Performance gain in computer-assisted detection schemes by averaging scores generated from artificial neural networks with adaptive filtering

Computer-aided classification of mammographic masses using visually sensitive image features

PURPOSE

To develop a new computer-aided diagnosis (CAD) scheme that computes visually sensitive image features routinely used by radiologists to develop a machine learning classifier and distinguish between the malignant and benign breast masses detected from digital mammograms.

METHODS

An image dataset including 301 breast masses was retrospectively selected. From each segmented mass region, we computed image features that mimic five categories of visually sensitive features routinely used by radiologists in reading mammograms. We then selected five optimal features in the five feature categories and applied logistic regression models for classification. A new CAD interface was also designed to show lesion segmentation, computed feature values and classification score.

RESULTS

Areas under ROC curves (AUC) were 0.786±0.026 and 0.758±0.027 when to classify mass regions depicting on two view images, respectively. By fusing classification scores computed from two regions, AUC increased to 0.806±0.025.

CONCLUSION

This study demonstrated a new approach to develop CAD scheme based on 5 visually sensitive image features. Combining with a "visual aid" interface, CAD results may be much more easily explainable to the observers and increase their confidence to consider CAD generated classification results than using other conventional CAD approaches, which involve many complicated and visually insensitive texture features.

A new and fast image feature selection method for developing an optimal mammographic mass detection scheme

PURPOSE

Selecting optimal features from a large image feature pool remains a major challenge in developing computer-aided detection (CAD) schemes of medical images. The objective of this study is to investigate a new approach to significantly improve efficacy of image feature selection and classifier optimization in developing a CAD scheme of mammographic masses.

METHODS

An image dataset including 1600 regions of interest (ROIs) in which 800 are positive (depicting malignant masses) and 800 are negative (depicting CAD-generated false positive regions) was used in this study. After segmentation of each suspicious lesion by a multilayer topographic region growth algorithm, 271 features were computed in different feature categories including shape, texture, contrast, isodensity, spiculation, local topological features, as well as the features related to the presence and location of fat and calcifications. Besides computing features from the original images, the authors also computed new texture features from the dilated lesion segments. In order to select optimal features from this initial feature pool and build a highly performing classifier, the authors examined and compared four feature selection methods to optimize an artificial neural network (ANN) based classifier, namely: (1) Phased Searching with NEAT in a Time-Scaled Framework, (2) A sequential floating forward selection (SFFS) method, (3) A genetic algorithm (GA), and (4) A sequential forward selection (SFS) method. Performances of the four approaches were assessed using a tenfold cross validation method.

RESULTS

Among these four methods, SFFS has highest efficacy, which takes 3%-5% of computational time as compared to GA approach, and yields the highest performance level with the area under a receiver operating characteristic curve (AUC) = 0.864 ± 0.034. The results also demonstrated that except using GA, including the new texture features computed from the dilated mass segments improved the AUC results of the ANNs optimized using other three feature selection methods. In addition, among 271 features, the shape, local morphological features, fat and calcification based features were the most frequently selected features to build ANNs.

CONCLUSIONS

Although conventional GA is a powerful tool in optimizing classifiers used in CAD schemes of medical images, it is very computationally intensive. This study demonstrated that using a new SFFS based approach enabled to significantly improve efficacy of image feature selection for developing CAD schemes.

Optimization of Network Topology in Computer-Aided Detection Schemes Using Phased Searching with NEAT in a Time-Scaled Framework

In the field of computer-aided mammographic mass detection, many different features and classifiers have been tested. Frequently, the relevant features and optimal topology for the artificial neural network (ANN)-based approaches at the classification stage are unknown, and thus determined by trial-and-error experiments. In this study, we analyzed a classifier that evolves ANNs using genetic algorithms (GAs), which combines feature selection with the learning task. The classifier named "Phased Searching with NEAT in a Time-Scaled Framework" was analyzed using a dataset with 800 malignant and 800 normal tissue regions in a 10-fold cross-validation framework. The classification performance measured by the area under a receiver operating characteristic (ROC) curve was 0.856 ± 0.029. The result was also compared with four other well-established classifiers that include fixed-topology ANNs, support vector machines (SVMs), linear discriminant analysis (LDA), and bagged decision trees. The results show that Phased Searching outperformed the LDA and bagged decision tree classifiers, and was only significantly outperformed by SVM. Furthermore, the Phased Searching method required fewer features and discarded superfluous structure or topology, thus incurring a lower feature computational and training and validation time requirement. Analyses performed on the network complexities evolved by Phased Searching indicate that it can evolve optimal network topologies based on its complexification and simplification parameter selection process. From the results, the study also concluded that the three classifiers - SVM, fixed-topology ANN, and Phased Searching with NeuroEvolution of Augmenting Topologies (NEAT) in a Time-Scaled Framework - are performing comparably well in our mammographic mass detection scheme.

Association between computed tissue density asymmetry in bilateral mammograms and near-term breast cancer risk

This study investigated association between bilateral mammographic density asymmetry and near-term breast cancer risk. A data base of digital mammograms acquired from 690 women was retrospectively collected. All images were originally interpreted as negative by radiologists. During the next subsequent screening examinations (between 12 and 36 months later), 230 women were diagnosed positive for cancer, 230 were recalled for additional diagnostic workups and proved to be benign, and 230 remained negative (not recalled). We applied a computerized scheme to compute the differences of five image features between the left and right mammograms, and trained an artificial neural network (ANN) to compute a bilateral mammographic density asymmetry score. Odds ratios (ORs) were used to assess associations between the ANN-generated scores and risk of women having detectable cancers during the next screening examinations. A logistic regression method was applied to test for trend as a function of the increase in ANN-generated scores. The results were also compared with ORs computed using other existing cancer risk factors. The ORs showed an increasing risk trend with the increase in ANN-generated scores (from 1.00 to 9.07 between positive and negative case groups). The regression analysis also showed a significant increase trend in slope (p < 0.05). No significant increase trends of the ORs were found when using woman's age, subjectively rated breast density, or family history of breast cancer. This study demonstrated that the computed bilateral mammographic density asymmetry had potential to be used as a new risk factor to improve discriminatory power in predicting near-term risk of women developing breast cancer.

A GMM-based breast cancer risk stratification using a resonance-frequency electrical impedance spectroscopy

PURPOSE

The authors developed and tested a multiprobe-based resonance-frequency-based electrical impedance spectroscopy (REIS) system. The purpose of this study was to preliminarily assess the performance of this system in classifying younger women into two groups, those ultimately recommended for biopsy during imaging-based diagnostic workups that followed screening and those rated as negative during mammography.

METHODS

A seven probe-based REIS system was designed, assembled, and is currently being tested in the breast imaging facility. During an examination, contact is made with the nipple and six concentric points on the breast skin. For each measurement channel between the center probe and one of the six external probes, a set of electrical impedance spectroscopy (EIS) signal sweeps is performed and signal outputs ranging from 200 to 800 kHz at 5 kHz interval are recorded. An initial subset of 174 examinations from an ongoing prospective clinical study was selected for this preliminary analysis. An initial set of 35 features, 33 of which represented the corresponding EIS signal differences between the left and right breasts, was established. A Gaussian mixture model (GMM) classifier was developed to differentiate between "positive" (biopsy recommended) cases and "negative" (nonbiopsy) cases. Selecting an optimal feature set was performed using genetic algorithms with an area under a receiver operating characteristic curve (AUC) as the fitness criterion.

RESULTS

The recorded EIS signal sweeps showed that, in general, negative (nonbiopsy) examinations have a higher level of electrical impedance symmetry between the two breasts than positive (biopsy) examinations. Fourteen features were selected by genetic algorithm and used in the optimized GMM classifier. Using a leave-one-case-out test, the GMM classifier yielded a performance level of AUC = 0.78, which compared favorably to other three widely used classifiers including support vector machine, classification tree, and linear discriminant analysis. These results also suggest that the REIS signal based GMM classifier could be used as a prescreening tool to correctly identify a fraction of younger women at higher risk of developing breast cancer (i.e., 47% sensitivity at 90% specificity).

CONCLUSIONS

The study confirms that asymmetry in electrical impedance characteristics between two breasts provides valuable information regarding the presence of a developing breast abnormality; hence, REIS data may be useful in classifying younger women into two groups of "average" and "significantly higher than average" risk of having or developing a breast abnormality that would ultimately result in a later imaging-based recommendation for biopsy.

Computer-aided detection of breast masses depicted on full-field digital mammograms: a performance assessment

OBJECTIVES

To investigate the feasibility of converting a computer-aided detection (CAD) scheme for digitised screen-film mammograms to full-field digital mammograms (FFDMs) and assessing CAD performance on a large database.

METHODS

The database included 6478 FFDM images acquired on 1120 females, with 525 cancer cases and 595 negative cases. The database was divided into five case groups: (1) cancer detected during screening, (2) interval cancers, (3) "high-risk" recommended for surgical excision, (4) recalled but negative and (5) negative (not recalled). A previously developed CAD scheme for masses depicted on digitised images was converted and re-optimised for FFDM images while keeping the same image-processing structure. CAD performance was analysed on the entire database.

RESULTS

The case-based sensitivity was 75.6% (397/525) for the current mammograms and 40.8% (42/103) for the prior mammograms deemed negative during clinical interpretation but "visible" during retrospective review. The region-based sensitivity was 58.1% (618/1064) for the current mammograms and 28.4% (57/201) for the prior mammograms. The CAD scheme marked 55.7% (221/397) and 35.7% (15/42) of the masses on both views of the current and the prior examinations, respectively. The overall CAD-cued false-positive rate was 0.32 per image, ranging from 0.29 to 0.51 for the five case groups.

CONCLUSION

This study indicated that (1) digitised image-based CAD can be converted for FFDMs while performing at a comparable, or better, level; (2) CAD detects a substantial fraction of cancers depicted on prior examinations, albeit most having been marked only on one view; and (3) CAD tends to mark more false-positive results on "difficult" negative cases that are more visually difficult for radiologists to interpret.

A preliminary evaluation of multi-probe resonance-frequency electrical impedance based measurements of the breast

RATIONALE AND OBJECTIVES

The aim of this study was to preliminarily assess the performance of a new, resonance-frequency electrical impedance spectroscopy (REIS) system in identifying young women who were recommended to undergo breast biopsy following imaging.

MATERIALS AND METHODS

A seven-probe REIS system was designed and assembled and is currently being prospectively tested. During examination, contact is made with the nipple and six concentric points on the breast skin. Signal sweeps are performed, and outputs ranging from 200 to 800 kHz at 5-kHz intervals are recorded. An initial set of 140 patients, including 56 who eventually had biopsies, 63 who had negative results on screening mammography, and 21 recalled for additional imaging but later determined to have negative results, was used. An initial set of 35 features, 33 representing impedance signal differences between breasts and two representing participant age and average breast density, was assembled and reduced by a genetic algorithm to 14. The performance of an artificial neural network-based classifier was assessed using a case-based leave-one-out method.

RESULTS

The substantially greater asymmetry between signals of mirror-matched regions ascertained from biopsy ("positive") compared to nonbiopsy ("negative") cases resulted in an artificial neural network classifier performance (area under the curve) of 0.830 ± 0.023. At 90% specificity, this classifier, optimized for "recommendation for biopsy" rather than "cancer," detected 30 REIS-positive cases (54%), including six of nine (67%) actual cancer cases and six of nine women (67%) recommended for surgical excision of high-risk lesions.

CONCLUSIONS

Asymmetry in impedance measurements between bilateral breasts may provide valuable discriminatory information regarding the presence of highly suspicious imaging-based findings.

Computer-aided detection of early interstitial lung diseases using low-dose CT images

This study aims to develop a new computer-aided detection (CAD) scheme to detect early interstitial lung disease (ILD) using low-dose computed tomography (CT) examinations. The CAD scheme classifies each pixel depicted on the segmented lung areas into positive or negative groups for ILD using a mesh-grid-based region growth method and a multi-feature-based artificial neural network (ANN). A genetic algorithm was applied to select optimal image features and the ANN structure. In testing each CT examination, only pixels selected by the mesh-grid region growth method were analyzed and classified by the ANN to improve computational efficiency. All unselected pixels were classified as negative for ILD. After classifying all pixels into the positive and negative groups, CAD computed a detection score based on the ratio of the number of positive pixels to all pixels in the segmented lung areas, which indicates the likelihood of the test case being positive for ILD. When applying to an independent testing dataset of 15 positive and 15 negative cases, the CAD scheme yielded the area under receiver operating characteristic curve (AUC = 0.884 ± 0.064) and 80.0% sensitivity at 85.7% specificity. The results demonstrated the feasibility of applying the CAD scheme to automatically detect early ILD using low-dose CT examinations.

Computer-aided detection; the effect of training databases on detection of subtle breast masses

RATIONALE AND OBJECTIVES

Lesion conspicuity is typically highly correlated with visual difficulty for lesion detection, and computer-aided detection (CAD) has been widely used as a "second reader" in mammography. Hence, increasing CAD sensitivity in detecting subtle cancers without increasing false-positive rates is important. The aim of this study was to investigate the effect of training database case selection on CAD performance in detecting low-conspicuity breast masses.

MATERIALS AND METHODS

A full-field digital mammographic image database that included 525 cases depicting malignant masses was randomly partitioned into three subsets. A CAD scheme was applied to detect all initially suspected mass regions and compute region conspicuity. Training samples were iteratively selected from two of the subsets. Four types of training data sets-(1) one including all available true-positive mass regions in the two subsets ("all"), (2) one including 350 randomly selected mass regions ("diverse"), (3) one including 350 high-conspicuity mass regions ("easy"), and (4) one including 350 low-conspicuity mass regions ("difficult")-were assembled. In each training data set, the same number of randomly selected false-positive regions as the true-positives were also included. Two classifiers, an artificial neural network (ANN) and a k-nearest neighbor (KNN) algorithm, were trained using each of the four training data sets and tested on all suspected regions in the remaining data set. Using a threefold cross-validation method, the performance changes of the CAD schemes trained using one of the four training data sets were computed and compared.

RESULTS

CAD initially detected 1025 true-positive mass regions depicted on 507 cases (97% case-based sensitivity) and 9569 false-positive regions (3.5 per image) in the entire database. Using the all training data set, CAD achieved the highest overall performance on the entire testing database. However, CAD detected the highest number of low-conspicuity masses when the difficult training data set was used for training. Results did agree for both ANN-based and KNN-based classifiers in all tests. Compared to the use of the all training data set, the sensitivity of the schemes trained using the difficult data set decreased by 8.6% and 8.4% for the ANN and KNN algorithm on the entire database, respectively, but the detection of low-conspicuity masses increased by 7.1% and 15.1% for the ANN and KNN algorithm at a false-positive rate of 0.3 per image.

CONCLUSIONS

CAD performance depends on the size, diversity, and difficulty level of the training database. To increase CAD sensitivity in detecting subtle cancer, one should increase the fraction of difficult cases in the training database rather than simply increasing the training data set size.

An adaptive incremental approach to constructing ensemble classifiers: application in an information-theoretic computer-aided decision system for detection of masses in mammograms

Ensemble classifiers have been shown efficient in multiple applications. In this article, the authors explore the effectiveness of ensemble classifiers in a case-based computer-aided diagnosis system for detection of masses in mammograms. They evaluate two general ways of constructing subclassifiers by resampling of the available development dataset: Random division and random selection. Furthermore, they discuss the problem of selecting the ensemble size and propose two adaptive incremental techniques that automatically select the size for the problem at hand. All the techniques are evaluated with respect to a previously proposed information-theoretic CAD system (IT-CAD). The experimental results show that the examined ensemble techniques provide a statistically significant improvement (AUC = 0.905 +/- 0.024) in performance as compared to the original IT-CAD system (AUC = 0.865 +/- 0.029). Some of the techniques allow for a notable reduction in the total number of examples stored in the case base (to 1.3% of the original size), which, in turn, results in lower storage requirements and a shorter response time of the system. Among the methods examined in this article, the two proposed adaptive techniques are by far the most effective for this purpose. Furthermore, the authors provide some discussion and guidance for choosing the ensemble parameters.

Automated detection and analysis of fluorescent in situ hybridization spots depicted in digital microscopic images of Pap-smear specimens

Fluorescence in situ hybridization (FISH) technology has been widely recognized as a promising molecular and biomedical optical imaging tool to screen and diagnose cervical cancer. However, manual FISH analysis is time-consuming and may introduce large inter-reader variability. In this study, a computerized scheme is developed and tested. It automatically detects and analyzes FISH spots depicted on microscopic fluorescence images. The scheme includes two stages: (1) a feature-based classification rule to detect useful interphase cells, and (2) a knowledge-based expert classifier to identify splitting FISH spots and improve the accuracy of counting independent FISH spots. The scheme then classifies detected analyzable cells as normal or abnormal. In this study, 150 FISH images were acquired from Pap-smear specimens and examined by both an experienced cytogeneticist and the scheme. The results showed that (1) the agreement between the cytogeneticist and the scheme was 96.9% in classifying between analyzable and unanalyzable cells (Kappa=0.917), and (2) agreements in detecting normal and abnormal cells based on FISH spots were 90.5% and 95.8% with Kappa=0.867. This study demonstrated the feasibility of automated FISH analysis, which may potentially improve detection efficiency and produce more accurate and consistent results than manual FISH analysis.

Development and Assessment of an Integrated Computer-Aided Detection Scheme for Digital Microscopic Images of Metaphase Chromosomes

The authors developed an integrated computer-aided detection (CAD) scheme for detecting and classifying metaphase chromosomes as well as assessing its performance and robustness. This scheme includes an automatic metaphase-finding module and a karyotyping module and it was applied to a testing database with 200 digital microscopic images. The automatic metaphase-finding module detects analyzable metaphase cells using a feature-based artificial neural network (ANN). The ANN-generated outputs are analyzed by a receiver operating characteristics (ROC) method and an area under the ROC curve is 0.966. Then, the automatic karyotyping module classifies individual chromosomes of this cell into 24 types. In this module, a two-layer decision tree-based classifier with eight ANNs established in its connection nodes was optimized by a genetic algorithm. Chromosomes are first classified into seven groups by the ANN in the first layer. The chromosomes in these groups are then separately classified by seven ANNs into 24 types in the second layer. The classification accuracy is 94.5% in the first layer. Six ANNs achieved the accuracy above 95% and only one had lessened performance (80.6%) in the second layer. The overall classification accuracy is 91.5% as compared to 86.7% in the previous study using two independent datasets randomly acquired from our genetic laboratory. The results demonstrate that our automated scheme achieves high and robust performance in identification and classification of metaphase chromosomes.

Automated classification of metaphase chromosomes: optimization of an adaptive computerized scheme

We developed and tested a new automated chromosome karyotyping scheme using a two-layer classification platform. Our hypothesis is that by selecting most effective feature sets and adaptively optimizing classifiers for the different groups of chromosomes with similar image characteristics, we can reduce the complexity of automated karyotyping scheme and improve its performance and robustness. For this purpose, we assembled an image database involving 6900 chromosomes and implemented a genetic algorithm to optimize the topology of multi-feature based artificial neural networks (ANN). In the first layer of the scheme, a single ANN was employed to classify 24 chromosomes into seven classes. In the second layer, seven ANNs were adaptively optimized for seven classes to identify individual chromosomes. The scheme was optimized and evaluated using a "training-testing-validation" method. In the first layer, the classification accuracy for the validation dataset was 92.9%. In the second layer, classification accuracy of seven ANNs ranged from 67.5% to 97.5%, in which six ANNs achieved accuracy above 93.7% and only one had lessened performance. The maximum difference of classification accuracy between the testing and validation datasets is <1.7%. The study demonstrates that this new scheme achieves higher and robust performance in classifying chromosomes.

Multiview-based computer-aided detection scheme for breast masses

In this study, we developed and tested a new multiview-based computer-aided detection (CAD) scheme that aims to maintain the same case-based sensitivity level as a single-image-based scheme while substantially increasing the number of masses being detected on both ipsilateral views. An image database of 450 four-view examinations (1800 images) was assembled. In this database, 250 cases depicted malignant masses, of which 236 masses were visible on both views and 14 masses were visible only on one view. First, we detected suspected mass regions depicted on each image in the database using a single-image-based CAD. For each identified region (with detection score > or = 0.55), we then identified a matching strip of interest on the ipsilateral view based on the projected distance to the nipple along the centerline. By lowering CAD operating threshold inside the matching strip, we searched for a region located inside the strip and paired it with the original region. A multifeature-based artificial neural network scored the likelihood of the paired "matched" regions representing true-positive masses. All single (unmatched) regions except for those either with very high detection scores (> or = 0.85) or those located near the chest wall that cannot be matched on the other view were discarded. The original single-image-based CAD scheme detected 186 masses (74.4% case-based sensitivity) and 593 false-positive regions. Of the 186 identified masses, 91 were detected on two views (48.9%) and 95 were detected only on one view (51.1%). Of the false-positive detections, 54 were paired on the ipsilateral view inside the corresponding matching strips and the remaining 485 were not, which represented 539 case-based false-positive detections (0.3 per image). Applying the multiview-based CAD scheme, the same case-based sensitivity was maintained while cueing 169 of 186 masses (90.9%) on both views and at the same time reducing the case-based false-positive detection rate by 23.7% (from 539 to 411). The study demonstrated that the new multiview-based CAD scheme could substantially increase the number of masses being cued on two ipsilateral views while reducing the case-based false-positive detection rate.

A completely automated CAD system for mass detection in a large mammographic database

Mass localization plays a crucial role in computer-aided detection (CAD) systems for the classification of suspicious regions in mammograms. In this article we present a completely automated classification system for the detection of masses in digitized mammographic images. The tool system we discuss consists in three processing levels: (a) Image segmentation for the localization of regions of interest (ROIs). This step relies on an iterative dynamical threshold algorithm able to select iso-intensity closed contours around gray level maxima of the mammogram. (b) ROI characterization by means of textural features computed from the gray tone spatial dependence matrix (GTSDM), containing second-order spatial statistics information on the pixel gray level intensity. As the images under study were recorded in different centers and with different machine settings, eight GTSDM features were selected so as to be invariant under monotonic transformation. In this way, the images do not need to be normalized, as the adopted features depend on the texture only, rather than on the gray tone levels, too. (c) ROI classification by means of a neural network, with supervision provided by the radiologist's diagnosis. The CAD system was evaluated on a large database of 3369 mammographic images [2307 negative, 1062 pathological (or positive), containing at least one confirmed mass, as diagnosed by an expert radiologist]. To assess the performance of the system, receiver operating characteristic (ROC) and free-response ROC analysis were employed. The area under the ROC curve was found to be Az = 0.783 +/- 0.008 for the ROI-based classification. When evaluating the accuracy of the CAD against the radiologist-drawn boundaries, 4.23 false positives per image are found at 80% of mass sensitivity.

A method to improve visual similarity of breast masses for an interactive computer-aided diagnosis environment

The purpose of this study was to develop and test a method for selecting "visually similar" regions of interest depicting breast masses from a reference library to be used in an interactive computer-aided diagnosis (CAD) environment. A reference library including 1000 malignant mass regions and 2000 benign and CAD-generated false-positive regions was established. When a suspicious mass region is identified, the scheme segments the region and searches for similar regions from the reference library using a multifeature based k-nearest neighbor (KNN) algorithm. To improve selection of reference images, we added an interactive step. All actual masses in the reference library were subjectively rated on a scale from 1 to 9 as to their "visual margins speculations". When an observer identifies a suspected mass region during a case interpretation he/she first rates the margins and the computerized search is then limited only to regions rated as having similar levels of spiculation (within +/-1 scale difference). In an observer preference study including 85 test regions, two sets of the six "similar" reference regions selected by the KNN with and without the interactive step were displayed side by side with each test region. Four radiologists and five nonclinician observers selected the more appropriate ("similar") reference set in a two alternative forced choice preference experiment. All four radiologists and five nonclinician observers preferred the sets of regions selected by the interactive method with an average frequency of 76.8% and 74.6%, respectively. The overall preference for the interactive method was highly significant (p < 0.001). The study demonstrated that a simple interactive approach that includes subjectively perceived ratings of one feature alone namely, a rating of margin "spiculation," could substantially improve the selection of "visually similar" reference images.

Effect of correlation on combining diagnostic information from two images of the same patient

We have shown previously, in the context of computer-aided diagnosis (CAD), that information derived from multiple images of the same patient can be used to improve diagnostic performance. In that work, we ignored the correlation among multiple images of the same patient. In the present study, we investigate theoretically, within the framework of receiver operating characteristic (ROC) analysis, the effect of correlation on three methods for combining quantitative diagnostic information from two images: taking the average, the maximum, and the minimum of a pair of normally distributed decision variables. We assume, as in our previous work, that the quantitative diagnostic information obtained from the two images of a given patient can be transformed monotonically to two latent decision variables that are normally distributed. Similar to the situation of uncorrelated images, we found that (1) the average always improves the area under the ROC curve (AUC) compared to the single-view image; (2) the maximum and the minimum can also, but not always, improve the AUC; and (3) each method can be the best method in certain situations. In addition, as the correlation strength increases, the average performs the best less often, whereas the maximum and the minimum perform the best more often. These theoretical results are illustrated with analysis of a mammography study.

CAD for mammography: the technique, results, current role and further developments

CAD systems, developed to assist the radiologist in the detection of suspicious lesions on mammograms, are currently controversially discussed. The highly sensitive detection of malignant structures including priors by CAD is linked with a low specific performance and a high rate of falsely positive markings. This causes controversial results regarding the effect of CAD systems for the diagnosing radiologist. This review aims to give an overview of the current literature, to state the currently discussed controversial results of CAD and to give an outlook on the next developments, which are not limited to senology, but include many other applications of CAD systems in radiology.

Computer-aided detection performance in mammographic examination of masses: assessment

PURPOSE

To compare performance of two computer-aided detection (CAD) systems and an in-house scheme applied to five groups of sequentially acquired screening mammograms.

MATERIALS AND METHODS

Two hundred nineteen film-based mammographic examinations, classified into five groups, were included in this study. Group 1 included 58 examinations in which verified malignant masses were detected during screening; group 2, 39 in which all available latest examinations were performed prior to diagnosis of these malignant masses (subset of 39 women from group 1); group 3, 22 in which findings were interpreted as negative but were verified as cancer within 1 year from the negative interpretation (missed cancers); group 4, 50 in which findings were negative and patients were not recalled for additional procedures; and group 5, 50 in which patients were recalled for additional procedures and findings were negative for cancer. In all examinations, images were processed with two Food and Drug Administration-approved commercially available CAD systems and an in-house scheme. Performance levels in terms of true-positive detection rates and number of false-positive identifications per image and per examination were compared.

RESULTS

Mass detection rates in positive examinations (group 1) were 67%-72%. Detection rates among three systems were not significantly different (P > .05). In 50 negative screening examinations (group 4), false-positive rates ranged from 1.08 to 1.68 per four-view examination. Performance level differences among systems were significant for false-positive rates (P = .008). Performance of all systems was at levels lower than publicly suggested in some retrospective studies. False-positive CAD cueing rates were significantly higher for negative examinations in which patients were recalled (group 5) than they were for those in which patients were not recalled (group 4) (P < or = .002).

CONCLUSION

Performance of CAD systems for mass detection at mammography varies significantly, depending on examination and system used. Actual performance of all systems in clinical environment can be improved.

Tolerance to missing data using a likelihood ratio based classifier for computer-aided classification of breast cancer

While mammography is a highly sensitive method for detecting breast tumours, its ability to differentiate between malignant and benign lesions is low, which may result in as many as 70% of unnecessary biopsies. The purpose of this study was to develop a highly specific computer-aided diagnosis algorithm to improve classification of mammographic masses. A classifier based on the likelihood ratio was developed to accommodate cases with missing data. Data for development included 671 biopsy cases (245 malignant), with biopsy-proved outcome. Sixteen features based on the BI-RADS lexicon and patient history had been recorded for the cases, with 1.3 +/- 1.1 missing feature values per case. Classifier evaluation methods included receiver operating characteristic and leave-one-out bootstrap sampling. The classifier achieved 32% specificity at 100% sensitivity on the 671 cases with 16 features that had missing values. Utilizing just the seven features present for all cases resulted in decreased performance at 100% sensitivity with average 19% specificity. No cases and no feature data were omitted during classifier development, showing that it is more beneficial to utilize cases with missing values than to discard incomplete cases that cannot be handled by many algorithms. Classification of mammographic masses was commendable at high sensitivity levels, indicating that benign cases could be potentially spared from biopsy.

Incorporation of an iterative, linear segmentation routine into a mammographic mass CAD system

In previous research, we have developed a computer-aided detection (CAD) system designed to detect masses in mammograms. The previous version of our system employed a simple but imprecise method to localize the masses. In this research, we present a more robust segmentation routine for use with mammographic masses. Our hypothesis is that by more accurately describing the morphology of the masses, we can improve the CAD system's ability to distinguish masses from other mammographic structures. To test this hypothesis, we incorporated the new segmentation routine into our CAD system and examined the change in performance. The developed iterative, linear segmentation routine is a gray level-based procedure. Using the identified regions from the previous CAD system as the initial seeds, the new segmentation algorithm refines the suspicious mass borders by making estimates of the interior and exterior pixels. These estimates are then passed to a linear discriminant, which determines the optimal threshold between the interior and exterior pixels. After applying the threshold and identifying the object's outline, two constraints on the border are applied to reduce the influence of background noise. After the border is constrained, the process repeats until a stopping criterion is reached. The segmentation routine was tested on a study database of 183 mammographic images extracted from the Digital Database for Screening Mammography. Eighty-three of the images contained 50 malignant and 50 benign masses; 100 images contained no masses. The previously developed CAD system was used to locate a set of suspicious regions of interest (ROIs) within the images. To assess the performance of the segmentation algorithm, a set of 20 features was measured from the suspicious regions before and after the application of the developed segmentation routine. Receiver operating characteristic (ROC) analysis was employed on the ROIs to examine the discriminatory capabilities of each individual feature before and after the segmentation routine. A statistically significant performance increase was found in many of the individual features, particularly those describing the mass borders. To examine how the incorporation of the segmentation routine affected the performance of the overall CAD system, free-response ROC (FROC) analysis was employed. When considering only malignant masses, the FROC performance of the system with the segmentation routine appeared better than the previous system. When detecting 90% of the malignant masses, the previous system achieved 4.9 false positives per image (FPpI) compared to the post-segmentation system's 4.2 FPpI. At 80% sensitivity, the respective FPpI were 3.5 and 1.6.

Performance change of mammographic CAD schemes optimized with most-recent and prior image databases

RATIONALE AND OBJECTIVES

The authors evaluated performance changes in the detection of masses on "current" (latest) and "prior" images by computer-aided diagnosis (CAD) schemes that had been optimized with databases of current and prior mammograms.

MATERIALS AND METHODS

The authors selected 260 pairs of matched consecutive mammograms. Each current image depicted one or two verified masses. All prior images had been interpreted originally as negative or probably benign. A CAD scheme initially detected 261 mass regions and 465 false-positive regions on the current images, and 252 corresponding mass regions (early signs) and 471 false-positive regions on prior images. These regions were divided into two training and two testing databases. The current and prior training databases were used to optimize two CAD schemes with a genetic algorithm. These schemes were evaluated with two independent testing databases.

RESULTS

The scheme optimized with current images produced areas under the receiver operating characteristic curve of (0.89 +/- 0.01 and 0.65 +/- 0.02 when tested with current images and prior images, respectively. The scheme optimized with prior images produced areas under the receiver operating characteristic curve of 0.81 +/- 0.02 and 0.71 +/- 0.02 when tested with current images and prior images, respectively. Performance changes for both current and prior testing databases were significant (P < .01) for the two schemes.

CONCLUSION

CAD schemes trained with current images do not perform optimally in detecting masses depicted on prior images. To optimize CAD schemes for early detection, it may be important to include in the training database a large fraction of prior images originally reported as negative and later proven to be positive.

Self-organizing map for cluster analysis of a breast cancer database

The purpose of this study was to identify and characterize clusters in a heterogeneous breast cancer computer-aided diagnosis database. Identification of subgroups within the database could help elucidate clinical trends and facilitate future model building. A self-organizing map (SOM) was used to identify clusters in a large (2258 cases), heterogeneous computer-aided diagnosis database based on mammographic findings (BI-RADS) and patient age. The resulting clusters were then characterized by their prototypes determined using a constraint satisfaction neural network (CSNN). The clusters showed logical separation of clinical subtypes such as architectural distortions, masses, and calcifications. Moreover, the broad categories of masses and calcifications were stratified into several clusters (seven for masses and three for calcifications). The percent of the cases that were malignant was notably different among the clusters (ranging from 6 to 83%). A feed-forward back-propagation artificial neural network (BP-ANN) was used to identify likely benign lesions that may be candidates for follow up rather than biopsy. The performance of the BP-ANN varied considerably across the clusters identified by the SOM. In particular, a cluster (#6) of mass cases (6% malignant) was identified that accounted for 79% of the recommendations for follow up that would have been made by the BP-ANN. A classification rule based on the profile of cluster #6 performed comparably to the BP-ANN, providing approximately 25% specificity at 98% sensitivity. This performance was demonstrated to generalize to a large (2177) set of cases held-out for model validation.

Computer-aided detection in mammography: an assessment of performance on current and prior images

RATIONALE AND OBJECTIVES

The authors assessed and compared the performance of a computer-aided detection (CAD) scheme for the detection of masses and microcalcification clusters on a set of images collected from two consecutive ("current" and "prior") mammographic examinations.

MATERIALS AND METHODS

A previously developed CAD scheme was used to assess two consecutive screening mammograms from 200 cases in which the current mammogram showed a mass or cluster of microcalcifications that resulted in breast biopsy. The latest prior examinations had been initially interpreted as negative or definitely benign findings (Breast Imaging Reporting and Data System rating, 1 or 2). The study involved images of 400 examinations acquired in 200 patients. Radiologists identified 172 masses and 128 clusters of microcalcifications on the current images. The performance of the CAD scheme was analyzed and compared for the current and latest prior images.

RESULTS

There were significant differences (P < .01) between current and prior images in many feature values. The performance of the CAD scheme was significantly lower for prior than for current images (P < .01). At 0.5 and 0.2 false-positive mass and cluster cues per image, the scheme detected 78 malignant masses (78%) and 63 malignant clusters (80%) on current images. Only 42% of malignant cases were detected on prior images, including 40 masses (40%) and 36 microcalcification clusters (46%).

CONCLUSION

CAD schemes can detect a substantial fraction of masses and microcalcification clusters depicted on prior images. To improve performance with prior images, the scheme may have to be adaptively reoptimized with increasingly more subtle abnormalities.