(英) |
In this paper, we propose a novel segmentation method of breast tumors,
namely carcinoma, fibroadenoma and cyst, to construct an fully automated diagnosis algorithm of breast tumors in ultrasonic images. We developed a discrimination algorithm of breast tumors whose boundaries were delineated by a human observer from original ultrasonic images. By combining a roposed segmentation process with the discrimination algorithm, we can realize a fully automated diagnosis algorithm. We propose a novel method that concatenates an ensemble segmentation trained by the AdaBoost with a geodesic active contour. When applying the proposed system to 200 carcinomas, 50 fibroadenomas and 50 cycts, it was confirmed that an average Jaccard index between the extracted tumors and manually segmented regions is over 97$\%$. Furthermore, the discrimination performance of the proposed fully automated process was almost comparable to that of the previous discrimination algorithm that needs manually segmented breast tumors. |