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

**Authors:** Pietro Antonio Cicalese, Aryan Mobiny, Pengyu Yuan, Jan Becker, Chandra Mohan, Hien Van Nguyen  
**arXiv:** [2007.05008](https://arxiv.org/abs/2007.05008)  
**Date:** 2020-07-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2007.05008v1/

## Key points

- Antibody Mediated Rejection (AMR) classification is a significant challenge in kidney transplant pathology
- Low inter-observer agreement and poor reproducibility in histological tissue stain analysis contribute to the challenge
- Variability in tissue stain quality among labs and fading archival sections hinder accurate morphological feature description by pathologists
- A novel pipeline proposes using StyPath, a histological data augmentation technique, to develop robust deep neural networks for AMR classification
- StyPath outperforms other popular histological data augmentation techniques, generating images efficiently and reducing error rates from 14.8% to 11.5%
- The model's performance was evaluated using Bayesian performance estimates and Grad-CAM representations, showing improved histological classification performance and generalization ability

## Summary

In the field of kidney transplant pathology, the classification of Antibody Mediated Rejection (AMR) poses a significant challenge for even experienced nephropathologists. This is due to low inter-observer agreement and poor reproducibility in histological tissue stain analysis. The variability in tissue stain quality among pathology labs, as well as the gradual fading of archival sections, contribute to this difficulty. These variations in stain colors and intensities hinder pathologists' ability to accurately describe relevant morphological features, impacting patient treatment and care. To address these challenges, a novel pipeline has been proposed to develop robust deep neural networks for AMR classification using StyPath, a histological data augmentation technique that utilizes a lightweight style-transfer algorithm to reduce sample-specific bias. This approach has shown promising results, with each image being generated efficiently in 1.84 +- 0.03 seconds using a single GTX TITAN V GPU and PyTorch, outperforming other popular histological data augmentation techniques. The model's performance was evaluated using a Monte Carlo estimate of Bayesian performance, generating an epistemic measure of uncertainty to compare both the baseline and StyPath augmented models. Additionally, Grad-CAM representations of the results were assessed by an experienced nephropathologist through qualitative analysis to elucidate on the assumptions made by each model. The findings suggest that the style-transfer augmentation technique significantly improves histological classification performance by reducing error rates from 14.8% to 11.5% while enhancing generalization ability. This advancement holds promise for improving AMR classification accuracy based on kidney histology images, ultimately benefiting patient outcomes and advancing the field of transplant pathology research.