Magnetic resonance in medicine · 2026
FlowVN Trained on a Single Dataset Enables Rapid Reconstruction of Highly Accelerated 4D Flow MRI Across Multiple Sites.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Qazi Sohaib Ayaz, Bianchessi Tamara, Viola Federica, Trenti Chiara, Ylipää Erik, Ebbers Tino, Dyverfeldt Petter
Аннотация
PURPOSE: The aim of this study is to evaluate a deep variational network, FlowVN, for the reconstruction of heavily undersampled 4D Flow MRI across multiple sites.
METHODS: FlowVN was trained on fully sampled 4D Flow MRI datasets of healthy volunteers from one site. The model was tested on retrospective undersampled data (R = 6-22) of six normal volunteers from the same site and six from another site, prospectively undersampled data (R = 12.4-13.8) of six healthy volunteers from the second site and six patients with aortic stenosis from first Site. Performance was evaluated using nRMSE, relative and angular error, average and maximum velocity, flow rate, volumetric flow, and turbulent kinetic energy (TKE), with a Wilcoxon signed-rank test to assess the difference from ground truth.
RESULTS: FlowVN showed minimal sensitivity to the number of training datasets and performed well even when trained on a single dataset. FlowVN also demonstrated good generalizability across sites. No significant difference in average and maximum velocity was observed up to R = 16. Total TKE was preserved up to R = 10 in the normal volunteers, but was well preserved even at higher acceleration factors in the aortic stenosis patients. Flow volumes through the ascending and descending aorta were well preserved for all acceleration factors, although ascending aorta flow volumes for data from one site were significantly different from ground truth for AF > 16.
CONCLUSION: FlowVN accurately reconstructs highly undersampled 4D Flow MRI from multiple sites using a model trained on a single dataset, maintaining excellent quantitative image quality even at very high acceleration factors.
METHODS: FlowVN was trained on fully sampled 4D Flow MRI datasets of healthy volunteers from one site. The model was tested on retrospective undersampled data (R = 6-22) of six normal volunteers from the same site and six from another site, prospectively undersampled data (R = 12.4-13.8) of six healthy volunteers from the second site and six patients with aortic stenosis from first Site. Performance was evaluated using nRMSE, relative and angular error, average and maximum velocity, flow rate, volumetric flow, and turbulent kinetic energy (TKE), with a Wilcoxon signed-rank test to assess the difference from ground truth.
RESULTS: FlowVN showed minimal sensitivity to the number of training datasets and performed well even when trained on a single dataset. FlowVN also demonstrated good generalizability across sites. No significant difference in average and maximum velocity was observed up to R = 16. Total TKE was preserved up to R = 10 in the normal volunteers, but was well preserved even at higher acceleration factors in the aortic stenosis patients. Flow volumes through the ascending and descending aorta were well preserved for all acceleration factors, although ascending aorta flow volumes for data from one site were significantly different from ground truth for AF > 16.
CONCLUSION: FlowVN accurately reconstructs highly undersampled 4D Flow MRI from multiple sites using a model trained on a single dataset, maintaining excellent quantitative image quality even at very high acceleration factors.