Automated Segmentation of Malignant Mass in Mammography Using the Principal Component Analysis Network Based Deep Learning Mode l

摘要

The accurate segmentation of breast tumor is essential for early diagnosis of breast cancer and targeted therapy. X-ray mammography is the standard method for breast cancer detection. However, the manual segmentation of the suspected lesion is an error-prone and time-consuming work for the physicians. The traditional human-designed segmentation method cannot identify the breast tumor accurately due to the fuzziness of the breast boundary and the complicated structural information in the breast image. To address this problem, the deep learning based two-stage segmentation method is proposed by combining the super-pixel segmentation with the principal component analysis network (PCANet) based segmentation. The first stage aims at the coarse segmentation (CS) of the breast image. At this stage, the super-pixel segmentation is implemented to divide the whole image into several homogeneous regions. Then the image patches centered at the center of the homogeneous regions are input into the trained two-layer PCANet for feature extracting. These features are fed into the support vector machine (SVM) for the classification of each central pixel. The second stage aims at the fine segmentation of image pixels by only using the suspected pixels within the CS results for the PCANet and SVM. Experiments have been conducted on the digital database for screening mammography (DDSM) and some clinical data. The segmentation performance of the proposed method is appreciated using such indexes as Accuracy, Sensitivity, Specificity and Recall and it is compared with that of the traditional level-set segmentation method, and PCANet-based segmentation methods. The statistical results demonstrate that the proposed method is provided with much better segmentation performance than the compared methods in terms of the above indexes and human vision.