Pooled transfer learning improves brain vessel MRI image resolution
Toward Clinical Translation of Deep Learning-Based Intracranial Super-Resolution 4D Flow MRI Using Multi-Cohort Transfer Learning
Computer Vision and Pattern Recognition
Summary
High-resolution MRI scans of brain blood vessels can help doctors see details better but are hard to get. The authors tested a computer program that boosts lower-quality MRI images to high resolution using smart learning from different patient groups. They found training the program on combined data from multiple groups made it more accurate than training on single groups or synthetic data alone. This improved detail helps show blood vessels more clearly, which could assist doctors in diagnosing brain diseases. The results suggest this approach should be tested more widely before clinical use.
What this means in practice
- •For medical imaging teams: Improve the quality of intracranial 4D Flow MRI images for more accurate vessel visualization by using multi-cohort transfer learning in deep learning models.
- •For clinical radiology departments: Enhance non-invasive angiographic imaging precision and visualization in neurologic patients using super-resolution techniques trained on diverse patient data.
Authors
Oliver Welin Odeback, Susanne Schnell, Patrick Winter, Johan Lundberg, Fredrik Piehl, Tobias Granberg, Alexander Fyrdahl, David Marlevi
Abstract
We evaluated synthetic pretraining and multi-cohort transfer learning for deep learning-based super-resolution of intracranial 4D Flow MRI across heterogeneous neurologic cohorts. This mixed prospective-retrospective study analyzed 18 participants (12 healthy participants, 3 with moyamoya disease, and 3 with multiple sclerosis) with paired 1.2-mm and 0.6-mm 4D Flow MRI scans. A convolutional neural network pretrained on synthetic cerebrovascular data was fine-tuned using cohort-specific or pooled multi-cohort in vivo data. High-resolution-derived inputs enabled voxel-wise evaluation; native standard-resolution scans were used for angiographic and hemodynamic analyses. Synthetic pretraining reduced mean artery-wise velocity-normalized root mean square error (vNRMSE) versus interpolation (0.147 vs 0.166). Pooled transfer learning yielded the lowest mean vNRMSE (0.119 vs 0.127 for cohort-specific models and 0.147 for synthetic pretraining), with narrower 95% limits-of-agreement widths than synthetic pretraining across all test cases (0.109-0.222 vs 0.154-0.278 m/s). Super-resolved phase-contrast angiography achieved higher precision-recall area under the curve than native standard-resolution imaging in all cohorts (0.848-0.896 vs 0.486-0.805). These results indicate that pooled transfer learning improves cross-cohort reconstruction agreement and that super-resolution can improve vessel depiction, supporting further multi-center validation for cerebrovascular imaging.