CataNet A Deep Learning Model to detect Breast Cancer Metastasis

Abstract

Pathologists invest hours to distinguish metastasis in hematoxylin and eosin (H&E) stained whole-slide images of lymph node sections. Digital pathology is a recent and fast growing field of medical imaging. In this field, whole-slide scanners are utilized to scan glass slides which contain tissue samples at high resolution (up to 160nm per pixel). Lymph node metastases occur in many different types of cancers such as breast, prostate, colon etc. The lymph nodes in the underarm are the most likely place for cancer to first spread to. ‘CataNet’ focuses on the detection of micro and macro metastasis in digitized images of sentinel lymph nodes. The project is divided into two phases. For Slide-Level classification, we detect and localize the cancer cells in whole slide images. For this purpose, we first extract Region of Interest with Image processing, then construct Training Data tiles from ROI, Train Deep Convolutional model for tile-based classification, building tumor probability heat-maps using trained model and report detected tumor. At the end of phase 1, given a hematoxylin and eosin (H&E) stained whole-slide images of lymph node, we are able to tell whether the image contains tumor or not, and if present then localize it within the image. For Patient-Level classification, we determine the pathologic N-stage (pN-stage) label per patient by extracting the features from the probability maps generated by Phase 1, training a classifier on these features to determine slide level metastasis label and the determining the pN stage of patient using predetermined criteria. At the end of phase 2, given at least 5 whole-slide images of a patient, the product shall be able to determine the pN-stage as standardized by the Union for International Cancer Control (UICC). The aim is to contribute our minor share for the service of this country by building this tool indigenously and helping the local research community.