LSRML: A latent space regularization based meta-learning framework for MR image segmentation

Abstract

Data sources for medical image segmentation can be quite extensive, and models trained with data from a source domain may perform poorly on data from the target domain owing to domain shift issues. To overcome the impact of domain shift, we propose a novel meta-learning-based multi-source domain adaptation framework for medical image segmentation. Specifically, we designed a domain discriminator module to produce category prediction over the latent features, and an image reconstruction module to reconstruct the foreground and background of the target domain image separately. Furthermore, we constructed a large-scale multi-modal prostate dataset, which contained 495,902 magnetic resonance images of 419 cases, with prostate and lesion masks, as well as diagnostic descriptions for each patient. We evaluated our proposed method through extensive experiments using the proposed and the benchmark datasets. Experimental results show that our model achieves better segmentation and generalization performance compared to state-of-the-art approaches.