ROI-SSIM Guided Framework for 3D Knee MRI Reconstruction under Artifact Degradation

Magnetic resonance imaging (MRI) reconstruction is highly sensitive to undersampling and acquisition artifacts. Although deep learning-based approaches have significantly improved reconstruction quality, they are typically optimized using global fidelity metrics that may not fully reflect reconstruction quality within clinically important structures such as the anterior cruciate ligament (ACL), meniscus, and articular cartilage. This study proposes a region-aware MRI reconstruction framework based on a Vision Transformer autoencoder (ViT-AE) guided by an ROI Structural Similarity (ROI-SSIM) loss. The method leverages fastMRI+ bounding-box annotations and incorporates a hybrid artifact simulation framework spanning k-space and image-domain degradations. Experiments on the fastMRI knee single-coil validation set (PD contrast, R=4) show that the proposed ViT-AE achieves reconstruction fidelity competitive with established baselines (SSIM 0.823, Peak Signal-to-Noise Ratio (PSNR) 32.4 dB , Normalized Mean Squared Error (NMSE) 0.0147), while ROI-SSIM supervision improves structural fidelity within clinically relevant annotated regions. Under mixed artifact conditions, region-specific structural similarity decreased by only 2.9 % , compared with a 6.8 % decrease in global SSIM. These findings support the use of region-aware supervision for artifact-robust MRI reconstruction. Overall, the method presented in this paper: Introduces ROI-SSIM supervision for region-aware MRI reconstruction. Develops a hybrid artifact simulation framework spanning k-space and image-domain degradations. Demonstrates improved anatomical preservation under artifact conditions.