https://atlas.rsna.org/schemas/2025-11/model.json

MRI-based clinically significant prostate cancer diagnosis

https://doi.org/10.1148/ryai.230362

1.0

Corresponding author: Zaiyi Liu

Key Area Research and Development Program of Guangdong Province, China (2021B0101420006); Regional Innovation and Development Joint Fund of National Natural Science Foundation of China (U22A20345); Guangdong Provincial Key Laboratory of Artificial Intelligence in Medical Image Analysis and Application (2022B1212010011); High-Level Hospital Construction Project (DFJHBF202105); National Science Foundation for Young Scientists of China (82302130, 82102034); Natural Science Foundation for Distinguished Young Scholars of Guangdong Province (2023B1515020043); National Natural Science Foundation of China Excellent Young Scientists Fund (Overseas) (22HAA01598).

Retrospective multicenter study approved by local ethics committee; informed consent waived due to retrospective, anonymous analysis (approval no. KY2023–146–01). Registered at ChiCTR (ChiCTR2300069832).

Automated 3D detection/diagnosis pipeline based on 3D nnU-Net. Prostate gland segmentation on T2-weighted images to derive central and peripheral gland masks; lesion detection uses five inputs (segmented gland masks, T2WI, ADC, DWI). TPAS training incorporates adversarial samples generated using style-transfer features from a pretrained VGG-16 and adversarial attack optimization. Implemented in Python with PyTorch 1.13.1.

Baseline model code available at https://github.com/DIAGNijmegen/picai_baseline. Data sharing on request from corresponding author; public training dataset PI-CAI v1.1 at https://zenodo.org/record/6667655.

Improves accuracy and robustness of MRI-based diagnosis of clinically significant prostate cancer, particularly in presence of rectal artifacts.

Clinical decision support systems; assists image interpretation and lesion detection on prostate MRI.

Example working point used in figure: 0.45 probability threshold (defined by ROC curve) for lesion-level decision; study primarily reports threshold-independent AUC/AUPRC.

Fully automated segmentation of prostate zones and automated lesion detection with patient-level csPCa probability; provides decision support without replacing clinician judgment.

Evaluation of adult male patients with suspected prostate cancer undergoing biparametric prostate MRI (T2WI, DWI, ADC) in radiology departments (3T scanners in external validation; 1.5T/3T in training).

Biparametric MRI volumes: T2-weighted imaging, high-b-value DWI, and ADC; plus internally generated central and peripheral gland segmentation masks from T2WI.

Apply preprocessing (cropping, resampling, normalization) consistent with study; generate prostate gland masks on T2WI; input T2WI, DWI, ADC, and masks to the 3D nnU-Net–based detector to obtain lesion confidence maps and patient-level csPCa probability. Training used fivefold cross-validation; TPAS alternates adversarial sample generation with model optimization.

Reference standard based on targeted plus systematic biopsies (no whole-mount histopathology). Training/evaluation limited to MRI-visible csPCa; MRI-invisible lesions may be missed. TPAS designed to address rectal artifacts; effects of other MRI artifacts not evaluated. Performance still declines under adversarial noise. External validation limited to three centers in China; TPAS code/weights not released in the article.

Consider simple bowel preparation to reduce artifacts and employ TPAS-trained model to mitigate rectal artifact interference for more stable csPCa detection on MRI.

Training performed with fivefold cross-validation on public PI-CAI dataset; internal and multicenter external testing reported with detailed metrics. Implementation details (PyTorch 1.13.1) provided; baseline code repository referenced.