Organizations: Nepal Applied Mathematics and Informatics Institute for Research (NAAMII), Lalitpur, Nepal · University of Exeter, Exeter, Devon, UK · Institute of Engineering, Purwanchal Campus, Dharan, Nepal · Institute of Engineering, Pulchowk Campus, Lalitpur, Nepal · Gandaki Medical College Teaching Hospital and Research Center, Pokhara, Nepal · Madan Bhandari University of Science and Technology, Chitlang, Nepal · Medical Artificial Intelligence Laboratory, Lagos, Nigeria · Department of Electrical and Computer Engineering, University of British Columbia, Vancouver, Canada · Montreal Neurological Institute, McGill University, Montreal, Canada · Department of Biomedical Engineering, McGill University, Montreal, Canada · Department of Medical Physics, College of Medicine, University of Lagos, Nigeria · Brain Health Imaging Institute, Department of Radiology, Weill Cornell Medicine, Weill Cornell Medicine, New York, United States
Diffusion-weighted imaging (DWI) is widely used in clinical settings but remains vulnerable to geometric distortion. Conventional correction methods often require additional acquisitions or vendor-specific solutions, limiting their feasibility in high-throughput, resource-constrained settings. This study investigates whether large-scale pretraining strategies can improve deep learning-based distortion correction for single-phase-encoding DWI. We formulate the task as image reconstruction, and compare a non-pretrained baseline against a self-supervised and a generative pretrained model, evaluated using both quantitative image-similarity metrics and qualitative expert assessment. The best-performing model was further tested for transferability on data collected in an LMIC setting with acquisition shift. Pretrained models outperformed the non-pretrained baseline, with cWDM achieving the strongest results across both quantitative and qualitative evaluation. However, application to LMIC data revealed transferability challenges, including contrast alteration and over-reliance on T1-weighted anatomical structure. Registering images to a common standard space improved predictions, suggesting that harmonized preprocessing may enhance cross-domain deployment.
Figures & tables
Figure 1: Block diagram of the proposed methodology. A) Large-scale MRI pretrained models were obtained from publicly available repositories. B) The pretrained models were fine-tuned for DWI distortion correction. C) Pretraining effectiveness was assessed through quantitative and qualitative evaluation. D) The best-performing model was tested on domain-shifted LMIC clinical data. E) Model transferability was evaluated through qualitative expert assessment.
Figure 2: Box plots representation of quantitative metrics across three models on HCP test set.
Figure 3: Radiographer ratings and podium plots [ 22 ] for the three models (cWDM, SwinBRAIN, Synb0-DISCO) across 159 test cases. (a) Distributions of ratings for Distortion Correction and Image Data Integrity. (b) Podium plots: top panel shows per-case metric values ranked 1st–3rd (best–worst), with lines connecting values from the same test case; bottom panel shows, for each algorithm, the relative frequency of attaining each rank.
Figure 4: Scatter plots showing the relationship between qualitative evaluation scores (Distortion and Integrity) and quantitative image quality metrics (MSE, PSNR, SSIM, and VIF) for the cWDM model. Spearman correlation coefficients (r) are reported for each metric pair.
Figure 5: Qualitative comparison of predicted DWI before and after registration training. For three subjects (A–C), original DWI, model prediction, and T1w reference are shown pre- and post-registration. Red circles mark information loss (A: frontal, C: occipital) that resolves after registration, and reduced over-reliance on T1w structure, particularly temporal lobe (B). White circle (C) shows improved distortion correction accuracy post-registration. Contrast alteration persists across all images.
UCL Hawkes Institute and Department of Medical Physics and Biomedical Engineering, University College London, UK · Division of Surgery and Interventional Science, University College London, UK · Centre for Medical Imaging, University College London, UK +5
Hawkes Institute & Department of Computer Science, University College London, London, UK · Institut Pasteur, Universit´e Paris Cit´e, Unit´e de Neuroanatomie Appliqu´ee et Th´eorique, France · AINOSTICS ltd., Manchester, UK +3
Center for Image Sciences, University Medical Center Utrecht, Utrecht, The Netherlands · Cerebriu, Copenhagen, Denmark · Department of Radiotherapy, University Medical Center Utrecht, Utrecht, The Netherlands +2