In the first stage, the mammogram was filtered so that only suspicious MC candidates remained. To do so, a hybrid filter consisting of a wavelet filter, a top-hat filter, and 15 Laws filters was applied to alleviate the problem of low contrast between MCs and surrounding breast tissue. This filtering not only lessened the low contrast problem but also reduced the tremendous computation time because only high-frequency components remained for further processing. In the second stage, all candidates were examined by a knowledge-based classifier to reduce the number of false positives (FPs). Furthermore, the remaining candidates were classified by support vector machines (SVM) via a set of features after an automatic feature selection, in which the optimal parameter sets for SVM were also determined.

Finally, we clustered individual MCs to MCCs and marked the identified MCCs on the images as a result.The rest of this paper is organized as follows: Section 2 introduces our mammogram database and methods of pre-processing, filtering, feature extraction, automatic feature selection, training, and classification. Our experimental results are shown in Section 3. We then discuss our method and methods from other groups in Section 4. Finally, the conclusions are provided in Section 5.2.?Methods2.1. Datasets and Ground TruthFifty-two patients (cases) with clinical reports were collected, from which a total of 111 digital mammograms were acquired. The image gray-level resolution was 14-bit per pixel. Each patient had at least one craniocaudal (CC) view and one mediolateral oblique (MLO) view.

All the mammograms, which were representative images containing MCCs, were acquired from China Medical University Hospital. The patient mammograms were selected by two radiologists, who selected mammograms that they both agreed contained precisely recognizable MCs. Patients whose mammograms were not able to be identified consistently by these two radiologists were excluded. To establish ground truth, all mammogram readings were performed by these two experienced radiologists independently. One radiologist was a senior clinician who has worked in this area for over ten years. The other radiologist was young and has worked more than two years. In each mammogram, a rectangle (or some rectangles) was drawn to enclose the MCCs, and a point was manually marked in the center of each MC. The rectangles were drawn as small as possible to cover the MCCs. The manually identified Drug_discovery MCs were set as the gold standard used as the ground truth to which the automated results were compared.To make a statistical analysis, we used 2-fold cross-validation [22,23] to test our algorithm. The dataset was randomly separated into two subsets.