Active Contour Model And Its Application To Prostate Cancer Grading

In clinical hospital, the Gleason grading of prostate pathology tissue is got by pathology doctors according to the experience through artificial analysis in different pathological markers. However the artificial method is time consuming and basically relies on the subjective opinion of the doctor, so it isn’t replication. Recent studies have shown that different pathology doctors with many years of experience got great difference for the same slice of pathology tissue. Even if the same doctor at different times with different mood, fatigue also got different opinion for the same slice. Therefore, computer-aided automatic pathology classification is very meaningful, because it can provide opinions with more objective analysis results, at the same time, it can avoid unnecessary error detection and missing detection. According to the current common prostate cancer in the Gleason classification systems, morphological structure of glands is closely related to the degree of cancer. In order to solve this problem, the goal of this thesis is to build a gland grading system automatically based on the segmentation and shape feature description of prostate glands. This system consists of two steps:glands segmentation and glands classification, and the specific content as follows:we use the local-based active contour model to extract the gland lumen and establish the medial axis shape of the glands during establishing automatic grading system. By medial axis transformations, we can realize the registration between medial axis shapes. And then, we take this as our shape feature after computing the similarity between them. Finally, we used the support vector machine to get the classification results of Gleason3 and Gleason4. From the experiment result, we can see the proposed computer automatic grading system which focus on Gleason3 tissue and Gleason4 tissue can get better segmentation accuracy and better classification performance when compared to the other systems which based on the classic segmentation method. Initial shape learning problem of active contour model, this paper puts forward a new initialization method to replace the artificial initialization or random initialization. As the global shapes or local shapes of different objects may be similar, we firstly extract the local shapes of the target image. Then we find the matched local shapes set in the exemplar database. The global shapes of objects from the exemplar database are subsequently transferred to the target image based on the size and relative location of the local shapes. Finally, we can obtain the initial shapes in accordance with the global shape coverage after their intersection operation. We regard the sign distance function of these initial shapes as the initial involution functional of the active contour model. From the experiment, we can see the proposed initialization method based on the hybrid active contour model can get higher segmentation accuracy, and reduce the segmentation time.