AI-Based Reconstruction for Fastmri

Magnetic Resonance Imaging (MRI) is the most used non-invasive modality across medical specialties. Despite its usefulness in the clinic, the slow process of acquiring k-space data significantly inhibits patient throughput and adds restrictions to which patients can be seen and which pathologies are examined. In this work, we pull together a comprehensive AI-based reconstruction framework for the fastMRI challenge into a reproducible implementation. Our approach brings together CS theories with deep CNNs to reconstruct diagnostic-quality MR images from heavily under-sampled k-space measurements. Using several different approaches, we explore” deep” P&P priors, agnostic supervised regression, and adversarial reconstruction in conjunction with baselines CS methods such as zero-filled IFFT, CS with total variation (CS-TV), SENSE, and GRAPPA reconstruction techniques. Our architecture builds off a U-Net backbone, augmented by iterative data consistency (DC) layers, and is trained end-to-end on the publicly available fastMRI dataset of knee MRIs (8,400 scans). At an acceleration factor of 4× our model achieves SSIM of 0.921, PSNR of 38.2 dB, and NMSE of 0.013 outperforming all the baselines we compared. We study ablations confirming the critical nature of DC layers, cross-domain processing, and loss function. Qualitatively we observe better recovery of fine anatomical details such as cartilage boundaries and ligament detail than classical reconstruction methods. The clinical implications of these results, risks of instability, and further deployment outlook are discussed.