Studies on Computerized Analysis of Clustered Microcalcifications on Mammograms for Diagnosis Aid

In order to overcome the increase in breast cancer death rate，mammography which is considered the most sensitive method for detection of early breast cancers has been introduced for breast cancer screening in many advanced nations. Clustered microcalcifications which are present in 30%-50% of all cancers found in a mammography are one of the important radiographic indications on mammograms. However，it is often difficult for radiologists to detect clustered microcalcifications correctly because they are very small and obscure. It is also difficult to differentiate between benign and malignant clustered microcalcifications. With the concept of CAD (Computer-aided Diagnosis)，it is expected that radiologists' performance in the diagnosis of medical images will improve by taking into account the analysis result of a lesion obtained from a computerized analysis as a“second opinion."" Therefore，the purpose of this dissertation research is to develop a computerized analysis for both a detection aid and a differentiation aid of clustered microcalcifications on mammograms. For a detection aid，we have developed a computerized detection method for indicating a potential region of clustered microcalcifications on mammograms for radiologists. In the computerized detection method, it is important to not only detect clustered microcalcifications with high sensitivity but also segment individual microcalcifications while maintaining their shapes because image features are extracted from the detected clustered microcalcifications in the computerized analysis for a differentiation aid. For the computerized detection method, therefore, we first constructed a novel filter bank with the requirements for a perfect reconstruction by introducing the concept of a Hessian matrix for classifying nodular structures and linear structures. A mammogram image is then decomposed into several subimages for a second difference at scales from 1 to 4 by this filter bank. The subimages for the nodular component (NC) and the subimages for the nodular and linear component(NLC)are then obtained from analysis of the Hessian matrix based on those subimages for second difference. Many regions of interest (ROIs) are selected from the mammogram image. In each ROI, eight objective features are determined from each of the subimages for NC at scales from 1 to 4 and the subimages for NLC at scales from 1 to 4. A Bayes discriminant function with the eight objective features is employed for distinguishing between abnormal ROIs with clustered microcalcifications and normal ROIs without clustered microcalcifications. The region connecting the ROIs classified as abnormal ROI is considered to be a potential region of clustered microcalcifications. With the proposed detection method, sensitivity and a false positive rate was 100.0% and 0.98 per image. respectively. For the differentiation aid, we developed a computerized classification method for providing radiologists the likelihood of histological classifications of clustered microcalcifications and a computerized retrieval method for providing radiologists images of lesions with known pathology similar to an unknown lesion. There are differences in both the image features and the growth speeds among histological classifications of clustered microcalcifications. In the computerized classification method, therefore，we extracted six objective features from clustered microcalcifications on each of the follow-up magnification mammograms (i.e. both current and previous magnification mammograms). In identifying histological classification of clustered microcalcifications, the histological classification of an unknown new case in question is assumed to be the same as that of the nearest neighbor case which has the shortest Euclidean distance in a feature-space. The feature-spaces for the nearest neighbor case consist of six objective features obtained from the previous magnification mammogram (previous features), six objective features obtained from the current magnification mammogram (current features), and the set of the six previous features and the six current features. The classification accuracies with the six current features were higher than those with the six previous features. These classification accuracies were improved substantially by using the set of the six previous features and the six current features. With the proposed classification method，the classification accuracies were 90.9% (10 of 11) for invasive carcinoma，89.5% (17 of 19) for noninvasive carcinoma of the comedo type，96.0% (24 of 25) for noninvasive carcinoma of the noncomedo type，82.6% (19 of 23) for mastopathy，and 93.3% (14 of 15) for fibroadenoma. In order to also present radiologists similar lesions as a differentiation aid，we investigated four objective similarity measures as an image-retrieval tool for selecting lesions similar to unknown lesions in terms of radiologists' visual perception. In the observer study，we confirmed that the presentation of similar images can improve radiologists' performance in the differential diagnosis of clustered microcalcifications on mammograms. The proposed computerized analysis for both detection aid and differentiation aid for clustered microcalcifications achieves high detection performance and high classification accuracies，and would help radiologists improve the diagnosis accuracy of clustered microcalcifications at mammography in clinical practice