WE-E-217A-03: Methodologies for Evaluation of Effects of CAD on Users.

CAD devices are specifically designed to aid the user in the diagnostic process. Therefore, the user (e.g., clinician) is an integral part of the CAD interpretation process and the impact of CAD on users needs to be accounted for in the assessment process. We use the term "reader performance assessment" to denote a test specifically designed to estimate the performance of the user (or reader) interpreting CAD as part of the diagnostic process. These reader assessment studies are generally more indicative of clinical performance compared with standalone testing. In this course, we will update the audience on the latest approaches and metrics being used to assess the impact of CAD on user performance, and discuss the opinions of the AAPM CAD subcommittee on the evaluation methodologies, Learning Objectives: 1. To review the most common reader performance assessment techniques and associated performance metrics including the multireader multi-case (MRMC) study design. 2. To review the AAPM CAD subcommittee's current thinking on best practices for performing and analyzing CAD reader performance assessment studies. 3. To understand the importance and limitations of various reader performance assessment approaches and metrics for various classes of CAD devices.

Data sets for the qualification of volumetric CT as a quantitative imaging biomarker in lung cancer.

The drug development industry is faced with increasing costs and decreasing success rates. New ways to understand biology as well as the increasing interest in personalized treatments for smaller patient segments requires new capabilities for the rapid assessment of treatment responses. Deployment of qualified imaging biomarkers lags apparent technology capabilities. The lack of consensus methods and qualification evidence needed for large-scale multi-center trials, as well as the standardization that allows them, are widely acknowledged to be the limiting factors. The current fragmentation in imaging vendor offerings, coupled with the independent activities of individual biopharmaceutical companies and their contract research organizations (CROs), may stand in the way of the greater opportunity were these efforts to be drawn together. A preliminary report, "Volumetric CT: a potential biomarker of response," of the Quantitative Imaging Biomarkers Alliance (QIBA) activity was presented at the Medical Imaging Continuum: Path Forward for Advancing the Uses of Medical Imaging in the Development of New Biopharmaceutical Products meeting of the Extended Pharmaceutical Research and Manufacturers of America (PhRMA) Imaging Group sponsored by the Drug Information Agency (DIA) in October 2008. The clinical context in Lung Cancer and a methodology for approaching the qualification of volumetric CT as a biomarker has since been reported [Acad. Radiol. 17, 100-106, 107-115 (2010)]. This report reviews the effort to collect and utilize publicly available data sets to provide a transparent environment in which to pursue the qualification activities in such a way as to allow independent peer review and verification of results. This article focuses specifically on our role as stewards of image sets for developing new tools.

Change in lung tumor volume as a biomarker of treatment response: a critical review of the evidence.

SPECIFIC AIM: To review the evidence indicating that volumetric image analysis of computed tomography scans meets specifications for qualification as a biomarker in clinical trials or the management of individual patients with lung cancer. METHODS: Claims of value were broken down into questions about technical feasibility, accuracy, the precision of measurement, sensitivity, the correlations with health outcomes, and the risks of producing misleading information. For each claim, the pertinent literature was reviewed. RESULTS: Technical feasibility has now been shown, but only in limited contexts. Accuracy has been demonstrated, but only for tumors with favorable anatomical features. Measurement error still makes the assessment of change in small nodules precarious in diagnostic settings unless rigorous image acquisition and analysis procedures are followed. Precision is sufficient in some larger masses to make volumetrics a sensitive biomarker. In a few trials, correlations with clinical outcomes have been higher for volumetric-based measures than for unidimensional or bidimensional diameters. Value in clinical practice settings and clinical trials has been suggested, but not proven. CONCLUSION: The weight of the evidence indicates there are circumstances in which volumetric image analysis adds value to clinical trial science and the practice of medicine.

Imaging as a tumor biomarker in oncology drug trials for lung cancer: the FDA perspective.

The US Food and Drug Administration (FDA) is committed to working with the oncology community to expedite the drug evaluation process in view of the many promising new oncology drugs under laboratory development and the time and expense required for such new drugs to reach the patient population. One significant advance would be to enable quantitative imaging as a tumor biomarker. The FDA is working with the pharmaceutical industry, academia, and sister stakeholders in the government, primarily through collaborative educational and research efforts, to identify how imaging can serve this function.

Selection of an optimal neural network architecture for computer-aided detection of microcalcifications--comparison of automated optimization techniques.

Many computer-aided diagnosis (CAD) systems use neural networks (NNs) for either detection or classification of abnormalities. Currently, most NNs are "optimized" by manual search in a very limited parameter space. In this work, we evaluated the use of automated optimization methods for selecting an optimal convolution neural network (CNN) architecture. Three automated methods, the steepest descent (SD), the simulated annealing (SA), and the genetic algorithm (GA), were compared. We used as an example the CNN that classifies true and false microcalcifications detected on digitized mammograms by a prescreening algorithm. Four parameters of the CNN architecture were considered for optimization, the numbers of node groups and the filter kernel sizes in the first and second hidden layers, resulting in a search space of 432 possible architectures. The area Az under the receiver operating characteristic (ROC) curve 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. An available data set was split into two groups with approximately equal number of samples. By using the two groups alternately for training and testing, two different cost surfaces were evaluated. For the first cost surface, the SD method was trapped in a local minimum 91% (392/432) of the time. The SA using the Boltzman schedule selected the best architecture after evaluating, on average, 167 architectures. The GA achieved its best performance with linearly scaled roulette-wheel parent selection; however, it evaluated 391 different architectures, on average, to find the best one. The second cost surface contained no local minimum. For this surface, a simple SD algorithm could quickly find the global minimum, but the SA with the very fast reannealing schedule was still the most efficient. The same SA scheme, however, was trapped in a local minimum on the first cost surface. Our CNN study demonstrated that, if optimization is to be performed on a cost surface whose characteristics are not known a priori, it is advisable that a moderately fast algorithm such as a SA using a Boltzman cooling schedule be used to conduct an efficient and thorough search, which may offer a better chance of reaching the global minimum.

Improvement of mammographic mass characterization using spiculation meausures and morphological features.

We are developing new computer vision techniques for characterization of breast masses on mammograms. We had previously developed a characterization method based on texture features. The goal of the present work was to improve our characterization method by making use of morphological features. Toward this goal, we have developed a fully automated, three-stage segmentation method that includes clustering, active contour, and spiculation detection stages. After segmentation, morphological features describing the shape of the mass were extracted. Texture features were also extracted from a band of pixels surrounding the mass. Stepwise feature selection and linear discriminant analysis were employed in the morphological, texture, and combined feature spaces for classifier design. The classification accuracy was evaluated using the area Az under the receiver operating characteristic curve. A data set containing 249 films from 102 patients was used. When the leave-one-case-out method was applied to partition the data set into trainers and testers, the average test Az for the task of classifying the mass on a single mammographic view was 0.83 +/- 0.02, 0.84 +/- 0.02, and 0.87 +/- 0.02 in the morphological, texture, and combined feature spaces, respectively. The improvement obtained by supplementing texture features with morphological features in classification was statistically significant (p = 0.04). For classifying a mass as malignant or benign, we combined the leave-one-case-out discriminant scores from different views of a mass to obtain a summary score. In this task, the test Az value using the combined feature space was 0.91 +/- 0.02. Our results indicate that combining texture features with morphological features extracted from automatically segmented mass boundaries will be an effective approach for computer-aided characterization of mammographic masses.

Computerized image analysis: estimation of breast density on mammograms.

An automated image analysis tool is being developed for the estimation of mammographic breast density. This tool may be useful for risk estimation or for monitoring breast density change in prevention or intervention programs. In this preliminary study, a data set of 4-view mammograms from 65 patients was used to evaluate our approach. Breast density analysis was performed on the digitized mammograms in three stages. First, the breast region was segmented from the surrounding background by an automated breast boundary-tracking algorithm. Second, an adaptive dynamic range compression technique was applied to the breast image to reduce the range of the gray level distribution in the low frequency background and to enhance the differences in the characteristic features of the gray level histogram for breasts of different densities. Third, rule-based classification was used to classify the breast images into four classes according to the characteristic features of their gray level histogram. For each image, a gray level threshold was automatically determined to segment the dense tissue from the breast region. The area of segmented dense tissue as a percentage of the breast area was then estimated. To evaluate the performance of the algorithm, the computer segmentation results were compared to manual segmentation with interactive thresholding by five radiologists. A "true" percent dense area for each mammogram was obtained by averaging the manually segmented areas of the radiologists. We found that the histograms of 6% (8 CC and 8 MLO views) of the breast regions were misclassified by the computer, resulting in poor segmentation of the dense region. For the images with correct classification, the correlation between the computer-estimated percent dense area and the "truth" was 0.94 and 0.91, respectively, for CC and MLO views, with a mean bias of less than 2%. The mean biases of the five radiologists' visual estimates for the same images ranged from 0.1% to 11%. The results demonstrate the feasibility of estimating mammographic breast density using computer vision techniques and its potential to improve the accuracy and reproducibility of breast density estimation in comparison with the subjective visual assessment by radiologists.

Automated registration of breast lesions in temporal pairs of mammograms for interval change analysis--local affine transformation for improved localization.

Analysis of interval change is important for mammographic interpretation. The aim of this study is to evaluate the use of an automated registration technique for computer-aided interval change analysis in mammography. Previously we developed a regional registration technique for identifying masses on temporal pairs of mammograms. In the current study, we improved lesion registration by including a local alignment step. Initially, the lesion position on the prior mammogram was estimated based on the breast geometry. An initial fan-shaped search region was then defined on the prior mammogram. In the second stage, the location of the fan-shaped region on the prior mammogram was refined by warping, based on an affine transformation and simplex optimization in a local region. In the third stage, a search for the best match between the lesion template from the current mammogram and a structure on the prior mammogram was carried out within the search region. This technique was evaluated on 124 temporal pairs of mammograms containing biopsyproven masses. Eighty-seven percent of the estimated lesion locations resulted in an area overlap of at least 50% with the true lesion locations and an average distance of 2.4 +/- 2.1 mm between their centroids. The average distance between the estimated and the true centroid of the lesions on the prior mammogram over all 124 temporal pairs was 4.2 +/- 5.7 mm. The registration accuracy was improved in comparison with our previous study that used a data set of 74 temporal pairs of mammograms. This improvement in accuracy resulted from the improved geometry estimation and the local affine transformation.

Digital mammography: observer performance study of the effects of pixel size on the characterization of malignant and benign microcalcifications.

RATIONALE AND OBJECTIVES: The authors performed this study to evaluate the effects of pixel size on the characterization of mammographic microcalcifications by radiologists. MATERIALS AND METHODS: Two-view mammograms of 112 microcalcification clusters were digitized with a laser scanner at a pixel size of 35 microm. Images with pixel sizes of 70, 105, and 140 microm were derived from the 35-microm-pixel size images by averaging neighboring pixels. The malignancy or benignity of the microcalcifications had been determined with findings at biopsy or 2-year follow-up. Region-of-interest images containing the microcalcifications were printed with a laser imager. Seven radiologists participated in a receiver operating characteristic (ROC) study to estimate the likelihood of malignancy. The classification accuracy was quantified with the area under the ROC curve (Az). The statistical significance of the differences in the Az values for different pixel sizes was estimated with the Dorfman-Berbaum-Metz method and the Student paired t test. The variance components were analyzed with a bootstrap method. RESULTS: The higher-resolution images did not result in better classification; the average Az with a pixel size of 35 microm was lower than that with pixel sizes of 70 and 105 microm. The differences in Az between different pixel sizes did not achieve statistical significance. CONCLUSION: Pixel sizes in the range studied do not have a strong effect on radiologists' accuracy in the characterization of microcalcifications. The low specificity of the image features of microcalcifications and the large interobserver and intraobserver variabilities may have prevented small advantages in image resolution from being observed.

Computer-aided characterization of mammographic masses: accuracy of mass segmentation and its effects on characterization.

Mass segmentation is used as the first step in many computer-aided diagnosis (CAD) systems for classification of breast masses as malignant or benign. The goal of this paper was to study the accuracy of an automated mass segmentation method developed in our laboratory, and to investigate the effect of the segmentation stage on the overall classification accuracy. The automated segmentation method was quantitatively compared with manual segmentation by two expert radiologists (R1 and R2) using three similarity or distance measures on a data set of 100 masses. The area overlap measures between R1 and R2, the computer and R1, and the computer and R2 were 0.76 +/- 0.13, 0.74 +/- 0.11, and 0.74 +/- 0.13, respectively. The interobserver difference in these measures between the two radiologists was compared with the corresponding differences between the computer and the radiologists. Using three similarity measures and data from two radiologists, a total of six statistical tests were performed. The difference between the computer and the radiologist segmentation was significantly larger than the interobserver variability in only one test. Two sets of texture, morphological, and spiculation features, one based on the computer segmentation, and the other based on radiologist segmentation, were extracted from a data set of 249 films from 102 patients. A classifier based on stepwise feature selection and linear discriminant analysis was trained and tested using the two feature sets. The leave-one-case-out method was used for data sampling. For case-based classification, the area Az under the receiver operating characteristic (ROC) curve was 0.89 and 0.88 for the feature sets based on the radiologist segmentation and computer segmentation, respectively. The difference between the two ROC curves was not statistically significant.

Analysis of temporal changes of mammographic features: computer-aided classification of malignant and benign breast masses.

A new classification scheme was developed to classify mammographic masses as malignant and benign by using interval change information. The masses on both the current and the prior mammograms were automatically segmented using an active contour method. From each mass, 20 run length statistics (RLS) texture features, 3 speculation features, and 12 morphological features were extracted. Additionally, 20 difference RLS features were obtained by subtracting the prior RLS features from the corresponding current RLS features. The feature space consisted of the current RLS features, the difference RLS features, the current and prior speculation features, and the current and prior mass sizes. Stepwise feature selection and linear discriminant analysis classification were used to select and merge the most useful features. A leave-one-case-out resampling scheme was used to train and test the classifier using 140 temporal image pairs (85 malignant, 55 benign) obtained from 57 biopsy-proven masses (33 malignant, 24 benign) in 56 patients. An average of 10 features were selected from the 56 training subsets: 4 difference RLS features, 4 RLS features, and 1 speculation feature from the current image, and 1 speculation feature from the prior, were most often chosen. The classifier achieved an average training Az of 0.92 and a test Az of 0.88. For comparison, a classifier was trained and tested using features extracted from the 120 current single images. This classifier achieved an average training Az of 0.90 and a test Az of 0.82. The information on the prior image significantly (p = 0.015) improved the accuracy for classification of the masses.

Feature selection and classifier performance in computer-aided diagnosis: the effect of finite sample size.

In computer-aided diagnosis (CAD), a frequently used approach for distinguishing normal and abnormal cases is first to extract potentially useful features for the classification task. Effective features are then selected from this entire pool of available features. Finally, a classifier is designed using the selected features. In this study, we investigated the effect of finite sample size on classification accuracy when classifier design involves stepwise feature selection in linear discriminant analysis, which is the most commonly used feature selection algorithm for linear classifiers. The feature selection and the classifier coefficient estimation steps were considered to be cascading stages in the classifier design process. We compared the performance of the classifier when feature selection was performed on the design samples alone and on the entire set of available samples, which consisted of design and test samples. The area Az under the receiver operating characteristic curve was used as our performance measure. After linear classifier coefficient estimation using the design samples, we studied the hold-out and resubstitution performance estimates. The two classes were assumed to have multidimensional Gaussian distributions, with a large number of features available for feature selection. We investigated the dependence of feature selection performance on the covariance matrices and means for the two classes, and examined the effects of sample size, number of available features, and parameters of stepwise feature selection on classifier bias. Our results indicated that the resubstitution estimate was always optimistically biased, except in cases where the parameters of stepwise feature selection were chosen such that too few features were selected by the stepwise procedure. When feature selection was performed using only the design samples, the hold-out estimate was always pessimistically biased. When feature selection was performed using the entire finite sample space, the hold-out estimates could be pessimistically or optimistically biased, depending on the number of features available for selection, the number of available samples, and their statistical distribution. For our simulation conditions, these estimates were always pessimistically (conservatively) biased if the ratio of the total number of available samples per class to the number of available features was greater than five.

Combined adaptive enhancement and region-growing segmentation of breast masses on digitized mammograms.

As an ongoing effort to develop a computer aid for detection of masses on mammograms, we recently designed an object-based region-growing technique to improve mass segmentation. This segmentation method utilizes the density-weighted contrast enhancement (DWCE) filter as a pre-processing step. The DWCE filter adaptively enhances the contrast between the breast structures and the background. Object-based region growing was then applied to each of the identified structures. The region-growing technique uses gray-scale and gradient information to adjust the initial object borders and to reduce merging between adjacent or overlapping structures. Each object is then classified as a breast mass or normal tissue based on extracted morphological and texture features. In this study we evaluated the sensitivity of this combined segmentation scheme and its ability to reduce false positive (FP) detections on a data set of 253 digitized mammograms, each of which contained a biopsy-proven breast mass. It was found that the segmentation scheme detected 98% of the 253 biopsy-proven breast masses in our data set. After final FP reduction, the detection resulted in 4.2 FP per image at a 90% true positive (TP) fraction and 2.0 FPs per image at an 80% TP fraction. The combined DWCE and object-based region growing technique increased the initial detection sensitivity, reduced merging between neighboring structures, and reduced the number of FP detections in our automated breast mass detection scheme.

Improvement of radiologists' characterization of mammographic masses by using computer-aided diagnosis: an ROC study.

PURPOSE: To evaluate the effects of computer-aided diagnosis (CAD) on radiologists' classification of malignant and benign masses seen on mammograms. MATERIALS AND METHODS: The authors previously developed an automated computer program for estimation of the relative malignancy rating of masses. In the present study, the authors conducted observer performance experiments with receiver operating characteristic (ROC) methodology to evaluate the effects of computer estimates on radiologists' confidence ratings. Six radiologists assessed biopsy-proved masses with and without CAD. Two experiments, one with a single view and the other with two views, were conducted. The classification accuracy was quantified by using the area under the ROC curve, Az. RESULTS: For the reading of 238 images, the Az value for the computer classifier was 0.92. The radiologists' Az values ranged from 0.79 to 0.92 without CAD and improved to 0.87-0.96 with CAD. For the reading of a subset of 76 paired views, the radiologists' Az values ranged from 0.88 to 0.95 without CAD and improved to 0.93-0.97 with CAD. Improvements in the reading of the two sets of images were statistically significant (P = .022 and .007, respectively). An improved positive predictive value as a function of the false-negative fraction was predicted from the improved ROC curves. CONCLUSION: CAD may be useful for assisting radiologists in classification of masses and thereby potentially help reduce unnecessary biopsies.

A regional registration technique for automated interval change analysis of breast lesions on mammograms.

Analysis of interval change is a useful technique for detection of abnormalities in mammographic interpretation. Interval change analysis is routinely used by radiologists and its importance is well-established in clinical practice. As a first step to develop a computerized method for interval change analysis on mammograms, we are developing an automated regional registration technique to identify corresponding lesions on temporal pairs of mammograms. In this technique, the breast is first segmented from the background on the current and previous mammograms. The breast edges are then aligned using a global alignment procedure based on the mutual information between the breast regions in the two images. Using the nipple location and the breast centroid estimated independently on both mammograms, a polar coordinate system is defined for each image. The polar coordinate of the centroid of a lesion detected on the most recent mammogram is used to obtain an initial estimate of its location on the previous mammogram and to define a fan-shaped search region. A search for a matching structure to the lesion is then performed in the fan-shaped region on the previous mammogram to obtain a final estimate of its location. In this study, a quantitative evaluation of registration accuracy has been performed with a data set of 74 temporal pairs of mammograms and ground-truth correspondence information provided by an experienced radiologist. The most recent mammogram of each temporal pair exhibited a biopsy-proven mass. We have investigated the usefulness of correlation and mutual information as search criteria for determining corresponding regions on mammograms for the biopsy-proven masses. In 85% of the cases (63/74 temporal pairs) the region on the previous mammogram that corresponded to the mass on the current mammogram was correctly identified. The region centroid identified by the registration technique had an average distance of 2.8+/-1.9 mm from the centroid of the radiologist-identified region. These results indicate that our new registration technique may be useful for establishing correspondence between structures on current and previous mammograms. Once such a correspondence is established an interval change analysis could be performed to aid in both detection as well as classification of abnormal breast densities.

Classifier design for computer-aided diagnosis: effects of finite sample size on the mean performance of classical and neural network classifiers.

Classifier design is one of the key steps in the development of computer-aided diagnosis (CAD) algorithms. A classifier is designed with case samples drawn from the patient population. Generally, the sample size available for classifier design is limited, which introduces variance and bias into the performance of the trained classifier, relative to that obtained with an infinite sample size. For CAD applications, a commonly used performance index for a classifier is the area, Az, under the receiver operating characteristic (ROC) curve. We have conducted a computer simulation study to investigate the dependence of the mean performance, in terms of Az, on design sample size for a linear discriminant and two nonlinear classifiers, the quadratic discriminant and the backpropagation neural network (ANN). The performances of the classifiers were compared for four types of class distributions that have specific properties: multivariate normal distributions with equal covariance matrices and unequal means, unequal covariance matrices and unequal means, and unequal covariance matrices and equal means, and a feature space where the two classes were uniformly distributed in disjoint checkerboard regions. We evaluated the performances of the classifiers in feature spaces of dimensionality ranging from 3 to 15, and design sample sizes from 20 to 800 per class. The dependence of the resubstitution and hold-out performance on design (training) sample size (Nt) was investigated. For multivariate normal class distributions with equal covariance matrices, the linear discriminant is the optimal classifier. It was found that its Az-versus-1/Nt curves can be closely approximated by linear dependences over the range of sample sizes studied. In the feature spaces with unequal covariance matrices where the quadratic discriminant is optimal, the linear discriminant is inferior to the quadratic discriminant or the ANN when the design sample size is large. However, when the design sample is small, a relatively simple classifier, such as the linear discriminant or an ANN with very few hidden nodes, may be preferred because performance bias increases with the complexity of the classifier. In the regime where the classifier performance is dominated by the 1/Nt term, the performance in the limit of infinite sample size can be estimated as the intercept (1/Nt= 0) of a linear regression of Az versus 1/Nt. The understanding of the performance of the classifiers under the constraint of a finite design sample size is expected to facilitate the selection of a proper classifier for a given classification task and the design of an efficient resampling scheme.

Classification of malignant and benign masses based on hybrid ART2LDA approach.

A new type of classifier combining an unsupervised and a supervised model was designed and applied to classification of malignant and benign masses on mammograms. The unsupervised model was based on an adaptive resonance theory (ART2) network which clustered the masses into a number of separate classes. The classes were divided into two types: one containing only malignant masses and the other containing a mix of malignant and benign masses. The masses from the malignant classes were classified by ART2. The masses from the mixed classes were input to a supervised linear discriminant classifier (LDA). In this way, some malignant masses were separated and classified by ART2 and the less distinguishable benign and malignant masses were classified by LDA. For the evaluation of classifier performance, 348 regions of interest (ROI's) containing biopsy proven masses (169 benign and 179 malignant) were used. Ten different partitions of training and test groups were randomly generated using an average of 73% of ROI's for training and 27% for testing. Classifier design, including feature selection and weight optimization, was performed with the training group. The test group was kept independent of the training group. The performance of the hybrid classifier was compared to that of an LDA classifier alone and a backpropagation neural network (BPN). Receiver operating characteristics (ROC) analysis was used to evaluate the accuracy of the classifiers. The average area under the ROC curve (A(z)) for the hybrid classifier was 0.81 as compared to 0.78 for the LDA and 0.80 for the BPN. The partial areas above a true positive fraction of 0.9 were 0.34, 0.27 and 0.31 for the hybrid, the LDA and the BPN classifier, respectively. These results indicate that the hybrid classifier is a promising approach for improving the accuracy of classification in CAD applications.

Computerized analysis of mammographic microcalcifications in morphological and texture feature spaces.

We are developing computerized feature extraction and classification methods to analyze malignant and benign microcalcifications on digitized mammograms. Morphological features that described the size, contrast, and shape of microcalcifications and their variations within a cluster were designed to characterize microcalcifications segmented from the mammographic background. Texture features were derived from the spatial gray-level dependence (SGLD) matrices constructed at multiple distances and directions from tissue regions containing microcalcifications. A genetic algorithm (GA) based feature selection technique was used to select the best feature subset from the multi-dimensional feature spaces. The GA-based method was compared to the commonly used feature selection method based on the stepwise linear discriminant analysis (LDA) procedure. Linear discriminant classifiers using the selected features as input predictor variables were formulated for the classification task. The discriminant scores output from the classifiers were analyzed by receiver operating characteristic (ROC) methodology and the classification accuracy was quantified by the area, Az, under the ROC curve. We analyzed a data set of 145 mammographic microcalcification clusters in this study. It was found that the feature subsets selected by the GA-based method are comparable to or slightly better than those selected by the stepwise LDA method. The texture features (Az = 0.84) were more effective than morphological features (Az = 0.79) in distinguishing malignant and benign microcalcifications. The highest classification accuracy (Az = 0.89) was obtained in the combined texture and morphological feature space. The improvement was statistically significant in comparison to classification in either the morphological (p = 0.002) or the texture (p = 0.04) feature space alone. The classifier using the best feature subset from the combined feature space and an appropriate decision threshold could correctly identify 35% of the benign clusters without missing a malignant cluster. When the average discriminant score from all views of the same cluster was used for classification, the Az value increased to 0.93 and the classifier could identify 50% of the benign clusters at 100% sensitivity for malignancy. Alternatively, if the minimum discriminant score from all views of the same cluster was used, the Az value would be 0.90 and a specificity of 32% would be obtained at 100% sensitivity. The results of this study indicate the potential of using combined morphological and texture features for computer-aided classification of microcalcifications.

Design of a high-sensitivity classifier based on a genetic algorithm: application to computer-aided diagnosis.

A genetic algorithm (GA) based feature selection method was developed for the design of high-sensitivity classifiers, which were tailored to yield high sensitivity with high specificity. The fitness function of the GA was based on the receiver operating characteristic (ROC) partial area index, which is defined as the average specificity above a given sensitivity threshold. The designed GA evolved towards the selection of feature combinations which yielded high specificity in the high-sensitivity region of the ROC curve, regardless of the performance at low sensitivity. This is a desirable quality of a classifier used for breast lesion characterization, since the focus in breast lesion characterization is to diagnose correctly as many benign lesions as possible without missing malignancies. The high-sensitivity classifier, formulated as the Fisher's linear discriminant using GA-selected feature variables, was employed to classify 255 biopsy-proven mammographic masses as malignant or benign. The mammograms were digitized at a pixel size of 0.1 mm x 0.1 mm, and regions of interest (ROIs) containing the biopsied masses were extracted by an experienced radiologist. A recently developed image transformation technique, referred to as the rubber-band straightening transform, was applied to the ROIs. Texture features extracted from the spatial grey-level dependence and run-length statistics matrices of the transformed ROIs were used to distinguish malignant and benign masses. The classification accuracy of the high-sensitivity classifier was compared with that of linear discriminant analysis with stepwise feature selection (LDAsfs). With proper GA training, the ROC partial area of the high-sensitivity classifier above a true-positive fraction of 0.95 was significantly larger than that of LDAsfs, although the latter provided a higher total area (Az) under the ROC curve. By setting an appropriate decision threshold, the high-sensitivity classifier and LDAsfs correctly identified 61% and 34% of the benign masses respectively without missing any malignant masses. Our results show that the choice of the feature selection technique is important in computer-aided diagnosis, and that the GA may be a useful tool for designing classifiers for lesion characterization.

Classification of compressed breast shapes for the design of equalization filters in x-ray mammography.

We are developing an external filter method for equalizing the x-ray exposure in mammography. Each filter is specially designed to match the shape of the compressed breast border and to preferentially attenuate the x-ray beam in the peripheral region of the breast. To be practical, this method should require the use of only a limited number of custom built filters. It is hypothesized that this would be possible if compressed breasts can be classified into a finite number of shapes. A study was performed to determine the number of shapes. Based on the parabolic appearance of the outer borders of compressed breasts in mammograms, the borders were fit with the polynomial equations y = ax2 + bx3 and y = ax2 + bx3 + cx4. The goodness-of-fit of these equations was compared. The a,b and a,b,c coefficients were employed in a K-Means clustering procedure to classify 470 CC-view and 484 MLO-view borders into 2-10 clusters. The mean coefficients of the borders within a given cluster defined the "filter" shape, and the individual borders were translated and rotated to best match that filter shape. The average rms differences between the individual borders and the "filter" were computed as were the standard deviations of those differences. The optimally shifted and rotated borders were refit with the above polynomial equations, and plotted for visual evaluation of clustering success. Both polynomial fits were adequate with rms errors of about 2 mm for the 2-coefficient equation, and about 1 mm for the 3-coefficient equation. Although the fits to the original borders were superior for the 3-coefficient equation, the matches to the "filter" borders determined by clustering were not significantly improved. A variety of modified clustering methods were developed and utilized, but none produced major improvements in clustering. Results indicate that 3 or 4 filter shapes may be adequate for each mammographic projection (CC- and MLO-view). To account for the wide variations in exposures observed at the peripheral regions of breasts classified to be of a particular shape, it may be necessary to employ different filters for thin, medium and thick breasts. Even with this added requirement, it should be possible to use a small number of filters as desired.