Feature selection with limited datasets

Exploring determinant factors influencing muscle quality and sarcopenia in Bilbao's older adult population through machine learning: A comprehensive analysis approach

BACKGROUND

Sarcopenia and reduced muscle quality index have garnered special attention due to their prevalence among older individuals and the adverse effects they generate. Early detection of these geriatric pathologies holds significant potential, enabling the implementation of interventions that may slow or reverse their progression, thereby improving the individual's overall health and quality of life. In this context, artificial intelligence opens up new opportunities to identify the key identifying factors of these pathologies, thus facilitating earlier intervention and personalized treatment approaches.

OBJECTIVES

investigate anthropomorphic, functional, and socioeconomic factors associated with muscle quality and sarcopenia using machine learning approaches and identify key determinant factors for their potential future integration into clinical practice.

METHODS

A total of 1253 older adults (89.5% women) with a mean age of 78.13 ± 5.78 voluntarily participated in this descriptive cross-sectional study, which examines determining factors in sarcopenia and MQI using machine learning techniques. Feature selection was completed using a variety of techniques and feature datasets were constructed according to feature selection. Three machine learning classification algorithms classified sarcopenia and MQI in each dataset, and the performance of classification models was compared.

RESULTS

The predictive models used in this study exhibited AUC scores of 0.7671 for MQI and 0.7649 for sarcopenia, with the most successful algorithms being SVM and MLP. Key factors in predicting both conditions have been shown to be relative power, age, weight, and the 5STS. No single factor is sufficient to predict either condition, and by comprehensively considering all selected features, the study underscores the importance of a holistic approach in understanding and addressing sarcopenia and MQI among older adults.

CONCLUSIONS

Exploring the factors that affect sarcopenia and MQI in older adults, this study highlights that relative power, age, weight, and the 5STS are significant determinants. While considering these clinical markers and using a holistic approach, this can provide crucial information for designing personalized and effective interventions to promote healthy aging.

Magnetization-prepared rapid acquisition with gradient echo magnetic resonance imaging signal and texture features for the prediction of mild cognitive impairment to Alzheimer's disease progression

Early diagnoses of Alzheimer's disease (AD) would confer many benefits. Several biomarkers have been proposed to achieve such a task, where features extracted from magnetic resonance imaging (MRI) have played an important role. However, studies have focused exclusively on morphological characteristics. This study aims to determine whether features relating to the signal and texture of the image could predict mild cognitive impairment (MCI) to AD progression. Clinical, biological, and positron emission tomography information and MRI images of 62 subjects from the AD neuroimaging initiative were used in this study, extracting 4150 features from each MRI. Within this multimodal database, a feature selection algorithm was used to obtain an accurate and small logistic regression model, generated by a methodology that yielded a mean blind test accuracy of 0.79. This model included six features, five of them obtained from the MRI images, and one obtained from genotyping. A risk analysis divided the subjects into low-risk and high-risk groups according to a prognostic index. The groups were statistically different ([Formula: see text]). These results demonstrated that MRI features related to both signal and texture add MCI to AD predictive power, and supported the ongoing notion that multimodal biomarkers outperform single-modality ones.

Potential of computer-aided diagnosis of high spectral and spatial resolution (HiSS) MRI in the classification of breast lesions

PURPOSE

To compare the performance of computer-aided diagnosis (CADx) analysis of precontrast high spectral and spatial resolution (HiSS) MRI to that of clinical dynamic contrast-enhanced MRI (DCE-MRI) in the diagnostic classification of breast lesions.

MATERIALS AND METHODS

Thirty-four malignant and seven benign lesions were scanned using two-dimensional (2D) HiSS and clinical 4D DCE-MRI protocols. Lesions were automatically segmented. Morphological features were calculated for HiSS, whereas both morphological and kinetic features were calculated for DCE-MRI. After stepwise feature selection, Bayesian artificial neural networks merged selected features, and receiver operating characteristic (ROC) analysis evaluated the performance with leave-one-lesion-out validation.

RESULTS

AUC (area under the ROC curve) values of 0.92 ± 0.06 and 0.90 ± 0.05 were obtained using CADx on HiSS and DCE-MRI, respectively, in the task of classifying benign and malignant lesions. While we failed to show that the higher HiSS performance was significantly better than DCE-MRI, noninferiority testing confirmed that HiSS was not worse than DCE-MRI.

CONCLUSION

CADx of HiSS (without contrast) performed similarly to CADx on clinical DCE-MRI; thus, computerized analysis of HiSS may provide sufficient information for diagnostic classification. The results are clinically important for patients in whom contrast agent is contra-indicated. Even in the limited acquisition mode of 2D single slice HiSS, by using quantitative image analysis to extract characteristics from the HiSS images, similar performance levels were obtained as compared with those from current clinical 4D DCE-MRI. As HiSS acquisitions become possible in 3D, CADx methods can also be applied. Because HiSS and DCE-MRI are based on different contrast mechanisms, the use of the two protocols in combination may increase diagnostic accuracy.

Breast image analysis for risk assessment, detection, diagnosis, and treatment of cancer

The role of breast image analysis in radiologists' interpretation tasks in cancer risk assessment, detection, diagnosis, and treatment continues to expand. Breast image analysis methods include segmentation, feature extraction techniques, classifier design, biomechanical modeling, image registration, motion correction, and rigorous methods of evaluation. We present a review of the current status of these task-based image analysis methods, which are being developed for the various image acquisition modalities of mammography, tomosynthesis, computed tomography, ultrasound, and magnetic resonance imaging. Depending on the task, image-based biomarkers from such quantitative image analysis may include morphological, textural, and kinetic characteristics and may depend on accurate modeling and registration of the breast images. We conclude with a discussion of future directions.

Evaluating feature selection strategies for high dimensional, small sample size datasets

In this work, we analyze and evaluate different strategies for comparing Feature Selection (FS) schemes on High Dimensional (HD) biomedical datasets (e.g. gene and protein expression studies) with a small sample size (SSS). Additionally, we define a new feature, Robustness, specifically for comparing the ability of an FS scheme to be invariant to changes in its training data. While classifier accuracy has been the de facto method for evaluating FS schemes, on account of the curse of dimensionality problem, it might not always be the appropriate measure for HD/SSS datasets. SSS lends the dataset a higher probability of containing data that is not representative of the true distribution of the whole population. However, an ideal FS scheme must be robust enough to produce the same results each time there are changes to the training data. In this study, we employed the robustness performance measure in conjunction with classifier accuracy (measured via the K-Nearest Neighbor and Random Forest classifiers) to quantitatively compare five different FS schemes (T-test, F-test, Kolmogorov-Smirnov Test, Wilks Lambda Test and Wilcoxon Rand Sum Test) on 5 HD/SSS gene and protein expression datasets corresponding to ovarian cancer, lung cancer, bone lesions, celiac disease, and coronary heart disease. Of the five FS schemes compared, the Wilcoxon Rand Sum Test was found to outperform other FS schemes in terms of classification accuracy and robustness. Our results suggest that both classifier accuracy and robustness should be considered when deciding on the appropriate FS scheme for HD/SSS datasets.

Automated detection of mass lesions in dedicated breast CT: a preliminary study

PURPOSE

To develop an automated method to detect breast masses on dedicated breast CT (BCT) volumes and to conduct a preliminary evaluation of its performance. This method can be used in a computer-aided detection (CADe) system for noncontrast enhanced BCT.

METHODS

The database included patient images, which were acquired under an IRB-approved protocol. The database in this study consisted of 132 cases. 50 cases contained 58 malignant masses, and 23 cases contained 24 benign masses. 59 cases did not contain any biopsy-proven lesions. Each case consisted of an unenhanced CT volume of a single breast. First, each breast was segmented into adipose and glandular tissues using a fuzzy c-means clustering algorithm. The glandular breast regions were then sampled at a resolution of 2 mm. At each sampling step, a 3.5-cm(3) volume-of-interest was subjected to constrained region segmentation and 17 characteristic features were extracted, yielding 17 corresponding feature volumes. Four features were selected using step-wise feature selection and merged with linear discriminant analysis trained in the task of distinguishing between normal breast glandular regions and masses. Detection performance was measured using free-response receiver operating characteristic analysis (FROC) with leave-one-case-out evaluation.

RESULTS

The feature selection stage selected features that characterized the shape and margin strength of the segmented region. CADe sensitivity per case was 84% (std = 4.2%) at 2.6 (std = 0.06) false positives per volume, or 6 × 10(-3) per slice (at an average of 424 slices per volume in this data set).

CONCLUSIONS

This preliminary study demonstrates the feasibility of our approach for CADe for BCT.

Evaluating imaging and computer-aided detection and diagnosis devices at the FDA

This report summarizes the Joint FDA-MIPS Workshop on Methods for the Evaluation of Imaging and Computer-Assist Devices. The purpose of the workshop was to gather information on the current state of the science and facilitate consensus development on statistical methods and study designs for the evaluation of imaging devices to support US Food and Drug Administration submissions. Additionally, participants expected to identify gaps in knowledge and unmet needs that should be addressed in future research. This summary is intended to document the topics that were discussed at the meeting and disseminate the lessons that have been learned through past studies of imaging and computer-aided detection and diagnosis device performance.

Fissures segmentation using surface features: content-based retrieval for mammographic mass using ensemble classifier

RATIONALE AND OBJECTIVES

Accurate classification is critical in mammography computer-aided diagnosis using content-based image retrieval approaches (CBIR CAD). The objectives of this study were to: 1) develop an accurate ensemble classifier based on domain knowledge and a robust feature selection method for CBIR CAD; 2) propose three new features; and 3) assess the performance of the proposed method and new features by using a relatively large imaging data set.

MATERIALS AND METHODS

The data set used in this study consisted of 2114 regions of interest (ROI) extracted from a publicly available image database. The proposed ensemble classifier method we called E-DGA-KNN included four steps. In the first step, 804 ROIs depict masses were divided into five classes according to their boundary types. Then, each class of ROI with an equal number of negative ROIs were put together to create a sub-database. Second, a dual-stage genetic algorithm, which was called DGA, was applied on those five sub-databases for feature selection and weights determination respectively. In the third step, five base K-nearest neighbor (KNN) classifiers were created by using the results of the second step on 2114 ROIs, and five detection scores for a given queried ROI were obtained. Finally, these classifiers are combined to yield a final classification. The performances of the proposed methods were evaluated by using receiver operating characteristic (ROC) analysis. A comparison with eight different methods on the data set was provided which include the stepwise linear discriminative analysis algorithm (SLDA) and particle swarm optimization (PSO) algorithm with KNN classifier.

RESULTS

When four hybrid feature selection methods were applied with single KNN classifier (ie, DGA-KNN, SLDA-WGA-KNN, SLDA-PSO-KNN, GA-PSO-KNN) and the proposed E-DGA-KNN method to the data set, the computed areas under the ROC curve (Az) were 0.8782 ± 0.0080, 0.8675 ± 0.0081, 0.8623 ± 0.0083, 0.8725 ± 0.0079, and 0.8927 ± 0.0073, respectively. If all features and single KNN classifier were used, the Az value was 0.8478 ± 0.0088. Az values were 0.8592 ± 0.0083 and 0.8632 ± 0.0081 when SLDA or GA algorithm used alone.

CONCLUSIONS

In this study, an ensemble classifier based on domain knowledge and a dual-stage feature selection method was proposed. Evaluation results indicated that the proposed method achieved largest value of ROC compared to other algorithms. The proposed method shows better performance and has the potential to improve the performance of CBIR CAD in interpreting and analyzing mammograms.

Enhancement of breast CADx with unlabeled data

PURPOSE

Unlabeled medical image data are abundant, yet the process of converting them into a labeled ("truth-known") database is time and resource expensive and fraught with ethical and logistics issues. The authors propose a dual-stage CADx scheme in which both labeled and unlabeled (truth-known and "truth-unknown") data are used. This study is an initial exploration of the potential for leveraging unlabeled data toward enhancing breast CADx.

METHODS

From a labeled ultrasound image database consisting of 1126 lesions with an empirical cancer prevalence of 14%, 200 different randomly sampled subsets were selected and the truth status of a variable number of cases was masked to the algorithm to mimic different types of labeled and unlabeled data sources. The prevalence was fixed at 50% cancerous for the labeled data and 5% cancerous for the unlabeled. In the first stage of the dual-stage CADx scheme, the authors term "transductive dimension reduction regularization" (TDR-R), both labeled and unlabeled images characterized by extracted lesion features were combined using dimension reduction (DR) techniques and mapped to a lower-dimensional representation. (The first stage ignored truth status therefore was an unsupervised algorithm.) In the second stage, the labeled data from the reduced dimension embedding were used to train a classifier toward estimating the probability of malignancy. For the first CADx stage, the authors investigated three DR approaches: Laplacian eigen-maps, t-distributed stochastic neighbor embedding (t-SNE), and principal component analysis. For the TDR-R methods, the classifier in the second stage was a supervised (i.e., utilized truth) Bayesian neural net. The dual-stage CADx schemes were compared to a single-stage scheme based on manifold regularization (MR) in a semisupervised setting via the LapSVM algorithm. Performance in terms of areas under the ROC curve (AUC) of the CADx schemes was evaluated in leave-one-out and .632+ bootstrap analyses on a by-lesion basis. Additionally, the trained algorithms were applied to an independent test data set consisting of 101 lesions with approximately 50% cancer prevalence. The difference in AUC (deltaAUC) between with and without the use of unlabeled data was computed.

RESULTS

Statistically significant differences in the average AUC value (deltaAUC) were found in many instances between training with and without unlabeled data, based on the sample set distributions generated from this particular ultrasound data set during cross-validation and using independent test set. For example, when using 100 labeled and 900 unlabeled cases and testing on the independent test set, the TDR-R methods produced average deltaAUC=0.0361 with 95% intervals [0.0301; 0.0408] (p-value < 0.0001, adjusted for multiple comparisons, but considering the test set fixed) using t-SNE and average deltaAUC=.026 [0.0227, 0.0298] (adjusted p-value < 0.0001) using Laplacian eigenmaps, while the MR-based LapSVM produced an average deltaAUC=.0381 [0.0351; 0.0405] (adjusted p-value < 0.0001). The authors also found that schemes initially obtaining lower than average performance when using labeled data only showed the most prominent increase in performance when unlabeled data were added in the first CADx stage, suggesting a regularization effect due to the injection of unlabeled data.

CONCLUSION

The findings reveal evidence that incorporating unlabeled data information into the overall development of CADx methods may improve classifier performance by non-negligible amounts and warrants further investigation.

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.

Computerized assessment of breast lesion malignancy using DCE-MRI robustness study on two independent clinical datasets from two manufacturers

RATIONALE AND OBJECTIVES

To conduct a preclinical evaluation of the robustness of our computerized system for breast lesion characterization on two breast magnetic resonance imaging (MRI) databases that were acquired using scanners from two different manufacturers.

MATERIALS AND METHODS

Two clinical breast MRI databases were acquired from a Siemens scanner and a GE scanner, which shared similar imaging protocols and retrospectively collected under an institutional review board-approved protocol. In our computerized analysis system, after a breast lesion is identified by the radiologist, the computer performs automatic lesion segmentation and feature extraction and outputs an estimated probability of malignancy. We used a Bayesian neural network with automatic relevance determination for joint feature selection and classification. To evaluate the robustness of our classification system, we first used Database 1 for feature selection and classifier training, and Database 2 to test the trained classifier. Then, we exchanged the two datasets and repeated the process. Area under the receiver operating characteristic curve (AUC) was used as a performance figure of merit in the task of distinguishing between malignant and benign lesions.

RESULTS

We obtained an AUC of 0.85 (approximate 95% confidence interval [CI] 0.79-0.91) for (a) feature selection and classifier training using Database 1 and testing on Database 2; and an AUC of 0.90 (approximate 95% CI 0.84-0.96) for (b) feature selection and classifier training using Database 2 and testing on Database 1. We failed to observe statistical significance for the difference AUC of 0.05 between the two database conditions (P = .24; 95% confidence interval -0.03, 0.1).

CONCLUSION

These results demonstrate the robustness of our computerized classification system in the task of distinguishing between malignant and benign breast lesions on dynamic contrast-enhanced (DCE) MRI images from two manufacturers. Our study showed the feasibility of developing a computerized classification system that is robust across different scanners.

Exploring nonlinear feature space dimension reduction and data representation in breast Cadx with Laplacian eigenmaps and t-SNE

PURPOSE

In this preliminary study, recently developed unsupervised nonlinear dimension reduction (DR) and data representation techniques were applied to computer-extracted breast lesion feature spaces across three separate imaging modalities: Ultrasound (U.S.) with 1126 cases, dynamic contrast enhanced magnetic resonance imaging with 356 cases, and full-field digital mammography with 245 cases. Two methods for nonlinear DR were explored: Laplacian eigenmaps [M. Belkin and P. Niyogi, "Laplacian eigenmaps for dimensionality reduction and data representation," Neural Comput. 15, 1373-1396 (2003)] and t-distributed stochastic neighbor embedding (t-SNE) [L. van der Maaten and G. Hinton, "Visualizing data using t-SNE," J. Mach. Learn. Res. 9, 2579-2605 (2008)].

METHODS

These methods attempt to map originally high dimensional feature spaces to more human interpretable lower dimensional spaces while preserving both local and global information. The properties of these methods as applied to breast computer-aided diagnosis (CADx) were evaluated in the context of malignancy classification performance as well as in the visual inspection of the sparseness within the two-dimensional and three-dimensional mappings. Classification performance was estimated by using the reduced dimension mapped feature output as input into both linear and nonlinear classifiers: Markov chain Monte Carlo based Bayesian artificial neural network (MCMC-BANN) and linear discriminant analysis. The new techniques were compared to previously developed breast CADx methodologies, including automatic relevance determination and linear stepwise (LSW) feature selection, as well as a linear DR method based on principal component analysis. Using ROC analysis and 0.632+bootstrap validation, 95% empirical confidence intervals were computed for the each classifier's AUC performance.

RESULTS

In the large U.S. data set, sample high performance results include, AUC0.632+ = 0.88 with 95% empirical bootstrap interval [0.787;0.895] for 13 ARD selected features and AUC0.632+ = 0.87 with interval [0.817;0.906] for four LSW selected features compared to 4D t-SNE mapping (from the original 81D feature space) giving AUC0.632+ = 0.90 with interval [0.847;0.919], all using the MCMC-BANN.

CONCLUSIONS

Preliminary results appear to indicate capability for the new methods to match or exceed classification performance of current advanced breast lesion CADx algorithms. While not appropriate as a complete replacement of feature selection in CADx problems, DR techniques offer a complementary approach, which can aid elucidation of additional properties associated with the data. Specifically, the new techniques were shown to possess the added benefit of delivering sparse lower dimensional representations for visual interpretation, revealing intricate data structure of the feature space.

A novel hybrid linear/nonlinear classifier for two-class classification: theory, algorithm, and applications

Classifier design for a given classification task needs to take into consideration both the complexity of the classifier and the size of the dataset that is available for training the classifier. With limited training data, as often is the situation in computer-aided diagnosis of medical images, a classifier with simple structure (e.g., a linear classifier) is more robust and therefore preferred. We propose a novel two-class classifier, which we call a hybrid linear/nonlinear classifier (HLNLC), that involves two stages: the input features are linearly combined to form a scalar variable in the first stage and then the likelihood ratio of the scalar variable is used as the decision variable for classification. We first develop the theory of HLNLC by assuming that the feature data follow normal distributions. We show that the commonly used Fisher's linear discriminant function is generally not the optimal linear function in the first stage of the HLNLC. We formulate an optimization problem to solve for the optimal linear function in the first stage of the HLNLC, i.e., the linear function that maximizes the area under the receiver operating characteristic (ROC) curve of the HLNLC. For practical applications, we propose a robust implementation of the HLNLC by making a loose assumption that the two-class feature data arise from a pair of latent (rather than explicit) multivariate normal distributions. The novel hybrid classifier fills a gap between linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA) in the sense that both its theoretical performance and its complexity lie between those of the LDA and those of the QDA. Simulation studies show that the hybrid linear/nonlinear classifier performs better than LDA without increasing the classifier complexity accordingly. With a finite number of training samples, the HLNLC can perform better than that of the ideal observer due to its simplicity. Finally, we demonstrate the application of the HLNLC in computer-aided diagnosis of breast lesions in ultrasound images.

Computer-Aided Diagnosis in Mammography Using Content-based Image Retrieval Approaches: Current Status and Future Perspectives

As the rapid advance of digital imaging technologies, the content-based image retrieval (CBIR) has became one of the most vivid research areas in computer vision. In the last several years, developing computer-aided detection and/or diagnosis (CAD) schemes that use CBIR to search for the clinically relevant and visually similar medical images (or regions) depicting suspicious lesions has also been attracting research interest. CBIR-based CAD schemes have potential to provide radiologists with "visual aid" and increase their confidence in accepting CAD-cued results in the decision making. The CAD performance and reliability depends on a number of factors including the optimization of lesion segmentation, feature selection, reference database size, computational efficiency, and relationship between the clinical relevance and visual similarity of the CAD results. By presenting and comparing a number of approaches commonly used in previous studies, this article identifies and discusses the optimal approaches in developing CBIR-based CAD schemes and assessing their performance. Although preliminary studies have suggested that using CBIR-based CAD schemes might improve radiologists' performance and/or increase their confidence in the decision making, this technology is still in the early development stage. Much research work is needed before the CBIR-based CAD schemes can be accepted in the clinical practice.

Anniversary paper: History and status of CAD and quantitative image analysis: the role of Medical Physics and AAPM

The roles of physicists in medical imaging have expanded over the years, from the study of imaging systems (sources and detectors) and dose to the assessment of image quality and perception, the development of image processing techniques, and the development of image analysis methods to assist in detection and diagnosis. The latter is a natural extension of medical physicists' goals in developing imaging techniques to help physicians acquire diagnostic information and improve clinical decisions. Studies indicate that radiologists do not detect all abnormalities on images that are visible on retrospective review, and they do not always correctly characterize abnormalities that are found. Since the 1950s, the potential use of computers had been considered for analysis of radiographic abnormalities. In the mid-1980s, however, medical physicists and radiologists began major research efforts for computer-aided detection or computer-aided diagnosis (CAD), that is, using the computer output as an aid to radiologists-as opposed to a completely automatic computer interpretation-focusing initially on methods for the detection of lesions on chest radiographs and mammograms. Since then, extensive investigations of computerized image analysis for detection or diagnosis of abnormalities in a variety of 2D and 3D medical images have been conducted. The growth of CAD over the past 20 years has been tremendous-from the early days of time-consuming film digitization and CPU-intensive computations on a limited number of cases to its current status in which developed CAD approaches are evaluated rigorously on large clinically relevant databases. CAD research by medical physicists includes many aspects-collecting relevant normal and pathological cases; developing computer algorithms appropriate for the medical interpretation task including those for segmentation, feature extraction, and classifier design; developing methodology for assessing CAD performance; validating the algorithms using appropriate cases to measure performance and robustness; conducting observer studies with which to evaluate radiologists in the diagnostic task without and with the use of the computer aid; and ultimately assessing performance with a clinical trial. Medical physicists also have an important role in quantitative imaging, by validating the quantitative integrity of scanners and developing imaging techniques, and image analysis tools that extract quantitative data in a more accurate and automated fashion. As imaging systems become more complex and the need for better quantitative information from images grows, the future includes the combined research efforts from physicists working in CAD with those working on quantitative imaging systems to readily yield information on morphology, function, molecular structure, and more-from animal imaging research to clinical patient care. A historical review of CAD and a discussion of challenges for the future are presented here, along with the extension to quantitative image analysis.

Computer-aided detection systems for breast masses: comparison of performances on full-field digital mammograms and digitized screen-film mammograms

RATIONALE AND OBJECTIVES

To compare the performance of computer aided detection (CAD) systems on pairs of full-field digital mammogram (FFDM) and screen-film mammogram (SFM) obtained from the same patients.

MATERIALS AND METHODS

Our CAD systems on both modalities have similar architectures that consist of five steps. For FFDMs, the input raw image is first log-transformed and enhanced by a multiresolution preprocessing scheme. For digitized SFMs, the input image is smoothed and subsampled to a pixel size of 100 microm x 100 microm. For both CAD systems, the mammogram after preprocessing undergoes a gradient field analysis followed by clustering-based region growing to identify suspicious breast structures. Each of these structures is refined in a local segmentation process. Morphologic and texture features are then extracted from each detected structure, and trained rule-based and linear discriminant analysis classifiers are used to differentiate masses from normal tissues. Two datasets, one with masses and the other without masses, were collected. The mass dataset contained 131 cases with 131 biopsy proven masses, of which 27 were malignant and 104 benign. The true locations of the masses were identified by an experienced Mammography Quality Standards Act (MQSA) radiologist. The no-mass data set contained 98 cases. The time interval between the FFDM and the corresponding SFM was 0 to 118 days.

RESULTS

Our CAD system achieved case-based sensitivities of 70%, 80%, and 90% at 0.9, 1.5, and 2.6 false positive (FP) marks/image, respectively, on FFDMs, and the same sensitivities at 1.0, 1.4, and 2.6 FP marks/image, respectively, on SFMs.

CONCLUSIONS

The difference in the performances of our FFDM and SFM CAD systems did not achieve statistical significance.

Computerized mass detection for digital breast tomosynthesis directly from the projection images

Digital breast tomosynthesis (DBT) has recently emerged as a new and promising three-dimensional modality in breast imaging. In DBT, the breast volume is reconstructed from 11 projection images, taken at source angles equally spaced over an arc of 50 degrees. Reconstruction algorithms for this modality are not fully optimized yet. Because computerized lesion detection in the reconstructed breast volume will be affected by the reconstruction technique, we are developing a novel mass detection algorithm that operates instead on the set of raw projection images. Mass detection is done in three stages. First, lesion candidates are obtained for each projection image separately, using a mass detection algorithm that was initially developed for screen-film mammography. Second, the locations of a lesion candidate are backprojected into the breast volume. In this feature volume, voxel intensities are a combined measure of detection frequency (e.g., the number of projections in which a given lesion candidate was detected), and a measure of the angular range over which a given lesion was detected. Third, features are extracted after reprojecting the three-dimensional (3-D) locations of lesion candidates into projection images. Features are combined using linear discriminant analysis. The database used to test the algorithm consisted of 21 mass cases (13 malignant, 8 benign) and 15 cases without mass lesions. Based on this database, the algorithm yielded a sensitivity of 90% at 1.5 false positives per breast volume. Algorithm performance is positively biased because this dataset was used for development, training, and testing, and because the number of algorithm parameters was approximately the same as the number.of patient cases. Our results indicate that computerized mass detection in the sequence of projection images for DBT may be effective despite the higher noise level in those images.

Robustness of computerized lesion detection and classification scheme across different breast US platforms

PURPOSE

To evaluate the performance of a computerized detection and diagnosis method with breast ultrasonographic (US) images obtained with US equipment from two different manufacturers.

MATERIALS AND METHODS

Two independent clinical breast US databases were used in this performance study. Data collection and database use were HIPAA-compliant and followed institutional review board-approved protocols, with waiver of informed consent. One database consisted of 1740 images obtained in 458 women with Philips US equipment. The other database consisted of 151 images obtained in 151 women with Siemens US equipment. The testing protocols included independent testing and round-robin analysis. The computerized scheme detects potential lesions, calculates imaging features for all candidate lesions, and subsequently classifies candidate lesions into different categories. Two separate classification tasks were evaluated: distinction between all actual lesions and false-positive detections and distinction between actual cancers and all other detected lesion candidates. Statistical analysis was performed by using both receiver operating characteristic (ROC) and free-response ROC methods.

RESULTS

For the distinction between all actual lesions and false-positive detections, area under the ROC curve (A(z)) values ranged between 0.87 and 0.95 for different testing protocols. In two instances, the difference in performance between databases was significant (P < .01), but it was shown that this was due to the difference in size of the databases. In the distinction of cancer from all other detections, the A(z) values ranged between 0.80 and 0.86. No statistically significant difference was found among the different testing protocols in this instance.

CONCLUSION

These results indicate that the performance of this fully automated computerized lesion detection and classification method, which demonstrated robustness over the different US equipment used, is promising.

Multimodality computerized diagnosis of breast lesions using mammography and sonography

RATIONALE AND OBJECTIVES

The purpose of this study is to investigate the use of computer-extracted features of lesions imaged by means of two modalities, mammography and breast ultrasound, in the computerized classification of breast lesions.

MATERIAL AND METHODS

We performed computerized analysis on a database of 97 patients with a total of 100 lesions (40 malignant, 40 benign solid, and 20 cystic lesions). Mammograms and ultrasound images were available for these breast lesions. There was an average of three mammographic images and two ultrasound images per lesion. Based on seed points indicated by a radiologist, the computer automatically segmented lesions from the parenchymal background and automatically extracted a set of characteristic features for each lesion. For each feature, its value averaged over all images pertaining to a given lesion was input to a Bayesian neural network for classification. We also investigated different approaches to combine image-based features into this by-lesion analysis. In that analysis, mean, maximum, and minimum feature values were considered for all images representing a lesion. We considered performance by using a leave-one-lesion-out approach, based on image features from mammography alone (two to five features), ultrasound alone (three to four features), and a combination of features from both modalities (three to five features total).

RESULTS

For the classification task of distinguishing cancer from other abnormalities in a lesion-based analysis by using a single modality, areas under the receiver operating characteristic curves (A(z) values) increased significantly when the computer selected the manner (mean, minimum, or maximum) in which image-based features were combined into lesion-based features. The highest performance was found for lesion-based analysis and automated feature selection from mean, maximum, and minimum values of features from both modalities (resulting in a total of four features being used). That A(z) value for the task of distinguishing cancer was 0.92, showing a statistically significant increase over that achieved with features from either mammography or ultrasound alone.

CONCLUSION

Computerized classification of cancer significantly improved when lesion features from both modalities were combined. Classification performance depended on specific methods for combining features from multiple images per lesion. These results are encouraging and warrant further exploration of computerized methods for multimodality imaging.

Computerized analysis of images in the detection and diagnosis of breast cancer

Improvements in mammographic acquisition techniques have resulted in making the early signs of breast cancer more apparent on mammograms. However, the accuracy of the overall mammographic examination depends on both the quality of the mammographic images and the ability of the radiologist to interpret those images. While mammography is the best screening method for the early detection of breast cancer, radiologists do miss lesions on mammograms. Use of output, however, from a computerized analysis of an image by a radiologist may help him/her in the detection or diagnostic tasks, and potentially improve the overall interpretation of breast images and the subsequent patient care. Computer-aided detection and diagnosis (CAD) involves the application of computer technology to the process of medical image interpretation. CAD can be defined as a diagnosis made by a radiologist, who uses the output from a computerized analysis of medical images as a "second opinion" in detecting and diagnosing lesions, with the final diagnosis being made by the radiologist. The computer output must be at a sufficient performance level, and in addition, the output must be displayed in a user-friendly format for effective and efficient use by the radiologist. This chapter reviews CAD in breast cancer detection and diagnosis, including examples of image analyses, multi-modality approaches (i.e., special-view diagnostic mammography, ultrasound, and MRI), and means of communicating the computer output to the human.

Computerized detection and classification of cancer on breast ultrasound1

RATIONALE AND OBJECTIVES

To develop and evaluate a two-stage computerized method that first detects suspicious regions on ultrasound images, and subsequently distinguishes among different lesion types.

MATERIALS AND METHODS

The first stage of detecting potential lesions was based on expected lesion shape and margin characteristics. After the detection stage, all candidate lesions were classified by a Bayesian neural net based on computer-extracted lesion features. Two separate tasks were performed and evaluated at the classification stage: the first classification task was the distinction between all actual lesions and false-positive detections; the second classification task was the distinction between actual cancer and all other detected lesion candidates (including false-positive detections). The neural nets were trained on a database of 400 cases (757 images), consisting of complex cysts and benign and malignant lesions, and tested on an independent database of 458 cases (1,740 images including 578 normal images).

RESULTS

In the distinction between all actual lesions and false-positive detections, Az values of 0.94 and 0.91 were obtained with the training and testing data sets, respectively. Sensitivity by patient of 90% at 0.45 false-positive detections per image was achieved for this detection-plus-classification scheme for the testing data set. Distinguishing cancer from all other detections (false-positives plus all benign lesions) proved to be more challenging, and Az values of 0.87 and 0.81 were obtained during training and testing, respectively. Sensitivity by patient of 100% at 0.43 false-positive malignancies per image was achieved in the detection and classification of cancerous lesions for the testing dataset.

CONCLUSION

The results show promising performance of the computerized lesion detection and classification method, and indicate the potential of such a system for clinical breast ultrasound.

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.

Development and evaluation of a case-based reasoning classifier for prediction of breast biopsy outcome with BI-RADS lexicon

Approximately 70-85% of breast biopsies are performed on benign lesions. To reduce this high number of biopsies performed on benign lesions, a case-based reasoning (CBR) classifier was developed to predict biopsy results from BI-RADS findings. We used 1433 (931 benign) biopsy-proven mammographic cases. CBR similarity was defined using either the Hamming or Euclidean distance measure over case features. Ten features represented each case: calcification distribution, calcification morphology, calcification number, mass margin, mass shape, mass density, mass size, associated findings, special cases, and age. Performance was evaluated using Round Robin sampling, Receiver Operating Characteristic (ROC) analysis, and bootstrap. To determine the most influential features for the CBR, an exhaustive feature search was performed over all possible feature combinations (1022) and similarity thresholds. Influential features were defined as the most frequently occurring features in the feature subsets with the highest partial ROC areas (0.90AUC). For CBR with Hamming distance, the most influential features were found to be mass margin, calcification morphology, age, calcification distribution, calcification number, and mass shape, resulting in an 0.90AUC of 0.33. At 95% sensitivity, the Hamming CBR would spare from biopsy 34% of the benign lesions. At 98% sensitivity, the Hamming CBR would spare 27% benign lesions. For the CBR with Euclidean distance, the most influential feature subset consisted of mass margin, calcification morphology, age, mass density, and associated findings, resulting in 0.90AUC of 0.37. At 95% sensitivity, the Euclidean CBR would spare from biopsy 41% benign lesions. At 98% sensitivity, the Euclidean CBR would spare 27% benign lesions. The profile of cases spared by both distance measures at 98% sensitivity indicates that the CBR is a potentially useful diagnostic tool for the classification of mammographic lesions, by recommending short-term follow-up for likely benign lesions that is in agreement with final biopsy results and mammographer's intuition.

Application of the mutual information criterion for feature selection in computer-aided diagnosis

The purpose of this study was to investigate an information theoretic approach to feature selection for computer-aided diagnosis (CAD). The approach is based on the mutual information (MI) concept. MI measures the general dependence of random variables without making any assumptions about the nature of their underlying relationships. Consequently, MI can potentially offer some advantages over feature selection techniques that focus only on the linear relationships of variables. This study was based on a database of statistical texture features extracted from perfusion lung scans. The ultimate goal was to select the optimal subset of features for the computer-aided diagnosis of acute pulmonary embolism (PE). Initially, the study addressed issues regarding the approximation of MI in a limited dataset as it is often the case in CAD applications. The MI selected features were compared to those features selected using stepwise linear discriminant analysis and genetic algorithms for the same PE database. Linear and nonlinear decision models were implemented to merge the selected features into a final diagnosis. Results showed that the MI is an effective feature selection criterion for nonlinear CAD models overcoming some of the well-known limitations and computational complexities of other popular feature selection techniques in the field.

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

The authors investigated a new method to optimize artificial neural networks (ANNs) with adaptive filtering used in computer-assisted detection schemes in digitized mammograms and to assess performance changes when averaging classification scores from three sets of optimized schemes. Two independent training and testing image databases involving 978 and 830 digitized mammograms, respectively, were used in this study. In the training data set, initial filtering and subtraction resulted in the identification of 592 mass regions and 3790 suspicious, but actually negative regions. These regions (including both true-positive and negative regions) were segmented into three subsets three times based on the calculation of the values of three features as segmentation indices. The indices were "mass" size multiplied by their digital value contrast, conspicuity, and circularity. Nine ANN-based classifiers were separately optimized using a genetic algorithm for each subset of regions. Each region was assigned three classification scores after applying the three adaptive ANNs. The performance gain of the CAD scheme after averaging the three scores for each suspicious region was tested using an independent data set and a ROC methodology. The experimental results showed that the areas under ROC curves (Az) for the testing database using three sets of optimized ANNs individually were 0.84+/-0.01, 0.83+/-0.01, and 0.84+/-0.01, respectively. The between-index correlations of three A values were 0.013, -0.007, and 0.086. Similar to averaging diagnostic ratings from independent observers, by averaging three ANN-generated scores for each testing region, the performance of the CAD scheme was significantly improved (p<0.001) with Az value of 0.95+/-0.01.