Related Experiment Video
Updated: Jul 18, 2025
A Cognitive Fusion-guided Prostate Biopsy Using Multiparametric Magnetic Resonance Imaging and Transrectal Ultrasound
Published on: March 21, 2025
Detecting Adverse Pathology of Prostate Cancer With a Deep Learning Approach Based on a 3D Swin-Transformer Model and
Litao Zhao1,2,3, Jie Bao4, Ximing Wang4
1School of Engineering Medicine, Beihang University, Beijing, China.
Background:
Accurately detecting adverse pathology (AP) presence in prostate cancer patients is important for personalized clinical decision-making. Radiologists' assessment based on clinical characteristics showed poor performance for detecting AP presence.
Purpose:
To develop deep learning models for detecting AP presence, and to compare the performance of these models with those of a clinical model (CM) and radiologists' interpretation (RI).
Study Type:
Retrospective.
Population:
Totally, 616 men from six institutions who underwent radical prostatectomy, were divided into a training cohort (508 patients from five institutions) and an external validation cohort (108 patients from one institution).
Field Strength/Sequences:
T2-weighted imaging with a turbo spin echo sequence and diffusion-weighted imaging with a single-shot echo plane-imaging sequence at 3.0 T.
Assessment:
The reference standard for AP was histopathological extracapsular extension, seminal vesicle invasion, or positive surgical margins. A deep learning model based on the Swin-Transformer network (TransNet) was developed for detecting AP. An integrated model was also developed, which combined TransNet signature with clinical characteristics (TransCL). The clinical characteristics included biopsy Gleason grade group, Prostate Imaging Reporting and Data System scores, prostate-specific antigen, ADC value, and the lesion maximum cross-sectional diameter.
Statistical Tests:
Model and radiologists' performance were assessed using area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. The Delong test was used to evaluate difference in AUC. P < 0.05 was considered significant.
Results:
The AUC of TransCL for detecting AP presence was 0.813 (95% CI, 0.726-0.882), which was higher than that of TransNet (0.791 [95% CI, 0.702-0.863], P = 0.429), and significantly higher than those of CM (0.749 [95% CI, 0.656-0.827]) and RI (0.664 [95% CI, 0.566-0.752]).
Data Conclusion:
TransNet and TransCL have potential to aid in detecting the presence of AP and some single adverse pathologic features.
Level Of Evidence:
4 TECHNICAL EFFICACY: Stage 4.
