Deep Transfer Learning for Texture Classification in Colorectal Cancer Histology
Microscopic examination of tissues or histopathology is one of the diagnostic procedures for detecting colorectal cancer. The pathologist involved in such an examination usually identifies tissue type based on texture analysis, especially focusing on tumour-stroma ratio. In this work, we automate the task of tissue classification within colorectal cancer histology samples using deep transfer learning. We use discriminative fine-tuning with one-cycle-policy and apply structure-preserving colour normalization to boost our results. We also provide visual explanations of the deep neural network's decision on texture classification. With achieving state-of-the-art test accuracy of 96.2% we also embark on using deployment friendly architecture called SqueezeNet for memory-limited hardware.