Contrastive Centroid Supervision Alleviates Domain Shift in Medical Image Classification

TL;DR

This paper introduces FCCL, a contrastive learning method that improves medical image classification across different domains without needing target domain data, outperforming existing methods.

Contribution

Proposes Feature Centroid Contrast Learning (FCCL), a novel contrastive supervision technique that enhances domain generalization in medical imaging with only source domain labels.

Findings

FCCL outperforms existing domain adaptation methods.

Effective across multiple imaging modalities.

Requires only labeled source data, no target data.

Abstract

Deep learning based medical imaging classification models usually suffer from the domain shift problem, where the classification performance drops when training data and real-world data differ in imaging equipment manufacturer, image acquisition protocol, patient populations, etc. We propose Feature Centroid Contrast Learning (FCCL), which can improve target domain classification performance by extra supervision during training with contrastive loss between instance and class centroid. Compared with current unsupervised domain adaptation and domain generalization methods, FCCL performs better while only requires labeled image data from a single source domain and no target domain. We verify through extensive experiments that FCCL can achieve superior performance on at least three imaging modalities, i.e. fundus photographs, dermatoscopic images, and H & E tissue images.