DiffM⁴RI: A Latent Diffusion Model With Modality Inpainting for Synthesizing Missing Modalities in MRI Analysis.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Multimodal data fusion

Background:

• Foundation Models (FMs) show promise for multimodal medical image analysis, particularly Magnetic Resonance Imaging (MRI).
• Missing MRI sequences due to constraints like scan time or patient comfort can limit FM performance in clinical settings.
• Existing imputation methods, such as Generative Adversarial Networks (GANs), often require scenario-specific training and suffer from limited synthesis diversity.

Purpose of the Study:

• To develop a robust method for imputing missing MRI modalities that can handle arbitrary missing scenarios without retraining.
• To enhance the applicability and reliability of Foundation Models in clinical practice by addressing data incompleteness.

Main Methods:

• Proposed DiffM4RI, a diffusion model designed for many-to-many missing modality imputation in MRI.
• Formulated missing modality imputation as a modality-level inpainting task, allowing for flexible handling of various missing data patterns.
• Evaluated performance on the BraTs datasets.

Main Results:

• DiffM4RI achieved significant improvements in imputation accuracy compared to existing methods.
• Demonstrated an average Structural Similarity Index Measure (SSIM) improvement of 0.15 over MustGAN, 0.1 over SynDiff, and 0.02 over VQ-VAE-2.
• The model's ability to handle diverse missing scenarios without retraining was validated.

Conclusions:

• DiffM4RI offers an effective solution for many-to-many missing modality imputation in MRI.
• The proposed approach enhances the robustness and clinical applicability of Foundation Models by addressing data gaps.
• This work paves the way for more reliable AI-driven medical image analysis in real-world scenarios.