Contrastive Centroid Supervision Alleviates Domain Shift in Medical Image Classification
This addresses domain shift for medical imaging applications, offering a method that outperforms existing approaches while requiring only labeled data from a single source domain.
The paper tackles the domain shift problem in medical image classification by proposing Feature Centroid Contrast Learning (FCCL), which improves target domain performance using contrastive loss between instances and class centroids, achieving superior results across three imaging modalities.
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.