraction image was obtained by the difference between the current image and the warped previous image.  The subtraction images were able to enhance various interval changes such as differences in the size of tumor masses, changes in heart size, and changes in pulmonary infiltrates or pleural effusions.  (415)
14.  Iterative Image Warping Technique for Temporal Subtraction to Detect Interval Change   Although the overall performance of the current temporal subtraction technique was relatively good, severe misregistration errors, mainly due to AP inclination and/or rotation, were observed in some cases.  In order to reduce these errors, we attempted to improve the subtraction scheme by applying an iterative image warping technique.  In cases obtained with the new temporal subtraction technique, 177 (97.8%) of 181 showed adequate, good, or excellent quality.  We also found that 156 (86.2%) of cases obtained with the new scheme showed improvements in the quality of the subtraction images compared with the previous scheme.  The results indicated that the performance of the temporal subtraction technique was greatly improved by use of the iterative image warping technique.  (664, 665, 673, 724)
15.  Effect of Temporal Subtraction Images on Radiologists’ Detection Accuracy   An observer test was performed to evaluate the effects of the temporal subtraction image technique on detection of interval change.  Fifty pairs of chest radiographs, including a baseline examination and a subsequent radiograph, were selected (25 cases in which potentially important new abnormalities had developed, and 25 in which there was no interval change).  The baseline examination was chosen from multiple prior radiographs to minimize initial misregistration.  By means of ROC analysis, the ability of 11 observers to detect pathologic change when viewing the paired digitized baseline and subsequent radiographs was compared with their ability when viewing the same paired radiographs together with temporal subtraction images.  Positive cases demonstrated focal new abnormalities that were greater than 1 cm in diameter.  The Az value increased from 0.89 without to 0.98 with the temporal subtraction images.  When the paired digitized previous and current chest radiographs were viewed in conjunction with temporal subtraction images, a significant improvement in detection of new abnormalities was achieved (P=0.00004), whereas the mean interpretation time was reduced by 19.3% (from 52 to 42 seconds, including the time to record the score and to move to next case) (P=0.0019).  The temporal subtraction technique can significantly improve sensitivity and specificity for detection of interval change in chest radiographs.  (552, 784, 814, 815, 818, 819)
16.  Contralateral Subtraction: A Novel Technique for Detection of Asymmetric Abnormalities   A novel contralateral subtraction technique was developed to assist radiologists in the detection of asymmetric abnormalities on a single chest radiograph.  With this method, the lateral inclination was first corrected by rotating and shifting the original chest image so that the midline of the thorax was aligned with the vertical centerline of the original chest image.  The rotated image was then flipped laterally to produce a reversed “mirror” image.  Finally, the mirror image was warped and subtracted from the original image for derivation of the contralateral subtraction image.  The three key techniques employed in this study were applied successively to the initial contralateral subtraction technique for acquisition of improved subtraction images.  One hundred PA chest radiographs, including 50 normals and 50 abnormals, were used as the database for this study.  The percentage of chest images, which were rated as being adequate, good, or excellent quality of subtraction images by employing a subjective evaluation method, was improved from 73% to 91% by use of the three key techniques.  The contralateral subtraction technique can be used for detection of any asymmetric abnormalities, such as lung nodules, pneumothorax, pneumonia, and emphysema, on a single chest radiograph, and therefore has potential utility in a high proportion of abnormal cases.  (722, 725)
17.  Potential Usefulness of a New Contralateral Subtraction Technique   To evaluate the potential usefulness of a contralateral subtraction technique in the detection of subtle lung nodules on chest radiograph, 50 chest radiographs (25 normal and 25 abnormal with a subtle lung nodule) were digitized with a 0.175-mm pixel size and 4,096 gray levels.  Twelve radiologists (10 attending and two residents) participated in observer tests and read both original and contralateral subtraction images with a sequential testing method.  Radiologists’ performance was evaluated by means of ROC analysis with use of a continuous rating scale.  The beneficial and detrimental effects of the contralateral subtraction technique on the radiologists’ performance were also evaluated.  Az values obtained without and with contralateral subtraction images were 0.926 and 0.962, respectively.  Results indicated that the contralateral subtraction images significantly (P <0.05) improved diagnostic accuracy, particularly for radiologists with limited experience.  The contralateral subtraction technique can assist radiologists in the correct identification of subtle lung nodules on chest radiographs.  (833)
18.  Automated Registration of Ventilation/Perfusion Images with Chest Radiographs   Many clinical situations require radiologists to visually correlate the functional information provided by radionuclide images with the anatomic detail of radiographs.  This is complicated by differences in size, resolution, and the format of images produced by these two modalities.  To assist radiologists in interpreting images obtained from these modalities, we developed an automated technique for registering radionuclide lung scan images with digital chest radiographs.  Threshold analysis was used to construct contours around the high activity regions of the ventilation/perfusion images.  Analogous contours were constructed around the aerated lung regions of the radiograph.  The contour dimensions and location of anatomic landmarks were then used to superimpose the radiographic image with the perfusion and ventilation images.  (544, 696)
19.  Automated Analysis of Costophrenic Angles in PA Chest Radiographs   We developed an automated method to identify abnormal blunting of the costophrenic (CP) angles due to pleural effusion, for example. CP angle location was identified during an automated method to segment the aerated lung fields from digital PA chest radiographs based on iterative global gray-level thresholding and local gray-level thresholding procedures. Initial costal and diaphragmatic margin points were then identified. These two sets of points were independently fit with parabolas, the convergence of which defines the angle measure used to identify an abnormality. ROC analysis was used to evaluate the ability of the scheme to detect abnormal CP angles. When applied to CP angles from the 600-image database, Az attained a value of 0.83.  (588)
20.  Computerized Analysis of Abnormal Asymmetry in PA Chest Radiographs   We developed a computerized method for the fully automated analysis of digital PA chest radiographs for abnormal asymmetry, which is present when pathology in one lung affects lung volume to such an extent that the projected area of aerated lung is considerably impacted. An automated lung segmentation method was used to identify the aerated lung regions in 600 chest radiographs. Minimal a priori lung morphology information was required for this gray-level thresholding-based segmentation scheme. Consequently, this approach to segmentation is applicable to grossly abnormal cases. The relative areas of segmented right and left lung regions in each image were compared with the corresponding area distributions of normal images to determine the presence of abnormal asymmetry.  Results were compared with image ratings by a radiologist. The ability of the automated method to distinguish normal from asymmetrically abnormal cases was evaluated by using ROC analysis. The performance of the method attained an Az value of 0.84. (403, 628)
21.  Automated Patient Recognition Method Based on an Image-Matching Technique   An automated patient recognition method for correcting “wrong” chest radiographs being stored in a picture archiving and communication system (PACS) environment was developed.  The method was based on an image-matching technique with previous chest radiographs.  For identification of a “wrong” patient, the correlation value was determined for a previous image of a patient and a new, current image of the presumed corresponding patient.  The correlation values between the current and previous images for the same, “correct” patients were generally greater than those for different, “wrong” patients.  Although the two histograms for the same patient and for different patients overlapped at correlation values greater than 0.80, most parts of the histograms were separated.  If the current image was considered potentially to belong to a “wrong” patient, then a warning sign with the probability for a “wrong” patient was provided to alert radiology personnel.  Our results indicated that at least half of the “wrong” images in our database could be identified correctly with this method.  (773, 820, 933, 934)
22.  Computerized Quantification of Pleural Disease   The development of novel treatment regimens for malignant pleural mesothelioma is an active area of research. Accurate quantification of mesothelioma tumor extent is essential to the conduct of clinical trials that seek to investigate these regimens. Although manual measurement of tumor thickness on computed tomography (CT) scans is the current standard for assessing response to treatment, this approach is a tedious and often incomplete process. We have developed a computer-assisted measurement method to quantify the extent of mesothelioma on CT scans in an efficient, reproducible manner. Gray-level thresholding techniques are used to segment automatically the lung regions in each section of a CT scan. For sections demonstrating mesothelioma, the boundaries of the lung segmentation regions represent the inner margin of the tumor. To identify the outer tumor margin at sites within a section for which tumor thickness measurements are desired, points along the chest wall or mediastinum are selected through a user interface. Six algorithms have been developed to automatically connect user-selected points with the most appropriate points along the boundary of the lung segmentation regions, with the distance between corresponding points along the outer and inner tumor margins representing the tumor thickness at that measurement site. These methods have been applied to 134 measurement sites in the CT scans of 22 mesothelioma patients. The computer-assisted measurement technique achieved a correlation coefficient of 0.98 with the average manual measurements of five clinical observers over 134 measurement sites. Paired t-tests for the difference in means showed that computer-generated measurements were not statistically significantly different from the average manual measurements (P>0.31), indicating that this method may yield results that are consistent with those of humans. Three of these methods have been validated through a study that allows for human interaction with the computer results to determine the extent to which, in the clinical setting, an observer would accept or modify the computer-assisted measurements. We expect these methods to become important components of mesothelioma treatment protocols by making the quantification of tumor extent more efficient, reproducible, and consistent.
23.  Automated Detection of Pathologic Change in Temporal Subtraction Images of Chest Radiographs   We have developed a computerized method for the automated detection of change in pulmonary pathology as demonstrated in temporal subtraction images created from radiographic chest image pairs. The lung regions, as segmented during the performance of temporal subtraction, are extracted from the subtraction image. A gray-level histogram is obtained from the pixels within the lung regions. The mode of this histogram is identified, and a gray-level threshold is established at a predefined fraction of this modal value. All pixels with gray levels less than this threshold that lie within the lung regions of the temporal subtraction image remain "on," while all other pixels are set to zero. Area and circularity requirements are imposed on the regions that remain to eliminate false-positive regions. Areas of pathologic change identified in this manner may be presented as outlines in the subtraction image or as highlighted regions in the original radiographic image so that, in effect, temporal subtraction becomes a "background" process for computer-aided diagnosis. As a pilot study, this method was applied to a preliminary database of 12 temporal subtraction images. Six of these images demonstrated no pathologic change between the constituent radiographic images and were considered normal. The other six temporal subtraction images were considered positive for change by an experienced chest radiologist and demonstrated a range of pathology including pleural effusion, interstitial disease, and lung cancer. The method correctly identified six of the eight foci of pathologic change (75%) in the six positive cases and generated no false positives in any of the 12 subtraction images. Such a computerized method may help radiologists assimilate the results of temporal subtraction in an intuitive way and aid in the identification of disease development and progression.
24.  Improved Chest Radiographs with Rib Suppression by Means of Massive Training Artificial Neural Network (MTANN)   We developed a novel image-processing technique for suppressing the contrast of ribs and clavicles in chest radiographs by means of an MTANN. The MTANN is a highly nonlinear filter that can be trained by use of input chest radiographs and the corresponding teacher images. The bone image obtained by use of a dual-energy subtraction technique was used as the teacher image for suppression of ribs and clavicles in chest radiographs. After training, the MTANN was able to produce a "bone-image-like" image, which was then subtracted from the corresponding chest radiograph. Thus our scheme can produce a "soft-tissue-image-like" image where ribs and clavicles were substantially suppressed; therefore, this image processing may be considered as a "rib suppression" technique. The database used in this study consisted of 137 conventional chest radiographs with solitary pulmonary nodules, which were acquired with conventional radiography systems. All nodules were confirmed by use of CT examinations. Most nodules were overlapped with ribs or clavicles. When the trained MTANN was applied to non-training chest radiographs, the ribs and clavicles in chest radiographs were suppressed substantially, while the visibility of nodules and lung vessels was maintained. The effect of the rib suppression was evaluated by measuring the contrast of ribs in both original chest radiographs and processed images. The results demonstrated that the contrast of ribs in chest radiographs almost disappeared, and was reduced to about 8% in the processed images, while the contrast of nodules was comparable in chest radiographs and processed images. Therefore, a new image-processing technique for rib suppression using the MTANN would be useful for radiologists as well as CAD schemes in detecting lung nodules in chest radiographs. (946, 968, 969)
25. False-Positive Reduction in CAD Scheme for Detection of Nodules on Chest Radiographs by Means of MTANN   We developed a technique that uses a multiple massive training artificial neural network (multi-MTANN) to reduce false positives in a computer-aided diagnostic (CAD) scheme for nodule detection on chest radiographs. Our database consisted of 91 solitary pulmonary nodules including 64 malignant nodules and 27 benign nodules in 91 chest radiographs. With our current CAD scheme based on a difference-image technique and linear discriminant analysis, a sensitivity of 82.4% with 4.5 false positives per image was achieved. We developed the multi-MTANN for further reduction of the false positives. An MTANN is a highly nonlinear filter that can be trained with input images and the corresponding teaching images. For reducing the effects of the background levels in chest radiographs, a background-trend-correction technique followed by contrast normalization was performed on the input images to the MTANN. For enhancement of nodules, the teaching image was designed to contain the distribution for a "likelihood of being a nodule." Six MTANNs in the multi-MTANN were trained by use of typical nodules and six different types of non-nodules (false positives). By use of the trained multi-MTANN, 68.3% of false positives were removed with a reduction of one true positive. The false-positive rate of our original CAD scheme was improved from 4.5 to 1.4 false positives per image at an overall sensitivity of 81.3%. By use of a multi-MTANN, the false-positive rate of our CAD scheme for lung nodule detection on chest radiographs was improved substantially, while a high sensitivity was maintained. (970)