MU-Diff: a mutual learning diffusion model for synthetic MRI with Application for brain lesions

This is the authors' abstract. We don't add key points for this paper.

Synthesizing brain MRI lesions is challenging due to the heterogeneity of lesion characteristics and the complexity of capturing fine-grained pathological information across MRI contrasts. Additionally, leveraging complementary information across multiple contrasts is difficult due to their diverse feature representations. To address these challenges, we propose a mutual learning-based framework with an adversarial diffusion approach. Our framework uses two denoising networks: one captures contrast-specific features to handle diverse representations, while the other emphasizes contrast-aware adaptation to model subtle pathological variations. A shared critic network ensures consistency, facilitates collaborative learning, and identifies critical lesion regions for focused synthesis. We benchmark our method on two public lesion datasets, treat each contrast as a missing target, and validate it on a brain tumor and multi-contrast healthy MRI dataset. Our approach outperforms state-of-the-art methods, delivering accurate lesion synthesis and superior downstream segmentation performance, highlighting the diagnostic value and accuracy of the proposed framework.