Deep Learning for Detecting Breast Cancer Metastases on WSI

Pathologists face a substantial increase in workload and complexity of histopathologic cancer diagnosis due to the advent of personalized medicine. Therefore, diagnostic protocols have to focus equally on efficiency and accuracy. In this paper, we proposed an improved Deep Learning based classification pipeline for detection of cancer metastases from histological images. The pipeline consists of five stages: 1. Region of Interest (ROI) detection with Image processing. 2. Tiling ROI. 3. Deep Convolutional Neural Network (CNN) for tile-based classification. 4. Building tumor probability heat-maps. 5. Post-processing of heat-maps for slide-based classification. Our system achieved the final AUC score of 90.23% which beats the winning method of Camelyon-16 grand challenge. Compared with common methods which pay attention to the training process, we lay emphasis on the data preprocessing and data quality. In order to reduce the patches without cells, we combined opening and closing operation with Otsu algorithm together. In addition, the hard negative method was also been used to remove false positives and balance positive and negative samples. Our method yields progressive sensitivity on the challenging task of detecting small tumors in gigapixel pathology slides. Moreover, we could improve accuracy and consistency of evaluating breast cancer cases, and potentially improve patient outcomes.