Identifying Tumor in Whole-Slide Images of Breast Cancer Using Transfer Learning and Adaptive Sampling

Abstract

Deep learning is widely used in medical applications in view of the excellent performance it achieved in image processing. Early methods of diagnosis on whole slide images (WSIs) is usually based on dense sampling which is time-consuming and requires a lot of memory to handle it. In this paper, we propose an adaptive sampling method that classify WSI of breast biopsies into two categories (cancer area and normal area) complied by transfer learning. This method involves: i) an adaptive sampling method based on probability gradient map. ii) a classifier which contain feature extraction part and classifier part to divide WSI into two categories. We tried nine different transfer learning models based on TensorFlow and Keras platform and apply the model to execute classification in WSI under three different magnifications (x5, x20, x40). The results showed that (1) the transfer learning combined with SVM or NN is enough to detect the cancer area which achieved an average test accuracy of 97.07% under x20 magnification, and (2) the adaptive sampling method is an effective strategy to deal with WSI with good performance (achieve the Dice coefficient of 80%) and far fewer samples (less than 5% of samples when use uniform sampling method).