Towards Measuring Domain Shift in Histopathological Stain Translation in an Unsupervised Manner

TL;DR

This paper introduces a method using PixelCNN and a domain shift metric to detect and quantify domain shift in digital histopathology, enabling better assessment of model generalisation to unseen data.

Contribution

It demonstrates that PixelCNN combined with a domain shift metric effectively detects and measures domain shift in histopathology, correlating with model performance.

Findings

PixelCNN and domain shift metric detect domain shift effectively.

Strong correlation between domain shift measurement and model generalisation.

Enables inference of model performance on unseen, unlabelled data.

Abstract

Domain shift in digital histopathology can occur when different stains or scanners are used, during stain translation, etc. A deep neural network trained on source data may not generalise well to data that has undergone some domain shift. An important step towards being robust to domain shift is the ability to detect and measure it. This article demonstrates that the PixelCNN and domain shift metric can be used to detect and quantify domain shift in digital histopathology, and they demonstrate a strong correlation with generalisation performance. These findings pave the way for a mechanism to infer the average performance of a model (trained on source data) on unseen and unlabelled target data.