Data-efficient Cross-domain Medical Image Segmentation

Abstract

Deep learning-based segmentation methods have been widely employed for medical image analysis, especially for automatic disease diagnosis and prognosis. Nevertheless, existing deep-learning models benefit from large amounts of annotated data, bringing auxiliary data acquisition and annotation costs. In practice, privacy, security, and storage concerns often impede the availability of medical images for model training.

On the other side, most of the deep learning models suffer from performance drops when validated on unseen datasets with distribution shifts. Unsupervised domain adaptation (UDA) has been developed to address this issue by transferring the knowledge from the labeled source data to the unlabeled target data. To further facilitate the data efficiency of the cross-domain segmentation methods, we explore UDA medical image segmentation problems using a few labeled source data and under a multi-source data-free situation in this work.

For UDA image segmentation with few labeled source data, we first create a searching-based multi-style invariant mechanism to expand the data distribution with style diversity. A prototype consistency mechanism for the foreground objects is then developed to enable the alignment of the features for each kind of tissue in various image styles. The segmentation performance on the target photos is further improved by a cross-style self-supervised learning strategy.

For a multi-source data-free UDA problem, a single student multiple teach network is initially established to distill knowledge from several pre-trained source models. The pre-trained model is then sorted to remove domain biases from the various source domains using a weighted transfer learning module. A Cross-domain averaging module also preserves overall consistency by accounting for model parameters. Our methods have outperformed several state-of-the-art UDA segmentation methods on both retinal fundus segmentation and MRI prostate segmentation tasks.