DeepFLAIR*: Improving Multiple Sclerosis Diagnostic Imaging Workflow Using Deep Learning

Background: Magnetic resonance imaging (MRI) plays a central role in diagnosing multiple sclerosis (MS), yet conventional T2-FLAIR imaging provides limited specificity for distinguishing MS lesions from other white matter abnormalities. The Central Vein Sign (CVS) is a sensitive and specific imaging biomarker which was recently included in the 2024 McDonald criteria for MS diagnosis. FLAIR*, which combines T2-FLAIR and T2* 3D EPI acquisitions, provides optimal detection of the CVS; however, this post-processing workflow requires two separate scans which increases scan time, susceptibility to motion artifacts, and registration error, thus limiting clinical deployment. This study aims to address this issue using a novel deep learning methodology called DeepFLAIR*. Methods: Retrospective analysis was performed on multicenter 3-Tesla brain MRI data as part of the Central Vein Sign in Multiple Sclerosis (CAVS-MS) study. The dataset included 315 participants scanned on Siemens and Philips 3T systems using standardized protocols incorporating 3D T2-FLAIR and 3D T2*-weighted EPI acquisitions (0.65-mm isotropic resolution; scan times ≈ 6-7 minutes per sequence). A 3D U-Net-based conditional generative model, DeepFLAIR*, was developed to synthesize FLAIR* contrast directly from single-sequence T2* 3D EPI images. The model was trained and validated using 89 subjects and tested on an independent cohort of 226 subjects. Quantitative evaluation included structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), mean squared error (MSE), and contrast-to-noise ratio (CNR) across lesion-vein, lesion-white matter, vein-white matter, and white matter-cerebrospinal fluid regions. Statistical comparisons between real-world and synthetic FLAIR* images were performed using paired Wilcoxon signed-rank tests with false discovery rate correction (α = 0.05). Results: Quantitative metrics confirmed that DeepFLAIR* achieved significantly improved contrast-to-noise ratios and comparable global similarity measures relative to real-world FLAIR* (P < 0.001). Synthetic FLAIR* images demonstrated high structural fidelity to real-world FLAIR* (SSIM = 0.78 ± 0.03, PSNR = 23.6 ± 1.35 dB, MSE = 0.0045 ± 0.0015). CNR analyses revealed enhanced lesion-vein and vein-white matter contrast, confirming preservation of perivenular morphology relevant to CVS detection. Lesion morphology and vein-lesion spatial relationships were consistently preserved across subjects. Conclusions: This study demonstrates feasibility of our novel DeepFLAIR* methodology for generating diagnostically relevant FLAIR* contrast from a single T2* 3D EPI input, thereby eliminating the need for dual acquisitions and offline post-processing. This approach could streamline MRI workflows, expand clinical access to CVS based MS evaluation, and facilitate automated biomarker detection in future diagnostic pipelines.