StyPath: Style-Transfer Data Augmentation For Robust Histology Image Classification

The classification of Antibody Mediated Rejection (AMR) in kidney transplant\nremains challenging even for experienced nephropathologists; this is partly\nbecause histological tissue stain analysis is often characterized by low\ninter-observer agreement and poor reproducibility. One of the implicated causes\nfor inter-observer disagreement is the variability of tissue stain quality\nbetween (and within) pathology labs, coupled with the gradual fading of\narchival sections. Variations in stain colors and intensities can make tissue\nevaluation difficult for pathologists, ultimately affecting their ability to\ndescribe relevant morphological features. Being able to accurately predict the\nAMR status based on kidney histology images is crucial for improving patient\ntreatment and care. We propose a novel pipeline to build robust deep neural\nnetworks for AMR classification based on StyPath, a histological data\naugmentation technique that leverages a light weight style-transfer algorithm\nas a means to reduce sample-specific bias. Each image was generated in 1.84 +-\n0.03 seconds using a single GTX TITAN V gpu and pytorch, making it faster than\nother popular histological data augmentation techniques. We evaluated our model\nusing a Monte Carlo (MC) estimate of Bayesian performance and generate an\nepistemic measure of uncertainty to compare both the baseline and StyPath\naugmented models. We also generated Grad-CAM representations of the results\nwhich were assessed by an experienced nephropathologist; we used this\nqualitative analysis to elucidate on the assumptions being made by each model.\nOur results imply that our style-transfer augmentation technique improves\nhistological classification performance (reducing error from 14.8% to 11.5%)\nand generalization ability.\n