Unsupervised Cross-Modality MR Image Segmentation via Prompt-Driven Foundation Model

Obtaining pixel-level expert annotations is expensive and labor-intensive in the medical imaging field, especially for multi-modality imaging data like MR. Most conventional cross-modality segmentation methods rely on unsupervised domain adaptation to achieve efficient cross-domain segmentation. However, these methods are often hindered by discrepancies between the source and target domains. In this paper, we propose a new scheme for cross-modality segmentation based on foundation models, which uses spatial consistency across multiple modalities and is not affected by discrepancies between the source and target domains. This scheme allows us to use annotated data from one imaging modality to train a network capable of performing accurate segmentation on other target imaging modalities, without the need for labels or registration processes. Specifically, we propose using a SAM-based model that uses segmentation results from one imaging modality as pseudo labels and prompts to guide training and testing in the target imaging modality. Moreover, we introduce consistency-based prompt tuning and hybrid representation learning to address potential unregistered issues and noisy label problems that may arise in cross-modality segmentation. We conducted extensive validation experiments on two internal datasets and one public dataset, including liver lesion segmentation and liver segmentation. Our method demonstrates significant improvement compared to current state-of-the-art approaches.