Predicting clinically significant prostate cancer with a deep learning approach: a multicentre retrospective study

Area of Science:

• Radiology and medical imaging
• Artificial intelligence in healthcare
• Oncology and cancer diagnostics

Background:

• Accurate diagnosis of clinically significant prostate cancer (csPCa) is crucial for effective patient management.
• Multiparametric magnetic resonance imaging (mpMRI) with Prostate Imaging and Reporting and Data System (PI-RADS) is a standard tool, but its performance can be limited.
• Developing advanced computational tools can enhance diagnostic accuracy and reduce unnecessary procedures.

Purpose of the Study:

• To develop and validate deep learning (DL) models using multicentre biparametric MRI (bpMRI) for csPCa diagnosis.
• To compare the diagnostic performance of these DL models against expert PI-RADS assessments.
• To evaluate an integrated model combining DL and PI-RADS for improved csPCa detection.

Main Methods:

• 1861 male patients from seven hospitals underwent mpMRI and subsequent radical prostatectomy or biopsy.
• DL models (DL-BM for benign/malignant, DL-CS for csPCa/non-csPCa) were developed using training data (1216 patients).
• Model performance was externally validated on 645 patients, comparing DL models and an integrated PIDL-CS model against radiologist PI-RADS assessments.

Main Results:

• DL-BM and DL-CS models showed comparable performance to PI-RADS in external validation cohorts (AUC, P > 0.05).
• The integrated PIDL-CS model demonstrated superior AUC compared to PI-RADS in most cohorts (P < 0.05).
• PIDL-CS significantly increased the specificity for csPCa detection compared to PI-RADS (P < 0.05).

Conclusions:

• Deep learning models offer a potential non-invasive auxiliary tool for predicting csPCa.
• The PIDL-CS model significantly enhances the specificity of csPCa detection, potentially reducing unnecessary biopsies.
• These AI-driven tools can assist radiologists in achieving more precise csPCa diagnoses.