Search

Previously, we developed a simple Laguerre-Gauss (LG) channelized Hotelling observer (CHO) for incorporation into our mass computer-aided detection (CAD) system. This LG-CHO was trained using initial detection suspicious region data and was empirically optimized for free parameters. For the study presented in this paper, we wish to create a more optimal mass detection observer based on a novel combination of LG channels. A large set of LG channels with differing free parameters was created. Each of these channels was applied to the suspicious regions, and an output test statistic was determined. A stepwise feature selection algorithm was used to determine which LG channels would combine best to detect masses. These channels were combined using a HO to create a single template for the mass CAD system. Results from free-response receiver operating characteristic curves demonstrated that the incorporation of the novel LG-CHO into the CAD system slightly improved performance in high-sensitivity regions.

The purpose of this study was to evaluate image similarity measures employed in an information-theoretic computer-assisted detection (IT-CAD) scheme. The scheme was developed for content-based retrieval and detection of masses in screening mammograms. The study is aimed toward an interactive clinical paradigm where physicians query the proposed IT-CAD scheme on mammographic locations that are either visually suspicious or indicated as suspicious by other cuing CAD systems. The IT-CAD scheme provides an evidence-based, second opinion for query mammographic locations using a knowledge database of mass and normal cases. In this study, eight entropy-based similarity measures were compared with respect to retrieval precision and detection accuracy using a database of 1820 mammographic regions of interest. The IT-CAD scheme was then validated on a separate database for false positive reduction of progressively more challenging visual cues generated by an existing, in-house mass detection system. The study showed that the image similarity measures fall into one of two categories; one category is better suited to the retrieval of semantically similar cases while the second is more effective with knowledge-based decisions regarding the presence of a true mass in the query location. In addition, the IT-CAD scheme yielded a substantial reduction in false-positive detections while maintaining high detection rate for malignant masses.

In this article, we present a characterization of the effect of difference of Gaussians (DoG) filters in the detection of mammographic regions. DoG filters have been used previously in mammographic mass computer-aided detection (CAD) systems. As DoG filters are constructed from the subtraction of two bivariate Gaussian distributions, they require the specification of three parameters: the size of the filter template and the standard deviations of the constituent Gaussians. The influence of these three parameters in the detection of mammographic masses has not been characterized. In this work, we aim to determine how the parameters affect (1) the physical descriptors of the detected regions, (2) the true and false positive rates, and (3) the classification performance of the individual descriptors. To this end, 30 DoG filters are created from the combination of three template sizes and four values for each of the Gaussians' standard deviations. The filters are used to detect regions in a study database of 181 craniocaudal-view mammograms extracted from the Digital Database for Screening Mammography. To describe the physical characteristics of the identified regions, morphological and textural features are extracted from each of the detected regions. Differences in the mean values of the features caused by altering the DoG parameters are examined through statistical and empirical comparisons. The parameters' effects on the true and false positive rate are determined by examining the mean malignant sensitivities and false positives per image (FPpI). Finally, the effect on the classification performance is described by examining the variation in FPpI at the point where 81% of the malignant masses in the study database are detected. Overall, the findings of the study indicate that increasing the standard deviations of the Gaussians used to construct a DoG filter results in a dramatic decrease in the number of regions identified at the expense of missing a small number of malignancies. The sharp reduction in the number of identified regions allowed the identification of textural differences between large and small mammographic regions. We find that the classification performances of the features that achieve the lowest average FPpI are influenced by all three of the parameters.

As more diagnostic testing options become available to physicians, it becomes more difficult to combine various types of medical information together in order to optimize the overall diagnosis. To improve diagnostic performance, here we introduce an approach to optimize a decision-fusion technique to combine heterogeneous information, such as from different modalities, feature categories, or institutions. For classifier comparison we used two performance metrics: The receiving operator characteristic (ROC) area under the curve [area under the ROC curve (AUC)] and the normalized partial area under the curve (pAUC). This study used four classifiers: Linear discriminant analysis (LDA), artificial neural network (ANN), and two variants of our decision-fusion technique, AUC-optimized (DF-A) and pAUC-optimized (DF-P) decision fusion. We applied each of these classifiers with 100-fold cross-validation to two heterogeneous breast cancer data sets: One of mass lesion features and a much more challenging one of microcalcification lesion features. For the calcification data set, DF-A outperformed the other classifiers in terms of AUC (p < 0.02) and achieved AUC=0.85 +/- 0.01. The DF-P surpassed the other classifiers in terms of pAUC (p < 0.01) and reached pAUC=0.38 +/- 0.02. For the mass data set, DF-A outperformed both the ANN and the LDA (p < 0.04) and achieved AUC=0.94 +/- 0.01. Although for this data set there were no statistically significant differences among the classifiers' pAUC values (pAUC=0.57 +/- 0.07 to 0.67 +/- 0.05, p > 0.10), the DF-P did significantly improve specificity versus the LDA at both 98% and 100% sensitivity (p < 0.04). In conclusion, decision fusion directly optimized clinically significant performance measures, such as AUC and pAUC, and sometimes outperformed two well-known machine-learning techniques when applied to two different breast cancer data sets.

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.

We propose to investigate the use of the subregion Hotelling observer for the basis of a computer aided detection scheme for masses in mammography. A database of 1320 regions of interest (ROIs) was selected from the DDSM database collected by the University of South Florida using the Lumisys scanner cases. The breakdown of the cases was as follows: 656 normal ROIs, 307 benign ROIs, and 357 cancer ROIs. Each ROI was extracted at a size of 1024 x 1024 pixels and sub-sampled to 128 x 128 pixels. For the detection task, cancer and benign cases were considered positive and normal was considered negative. All positive cases had the lesion centered in the ROI. We chose to investigate the subregion Hotelling observer as a classifier to detect masses. The Hotelling observer incorporates information about the signal, the background, and the noise correlation for prediction of positive and negative and is the optimal detector when these are known. For our study, 225 subregion Hotelling observers were set up in a 15 x 15 grid across the center of the ROIs. Each separate observer was designed to "observe," or discriminate, an 8 x 8 pixel area of the image. A leave one out training and testing methodology was used to generate 225 "features," where each feature is the output of the individual observers. The 225 features derived from separate Hotelling observers were then narrowed down by using forward searching linear discriminants (LDs). The reduced set of features was then analyzed using an additional LD with receiver operating characteristic (ROC) analysis. The 225 Hotelling observer features were searched by the forward searching LD, which selected a subset of 37 features. This subset of 37 features was then analyzed using an additional LD, which gave a ROC area under the curve of 0.9412 +/- 0.006 and a partial area of 0.6728. Additionally, at 98% sensitivity the overall classifier had a specificity of 55.9% and a positive predictive value of 69.3%. Preliminary results suggest that using subregion Hotelling observers in combination with LDs can provide a strong backbone for a CAD scheme to help radiologists with detection. Such a system could be used in conjunction with CAD systems for false positive reduction.

The likelihood ratio (LR) is an optimal approach for deciding which of two alternate hypotheses best describes a given situation. We adopted this formalism for predicting whether biopsy results of mammographic masses will be benign or malignant, aiming to reduce the number of biopsies performed on benign lesions. We compared the performance of this LR-based algorithm (LRb) to a case-based reasoning (CBR) classifier, which provides a solution to a new problem using past similiar cases. Each classifier used mammographers' BI-RADS descriptions of mammographic masses as input. The database consisted of 646 biopsy-proven mammography cases. Performance was evaluated using Receiver Operating Characteristic (ROC) analysis, Round Robin sampling, and bootstrap. The ROC areas (AUC) for the LRb and CBR were 0.91+/- 0.01 and 0.92 +/- 0.01, respectively. The partial ROC area index (0.90AUC) was the same for both classifiers, 0.59 +/- 0.05. At a sensitivity of 98%, the CBR would spare 204 (49%) of benign lesions from biopsy; the LRb would spare 209 (51%) benign lesions. The performance of the two classifiers was very similar, with no statistical differences in AUC or 0.90AUC. Although the CBR and LRb originate from different fields of study, their implementations differ only in the estimation of the probability density functions (PDFs) of the feature distributions. The CBR performs this estimation implicitly, while using various similarity metrics. On the other hand, the estimation of the PDFs is specified explicitly in the LRb implementation. This difference in the estimation of the PDFs results in the very small difference in performance, and at 98% sensitivity, both classifiers would spare about half of the benign mammographic masses from biopsy. The CBR and LRb are equivalent methods in implementation and performance.

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.