Diagnostically Calibrated AI Image Quality Framework Applied to Prostate Cancer Detection at MRI.

Authors

Abstract

Background Image quality affects diagnostic performance; however, current quality metrics rely on subjective reader labels rather than diagnostic outcomes. Purpose To develop and evaluate a diagnostically calibrated artificial intelligence (AI) framework using deep learning (DL) for image quality assessment, linking prostate MRI quality scores to diagnostic performance. Materials and Methods This single-center retrospective study analyzed 12 496 consecutive prostate multiparametric MRI examinations performed between January 2014 and December 2023 at Radboud University Medical Center. A two-step framework was implemented: First, a DL model was trained on the most reliable low-quality (artifact-degraded structures) and high-quality (clear delineated zones) axial T2-weighted images to produce a continuous image quality score. Second, this score was applied in an independent internal test set and calibrated with diagnostic performance measured as <i>(a)</i> area under the receiver operating characteristic curve (AUC) for clinically significant prostate cancer (csPCa) detection by an AI model and <i>(b)</i> accuracy of csPCa detection by radiologists in routine clinical practice. The reference standard was histopathologic findings at biopsy. Diagnostic performance was calculated for quality score 10-percentile thresholds from 10% to 80%. Linear trend between performance and threshold was assessed using a permutation test on the regression slope. Results A total of 1229 T2-weighted series were used for training, and 568 multiparametric MRI examinations for diagnostic testing (1638 male patients; median age, 67 years [IQR, 62-71 years]). The DL model achieved a mean AUC of 0.99 for distinguishing low- versus high-quality images in the training set during fivefold cross-validation. For csPCa detection in the test set, the AI model AUC improved from 0.92 (95% CI: 0.89, 0.94) to 0.99 (95% CI: 0.97, 1.00) (<i>P</i> = .005) across thresholds of DL model image quality score; radiologist accuracy improved from 77% (95% CI: 73, 80) to 85% (95% CI: 78, 91) (<i>P</i> = .01). Conclusion The AI framework provided a standardized method to derive an objective image quality score that correlated with improved AI model diagnostic performance and radiologist accuracy. © RSNA, 2026 <i>Supplemental material is available for this article.</i> See also the editorial by Barrett in this issue.

Tags

Topics