Navigating Uncertainty in MRI Diagnosis: A Human-AI Collaborative Strategy for Stratifying Clinically Significant Prostate Cancer.

BACKGROUND: Deep learning (DL) methods have shown potential for predicting clinically significant prostate cancer (csPCa), but radiologists often face challenges in effectively leveraging these techniques for csPCa prediction. PURPOSE: To develop an automated DL model based on biparametric-MRI (bpMRI) and propose a human-machine collaborative strategy for predicting csPCa. STUDY TYPE: Retrospective. POPULATION: A total of 4305 patients were enrolled. Centers 1-2 and 4-7 comprised the training (2437 patients, mean age 68 ± 8) and the internal validation (581 patients, mean age 67 ± 8) cohorts; Centers 8-10 comprised the external validation cohort 1 (622 patients, mean age 71 ± 8), and Center 3 comprised the external validation cohort 2 (665 patients, age not available). FIELD STRENGTH/SEQUENCE: T2-weighted imaging (T2WI) using fast or turbo spin echo and diffusion-weighted imaging (DWI) using single-shot echo planar imaging were acquired at 1.5 and 3 T. ASSESSMENT: A DL model (UFormer) including prostate segmentation and csPCa prediction was constructed using bpMRI. Its performance was evaluated in two external validation cohorts (EVCs) and compared with that of radiologists. Further, a UFormer-radiologist collaborative predictive strategy was proposed. STATISTICAL TESTS: Area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, DeLong test, and McNemar test. p < 0.05 was considered significant. RESULTS: Compared with radiologists' Prostate Imaging Reporting and Data System (PI-RADS) assessment, UFormer-combined radiologists showed significantly higher AUC and accuracy of 0.918 ± 0.012 and 0.857 ± 0.014 for the less-experienced radiologists, 0.931 ± 0.010 and 0.870 ± 0.014 for the more-experienced radiologists, respectively, due to greatly increasing specificity by 121.7% for the less-experienced radiologists and 60.2% for the more-experienced radiologists in EVC1. Additionally, UFormer identified 86.5% and 93.9% of non-csPCa patients, who had been interpreted originally as PI-RADS 3 by more- and less-experienced radiologists, respectively. DATA CONCLUSIONS: UFormer enhanced the predictive performance of radiologists and narrowed performance gaps between experience levels. The UFormer-radiologist collaborative paradigm combined model advantages with PI-RADS assessment, providing a strategy for clinical application. EVIDENCE LEVEL: 4. TECHNICAL EFFICACY: Stage 2.