A DEEP LEARNING BASED PIPELINE FOR METASTATIC BREAST CANCER CLASSIFICATION FROM WHOLE SLIDE IMAGES (WSI)

Abstract

Pathology is a 150-year-old medical specialty that has seen a paradigm shift over the past few years with the advent of Digital Pathology. Digital Pathology is a very promising approach to diagnostic medicine to accomplish better, faster and cheaper diagnosis, prognosis and prediction of cancer and other important diseases. Historical approaches in Digital Pathology have focused primarily on low-level image analysis tasks (e.g., color normalization, nuclear segmentation, and feature extraction) hence they are not generalized, thus not useful for practical use in clinical practices. In this thesis, a general Deep Learning based classification pipeline for identifying cancer metastases from histological images is proposed. GoogLeNet, a deep 27 layer Convolution Neural Network (ConvNet) is used to distinguish positive tumor areas from negative ones. The key challenge of detecting hard negative areas (areas surrounding tumor region) is tackled with ensemble learning method using two deep ConvNet models. Using dataset of the Camelyon'16 grand challenge, proposed pipeline achieved an area under the receiver operating curve (ROC) of 92.57%, which beats the winning method of Camelyon'16 grand challenge, developed together by Harvard & MIT research labs. These results demonstrate the power of using deep learning to produce significant improvements in the accuracy of pathological diagnoses.