A Self-Consistent Diffusion Schrodinger Bridge for Multi-Modal Medical Image Translation

Motivation: In medical image translation, denoising diffusion models (DDM) learn a task-irrelevant denoising transformation that maps Gaussian-noise onto a target-modality image, while receiving a source-modality image as a static-input channel. This causes suboptimal source-modality guidance due to a compromise between denoising and source-to-target transformations. Goal(s): Our goal was to devise a new diffusion-based method that learns a task-relevant source-to-target transformation to improve translation fidelity. Approach: We introduced a novel self-consistent recursive diffusion bridge (SelfRDB) that performs a gradual mapping from source onto target images. Results: Higher performance was obtained with SelfRDB over previous state-of-the-art in multi-contrast MRI and MRI-CT translation. Impact: The enhanced image fidelity in multi-modal protocols achieved by SelfRDB can extend the scope of imaging-based assessments, while maintaining relatively low scan budgets and minimizing exposure to invasive agents or radiation, particularly benefiting at-risk pediatric and elderly populations.