Mutual Information Guided Diffusion for Zero-Shot Cross-Modality Medical Image Translation.

Area of Science:

• Medical image computing
• Deep generative models
• Artificial intelligence in healthcare

Background:

• Cross-modality image translation is crucial in medical imaging but faces challenges, particularly in zero-shot learning scenarios.
• Existing deep generative models struggle with fidelity and adaptability in unseen domains.
• The need for methods that can translate images between modalities without direct source-target domain mapping is significant.

Purpose of the Study:

• To address the challenge of unsupervised zero-shot cross-modality image translation with high fidelity.
• To propose a novel method that leverages statistical consistency between modalities for image translation.
• To enable adaptation to new source domains without requiring retraining.

Main Methods:

• Proposed a novel unsupervised zero-shot learning method named Mutual Information guided Diffusion Model (MIDiffusion).
• Introduced a differentiable local-wise mutual information layer to condition the diffusion model's denoising process.
• Utilized local-wise mutual information to capture cross-modality features statistically, guiding the diffusion process without direct mappings.

Main Results:

• Demonstrated superior performance of MIDiffusion in zero-shot cross-modality translation tasks compared to adversarial and diffusion-based models.
• Showcased the method's ability to adapt to changing source domains without retraining.
• Validated the practical application of MIDiffusion in 3D zero-shot learning-based cross-modality image segmentation.

Conclusions:

• MIDiffusion effectively overcomes the limitations of current methods in unsupervised zero-shot cross-modality image translation.
• The proposed local-wise mutual information layer enables robust and adaptable image translation.
• The method holds significant potential for real-world applications in medical image analysis and segmentation.