Artificial intelligence-based prostate cancer risk stratification via large-scale ultrasound images paired with histopathology

Background: Prostate cancer (PCa) is primarily screened and diagnosed via prostate-specific antigen (PSA) testing, digital rectal examination (DRE), multiparametric magnetic resonance imaging (mpMRI), and transrectal ultrasound (TRUS). Nevertheless, an open research challenge remains to develop artificial intelligence (AI) tools that enable non-invasive, accurate diagnosis, and risk stratification of PCa. This study aimed to construct an Anatomy-Aware Transformer (AAT) model based on a large-scale dataset of paired TRUS images and corresponding histopathological findings, and to further validate the clinical value of TRUS-based AI for PCa diagnosis and risk stratification.

Methods: Sagittal TRUS images were prospectively acquired pre- and intra-biopsy. Based on the intra-biopsy images, the prostate contour and the location of biopsy tissue cores on pre-biopsy images were manually annotated. According to the characteristic features of the prostate and periprostatic tissues, we designed a feature screening module and a feature integration module, and further constructed an AAT model that deeply integrates prostate regional segmentation and PCa diagnosis tasks. The diagnostic performance of the AAT model in PCa detection and risk stratification was evaluated by comparing its results with those of 18 classic models and four ultrasonographers.

Results: A Primary Dataset (18,424 images; May 2022 to Dec 2023) supported training/internal validation, and an External Dataset (683 images; Jun to Jul 2024) external validation. The AAT model outperformed all comparison models across all tasks: PCa detection (internal accuracy 0.977 vs. 0.655–0.972; external accuracy 0.810 vs. 0.616–0.737), clinically significant PCa (CSPCa) detection (internal/external accuracy: 0.936/0.843 vs. 0.697–0.934/0.304–0.762), 3-category risk stratification (internal/external accuracy: 0.828/0.836 vs. 0.697–0.822/0.642–0.768), and 5-category stratification (internal/external accuracy: 0.760/0.722 vs. 0.655–0.760/0.690–0.715). AAT’s PCa diagnostic accuracy (0.943) exceeded ultrasonographers (0.567–0.757) and improved their performance by 0.036–0.304 as an aid.

Conclusions: We developed a novel AAT model that exhibits outstanding performance in diagnosing PCa and CSPCa, with reliable efficacy in PCa risk stratification. This non-invasive model effectively improves the diagnostic accuracy of ultrasonographers and enables individualized clinical management for PCa patients with varying risk levels, showing promising clinical translation potential.