Year: 2026
Type: dissertation
Abstract: Denoising diffusion models (DDMs) have recently become prominent in medical image translation owing to their favorable training stability and image fidelity.In conventional DDM-based translation, generation begins from random Gaussian noise and proceeds through a sequence of denoising steps towards the target modality, while the source image is provided separately as static guidance.The resulting generative process therefore primarily models a noise-to-target transformation rather than the source-to-target transformation required by the translation task, since source images follow a non-noise distribution.This mismatch can lead to suboptimal use of source-modality information and limit translation performance.This thesis presents SelfRDB, a self-consistent recursive diffusion bridge designed to more directly couple the diffusion process with the source-totarget mapping.Its forward process connects a target image at the start-point to an end-point defined from the corresponding source image.Intermediate samples follow a normal distribution whose mean is a convex combination of the two end-points and whose variance is determined through additive noise.Whereas regular diffusion bridges use zero noise variance at both end-points and high variance near the mid-point, SelfRDB employs a variance schedule that increases monotonically towards the source-modality end-point.This formulation facilitates information transfer between the modalities while improving robustness to measurement noise.Sampling accuracy during reverse diffusion is further enhanced through self-consistent recursive estimation, in which the network repeatedly refines a transient target-image estimate until convergence to a self-consistent solution.Comprehensive experiments on multi-contrast MRI and MRI-CT translation demonstrate that SelfRDB achieves state-of-the-art image quality.
Author: Arslan, Fuat
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7205481243