FMM-Diff: A Feature Mapping and Merging Diffusion Model for MRI Generation with Missing Modality

Wenjin Zhong, Cong Cong, Zihan Wang, Zeya Yan, Antonio Di Ieva, Sidong Liu

Abstract: Brain disease diagnosis and treatment planning rely on complementary information from multiple MRI modalities. Compared to routine modalities (\(\texttt{RM}\)) such as T1, T2, and FLAIR, modalities like DWI and T1ce provide unique diagnostic information but are less commonly used due to longer scan times, higher costs, or the need for contrast agents. To mitigate this, multi-modal MRI synthesis methods are proposed to generate advanced MRIs from routine MRIs. However, in clinical practice, missing modality is a known issue in MRI generation which degrades the synthesis quality. Existing methods typically use shared encoders and masking strategies to compensate for missing modality. However, as the number of missing modalities increases, it becomes harder to capture the inter-modal correlations, causing a sharp performance drop. To address this, we propose the Feature Mapping and Merging Diffusion Model (FMM-Diff). Instead of using a shared encoder, we introduce dedicated mapping encoders for each modality. When a modality is missing, its latent representation is inferred from the available ones via its dedicated encoder. This ensures complete latent representations, allowing the Merge Module to selectively extract and fuse inter-modal correlations, significantly improving synthesis performance. Evaluated on two public MRI datasets, including CGGA and BraTS2021, FMM-Diff not only outperforms the state-of-the-art models by 4.35% in terms of Structural Similarity Index Measure (SSIM) while demonstrating exceptional stability, with less than a 1.0% SSIM drop, which is significantly lower than the 2.0–3.45% drop observed with other methods, across various missing modality scenarios. The source code will be available at: https://github.com/ZJohnWenjin/FMMDIFF.git.

External IDs:dblp:conf/miccai/ZhongCWYIL25