Phương pháp chẩn đoán hình ảnh (Phần 2)

Phương pháp chẩn đoán hình ảnh (Phần 2)

2089_book.fm Page Tuesday, May 10, 2005 3:38 PM Computer-Aided Diagnosis of Breast Cancer Heang-Ping Chan, Berkman Sahiner, Nicholas Petrick, Lubomir Hadjiiski, and Sophie Paquerault CONTENTS 1.1 1.2 1.3 1.4 Introduction Computerized Detection of Microcalcifications 1.2.1 Methods 1.2.1.1 Preprocessing Technique 1.2.1.2 Microcalcification Segmentation 1.2.1.3 Rule-Based False-Positive Reduction 1.2.1.4 False-Positive Reduction Using Convolution Neural Network Classifier 1.2.1.5 False-Positive Reduction Using Clustering 1.2.2 FROC Analysis of Detection Accuracy 1.2.3 Effects of Computer-Aided Detection on Radiologists’ Performance Computerized Detection of Masses 1.3.1 Methods 1.3.1.1 Preprocessing and Segmentation 1.3.1.2 Object Refinement 1.3.1.3 Feature Extraction and Classification 1.3.2 FROC Analysis of Detection Accuracy 1.3.2.1 Data Sets 1.3.2.2 True Positive and False Positive 1.3.2.3 Training and Testing 1.3.2.4 Performance of Mass Detection Algorithm Mass Detection with Two-View Information 1.4.1 Methods 1.4.1.1 Geometrical Modeling 1.4.1.2 One-View Analysis 1.4.1.3 Two-View Analysis 1.4.1.4 Fusion Analysis Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page Tuesday, May 10, 2005 3:38 PM Medical Image Analysis 1.4.2 Results 1.4.2.1 Geometrical Modeling 1.4.2.2 Comparison of One-View and Two-View Analysis 1.5 Summary Acknowledgment References 1.1 INTRODUCTION Mammography is currently the only proven and cost-effective method to detect early breast cancer A mammographic examination generally contains four images, two views for each breast One is a craniocaudal (CC) view, and the other is a mediolateral oblique (MLO) view These two views are designed to include most of the breast tissues within the X-ray images Mammographic interpretation can be considered a two-step process A radiologist first screens the mammograms for abnormalities If a suspicious abnormality is detected, further diagnostic workup is then performed to estimate the likelihood that the abnormality is malignant Diagnostic workup might include mammograms of additional views such as lateromedial (LM) or exaggerated craniocaudal (XCC) views, magnification views, spot views, as well as ultrasound scanning of the suspicious area The main mammographic signs of breast cancer are clustered microcalcifications and masses Microcalcifications are calcium deposits in the breast tissue manifested as clusters of white specks of sizes from about 0.05 mm to 0.5 mm in diameter Masses have X-ray absorption similar to that of fibroglandular tissue and are manifested as focal low-optical-density regions on mammograms Some benign breast diseases also cause the formation of clustered microcalcifications and masses in the breast The mammographic features of the malignant microcalcifications or masses are nonspecific and have a large overlap with those from benign diseases Because of the nonspecific features of malignant lesions, mammographic interpretation is a very challenging task for radiologists Studies indicate that the sensitivity of breast cancer detection on mammograms is only about 70 to 90% [1–6] In a study that retrospectively reviewed prior mammograms taken of breast cancer patients before the exam in which the cancer was detected, it was found that 67% (286/427) of the cancers were visible on the prior mammograms and about 26% (112/427) were considered actionable by radiologists [7] Missed cancers can be caused by detection errors or characterization errors Detection errors can be attributed to factors such as oversight or camouflaging of the lesions by overlapping tissues Even if a lesion is detected, the radiologist may underestimate the likelihood of malignancy of the lesion so that no action is taken This corresponds to a characterization error On the other hand, the radiologist may overestimate the likelihood of malignancy and recommend benign lesions for biopsy It has been reported that of the lesions that radiologists recommended for biopsy, only about 15 to 30% are actually malignant [8] The large number of benign biopsies not only causes patient anxiety, but also increases health-care costs In addition, the scar tissue resulting from biopsy often makes it more difficult to interpret the patient’s Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page Tuesday, May 10, 2005 3:38 PM Computer-Aided Diagnosis of Breast Cancer mammograms in the future The sensitivity and specificity of mammography for detecting a lesion and differentiating the lesion as malignant or benign will need to be improved It can be expected that early diagnosis and treatment will further improve the chance of survival for breast cancer patients [9–12] Various methods are being developed to improve the sensitivity and specificity of breast cancer detection [13] Double reading can reduce the miss rate of radiographic reading [14, 15] However, double reading by radiologists is costly Computer-aided detection (CAD) is considered to be one of the promising approaches that may improve the efficacy of mammography [16, 17] Computer-aided lesion detection can be used during screening to reduce oversight of suspicious lesions that warrant further diagnostic workup Computer-aided lesion characterization can also be used during workup to provide additional information for making biopsy recommendation It has been shown that CAD can improve radiologists’ detection accuracy significantly [18–23] Receiver operating characteristic (ROC) studies [24, 25] showed that computer-aided characterization of lesions can improve radiologists’ ability in differentiating malignant and benign masses or microcalcifications CAD is thus a viable cost-effective alternative to double reading by radiologists The promise of CAD has stimulated research efforts in this area Many computer vision techniques have been developed in various areas of CAD for mammography Examples of work include: detection of microcalcifications [18, 26–38], characterization of microcalcifications [39–49], detection of masses [19, 40, 50–73], and characterization of masses [24, 74–78] Computerized classification of mammographic lesions using radiologist-extracted features has also been reported by a number of investigators [79–84] There are similarities and differences among the computer vision techniques used by researchers However, it is difficult to compare the performance of different detection programs because the performance strongly depends on the data set used for testing These studies generally indicate that an effective CAD system can be developed using properly designed computer vision techniques Efforts to evaluate the usefulness of CAD in reducing missed cancers are ongoing Results of a prospective study by Nishikawa et al [85] indicated that their CAD algorithms can detect 54% (9/16) of breast cancer in the prior year with four false positives (FPs) per image when the mammograms were called negative but the cancer was visible in retrospect In our recent study of detection on independent prior films [86], we found that 74% (20/27) of the malignant masses and 57% (4/7) of the malignant microcalcifications were detected with 2.2 mass marks/image and 0.8 cluster marks/image by our computer programs A commercial system also reported a sensitivity of 77% (88/115) in one study [7] and 61% (14/23) in another study [87] for detection of the cancers in the prior years that were considered actionable in retrospect by expert mammographers A prospective study of 12,860 patients in a community breast cancer center with a commercial CAD system that had about one mark per image reported a cancer detection rate of 81.6% (40/49), with eight of the cancers initially detected by computer only This corresponded to a 20% increase in the number of cancers detected (41 vs 49) when radiologists used CAD Similar gain in cancer detection has been observed in a premarket retrospective study of another commercial system [23] Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page Tuesday, May 10, 2005 3:38 PM Medical Image Analysis These results demonstrate that, even if a CAD system does not detect all cancers present and has some FPs, it can still reduce the missed cancer rate when used as a second opinion by radiologists This is consistent with the first laboratory ROC study in CAD reported by us in 1990 [18], which demonstrated that a CAD program with a sensitivity of 87% and an FP rate of 0.5 to per image could significantly improve radiologists' accuracy in detection of subtle microcalcifications In a recent prospective pilot clinical trial [88] of a CAD system developed by our group, a total of 11 cancers were detected in a screening patient cohort of about 2600 patients The radiologists detected 10 of the 11 cancers without our CAD system The CAD system also detected 10 of the 11 cancers However, one of the computer-detected cancers was different from those detected by the radiologists, and this additional cancer was diagnosed when the radiologist was alerted to the site by the CAD system In a 1-year follow-up of the cases, it was found that five more cancers were diagnosed in the patient cohort Our computer system marked two of the five cancers, although all five cancers were deemed not actionable in the year of the pilot study when the mammograms were reviewed retrospectively by an experienced radiologist For classification of malignant and benign masses, our ROC study [24] indicated that a classifier with an area under the ROC curve, Az, of 0.92 could significantly improve radiologists' classification accuracy with a predicted increase in the positive predictive value of biopsy Jiang et al [25] also found in an ROC study that their classifier with an Az of 0.80 could significantly improve radiologists' characterization of malignant and benign microcalcifications, with a predicted reduction in biopsies Recently, Hadjiiski et al [89, 90] performed an ROC study to evaluate the effects of a classifier based on interval-change analysis on radiologists’ classification accuracy of masses in serial mammograms They found that when the radiologists took into account the rating of the computer classifier, they reduced the biopsy recommendation of the benign masses in the data set while slightly increasing the biopsy recommendation of the malignant masses This result indicated that CAD improved radiologists’ accuracy in classifying malignant and benign masses based on serial mammograms and has the potential of reducing unnecessary biopsy In the last few years, full-field digital mammography (FFDM) technology has advanced rapidly because of the potential of digital imaging to improve breast cancer detection Four manufacturers have obtained clearance from the Food and Drug Administration (FDA) for clinical use It is expected that digital mammography detectors will provide higher signal-to-noise ratio (SNR) and detective quantum efficiency (DQE), wider dynamic range, and higher contrast sensitivity than digitized film mammograms Because of the higher SNR and linear response of digital detectors, there is a strong potential that more effective feature-extraction techniques can be designed to optimally extract signal content from the direct digital images and improve the accuracy of CAD The potential of improving CAD accuracy by exploiting the imaging properties of digital mammography is a subject of ongoing research In mammographic screening, it has been reported that taking two views of each breast, a CC and an MLO view, provides a higher sensitivity and specificity than one view for breast cancer detection [2, 91–93] Radiologists use the two views to Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page Tuesday, May 10, 2005 3:38 PM Computer-Aided Diagnosis of Breast Cancer confirm true positives (TPs) and to reduce FPs Current CAD algorithms detect lesions only on a single mammographic view New CAD algorithms that utilize the correlation of computer-detected lesions between the two views are being developed [69, 94–99] Our studies demonstrated that the correlated lesion information from two views could be used to reduce FPs and improve detection [100, 101] Although the development is still at the early stage and continued effort is needed to further improve the two-view correlation techniques, this promising development will be summarized here in the hope that it will stimulate research interests Another important technique that radiologists use in mammographic interpretation is to compare the current and prior mammograms and to evaluate the interval changes Interval-change analysis can be used to detect newly developed abnormality or to evaluate growth of existing lesions Hadjiiski et al [97, 98] developed a regional-registration technique to automatically identify the location of a corresponding lesion on the same view of a prior mammogram Feature extraction and classification techniques could then be developed to differentiate malignant and benign lesions using interval-change information Interval-change features were found to be useful in improving the classification accuracy In this chapter, we will concentrate on lesion detection, rather than characterization Computer vision methods for classification of malignant and benign lesions, including interval-change analysis, can be found in the literature [89, 90, 97, 98] 1.2 COMPUTERIZED DETECTION OF MICROCALCIFICATIONS Clustered microcalcifications are seen on mammograms in 30 to 50% of breast cancers [102–106] Because of the small sizes of microcalcifications and the relatively noisy mammographic background, subtle microcalcifications can be missed by radiologists Computerized methods for detection of microcalcifications have been developed by a number of investigators Chan et al [18, 26, 27] designed a difference-image technique to detect microcalcifications on digitized mammograms and to extract these features to distinguish true and false microcalcifications A convolution neural network was developed to further recognize true and false patterns [28] Wu et al [107] used the difference-image technique [26] for prescreening of microcalcification sites, and then classified their power-spectra features with an artificial neural network to differentiate true and false microcalcifications Zhang et al [36] further modified the detection system by using a shift-invariant neural network to reduce false-positive microcalcifications Fam et al [108] and Davies et al [29] detected microcalcifications using conventional image processing techniques Qian et al [30] developed a tree-structure filter and wavelet transform for enhancement of microcalcifications Other investigators trained classifiers to classify microcalcifications and false detections based on morphological features such as contrast, size, shape, and edge gradient [31–35, 109–112] Zheng et al [37] used a differenceof-Gaussian band-pass filter to enhance the microcalcifications and then used multilayer feature analysis to identify true and false microcalcifications Although the details of the various microcalcification-detection algorithms differ, many have similar major steps Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page Tuesday, May 10, 2005 3:38 PM Medical Image Analysis In the first step, the image is processed to enhance the signal-to-noise ratio (SNR) of the microcalcifications Second, microcalcification candidates are segmented from the image background In the third step, features of the candidate signals are extracted, and a feature classifier is trained or some rule-based methods are designed to distinguish true signals from false signals In the last step, a criterion is applied to the remaining signals to search for microcalcification clusters The computer vision methods used in our microcalcification-detection program are discussed in the following subsection as an example 1.2.1 METHODS 1.2.1.1 Preprocessing Technique Microcalcifications on mammograms are surrounded by breast tissues of varied densities The background gray levels thus vary over a wide range A preprocessing technique that can suppress the background and enhance the signals will facilitate segmentation of the microcalcifications from the image Chan et al [18, 26–28, 113] first demonstrated that a difference-image technique can effectively enhance microcalcifications on digitized mammograms In the difference-image technique, a signalenhancement filter enhances the microcalcifications and a signal-suppression filter suppresses the microcalcifications and smoothes the noise By taking the difference of the two filtered images, an SNR-enhanced image is obtained in which the lowfrequency structured background is removed and the high-frequency noise is suppressed When both the signal-enhancement filter and the signal-suppression filter are linear, the difference-image technique is equivalent to band-pass filtering with a frequency band adjusted to amplify that of the microcalcifications Nonlinear filters can also be designed for enhancement or suppression of the microcalcifications An example of a signal-suppression filter is a median filter, the kernel size of which can be chosen to remove microcalcifications and noise from the mammograms [26] Other investigators used preprocessing techniques such as wavelet filtering [30] and difference-ofGaussian filters [36] in the initial step of their microcalcification-detection programs These techniques can be considered variations of the difference-image technique 1.2.1.2 Microcalcification Segmentation After the SNR enhancement, the background gray level of the mammograms is relatively constant This facilitates the segmentation of the individual microcalcifications from the background Our approach is to first employ a gray-level thresholding technique to locate potential signal sites above a global threshold The global threshold is adapted to a given mammogram by an iterative procedure that automatically changes the threshold until the number of sites obtained falls within the chosen input maximum and minimum numbers At each potential site, a locally adaptive gray-level thresholding technique in combination with region growing is then performed to extract the connected pixels above a local threshold, which is calculated as the product of the local root-mean-square (RMS) noise and an input SNR threshold The features of the extracted signals — such as the size, maximum contrast, SNR, and its location — will also be extracted during segmentation Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page Tuesday, May 10, 2005 3:38 PM Computer-Aided Diagnosis of Breast Cancer 1.2.1.3 Rule-Based False-Positive Reduction In the false-positive reduction step, we combine rule-based classification with an artificial neural network to distinguish true microcalcifications from noise or artifacts The rule-based classification includes three rules: maximum and minimum number of pixels in a calcification, and contrast The two rules on the size exclude signals below a certain size, which are likely to be noise, and signals greater than a certain size, which are likely to be large benign calcifications The contrast rule sets an upper bound to exclude potential signals that have a contrast greater than an input number of standard deviations above the average contrast of all potential signals found with local thresholding This rule excludes the very-high-contrast signals that are likely to be image artifacts and large benign calcifications After rule-based classification, a convolution neural network (CNN) [28] was trained to further reduce false signals, as detailed in the next subsection 1.2.1.4 False-Positive Reduction Using Convolution Neural Network Classifier The CNN is based on the neocognitron structure [114] designed to simulate the human visual system It has been used for detection of lung nodules on chest radiographs, detection of microcalcifications on mammograms, and classification of mass and normal breast tissue on mammograms [28, 115, 116] The general architecture of the CNN used in this study is shown in Figure 1.1 The input to the CNN is a regionof-interest (ROI) image, extracted for each of the potential signal sites The nodes in the hidden layers are arranged in groups, as are the weights associated with each node; each weight group functions like a filter kernel The CNN is trained to classify the input ROI as containing a true microcalcification (TP) or a false signal (FP) In the implementation used in this study, the CNN had one input node, two hidden layers, and one output node All node groups in the two hidden layers were fully connected Training was performed with an error back propagation delta-bar-delta rule There were N1 node groups in the first hidden layer, and N2 node groups in the second hidden layer The kernel sizes of the first group of filters between the input node and the first hidden layer were K1 × K1, and those of the second group of filters between the first and second hidden layer were K2 × K2 For a CNN, learning is constrained such that forward signal propagation is similar to a spatially invariant convolution operation; the signals from the nodes in the lower layer are convolved with the weight kernel, and the resultant value of the convolution is collected into the corresponding node in the upper layer This value is further processed by the node through a sigmoidal activation function and produces an output signal that will, in turn, be forward propagated to the subsequent layer in a similar manner The convolution kernel incorporates the neighborhood information in the input image pattern and transfers the information to the receiving layers, thus providing the pattern-recognition capability of the CNN The neural-network architecture used in many studies was selected using a manual optimization technique [28] We evaluated the use of automated optimization methods for selecting an optimal CNN architecture [117] Briefly, three automated Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page Tuesday, May 10, 2005 3:38 PM Medical Image Analysis First Hidden Layer Second Hidden Layer K2 Output Node 1 Input ROI • • • K1 • • • N2 N1 FIGURE 1.1 Schematic diagram of the architecture of a convolution neural network The input to the CNN is a region-of-interest (ROI) image extracted for each of the detected signals The output is a scalar that is the relative rating by the CNN representing the likelihood that the input ROI contains a true microcalcification or a false-positive signal methods, the steepest descent (SD), the simulated annealing (SA), and the genetic algorithm (GA) were compared Four main parameters of the CNN architecture, N1, N2, K1, and K2, were considered for optimization The area under the ROC curve, Az, [118] was used to design a cost function The SA experiments were conducted with four different annealing schedules Three different parent selection methods were compared for the GA experiments The CNN was optimized with a set of ROI images extracted from 108 mammograms The suspected microcalcifications were detected after the initial steps of the microcalcification-detection program [28] The detected signals were labeled as TP or FP automatically based on the ground truth of the data set A 16 × 16-pixel ROI centered at the signal site was extracted for each of the detected locations, and these ROI images were used for training and testing the CNN The microcalcification-detection program detected more FP ROIs than TP ROIs at the prescreening stage For classifier training, it is more efficient to have approximately equal numbers of TP and FP ROIs Therefore, only a randomly selected subset of FP ROI images was used The selected ROIs were divided into two separate groups, one for training and the other for monitoring the classification accuracy of the trained CNN Each group contained more than 1000 ROIs Another data set of 152 mammograms, which was different from the set of 108 mammograms employed for optimization of the CNN, was used for validation of the detection program in combination with the CNN classifier The optimal architecture (N1-N2-K1-K2) was determined to be 14-10-5-7 using the training and validation Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page Tuesday, May 10, 2005 3:38 PM Computer-Aided Diagnosis of Breast Cancer sets This optimal CNN architecture was then compared with the CNN architecture of 12-8-5-3 determined by a manual search technique [28] For comparison of the performance of the CNN of different architectures, an independent data set of 472 digitized mammograms was used This test data set was selected from the University of South Florida (USF) digitized mammogram database, which is publicly available over the Internet [119] From the available cases in this database, only malignant cases that were digitized with the Lumisys 200 laser scanner were selected (volumes: cancer_01, cancer_02, cancer_05, cancer_09, and cancer_15) The data set contained 272 biopsy-proven microcalcification clusters, of which 253 were malignant and 19 were benign There were 184 mammograms free of microcalcifications [119] All mammograms in the training, validation, and test sets were digitized at a pixel resolution of 0.05 × 0.05 mm with 4096 gray levels The images were converted to 0.1 × 0.1-mm resolution by averaging adjacent × pixels and subsampling The detection was carried out on the 0.1 × 0.1-mm resolution images 1.2.1.5 False-Positive Reduction Using Clustering A final step to reduce false positives is clustering This approach is devised based on clinical experiences that the likelihood of malignancy for clustered microcalcifications is generally much greater than sparsely scattered microcalcifications [102106] Chan et al [28, 113] designed a dynamic clustering procedure to identify clustered microcalcifications The image is initially partitioned into regions and the number of potential signals in each region is determined A region with a higher concentration of potential signals is given a higher priority as a starting region to grow a cluster The cluster grows by searching for new members in its neighborhood one at a time A signal is included as a new member if it is within a threshold distance from the centroid of the current cluster The cluster centroid location is updated after each new member is added The cluster can grow across region boundaries without constraints Clustering stops when no more new members can be found to satisfy the inclusion criteria A cluster is considered to be true if the number of members in the cluster is greater than a preselected threshold The signals that are not found to be in the neighborhood of any clusters will be considered isolated noise points or insignificant calcifications and excluded The specific parameters or thresholds used in the various steps depend on the spatial and gray level resolutions of the digitized or digital mammograms [28, 113] It was found that having four detected signals within a clustering diameter of cm provided a high sensitivity for cluster detection 1.2.2 FROC ANALYSIS OF DETECTION ACCURACY The performance of a computer-aided detection system is generally evaluated by the free-response receiver operating characteristic (FROC) analysis [120] An FROC curve shows the sensitivity of lesion detection as a function of the number of FPs per image In this study, it was generated by varying the input SNR threshold over a range of values so that the detection criterion varied from lenient (low threshold) to stringent (high threshold) After passing the size and contrast criteria, screening by the trained CNN, and passing the regional-clustering criterion, the detected Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 10 Tuesday, May 10, 2005 3:38 PM 10 Medical Image Analysis individual microcalcifications and clusters are compared with the "truth" file of the input image The number of TP and FP microcalcifications and the number of TP and FP clusters are scored The scoring method varies among researchers In our study, the detected signal was scored as a TP microcalcification if it was within 0.5 mm from a true microcalcification in the "truth" file A detected cluster was scored as a TP if its centroid coordinate was within a cluster radius (5 mm) from the centroid of a true cluster and at least two of its member microcalcifications were scored as TP Once a true microcalcification or cluster was matched to a detected microcalcification or cluster, it would be eliminated from further matching Any detected microcalcifications or clusters that did not match to a true microcalcification or cluster were scored as FPs The trade-off between the TP and FP detection rates by the computer program was analyzed as an FROC curve A low SNR threshold corresponded to a lax criterion with high sensitivity and a large number of FP clusters A high SNR threshold corresponded to a stringent criterion with a small number of FP clusters and a loss in TP clusters The detection accuracy of the computer program with and without the CNN classifier could then be assessed by comparison of the FROC curves To test the performance of the selected optimal architecture, the detection program was run at seven SNR threshold values varying between 2.6 and 3.2 at increments of 0.1 Figure 1.2a shows the FROC curves of the microcalcificationdetection program using both the manually optimized and automatically optimized CNN architectures The FP rate was estimated from the computer marks on the 184 normal mammograms that were free of microcalcifications in the USF data set The automatically optimized architecture outperformed the manually optimized architecture At an FP rate of 0.7 cluster per image, the film-based sensitivity is 84.6% with the optimized CNN, in comparison with 77.2% for the manually selected CNN Figure 1.2b shows the FROC curves for the microcalcification-detection programs if clusters having images in both CC and MLO views are analyzed and a cluster is considered to be detected when it is detected in one or both views This “case-based” scoring has been adopted for the evaluation of some CAD systems [20] The rationale is that if the CAD system can bring the radiologist’s attention to the lesion on one of the views, it will be unlikely that the radiologist will miss the lesion For casebased scoring, the sensitivity at 0.7 FPs/image is 93.3% for the automatically optimized CNN and 87.0% for the manually selected CNN This study demonstrates that classification of true and false signals is an important step in the microcalcification-detection program and that an optimized CNN can effectively reduce FPs and improve the detection accuracy of the CAD system An automated optimization algorithm such as simulated annealing can find the optimum more efficiently [117, 121–123] than a manual search, which may find only a local optimum because it is difficult to explore adequately a high-dimensional parameter space The optimization described here is applied to one stage, FP reduction with the CNN, of the detection program The cost function was based on the Az of the CNN classifier for its performance in differentiating the TP and FP signals Ideally, one would prefer to optimize all parameters in the detection program together In such a case, optimizing the performance in terms of the FROC curve will be necessary The principle of optimizing the entire detection system is similar Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 35 Tuesday, May 10, 2005 3:38 PM Computer-Aided Diagnosis of Breast Cancer 35 using a 3:1 training/test ratio The training set was used to select a subset of useful morphological features using stepwise feature selection and to estimate the coefficients of an LDA classifier To reduce biases in the classifier, 50 random 3:1 partitions of the cases were employed A morphological score was obtained for each individual object by averaging the object’s test scores obtained from the different partitions The morphological score was then combined with the single-view texture score by averaging the two scores A single combined score thus characterized each prescreening object This one-view score is further fused with the discriminant score obtained by the two-view scheme, as described in the next subsection 1.4.1.3 Two-View Analysis The block diagram in Figure 1.14 illustrates our two-view mass-detection scheme and its relationship to the one-view approach The prescreening objects were further analyzed by the two-view method shown in the right branch of the diagram All possible pairings between the prescreening objects in the two views of the same breast were determined using the distance from the nipple to the centroid of each object and the previously described geometrical model Because the location of a given object detected in one view cannot be uniquely identified in the other view, as described in Section 1.4.1.1, an object was initially paired with all objects with centroids located within its defined annular region in the other view The geometric constraints reduced the number of object pairs that needed to be classified as true or false correspondences in the subsequent steps A true pair (TP-TP) was defined as the correspondence between the same true masses on the two mammographic views, and a false pair was defined as any other object pairing (TP-FP, FP-TP and FP-FP) For each object pair, the set of 15 texture and 31 morphological features (described previously) was used to form similarity measures In this study, two simple measures — the absolute difference and the mean — were used A total of 30 texture measures and 62 morphological measures were thus obtained for each object pair The absolute difference between the nipple-to-object distances in the CC and MLO views was also included as a feature for differentiating true from false object pairs Two separate LDA classifiers with stepwise feature selection were trained to classify the true and false pairs using the similarity features in the morphological- and texture-feature spaces The classifiers were trained by randomly dividing the data set into a training set and a test set, again using a 3:1 training/test ratio Fifty random 3:1 partitions of the cases were used to reduce bias Individual morphological and texture scores were obtained for each object by averaging the object’s test scores obtained from the different partitionings The two classification scores were then averaged to obtain one “correspondence” score for each object pair This score, along with the singleview prescreening score, was used in the fusion step described in the next subsection 1.4.1.4 Fusion Analysis The fusion of the single-view prescreening scores with the two-view correspondence scores was the final step in the two-view detection scheme All prescreening object Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 36 Tuesday, May 10, 2005 3:38 PM 36 Medical Image Analysis scores were first ranked within a given film from the largest to the smallest The correspondence scores were ranked in a similar way These two new rank scores were then merged into a single score for each object in each view Because an object could have more than one correspondence score, its two-view correspondence score was taken to be the maximum correspondence score among all object pairs in which this object was a member There can be many variations for the fusion step In this work, the final discriminant score for an object was obtained by averaging its twoview correspondence-score rank with its one-view prescreening-score rank The accuracy of the single-view detection scheme and the two-view approach are compared in the following subsection based on their FROC performance curves To demonstrate the effects of the number of the prescreening objects on the overall detection accuracy of the two-view scheme, the FROC curves obtained with 5, 10, and 15 prescreening objects per image are compared 1.4.2 RESULTS 1.4.2.1 Geometrical Modeling For the geometric modeling of object location on two views, the database consisted of 116 cases with masses, large benign calcifications, or clustered microcalcifications identifiable on both views of the same breast The mammograms were digitized with a Lumisys 85 film scanner with a pixel size of 50 µm and 12-bit gray levels Since the geometric modeling was not expected to have accuracy within mm, high-resolution processing was not needed To reduce processing time, the images were reduced to a pixel resolution of 800 × 800 µm by averaging 16 × 16 neighboring pixels and downsampling For each case, the two standard mammographic views were available A total of 177 objects were manually selected and marked by an expert radiologist on each of these two views The nipple location was also identified for each breast image In the geometrical analysis, we first estimated a prediction model of the radial distance of an object in a second view from its radial distance in the first view using the training set The model was then used to predict object location from one view to the other for the independent test cases Because the model did not provide an exact solution, a search region, Rx ± ∆R, where Rx was the predicted radial distance and ∆R the half-width of an annular region, was defined The percentage of the true object centroids enclosed within the search region was measured as a function of the size of 2∆R Figure 1.15 shows the prediction accuracy as a function of 2∆R for estimating the object radial distance in the MLO view from that in the CC view The results for predicting the object radial distance in the CC view from that in the MLO view are very similar and are not shown The training and test curves almost overlap in each case The difference in the accuracy between searching the object centers in the CC or MLO views is small About 83% of the object centers are within the search region when the radial width of the search region is ≈40 pixels (32 mm) for either the CC view or the MLO view The search region, although large, is much smaller than the entire area of the breast The limited size of the search region reduces the number of object pairs to be analyzed in the two-view detection scheme To avoid missing any pairs of true masses in the two-view scheme, we chose to set Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 37 Tuesday, May 10, 2005 3:38 PM Computer-Aided Diagnosis of Breast Cancer 37 100 Percentage of Object Centers within the Annular Region 90 80 70 60 50 40 30 20 Training set Test set 10 0 10 20 30 40 50 60 70 80 90 100 Number of Pixels for the Radial Width of the Annular Search Region FIGURE 1.15 Prediction of the center of an object in the MLO view from its location in the CC view Training and test performances are given as a function of the radial width of the annular search region the radial width of the annular search region to about 80 pixels This led to a larger number of false pairs, but it was substantially less than what would be detected if the entire breast area was considered 1.4.2.2 Comparison of One-View and Two-View Analysis For the comparison of the one-view and two-view mass detection schemes, a data set of 169 pairs of mammograms containing masses on both the CC and MLO views was used The mammograms were obtained from 117 patients, of which 128 pairs were current mammograms (defined as mammograms from the exam before biopsy), and 41 pairs were from exams to years prior to biopsy A malignant mass was observed in 58 of the 128 current and 26 of the 41 prior image pairs The 338 mammograms were also digitized with the Lumisys 85 film scanner The true mass locations on both views were identified by an MQSA radiologist Three different decision thresholds that retained a maximum of 5, 10, and 15 objects per image after the one-view prescreening stage were used to select mass candidates as inputs to the two-view detection scheme The FROC curves for the detection of malignant and benign masses on each image, using the two-view fusion technique, are similar for the three thresholds of 5, 10, and 15 prescreening objects per image This similarity also holds for the FROC curves for detection of malignant masses, as illustrated in Figure 1.16 The improvement in detection by the two-view fusion method therefore seems to be independent of the operating threshold when the maximum number of objects retained per image in the prescreening stage is between and 15 We therefore chose the condition of 10 prescreening objects per image for the following discussion Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 38 Tuesday, May 10, 2005 3:38 PM 38 Medical Image Analysis 1.0 0.8 TPF 0.6 0.4 objects per image 10 objects per image 15 objects per image 0.2 0.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 FPs/Image FIGURE 1.16 Film-based performances of the two-view detection scheme applied to the current malignant masses Three initial conditions — depending on the maximum number of retained objects per image (5, 10, and 15 objects per image) at the prescreening stage — were evaluated The performance of the single-view mass-detection algorithm and the two-view fusion-detection algorithm are compared The image-based FROC curves for the detection of malignant masses in the data set are shown in Figure 1.17 The corresponding case-based FROC curves are shown in Figure 1.18 The FROC curves for detection of the malignant masses on the current and prior mammograms are plotted separately for comparison It is apparent that the two-view fusion method can improve the detection sensitivity by 10 to 15% in the range of 0.5 to 1.5 FPs/image for the malignant masses on current mammograms For example, at FPs/image, the two-view algorithm achieved a case-based detection sensitivity of 91%, whereas the current single-view scheme had a 77% sensitivity at the same number of FPs per image in this data set For the case-based comparison, the detection of prior masses could be improved by more than 5% within the range of 0.5 to 1.2 FPs/image Alternatively, the two-view fusion can be used as an FP reduction technique The results indicate that the two-view fusion method is more effective in reducing FPs in the subset of cases containing malignant masses on current mammograms At a case-based detection sensitivity of 75% for all masses, the number of FPs per image was reduced from 1.5 FPs/image using the single-view detection technique to 1.13 FPs/image using the two-view fusion technique At a case-based sensitivity of 85% for malignant masses on current mammograms, the number of FPs per image was reduced from 1.5 FPs/image to 0.5 FPs/image (Figure 1.18) This study demonstrates that including correspondence information from two mammographic views is a promising approach to improving detection accuracy in a CAD system for detection of breast cancer Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 39 Tuesday, May 10, 2005 3:38 PM Computer-Aided Diagnosis of Breast Cancer 39 1.0 0.8 TPF 0.6 0.4 1-view, current malignant 1-view, prior malignant 2-view, current malignant 2-view, prior malignant 0.2 0.0 0.0 0.5 1.0 1.5 FPs/Image 2.0 2.5 3.0 FIGURE 1.17 Comparison of the image-based performance of the one-view and two-view detection methods for the detection of malignant masses on current mammograms and prior mammograms 1.0 0.8 TPF 0.6 0.4 1-view, current malignant 1-view, prior malignant 2-view, current malignant 2-view, prior malignant 0.2 0.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 FPs/Image FIGURE 1.18 Comparison of the case-based performance of the one-view and two-view detection methods for the detection of malignant masses on current mammograms and prior mammograms Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 40 Tuesday, May 10, 2005 3:38 PM 40 Medical Image Analysis 1.5 SUMMARY In this chapter, we discussed some of the computer vision techniques used for computer-aided detection (CAD) of breast cancer We used our studies in this area as examples to illustrate the various methods that may be useful for the development of CAD algorithms in mammography These examples are by no means exhaustive, and many variations of the methods used in the different stages of the automated detection process can be found in the literature Although several CAD systems are already commercially available for assisting radiologists in clinical practice, the performances of the CAD systems are not yet ideal Further investigation is needed to improve the sensitivity and the specificity of the systems One promising approach to improving the performance of computerized breast cancer detection systems is to incorporate multiple image information, including two views or three views of the same breast, comparison of current and prior mammograms, or comparison of bilateral mammograms, which has been practiced routinely by radiologists in mammographic interpretation The adaptation of the CAD systems to direct digital mammography may also improve lesion detectability We have focused our discussion on lesion detection Computer-aided characterization of breast lesions is another important CAD application CAD techniques for differentiation of malignant and benign lesions have been published in the literature ROC studies have also been performed to demonstrate the potential of CAD in reducing unnecessary biopsy For both detection and characterization of breast lesions, a promising direction of research is to combine information from multiple breast-imaging modalities Ultrasound imaging has been routinely used for diagnostic workup of suspicious masses Contrast-enhanced magnetic-resonance breast imaging is a new approach to differentiating malignant and benign breast lesions and detecting multifocal lesions A number of new breast-imaging techniques are under development, including three-dimensional ultrasound imaging, digital tomosynthesis, breast computed tomography, and single-energy or dual-energy contrastenhanced digital-subtraction mammography These new techniques hold the promise of improving breast cancer detection and diagnosis However, they can also drastically increase the amount of information that radiologists have to interpret for each case A CAD system that can analyze the multimodality images and merge the information will not only improve the accuracy of the computer system, but also provide radiologists with a useful second opinion that could improve the efficiency and effectiveness of breast cancer detection and management ACKNOWLEDGMENT This work is supported by USPHS Grants CA 48129 and CA 95153 and by U.S Army Medical Research and Materiel Command (USAMRMC) grants DAMD1796-1-6254, and DAMD17-02-1-0214 Berkman Sahiner is also supported by a USAMRMC grant DAMD17-01-1-0328 Lubomir Hadjiiski is also supported by a USAMRMC grant DAMD17-02-1-0489 Nicholas Petrick and Sophie Paquerault were at the University of Michigan when the work was performed Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 41 Tuesday, May 10, 2005 3:38 PM Computer-Aided Diagnosis of Breast Cancer 41 REFERENCES Hillman, B.J., Fajardo, L.L., Hunter, T.B et al., Mammogram interpretation by physician assistants, Am J Roentgenol., 149, 907, 1987 Bassett, L.W., Bunnell, D.H., Jahanshahi, R et al., Breast cancer detection: one vs two views, Radiology, 165, 95, 1987 Wallis, M.G., Walsh, M.T., and Lee, J.R., A review of false negative mammography in a symptomatic population, Clinical Radiol., 44, 13, 1991 Harvey, J.A., Fajardo, L.L., and Innis, C.A., Previous mammograms in patients with impalpable breast carcinomas: retrospective vs blinded interpretation, Am J Roentgenol., 161, 1167, 1993 Bird, R.E., Wallace, T.W., and Yankaskas, B.C., Analysis of cancers missed at screening mammography, Radiology, 184, 613, 1998 Beam, C.A., Layde, P.M., and Sullivan, D.C., Variability in the interpretation of screening mammograms by U.S radiologists: findings from a national sample, Arch Intern Medicine, 156, 209, 1996 Birdwell, R.L., Ikeda, D.M., O'Shaughnessy, K.F et al., Mammographic characteristics of 115 missed cancers later detected with screening mammography and the potential utility of computer-aided detection, Radiology, 219, 192, 2001 Kopans, D.B., The positive predictive value of mammography, Am J Roentgenol., 158, 521, 1992 Smart, C.R., Hendrick, R.E., Rutledge, J.H et al., Benefit of mammography screening in women ages 40 to 49 years: current evidence from randomized controlled trials, Cancer, 75, 1619, 1995 10 Byrne, C., Smart, C.R., Cherk, C et al., Survival advantage differences by age: Evaluation of the extended follow-up of the breast cancer detection demonstration project, Cancer, 74, 301, 1994 11 Feig, S.A and Hendrick, R.E., Risk, benefit, and controversies in mammographic screening, in Syllabus: A Categorical Course in Physics Technical Aspects of Breast Imaging, Haus, A.G and Yaffe, M.J., Eds., Radiological Society of North America, Oak Brook, IL, 1993, p 119 12 Seidman, H., Gelb, S.K., Silverberg, E et al., Survival experience in the breast cancer detection demonstration project, CA Cancer J Clin., 37, 258, 1987 13 Sabel, M and Aichinger, H., Recent developments in breast imaging, Phys Med Biol., 41, 315, 1996 14 Thurfjell, E.L., Lernevall, K.A., and Taube, A.A.S., Benefit of independent double reading in a population-based mammography screening program, Radiology, 191, 241, 1994 15 Anderson, E.D.C., Muir, B.B., Walsh, J.S et al., The efficacy of double reading mammograms in breast screening, Clinical Radiol., 49, 248, 1994 16 Shtern, F., Stelling, C., Goldberg, B et al., Novel Technologies in Breast Imaging: National Cancer Institute Perspective, Presented at Society of Breast Imaging Conference, Orlando, FL, 1995, p 153 17 Vyborny, C.J., Can computers help radiologists read mammograms? Radiology, 191, 315, 1994 18 Chan, H.P., Doi, K., Vyborny, C.J et al., Improvement in radiologists' detection of clustered microcalcifications on mammograms: the potential of computer-aided diagnosis, Acad Radiol., 25, 1102, 1990 19 Kegelmeyer, W.P., Pruneda, J.M., Bourland, P.D et al., Computer-aided mammographic screening for spiculated lesions, Radiology, 191, 331, 1994 Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 42 Tuesday, May 10, 2005 3:38 PM 42 Medical Image Analysis 20 Warren Burhenne, L.J., Wood, S.A., D'Orsi, C.J et al., Potential contribution of computer-aided detection to the sensitivity of screening mammography, Radiology, 215, 554, 2000 21 Freer, T.W and Ulissey, M.J., Screening mammography with computer-aided detection: prospective study of 12,860 patients in a community breast center, Radiology, 220, 781, 2001 22 Zheng, B., Ganott, M.A., Britton, C.A et al., Soft-copy mammographic readings with different computer-assisted detection cuing environments: preliminary findings, Radiology, 221, 633, 2001 23 Brem, R.F., Baum, J.K., Lechner, M et al., Improvement in sensitivity of screening mammography with computer-aided detection: a multi-institutional trial, Am J Roentgenol., 181, 687, 2003 24 Chan, H.-P., Sahiner, B., Helvie, M.A et al., Improvement of radiologists' characterization of mammographic masses by computer-aided diagnosis: an ROC study, Radiology, 212, 817, 1999 25 Jiang, Y., Nishikawa, R.M., Schmidt, R.A et al., Improving breast cancer diagnosis with computer-aided diagnosis, Acad Radiol., 6, 22, 1999 26 Chan, H.P., Doi, K., Galhotra, S et al., Image feature analysis and computer-aided diagnosis in digital radiography: 1, Automated detection of microcalcifications in mammography, Med Phys., 14, 538, 1987 27 Chan, H.P., Doi, K., Vyborny, C.J et al., Computer-aided detection of microcalcifications in mammograms: methodology and preliminary clinical study, Invest Radiol., 23, 664, 1988 28 Chan, H.P., Lo, S.C.B., Sahiner, B et al., Computer-aided detection of mammographic microcalcifications: pattern recognition with an artificial neural network, Med Phys., 22, 1555, 1995 29 Davies, D.H and Dance, D.R., Automatic computer detection of clustered calcifications in digital mammograms, Phys Med Biol., 35, 1111, 1990 30 Qian, W., Clarke, L.P., Kallergi, M et al., Tree-structured nonlinear filter and wavelet transform for microcalcification segmentation in mammography, Proc SPIE Medical Imaging, 1905, 509, 1993 31 Nishikawa, R.M., Giger, M.L., Doi, K et al., Computer-aided detection and diagnosis of masses and clustered microcalcifications from digital mammograms, Proc SPIE Medical Imaging, 1905, 422, 1993 32 Astley, S., Hutt, I., Adamson, S et al., Automation in mammography: computer vision and human perception, Proc SPIE Medical Imaging, 1905, 716, 1993 33 Bankman, I.N., Christens-Barry, W.A., Kim, D.W et al., Automated recognition of microcalcification clusters in mammograms, Proc SPIE Biomed Image Processing Biomed Visualization, 1905, 731, 1993 34 Karssemeijer, N., Recognition of clustered microcalcifications using a random field model, Proc SPIE Medical Imaging, 1905, 776, 1993 35 Shen, L., Rangayyan, R.M., and Desautels, J.E.L., Automatic detection and classification system for calcifications in mammograms, Proc SPIE Medical Imaging, 1905, 799, 1993 36 Zhang, W., Doi, K., Giger, M.L et al., Computerized detection of clustered microcalcifications in digital mammograms using a shift-invariant artificial neural network, Med Phys., 21, 517, 1994 37 Zheng, B., Chang, Y.S., Staiger, M et al., Computer-aided detection of clustered microcalcifications in digitized mammograms, Acad Radiol., 2, 655, 1995 38 Gavrielides, M.A., Lo, J.Y., Vargas-Voracek, R et al., Segmentation of suspicious clustered microcalcifications in mammograms, Med Phys., 27, 13, 2000 Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 43 Tuesday, May 10, 2005 3:38 PM Computer-Aided Diagnosis of Breast Cancer 43 39 Ackerman, L.V and Gose, E.E., Breast lesion classification by computer and xeroradiograph, Cancer, 30, 1025, 1972 40 Kimme, C., O'Laughlin, B.J., and Sklansky, J., Automatic Detection of Suspicious Abnormalities in Breast Radiographs, Academic Press, New York, 1975 41 Chan, H.P., Niklason, L.T., Ikeda, D.M et al., Computer-aided diagnosis in mammography: detection and characterization of microcalcifications, Med Phys., 19, 831, 1992 42 Chan, H.P., Sahiner, B., Lam, K.L et al., Classification of malignant and benign microcalcifications on mammograms using an artificial neural network, Proc World Congress Neural Networks, II, 889, 1995 43 Chan, H.P., Sahiner, B., Petrick, N et al., Computerized classification of malignant and benign microcalcifications on mammograms: texture analysis using an artificial neural network, Phys Med Biol., 42, 549, 1997 44 Chan, H.P., Sahiner, B., Lam, K.L et al., Computerized analysis of mammographic microcalcifications in morphological and texture feature space, Med Phys., 25, 2007, 1998 45 Shen, L., Rangayyan, R.M., and Desautels, J.E.L., Application of shape analysis to mammographic calcifications, IEEE Trans Medical Imaging, 13, 263, 1994 46 Wu, Y., Freedman, M.T., Hasegawa, A et al., Classification of microcalcifications in radiographs of pathologic specimens for the diagnosis of breast cancer, Acad Radiol., 2, 199, 1995 47 Jiang, Y., Nishikawa, R.M., Wolverton, D.E et al., Malignant and benign clustered microcalcifications: automated feature analysis and classification, Radiology, 198, 671, 1996 48 Thiele, D.L., Kimme-Smith, C., Johnson, T.D et al., Using tissue texture surrounding calcification clusters to predict benign vs malignant outcomes, Med Phys., 23, 549, 1996 49 Dhawan, A.P., Chitre, Y., KaiserBonasso, C et al., Analysis of mammographic microcalcifications using gray-level image structure features, IEEE Trans Med Imaging, 15, 246, 1996 50 Winsberg, F., Elkin, M., Macy, J et al., Detection of radiographic abnormalities in mammograms by means of optical scanning and computer analysis, Radiology, 89, 211, 1967 51 Semmlow, J.L., Shadagopappan, A., Ackerman, L.V et al., A fully automated system for screening mammograms, Comput Biomed Res., 13, 350, 1980 52 Lai, S.M., Li, X., and Bischof, W.F., On techniques for detecting circumscribed masses in mammograms, IEEE Trans Medical Imaging, 8, 377, 1989 53 Lau, T.K and Bischof, W.F., Automated detection of breast tumors using the asymmetry approach, Comput Biomed Res., 24, 273, 1991 54 Yin, F.F., Giger, M.L., Doi, K et al., Computerized detection of masses in digital mammograms: analysis of bilateral subtraction images, Med Phys., 18, 955, 1991 55 Ng, S.L and Bischof, W.F., Automated detection and classification of breast tumors, Comput Biomed Res., 25, 218, 1992 56 Yin, F.F., Giger, M.L., Vyborny, C.J et al., Comparison of bilateral subtraction and single-image processing techniques in the computerized detection of mammographic masses, Acad Radiol., 28, 473, 1993 57 Brzakovic, D., Vujovic, N., Neskovic, M et al., Mammogram analysis by comparison with previous screenings, in Digital Mammography, Gale, A.G., Astley, S.M., Dance, D.R et al., Eds Elsevier, Amsterdam, 1994, p 131 58 Laine, A.F., Schuler, S., Fan, J et al., Mammographic feature enhancement by multiscale analysis, IEEE Trans Medical Imaging, 13, 725, 1994 Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 44 Tuesday, May 10, 2005 3:38 PM 44 Medical Image Analysis 59 Li, H.D., Kallergi, M., Clarke, L.P et al., Markov random field for tumor detection in digital mammograms, IEEE Trans Medical Imaging, 14, 565, 1995 60 Laine, A.F., Huda, W., Steinbach, B.G et al., Mammographic image processing using wavelet processing techniques, Eur Radiol., 5, 518, 1995 61 Zheng, B., Chang, Y.H., and Gur, D., Computerized detection of masses in digitized mammograms using single-image segmentation and a multilayer topographic feature analysis, Acad Radiol., 2, 959, 1995 62 Petrick, N., Chan, H.P., Sahiner, B et al., An adaptive density-weighted contrast enhancement filter for mammographic breast mass detection, IEEE Trans Medical Imaging, 15, 59, 1996 63 Petrick, N., Chan, H.P., Wei, D et al., Automated detection of breast masses on mammograms using adaptive contrast enhancement and texture classification, Med Phys., 23, 1685, 1996 64 Karssemeijer, N and te Brake, G., Detection of stellate distortions in mammograms, IEEE Trans Medical Imaging, 15, 611, 1996 65 Kobatake, H and Yoshinaga, Y., Detection of spicules on mammogram based on skeleton analysis, IEEE Trans Medical Imaging, 15, 235, 1996 66 Guliato, D., Randayyan, R.M., Carnielli, W.A et al., Segmentation of breast tumors in mammograms by fuzzy region growing, in Proc 20th Annual International Conference of IEEE Engineering in Medicine and Biology Society, Hong Kong, 1998, p 1002 67 Kupinski, M.A and Giger, M.L., Automated seeded lesion segmentation on digital mammograms, IEEE Trans Medical Imaging, 17, 510, 1998 68 Kobatake, H., Murakami, M., Takeo, H et al., Computer detection of malignant tumors on digital mammograms, IEEE Trans Medical Imaging, 18, 369, 1999 69 Good, W.F., Zheng, B., Chang, Y.H et al., Multi-image CAD employing features derived from ipsilateral mammographic views, Proc SPIE, 3661, 474, 1999 70 Petrick, N., Chan, H.P., Sahiner, B et al., Combined adaptive enhancement and region-growing segmentation of breast masses on digitized mammograms, Med Phys., 26, 1642, 1999 71 te Brake, G.M and Karssemeijer, N., Single and multiscale detection of masses in digital mammograms, IEEE Trans Medical Imaging, 18, 628, 1999 72 te Brake, G.M., Karssemeijer, N., and Hendriks, J.H.C.L., An automatic method to discriminate malignant masses from normal tissue in digital mammograms, Phys Med Biol., 45, 2843, 2000 73 Petrick, N., Chan, H.P., Sahiner, B et al., Breast cancer detection: evaluation of a mass detection algorithm for computer-aided diagnosis: experience in 263 patients, Radiology, 224, 217, 2002 74 Kilday, J., Palmieri, F., and Fox, M.D., Classifying mammographic lesions using computer-aided image analysis, IEEE Trans Medical Imaging, 12, 664, 1993 75 Pohlman, S., Powell, K.A., Obuchowshi, N.A et al., Quantitative classification of breast tumors in digitized mammograms, Med Phys., 23, 1337, 1996 76 Huo, Z.M., Giger, M.L., Vyborny, C.J et al., Automated computerized classification of malignant and benign masses on digitized mammograms, Acad Radiol., 5, 155, 1998 77 Sahiner, B., Chan, H.P., Petrick, N et al., Design of a high-sensitivity classifier based on a genetic algorithm: application to computer-aided diagnosis, Phys Med Biol., 43, 2853, 1998 78 Sahiner, B., Chan, H.P., Petrick, N et al., Computerized characterization of masses on mammograms: the rubber band straightening transform and texture analysis, Med Phys., 25, 516, 1998 Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 45 Tuesday, May 10, 2005 3:38 PM Computer-Aided Diagnosis of Breast Cancer 45 79 Ackerman, L.V., Mucciardi, A.N., Gose, E.E et al., Classification of benign and malignant breast tumors on the basis of 36 radiographic properties, Cancer, 31, 342, 1973 80 Getty, D.J., Pickett, R.M., D'Orsi, C.J et al., Enhanced interpretation of diagnostic images, Invest Radiol., 23, 240, 1988 81 D'Orsi, C.J., Getty, D.J., Swets, J.A et al., Reading and decision aids for improved accuracy and standardization of mammographic diagnosis, Radiology, 184, 619, 1992 82 Wu, Y., Giger, M.L., Doi, K et al., Artificial neural networks in mammography: application to decision making in the diagnosis of breast cancer, Radiology, 187, 81, 1993 83 Baker, J.A., Kornguth, P.J., Lo, J.Y et al., Breast cancer: prediction with artificial neural network based on bi-rads standardized lexicon, Radiology, 196, 817, 1995 84 Lo, J.Y., Markey, M.K., Baker, J.A et al., Cross-institutional evaluation of bi-rads predictive model for mammographic diagnosis of breast cancer, Am J Roentgenol., 178, 457, 2002 85 Nishikawa, R.M., Giger, M.L., Wolverton, D.E et al., Prospective testing of a clinical CAD workstation for the detection of breast lesions on mammograms, in Proc First International Workshop on Computer-Aided Diagnosis, Chicago, IL, 1999, p 209 86 Petrick, N., Chan, H.-P., Sahiner, B et al., Evaluation of an automated computeraided diagnosis system for the detection of masses on prior mammograms, in Proc SPIE Medical Imaging, San Diego, 2000, p 967 87 Kass, D.A., Gabbay, R., and Siedler, D.E., Results of computer-aided detection (CAD) on the prior screening mammograms of patients with interval breast cancers, Radiology, 217(P), 400, 2000 88 Helvie, M.A., Hadjiiski, L.M., Makariou, E et al., Sensitivity of noncommercial computer-aided detection system for mammographic breast cancer detection: a pilot clinical trial, Radiology, 231, 208, 2004 89 Hadjiiski, L.M., Chan, H.P., Sahiner, B et al., A CAD system for characterization of malignant and benign breast masses in temporal pairs of mammograms and its effects on radiologists' performance: an ROC study, Radiology, 225(P), 683, 2002 90 Hadjiiski, L.M., Chan, H.P., Sahiner, B et al., ROC study: effects of computer-aided diagnosis on radiologists' characterization of malignant and benign breast masses in temporal pairs of mammograms, Proc SPIE Medical Imaging, 5032, 94, 2003 91 Hackshaw, A.K., Wald, N.J., Michell, M.J et al., An investigation into why two-view mammography is better than one-view in breast cancer screening, Clinical Radiol., 55, 454, 2000 92 Blanks, R.G., Given-Wilson, R.M., and Moss, S.M., Efficiency of cancer detection during routine repeat (incident) mammographic screening: two- vs one-view mammography, J Medical Screening, 5, 141, 1998 93 Blanks, R.G., Wallis, M.G., and Given-Wilson, R.M., Observer variability in cancer detection during routine repeat (incident) mammographic screening in a study of twovs one-view mammography, J Medical Screening, 6, 152, 1999 94 Kita, Y., Highnam, R.P., and Brady, J.M., Correspondence between different view breast X-rays using curved epipolar lines, Computer Vision Image Understanding, 83, 38, 2001 95 Gopal, S.S., Chan, H.-P., Petrick, N et al., A regional registration technique for automated analysis of interval changes of breast lesions, Proc SPIE, 3338, 118, 1998 96 Sanjay-Gopal, S., Chan, H.P., Wilson, T et al., A regional registration technique for automated interval change analysis of breast lesions on mammograms, Med Phys., 26, 2669, 1999 Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 46 Tuesday, May 10, 2005 3:38 PM 46 Medical Image Analysis 97 Hadjiiski, L.M., Chan, H.P., Sahiner, B et al., Automated identification of breast lesions in temporal pairs of mammograms for interval-change analysis, Radiology, 213(P), 229, 1999 98 Hadjiiski, L.M., Chan, H.P., Sahiner, B et al., Interval-change analysis in temporal pairs of mammograms using a local affine transformation, Proc SPIE, 3979, 847, 2000 99 Paquerault, S., Sahiner, B., Petrick, N et al., Prediction of object location in different views using geometrical models, in Proc 5th International Workshop on Digital Mammography, Toronto, 2001, p 748 100 Paquerault, S., Petrick, N., Chan, H.P et al., Improvement of computerized mass detection on mammograms: fusion of two-view information, Med Phys., 29, 238, 2002 101 Sahiner, B., Petrick, N., Chan, H.P et al., Recognition of lesion correspondence on two mammographic views: a new method of false-positive reduction for computerized mass detection, Proc SPIE, 4322, 649, 2001 102 Tabar, L and Dean, P.B., Teaching Atlas of Mammography, Thieme, New York, 1985 103 Wolfe, J.N., Analysis of 462 breast carcinomas, AJR, 121, 846, 1974 104 Murphy, W.A and DeSchryver-Kecskemeti, K., Isolated clustered microcalcification in the breast: radiologic-pathologic correlation, Radiology, 127, 335, 1978 105 Millis, R.R., Davis, R., and Stacey, A.J., The detection and significance of calcifications in the breast: a radiological and pathological study, Br J Radiol., 49, 12, 1976 106 Sickles, E.A., Mammographic features of 300 consecutive nonpalpable breast cancers, Am J Roentgenol., 146, 661, 1986 107 Wu, Y., Doi, K., Giger, M.L et al., Computerized detection of clustered microcalcifications in digital mammograms: applications of artificial neural network, Med Phys., 19, 555, 1992 108 Fam, B.W., Olson, S.L., Winter, P.F et al., Algorithm for the detection of fine clustered calcifications on film mammograms, Radiology, 169, 333, 1988 109 Mascio, L.N., Hernandez, J.M., and Logan, C.M., Automated analysis for microcalcifications in high-resolution digital mammograms, Proc SPIE Medical Imaging, 1898, 472, 1993 110 Brzakovic, D., Brzakovic, P., and Neskovic, M., Approach to automated screening of mammograms, Proc SPIE, 1905, 690, 1993 111 Dhawan, A.P., Chitre, Y.S., and Moskowitz, M., Artificial-neural-network-based classification of mammographic microcalcifications using image structure features, Proc SPIE Medical Imaging, 1905, 820, 1993 112 Woods, K.S., Solka, J.L., Priebe, C.E et al., Comparative evaluation of pattern recognition techniques for detection of microcalcifications, Proc SPIE Medical Imaging, 1905, 841, 1993 113 Chan, H.P., Niklason, L.T., Ikeda, D.M et al., Digitization requirements in mammography: effects on computer-aided detection of microcalcifications, Med Phys., 21, 1203, 1994 114 Fukushima, K., Miyake, S., and Ito, T., Neocognitron: a neural network model for a mechanism of visual pattern recognition, IEEE Trans Systems Man Cybernetics, SMC-13, 826, 1983 115 Lo, S.C.B., Chan, H.P., Lin, J.S et al., Artificial convolution neural network for medical-image pattern recognition, Neural Networks, 8, 1201, 1995 116 Sahiner, B., Chan, H.P., Petrick, N et al., Classification of mass and normal breast tissue: a convolution neural network classifier with spatial domain and texture images, IEEE Trans Medical Imaging, 15, 598, 1996 Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 47 Tuesday, May 10, 2005 3:38 PM Computer-Aided Diagnosis of Breast Cancer 47 117 Gurcan, M.N., Sahiner, B., Chan, H.P et al., Selection of an optimal neural network architecture for computer-aided detection of microcalcifications: comparison of automated optimization techniques, Med Phys., 28, 1937, 2001 118 Metz, C.E., Some practical issues of experimental design and data analysis in radiological ROC studies, Invest Radiol., 24, 234, 1989 119 Heath, M., Bowyer, K., Kopans, D et al., Current status of the digital database for screening mammography, in Digital Mammography, Karssemeijer, N., Thijssen, M., Hendriks, J et al., Eds., Kluwer Academic, Dordrecht, Netherlands, 1998, p 457 120 Bunch, P.C., Hamilton, J.F., Sanderson, G.K et al., A free response approach to the measurement and characterization of radiographic observer performance, Proc SPIE, 127, 124, 1977 121 Gurcan, M.N., Sahiner, B., Chan, H.P et al., Optimal selection of neural network architecture for CAD using simulated annealing, in Proc 22nd Annual International Conference of IEEE Engineering in Medicine and Biology Society, Chicago, 2000, p 3052 122 Gurcan, M.N., Sahiner, B., Chan, H.P et al., Selection of an optimal neural network architecture for computer-aided diagnosis: comparison of automated optimization techniques, Radiology, 217(P), 436, 2000 123 Gurcan, M.N., Chan, H.P., Sahiner, B et al., Improvement of Computerized Detection of Microcalcifications Using a Convolution Neural Network Architecture Selected by an Automated Optimization Algorithm, Presented at Medical Image Perception Conference IX, Airlie Conference Center, Warrenton, VA, 2001 124 Petrick, N., Chan, H.P., Sahiner, B et al., Automated detection of breast masses on digital mammograms using adaptive density-weighted contrast-enhancement filtering, Proc SPIE Medical Imaging, 2434, 590, 1995 125 Hadjiiski, L.M., Sahiner, B., Chan, H.P et al., Analysis of temporal change of mammographic features: computer-aided classification of malignant and benign breast masses, Med Phys., 28, 2309, 2001 126 Hadjiiski, L.M., Chan, H.P., Sahiner, B et al., Automated registration of breast lesions in temporal pairs of mammograms for interval change analysis: local affine transformation for improved localization, Med Phys., 28, 1070, 2001 127 Peli, T and Lim, J.S., Adaptive filtering for image enhancement, Optical Eng., 21, 108, 1982 128 Lunscher, W.H.H.J and Beddoes, M.P., Optimal edge detection: parameter selection and noise effects, IEEE Trans Pattern Anal Machine Intelligence, 8, 154, 1986 129 Marr, D and Hildreth, E., Theory of edge detection, Proc Royal Soc London, Series B, Biological Sci., 207, 187, 1980 130 Russ, J.C., The Image Processing Handbook, CRC Press, Boca Raton, FL, 1992 131 Chan, H.-P., Petrick, N., and Sahiner, B., Computer-aided breast cancer diagnosis, chap in Artificial Intelligence Techniques in Breast Cancer Diagnosis and Prognosis, Jain, A., Jain, A., Jain, S et al., Eds., World Scientific, River Edge, New Jersey, 2000, p 179 132 Sahiner, B., Chan, H.P., Petrick, N et al., Classification of mass and normal breast tissue: an artificial neural network with morphological features, Proc World Congress Neural Networks, II, 876, 1995 133 Sahiner, B., Chan, H.P., Petrick, N et al., Image feature selection by a genetic algorithm: application to classification of mass and normal breast tissue on mammograms, Med Phys., 23, 1671, 1996 134 Petrick, N., Sahiner, B., Chan, H.P et al., Preclinical evaluation of a CAD algorithm for early detection of breast cancer, in Proc 5th International Workshop on Digital Mammography, Toronto, 2001, p 328 Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 48 Tuesday, May 10, 2005 3:38 PM 48 Medical Image Analysis 135 Wei, D., Chan, H.P., Helvie, M.A et al., Classification of mass and normal breast tissue on digital mammograms: multiresolution texture analysis, Med Phys., 22, 1501, 1995 136 Wei, D., Chan, H.P., Petrick, N et al., False-positive reduction technique for detection of masses on digital mammograms: global and local multiresolution texture analysis, Med Phys., 24, 903, 1997 137 Chan, H.P., Wei, D., Helvie, M.A et al., Computer-aided classification of mammographic masses and normal tissue: linear discriminant analysis in texture feature space, Phys Med Biol., 40, 857, 1995 138 Haralick, R.M., Shanmugam, K., and Dinstein, I., Texture features for image classification, IEEE Trans Systems, Man, Cybernetics, SMC-3, 610, 1973 139 Wei, D., Chan, H.P., Helvie, M.A et al., Multiresolution texture analysis for classification of mass and normal breast tissue on digital mammograms, Proc SPIE Medical Imaging, 2434, 606, 1995 140 Norusis, M.J., SPSS for Windows, release 6, professional statistics software, SPSS Inc., Chicago, 1993 141 Birdwell, R.L., Ikeda, D.M., O'Shaughnessy, K.F et al., Mammographic characterization of 111 missed cancers later detected by screening mammography, Radiology, 213(P), 240, 1999 142 Brem, R.F., Schoonjans, J.M., Hoffmeister, J et al., Evaluation of breast cancer with a computer-aided detection system by mammographic appearance, histology and lesion size, Radiology, 217(P), 400, 2000 143 Castellino, R.A., Roehrig, J., and Zhang, W., Improved computer-aided detection (CAD) algorithm for screening mammography, Radiology, 217(P), 400, 2000 144 Thurfjell, E., Mammography screening: one vs two views and independent double reading, Acta Radiologica, 35, 345, 1994 145 Hadjiiski, L.M., Sahiner, B., Chan, H.P et al., Analysis of temporal change of mammographic features for computer-aided characterization of malignant and benign masses, Proc SPIE, 4322, 661, 2001 146 Highnam, R.P., Kita, Y., Brady, J.M et al., Determining correspondence between views, in Proc 4th International Workshop on Digital Mammography, Nijmegen, Netherlands, 1998 147 Kita, Y., Highnam, R.P., and Brady, J.M., Correspondence between two different views of X-ray mammograms using simulation of breast deformation, in Proc IEEE Conference on Computer Vision and Pattern Recognition, Santa Barbara, CA, 1998, p 700 148 Chang, Y.H., Good, W.F., Sumkin, J.H et al., Computerized localization of breast lesions from two views: an experimental comparison of two methods, Acad Radiol., 34, 585, 1999 149 Gopal, S.S., Chan, H.-P., Sahiner, B et al., Evaluation of interval change in mammographic features for computerized classification of malignant and benign masses, Radiology, 205(P), 216, 1997 150 Hadjiiski, L.M., Sahiner, B., Chan, H.P et al., Computer-aided classification of malignant and benign breast masses by analysis of interval change of features in temporal pairs of mammograms, Radiology, 217(P), 435, 2000 151 Sahiner, B., Gurcan, M.N., Chan, H.P et al., The use of joint two-view information for computerized lesion detection on mammograms: improvement of microcalcification detection accuracy, Proc SPIE Medical Imaging, 4684, 754, 2002 Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 49 Tuesday, May 10, 2005 3:38 PM Computer-Aided Diagnosis of Breast Cancer 49 152 Morton, A.R., Chan, H.P., and Goodsitt, M.M., Automated model-guided breast segmentation algorithm, Med Phys., 23, 1107, 1996 153 Zhou, C., Chan, H.P., Petrick, N et al., Computerized image analysis: breast segmentation and nipple identification on mammograms, in Proc Chicago 2000: World Congress on Medical Physics and Biomedical Engineering, Chicago, paper no TH, 2000 Copyright 2005 by Taylor & Francis Group, LLC [...]... nonlinear pixelwise transformation to generate the final “enhanced” image The two-step DWCE filtering is described as ( ) ( ( )) ( ) I W x, y = WD I D x, y ⋅ I C x, y ( ) ( ( )) I E x, y = W I W x, y (1.1) (1 .2) The multiplication factor and the nonlinear transformation function used in this application, WD(⋅) and W(⋅), can be found in the literature [62] The DWCE filter suppresses very low-contrast regions,... both independent test sets over a wide range of marker rates In the Group 1 database, 34% (49/146) of the malignant and 5% (8/159) of the benign masses were spiculated There were 33% (65/197) and 0% (0/1 32) spiculated Copyright 2005 by Taylor & Francis Group, LLC 2089_book.fm Page 26 Tuesday, May 10, 2005 3:38 PM 26 Medical Image Analysis 1.0 Individual Mass Scoring TP Fraction 0.8 0.6 0.4 Malignant (per-case)... preferable Good et al and Chang et al reported preliminary attempt of matching computer-detected objects in two views [69, 148] They demonstrated the feasibility of identifying corresponding objects (Az = 0. 82) in the two views by exhaustive pairing of the detected objects and feature classification None of these studies attempted to use the two-view correspondence information to improve lesion detection or... (correlation coefficient = 0.94) of the radial distances of the corresponding objects in the two views However, the angular coordinates in the two views are basically uncorrelated (correlation coefficient = 0. 42) A linear model with parameters ar and br was therefore chosen to predict the radial distance of an object in a second view from that in the first view: RM = ar ⋅ RC + br Copyright 2005 by Taylor & Francis