Toward safe deployment of deep learning in MRI: A physics-driven uncertainty framework for automated quality control and risk Tiering in virtual fat suppression.
Authors
Affiliations (4)
Affiliations (4)
- Wuhan Puren Hospital (Affiliated to Wuhan University of Science and Technology), 1 Benxi Street, Jianshe Fourth Road, Qingshan District, 430080 Wuhan, Hubei, China; Institute of Medical Innovation and Transformation, Puren Hospital Affiliated to Wuhan University of Science and Technology, 1 Benxi Street, Jianshe Fourth Road, Qingshan District, 430080 Wuhan, Hubei, China; Wuhan Liu Sanwu Traditional Chinese Medicine Orthopedic Hospital, No. 388, Xingkeli, Wenchang Avenue, Zhucheng Street, Xinzhou District, 431400 Wuhan, Hubei, China. Electronic address: [email protected].
- Wuhan Puren Hospital (Affiliated to Wuhan University of Science and Technology), 1 Benxi Street, Jianshe Fourth Road, Qingshan District, 430080 Wuhan, Hubei, China; Institute of Medical Innovation and Transformation, Puren Hospital Affiliated to Wuhan University of Science and Technology, 1 Benxi Street, Jianshe Fourth Road, Qingshan District, 430080 Wuhan, Hubei, China.
- Wuhan Puren Hospital (Affiliated to Wuhan University of Science and Technology), 1 Benxi Street, Jianshe Fourth Road, Qingshan District, 430080 Wuhan, Hubei, China.
- Institute of Medical Innovation and Transformation, Puren Hospital Affiliated to Wuhan University of Science and Technology, 1 Benxi Street, Jianshe Fourth Road, Qingshan District, 430080 Wuhan, Hubei, China; Wuhan Liu Sanwu Traditional Chinese Medicine Orthopedic Hospital, No. 388, Xingkeli, Wenchang Avenue, Zhucheng Street, Xinzhou District, 431400 Wuhan, Hubei, China.
Abstract
Deep generative models in MRI are hindered by "hallucinations" and a lack of safety mechanisms. This study introduces a physics-driven framework for trustworthy Virtual Fat Suppression (VFS), enabling automated quality control and active risk tiering. A differentiable Bloch layer was embedded into an RRDB-RaGAN architecture to enforce physics-based signal consistency during synthesis. An uncertainty-guided training strategy with artifact-enriched supervision was further introduced to support a three-tier risk model for auto-pass, human review, and rejection. The framework was evaluated using quantitative fidelity metrics, real-world severe artifact detection, and blinded clinical reader assessment. In real-world severe artifact detection, the image-level uncertainty score achieved an AUC of 0.9053 and a recall of 81.82% at a false-positive rate of 4.69%. In blinded clinical evaluation, agreement between AI-predicted and reader-mapped risk tiers was substantial (quadratic weighted Cohen's κ = 0.7938), and images assigned to the Low-Risk tier achieved a mean Likert score of 4.70 ± 0.61. By coupling physics-constrained synthesis with uncertainty-based risk governance, the proposed framework provides a practical and auditable safety layer for virtual fat suppression in clinical MRI workflows.