JUI-YO HSULian, Pin-HsunPin-HsunLianChuang, Tzu-YiTzu-YiChuangChang, Gary HanGary HanChang2025-07-012025-07-012025-04-16https://scholars.lib.ntu.edu.tw/handle/123456789/730387Background and Objective: Magnetic resonance imaging (MRI) in musculoskeletal applications often suffers from anisotropic resolution that complicates the detection and analysis of subtle pathologies. This study aims to address these limitations by developing a novel self-supervised super-resolution and diffusion inpainting framework to generate isotropic, high-fidelity 3D reconstructions without requiring paired high-resolution ground truth data.Methods: We integrated a generative adversarial network (GAN) for anisotropy-to-isotropy conversion with a probabilistic diffusion inpainting model to refine and remove pathological artifacts. Our pipeline employed PatchNCE contrastive loss to preserve structural integrity, multi-view adversarial training to enhance through-plane resolution, and denoising diffusion steps guided by anatomic priors for robust volumetric anomaly delineation.Results: Comparisons against existing unsupervised methods demonstrated superior image quality, validated by quantitative metrics such as FID, KID, and LPIPS. Expert reviews confirmed enhanced visualization and clearer boundaries of bone marrow lesions and effusion-synovitis, underscoring the framework’s ability to handle even high anisotropy ratios with minimal artifacts.Conclusion: Our self-supervised GAN-diffusion approach effectively produces high-resolution, isotropic MRI volumes and improves 3D pathological detection without relying on ground-truth isotropic data. This innovation has the potential to broaden clinical and research applications, particularly for retrospective analyses and longitudinal musculoskeletal imaging studies. © 2025, The Authors. All rights reserved.enAbnormality DetectionMusculoskeletal MRIself-supervised learningsuper-resolutionother10.2139/ssrn.52130392-s2.0-105004246469