This multicenter, retrospective, consecutive-cohort study included 120 multi-coil T2-weighted prostate MRI scans from the University Medical Center Groningen (UMCG) and 312 publicly available scans from New York University (NYU). An AI model trained on the NYU data was tested on retrospectively undersampled UMCG scans at acceleration factors R = 3 and R = 6 (i.e., data reduction in k-space). Eight experienced radiologists participated in a multi-reader multi-case PCa detection study. Diagnostic performance was assessed using the area under the receiver operating characteristic curve (AUROC). Histopathology and PI-RADS ≤ 2 findings served as reference standards. Multiple image quality metrics were subjectively evaluated using a 4-point Likert scale.
AI-driven reconstruction enabled a sixfold acceleration of T2-weighted prostate MRI (0:33-1:27 min) without a statistically significant reduction in PCa detection, while preserving perceived image quality. However, the decreasing diagnostic performance at higher accelerations warrants further prospective evaluation.
No statistically significant reduction in PCa detection was observed at an MRI acceleration up to R = 6 (p = 0.08). AUROC values were 0.86 (95% CI: 0.74-0.90) for R = 1, 0.82 (0.72-0.88) for R = 3, and 0.80 (0.70-0.86) for R = 6. Compared to R = 1, R = 3 scans were rated by radiologists to have significantly improved sharpness (+0.2, p < 0.05) and lower noise (+0.1, p < 0.05). Overall visual quality at R = 6 remained comparable to R = 1 (2.81 at R6 vs. 2.74 at R1).
To determine whether AI-reconstructed prostate MRI at reduced acquisition times maintains prostate cancer (PCa) detection performance comparable to conventional scans.
Question This study investigated whether deep learning reconstruction enables three- to sixfold acceleration without reducing radiologists' detection of clinically significant prostate cancer. Findings In a multi-reader multi-case study with eight radiologists, three- and sixfold acceleration showed no significant change in area under the receiver operating characteristic curve. Clinical relevance Deep learning reconstruction shortened T2-weighted acquisition times at sixfold acceleration while preserving perceived image quality and diagnostic performance across acceleration factors.
This website uses cookies to ensure you get the best experience on our website.