Deep Learning Reconstruction Enables Accelerated Prostate MRI Without Compromising Reader-Assessed Visual Quality

Motivation: MRI demand for prostate cancer detection is rising, necessitating reduced scan times. While deep learning (DL) shows promise for accelerating MRI, the perceived visual quality of DL-reconstructed images by expert radiologists remains underexplored. Goal(s): To assess expert radiologists' perceived visual quality of DL-reconstructed prostate MRI scans with various acceleration factors (3x, 6x). Approach: A retrospective study with 120 prostate MRIs was conducted. Expert radiologists evaluated image quality using a split-plot design and ordinal mixed-effects modeling. Results: Up to 6x acceleration did not significantly reduce perceived image quality. Interestingly, 3x accelerated images even improved perceived quality compared to non-accelerated images. Impact: DL-based reconstruction can generate high-quality T2-weighted prostate MRIs from accelerated acquisitions without any perceived loss in overall visual quality by expert radiologists compared with the original non-accelerated images.