Anna O. Bilska-Wolak

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 mammographers intuition.

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.

Investigating different similarity measures for a case-based reasoning classifier to predict breast cancer

This paper investigates the effects of using different similarity measures for a case-based reasoning (CBR) classifier to predict breast cancer. The CBR classifier used a mammographers BI-RADSTM description of a lesion to predict breast biopsy outcome. The classifier compared the case to be examined to a reference collection of cases and identified those that were similar. The decision variable was formed as the ratio of similar cases that were malignant to all similar cases. A reference collection of 1027 biopsy-proven cases from Duke University Medical Center was used as input. Both Euclidean and Hamming distance measures were compared using all possible combinations of nine BI-RADSTM features and age. Performance was evaluated using jackknife sampling and ROC analysis. For all combinations of features, it was found that Euclidean distance measure produced greater ROC areas and partial ROC areas than Hamming. The differences were significant at an alpha level of 0.05. The greatest ROC area of 0.82 +/- 0.01 was generated using six of the features and Euclidean distance measure. The results of both distance measures yielded greater ROC areas than previously reported values and were similar to results generated with an Artificial Neural Network using 10 features.

Incorporation of a multiscale texture-based approach to mutual information matching for improved knowledge-based detection of masses in screening mammograms

Mutual information is a popular intensity-based image similarity measure mainly used in image registration. This measure has been also very successful as the similarity metric in our knowledge-based computer-assisted detection (CADe) system for the detection of masses in screening mammograms. Our CADe system is designed to assess a new, query case based on its similarity with known cases stored in the knowledge database. However, intensity-based mutual information captures only relationships between the gray level values of corresponding pixels. This study presents a novel advancement of our CADe system by incorporating neighborhood textural information when estimating the mutual information of two images. Specifically, an entropy filter is applied to the images, effectively replacing each image pixel value with its neighborhood entropy. This pixel-based entropy is a localized measure of image texture. Then, the information-theoretic CAD system is asked to make a decision regarding the query case using the texture-based mutual information similarity metric. The entropy-based image enhancement and MI-based decision making processes are repeated at different neighborhood scales. Finally, an artificial network merges intensity-based and texture-based decisions to investigate possible improvements in mass detection performance. Given a database of 1,820 regions of interest (ROIs) extracted from screening mammograms (901 depicting a biopsy-proven mass and 919 depicting normal parenchyma) and a leave-one out sampling scheme, the study showed that our CADe system achieves an ROC area of 0.87±0.01 using the intensity-based ROC. The ROC performance for the texture-based CADe system ranges from 0.69±0.01 to 0.83±0.01 depending on the scale of analysis. The synergistic approach of the ANN using both intensity-based and texture-based information resulted in statistically significantly better performance with an ROC area index of 0.93±0.01.

Prediction of breast biopsy outcome using a likelihood ratio classifier and biopsy cases from two medical centers

Potential malignancy of a mammographic lesion can be assessed using the mathematically optimal likelihood ratio (LR) from signal detection theory. We developed a LR classifier for prediction of breast biopsy outcome of mammographic masses from BI-RADS findings. We used cases from Duke University Medical Center (645 total, 232 malignant) and University of Pennsylvania (496, 200). The LR was trained and tested alternatively on both subsets. Leave-one-out sampling was used when training and testing was performed on the same data set. When tested on the Duke set, the LR achieved a Received Operating Characteristic (ROC) area of 0.91± 0.01, regardless of whether Duke or Pennsylvania set was used for training. The LR achieved a ROC area of 0.85± 0.02 for the Pennsylvania set, again regardless of which set was used for training. When using actual case data for training, the LRs procedure is equivalent to case-based reasoning, and can explain the classifiers decisions in terms of similarity to other cases. These preliminary results suggest that the LR is a robust classifier for prediction of biopsy outcome using biopsy cases from different medical centers.