Abstract
Microscopic examination of tissues or histopathology is one of the diagnosticprocedures for detecting colorectal cancer. The pathologist involved in such anexamination usually identifies tissue type based on texture analysis,especially focusing on tumour-stroma ratio. In this work, we automate the taskof tissue classification within colorectal cancer histology samples using deeptransfer learning. We use discriminative fine-tuning with one-cycle-policy andapply structure-preserving colour normalization to boost our results. We alsoprovide visual explanations of the deep neural network's decision on textureclassification. With achieving state-of-the-art test accuracy of 96.2% we alsoembark on using deployment friendly architecture called SqueezeNet formemory-limited hardware.