Deep learning reconstruction for prostate MRI: Impact of field-of-view selection on global and regional image quality

Purpose: To compare global and region-specific image quality in deep learning-based prostate MRI reconstructions and determine whether global assessments mask localized degradations in diagnostically relevant areas. Methods: This retrospective study included 3.0 T T2-weighted prostate k-space data from the University Medical Center Groningen, with 63 lesions (PI-RADS ≥ 3) segmented by a radiologist. Scans were accelerated three-fold (R=3) and six-fold (R=6) using a state-of-the-art deep learning (DL) reconstruction model (vSHARP). Image quality was evaluated using the Structural Similarity Index Measure (SSIM) across the entire clinical FOV and in region-specific (diagnostic) FOVs, defined as normal prostate tissue and prostate lesion areas, as well as in non-diagnostic reference regions (muscle, femur, and subcutaneous fat). Nonparametric tests and linear mixed-effects models assessed differences in visual quality between FOVs and characterized degradation patterns with increased acceleration. Results: The clinical FOV maintained stable image quality, with SSIM values of 0.88 ± 0.06 at R= 3 and 0.86 ± 0.06 at R= 6. In contrast, region-specific (diagnostic) FOVs exhibited significant degradation, with prostate lesions SSIM of 0.67 ± 0.10 at R= 3 and 0.61 ± 0.10 at R= 6 (p < 0.001). Regression analysis confirmed that quality loss was significantly greater in the region-specific compared to the clinical FOV (p < 0.001). Conclusion: The current evaluation standard—SSIM over the entire clinical FOV—fails to capture region-specific degradations in diagnostic areas. Our findings reveal that assessments on diagnostic regions expose significantly greater quality loss with increased acceleration, highlighting the need to integrate region-specific quality metrics into DL reconstruction evaluation frameworks. • AI reconstruction shows faster quality loss in diagnostically relevant regions. • Prostate and lesion areas exhibit lower image quality than uniform tissues. • Global image quality metrics mask regional degradation in accelerated prostate MRI. • Segmentation standardizes region extraction, enabling objective quality comparisons.