Speaker: Dr. Baidya Nath Saha
Topic: “Automated Histopathological Image Analysis to Predict Breast Cancer Disease survival and Recurrency “
Time: 2:00 pm, Friday, October 17, 2014
Place: P 148 (refreshments will be served at 1:45 in P-145A)
Abstract:
Immunohistochemical staining of the proliferation antigen Ki-67, tumor epithelial marker Pan Cytokeratin and nuclear stain DAPI are routinely used in the diagnostic and prognostic assessment of breast cancer. In clinical practice, these histological slides are interpreted by a pathologist which are not only prone to inter-observer variation, but are also tedious and time-consuming. To overcome these, an automated image analysis is proposed which can be decomposed into three steps: (a) segmentation, (b) feature extraction and (c) classification. At first, a novel level set based energy functional is proposed to segment the cell boundaries from Ki-67, Pan Cytokeratin and DAPI images. The novel energy functional allows an individual energy functional for each cell and the individual energy functional provide repulsive forces among the neighboring cells that enable to separate the overlapping cells. Then a fast ellipse fitting algorithm is applied along each cell contour and we compute the histogram of the deviation of the cell boundaries from the best fitted ellipse to capture the inherent morphological complexity of epithelial architecture. It is assumed that the shapes of the normal cells are elliptical in nature and the abnormality of the epithelial cell architecture is also exploited for disease prognosis (distant recurrence / non-recurrence breast cancer, five year-survival or less). This histogram information is fed into a novel K-Nearest Neighbor (kNN) regularized SVM classifier. We propose a new k-Nearest Neighbor (kNN) based regularization term into SVM optimization framework which improves the classification accuracy of SVM by leveraging the strengths of both SVM and kNN. The intuition is that data points that are mostly surrounded by the data points belong to its own class are given more priority during training to increase generalization ability. Results of these proposed methods are demonstrated superiority over state-of-the-art techniques. Improving the performance of the proposed method by exploiting a pool of shape, gradient and intensity based features to capture the complex epithelial architecture for disease diagnosis and prognosis is the future work of this research