Open Access

ARTICLE

AI-based detection of MRI-invisible prostate cancer with nnU-Net

Jingcheng Lyu^1,2,#, Ruiyu Yue^1,2,#, Boyu Yang^1,2, Xuanhao Li^1,2, Jian Song^1,2,*

1 Department of Urology, Beijing Friendship Hospital, Capital Medical University, Beijing, 100050, China
2 Institute of Urology, Beijing Municipal Health Commission, Beijing, 100050, China

* Corresponding Author: Jian Song. Email:
# Jingcheng Lyu and Ruiyu Yue are co-first authors

(This article belongs to the Special Issue: Advancing Early Detection of Prostate Cancer: Innovations, Challenges, and Future Directions)

Canadian Journal of Urology 2025, 32(5), 445-456. https://doi.org/10.32604/cju.2025.068853

Received 08 June 2025; Accepted 15 August 2025; Issue published 30 October 2025

Abstract

Objectives: This study aimed to develop an artificial intelligence (AI)-based image recognition system using the nnU-Net adaptive neural network to assist clinicians in detecting magnetic resonance imaging (MRI)-invisible prostate cancer. The motivation stems from the diagnostic challenges, especially when MRI findings are inconclusive (Prostate Imaging Reporting and Data System [PI-RADS] score ≤ 3). Methods: We retrospectively included 150 patients who underwent systematic prostate biopsy at Beijing Friendship Hospital between January 2013 and January 2023. All were pathologically confirmed to have clinically significant prostate cancer, despite negative findings on preoperative MRI. A total of 1475 MRI images, including T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC) sequences, were collected. The nnU-Net was employed as the initial segmentation framework to delineate tumor regions in MRI images, based on histopathologically confirmed prostate cancer sites. A convolutional neural network-based deep learning model was subsequently designed and trained. Its performance was evaluated using five-fold cross-validation. Results: Among 150 patients with clinically significant prostate cancer diagnosed, all with PI-RADS ≤ 3 on MRI, the median age was 67 years (IQR: 62–72), and 105 patients (70.0%) had a Gleason score ≥ 7. A total of 1475 multiparametric MRI images were analyzed. Using five-fold cross-validation, the AI-based image recognition system achieved a mean Dice similarity coefficient of 55.0% (range: 51.6–56.5%), with a mean sensitivity of 50.5% and a mean specificity of 96.9%. The corresponding mean false-positive and false-negative rates were 3.1% and 49.5%, respectively. Conclusion: We successfully developed an AI-based image recognition system utilizing the nnU-Net adaptive neural network, demonstrating promising diagnostic performance in detecting MRI-invisible prostate cancer. This system has the potential to enhance early detection and management of prostate cancer.

---

Keywords

---