Related Experiment Videos
Seeing the Risk in Virtual Fat-Suppressed Spine MRI: External Benchmarking and Uncertainty-Guided Selective Risk
Zenghan Zhou1,2,3, Yingying Song2, Rong Liu4,5,6,7
1School of Medicine, Wuhan University of Science and Technology, No. 2, Huangjiahu West Road, Hongshan District, Wuhan, 430065, Hubei, China.
Abstract:
Virtual fat-suppressed spine MRI may recover clinically useful contrast from routine T1-weighted and T2-weighted images, but visually plausible synthesis can conceal local failure. We evaluated a reliability-centered framework that reframed this task from replacement imaging to uncertainty-guided selective review. This retrospective study included 5119 cases or examinations from two in-house and four external or public datasets. A multi-source encoder was transferred to a 2.5-dimensional dual-output generator for virtual STIR and water-dominant fat-suppressed synthesis, followed by frozen-generator post hoc Laplace uncertainty and exploratory Effective-EPG analyses. On locked test sets, body-mask SSIM was 0.602 for internal lumbar STIR, 0.567 for internal cervical STIR, 0.496 for IDEAL-derived water, and 0.376 in C43, a patient- and examination-independent held-out cohort drawn from the same public source as Dataset C. Restormer-Small achieved the strongest internal body-based fidelity, whereas the proposed model achieved the highest body-mask PSNR and SSIM and the lowest MAE in the held-out C43 stress test cohort. Protocol conditioning showed very small and inconsistent differences. Post hoc uncertainty provided the strongest reliability signal: the highest uncertainty decile contained 1.79-2.11 times the slice-average error, and retaining the lowest uncertainty 50% of pixels reduced normalized MAE to 0.537-0.565. Uncertainty retained incremental error-ranking information beyond input-derived proxies, but its raw scale did not transfer quantitatively to C43; Effective-EPG associations attenuated after adjustment for input image covariates. In a four-reader study, virtual image balanced accuracy was 0.813-0.975, and 79.1% of auxiliary acceptability ratings were at least 4. Virtual fat-suppressed MRI is best positioned as uncertainty-guided auxiliary information for selective review, not replacement imaging.