Self-supervised isotropic reconstruction for abnormality detection in anisotropic MRI.
Authors
Affiliations (4)
Affiliations (4)
- Program for Precision Health and Intelligent Medicine, Graduate School of Advanced Technology, National Taiwan University, Taipei, Taiwan; Department of Orthopedic Surgery, National Taiwan University Hospital, Taipei, Taiwan; Department of Orthopedic Surgery, College of Medicine, National Taiwan University, Taipei, Taiwan.
- Institute of Medical Device and Imaging, College of Medicine, National Taiwan University, Taipei, Taiwan; Department of Counselling & Clinical Psychology, National Dong Hwa University, Hualien, Taiwan.
- Institute of Medical Device and Imaging, College of Medicine, National Taiwan University, Taipei, Taiwan.
- Program for Precision Health and Intelligent Medicine, Graduate School of Advanced Technology, National Taiwan University, Taipei, Taiwan; Institute of Medical Device and Imaging, College of Medicine, National Taiwan University, Taipei, Taiwan. Electronic address: [email protected].
Abstract
Accelerating musculoskeletal magnetic resonance imaging (MRI) while preserving diagnostic detail remains challenging because acquiring fully‑isotropic ground‑truth volumes is clinically costly. In routine practice, anisotropic scans with reduced through-plane resolution degrade multiplanar visualization and slice-by-slice review in reformatted planes, obscure subtle abnormalities spanning only a few slices, and limit automated three-dimensional (3D) analyses that assume comparable spatial resolution across axes. We present a two‑stage, fully self‑supervised pipeline that learns directly from anisotropic scans-obviating any paired high‑resolution data-and converts highly anisotropic (8:1) turbo‑spin‑echo volumes into isotropic images and 3D abnormality maps. Unlike prior self-supervised super-resolution methods, Stage 1 uses a single forward multi-view generative adversarial network (GAN) with patch-based contrastive and adversarial objectives rather than a backward/cycle-consistency approach. Stage 2 leverages an anatomy-conditioned denoising-diffusion model for healthy counterfactual generation, yielding voxel-wise lesion maps without external annotations. On 2225 Osteoarthritis Initiative knee scans from five different imaging centres, the framework reduced Fréchet inception distance from 407.4 → 254.4 (coronal) and 429.9 → 266.9 (axial), achieved the best Kernel Inception Distance (KID) / Learned Perceptual Image Patch Similarity (LPIPS) scores among competing unsupervised methods, and was preferred in 65-67% of blinded orthopedist comparisons. Crucially, isotropic enhancement propagated to downstream tasks: femur-tibia segmentation F1 scores increased and previously confluent bone‑marrow lesions were separated into discrete entities, enabling precise volumetric quantification. Robustness experiments demonstrated consistent gains across five imaging centers, synthetic noise/contrast perturbations, and transfer of the resolution-enhancement module to two additional MRI protocols, supporting robustness across sites and acquisition protocols. By eliminating the need for ground‑truth isotropic images while surpassing state‑of‑the‑art unsupervised super‑resolution in both perceptual quality and clinical utility, our method may facilitate retrospective cohort studies and prospective scan-time reduction in heterogeneous knee MRI settings, with preliminary transferability to additional protocols.