Abstract
We have been developing a computer-aided diagnosis (CAD) system for mammograms. We developed a triple-ring filter for detecting microcalcifications, and the prototype detection system is nearly complete. However, our prototype system does not allow for the detection of subtle microcalcifications, which have a low contrast and can be confused with circumferences of almost the same density. The purpose of this study is to develop a new pattern recognition method using the higher-order autocorrelation features (HOAFs) specially created for subtle microcalcifications. The ROI (region of interest) for extracting the feature was experimentally determined as 9×9 in consideration of the size of microcalcifications in the mammography. We employed 120 ROIs including subtle microcalcifications and 120 normal ROIs for training. Forty-five features calculated from the triple-ring filter (8 features), the mean and the variance of the pixel values of the local area (2 features), and the HOAFs (35 features) were extracted from these ROIs. The various features of microcalcifications and FP (false-positive) shadows were extracted and trained using the multi-regression analysis. As a result of comparing those performances using a FROC (free-response receiver operating characteristic) curve, the proposed method always performed better than the current method. It seems that the HOAFs, which reflect local features, are effective in detecting subtle microcalcifications. As a result of applying the new clustered microcalcification system, which combined this method and the current method with 556 unknown images, sensitivity was 94.2%, and the number of FPs per image was 0.61. In our prototype system, the sensitivity was 92.4% and the number of FPs per image was 0.61. The results of our experiment proved that this method can identify subtle microcalcifications that cannot be detected using the current system, and detection performance might be improved by combining it with our prototype system.
