Stability and generalizability in learning-based medical image reconstruction
File(s)
Author(s)
Zhang, Weitong
Type
Thesis or dissertation
Abstract
Deep learning has improved medical image reconstruction, but clinical use is limited by poor reliability under domain shift (new scanners, protocols, motion, noise) and scarce expert labels. This thesis builds reconstruction methods that generalize beyond training data, need less supervision, and can indicate when outputs are unreliable.
It contributes: (1) unsupervised/self-supervised MRI reconstruction that models motion and acquisition degradation without paired ground truth; (2) test-time adaptation without source data or labels, using physics constraints and self-consistency to recalibrate to new domains; (3) a stable diffusion-based reconstruction framework based on measure-consistent stochastic dynamics, improving robustness and reducing sampling error while clarifying failure modes. Uncertainty and mismatch detection are integrated throughout to support safer deployment.
It contributes: (1) unsupervised/self-supervised MRI reconstruction that models motion and acquisition degradation without paired ground truth; (2) test-time adaptation without source data or labels, using physics constraints and self-consistency to recalibrate to new domains; (3) a stable diffusion-based reconstruction framework based on measure-consistent stochastic dynamics, improving robustness and reducing sampling error while clarifying failure modes. Uncertainty and mismatch detection are integrated throughout to support safer deployment.
Date Issued
2025-11-07
Date Awarded
2026-02-01
Copyright Statement
Attribution 4.0 International Licence (CC BY)
License URL
Advisor
Kainz, Bernhard
Karampinos, Dimitrios
Sponsor
UK Research and Innovation
Grant Number
W.Z. is supported by the JADS programme and the UKRI Centre for Doctoral Training in AI for Healthcare (EP/S023283/1).
High-performance computing resources were provided by the Erlangen National High Performance Computing Center (NHR@FAU) at Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), under the NHR projects b143dc and b180dc.
Isambard-AI is operated by the University of Bristol and is funded by the UK Government’s DSIT via UKRI; and the Science and Technology Facilities Council [ST/AIRR/I-A-I/1023].
NHR is funded by federal and Bavarian state authorities, and NHR@FAU hardware is partially funded by the German Research Foundation (DFG) – 440719683.
ERC - project MIA-NORMAL 101083647, DFG 512819079, 513220538, and by the state of Bavaria (HTA).
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)