Generating Synthetic MR Perfusion Maps From DWI and FLAIR in Acute Ischemic Stroke: Development and External Validation of a Deep Learning Model.
Authors
Affiliations (8)
Affiliations (8)
- Groupe d'Imagerie Neurofonctionnelle, Institut des Maladies Neurodégénératives- UMR 5293, CNRS, CEA, University of Bordeaux, France (A.M., M.T.S.).
- Brain Connectivity and Behaviour Laboratory, Bordeaux University, France (A.M., M.T.S., T.T.).
- CLAIM-Charité Lab for AI in Medicine, Charité-Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Germany (A.K., A.H., O.U.A., D.F.).
- Centre Hospitalier Universitaire (CHU) de Bordeaux, Neuroimagerie Diagnostique et Thérapeutique, France (C.M.-V., G.M., T.T.).
- Department of Neuroradiology, University Hospital Heidelberg, Germany (M. Bendszus).
- Centre Hospitalier Universitaire (CHU) de Bordeaux, Neurologie Vasculaire, France (M. Boullet, S.O., I.S.).
- Division of Neuroradiology, Clinic of Radiology and Nuclear Medicine, University Hospital Basel (M.A.M.).
- Université de Bordeaux, Neurocentre Magendie, INSERM U1215, France (T.T.).
Abstract
Magnetic resonance imaging (MRI) is critical for acute stroke triage, but it is time-consuming and often requires contrast injection for perfusion imaging. This study aimed to synthesize T-map perfusion maps from routinely available, noncontrast diffusion-weighted imaging and fluid-attenuated inversion recovery using deep generative models. We hypothesized that relevant perfusion information could be inferred from these modalities to streamline imaging and reduce reliance on dynamic susceptibility contrast perfusion. Acute magnetic resonance imaging data from 355 patients with anterior circulation stroke, including dynamic susceptibility contrast perfusion, were retrospectively collected from 2 European centers (Heidelberg: 2010-2018; Bordeaux: 2021-2022). Six versions of a denoising diffusion probabilistic model and a generative adversarial network architecture were trained to generate synthetic time-to-maximum (T-max) perfusion maps from diffusion-weighted imaging, fluid-attenuated inversion recovery, and infarct core mask as inputs. Performance was assessed by comparing synthetic and ground-truth T-max maps using image similarity metrics. Regions with T-max >6 s were compared using Dice coefficients, and mismatch volume distributions were analyzed. An ablation study quantified the contribution of each input. The best performance was achieved by a denoising diffusion probabilistic model with a 2.5D architecture using diffusion-weighted imaging, fluid-attenuated inversion recovery, infarct core mask, and a perfusion-weighted loss function. It produced synthetic perfusion T-max maps with high similarity to ground truth under 110 s. The model showed strong spatial overlap for T-max> 6 s regions in internal validation (average Dice, 0.82; SD, 0.08) and external validation (average Dice, 0.59; SD, 0.13), respectively. Synthetic maps closely matched ground truth mismatch distributions, capturing key perfusion patterns. The infarct core mask played a critical role in model performance, alongside diffusion-weighted imaging and fluid-attenuated inversion recovery inputs. We propose a noninvasive, scalable framework to generate synthetic T-max perfusion maps from noncontrast magnetic resonance imaging. This approach could expand access to perfusion data in acute stroke, shorten imaging protocols, and accelerate treatment decisions by eliminating the need for contrast-enhanced acquisition.