Improving Clinically Significant Prostate Cancer Detection with a Multimodal Machine Learning Approach: A Large-Scale...

Area of Science:

• Oncology and Genitourinary Medicine
• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Clinically significant prostate cancer (csPCa) detection relies on accurate imaging and risk stratification.
• Biparametric MRI (bpMRI) is increasingly used, but its predictive performance can be enhanced.
• Existing models like PI-RADS have limitations in specificity and sensitivity.

Purpose of the Study:

• To develop and prospectively validate a novel clinical and radiologic model for predicting csPCa.
• To integrate bpMRI radiomics with clinical data for improved prediction accuracy.
• To compare the performance of the new multimodal model against the PI-RADS scoring system.

Main Methods:

• A deep learning algorithm extracted radiomic features from bpMRI scans of 7157 patients (development) and 1629 patients (validation).
• A multimodal model was trained incorporating radiomics, age, prostate-specific antigen, PI-RADS, and lesion location.
• Prospective validation assessed the model's performance using AUC and specificity, with sensitivity analyses for imaging sequences and scanner vendors.

Main Results:

• The multimodal model demonstrated superior performance over PI-RADS in both retrospective (AUC 0.88 vs 0.80) and prospective (AUC 0.91 vs 0.85) validation.
• The model achieved higher specificity (71% vs 58% retrospective, 77% vs 66% prospective) and led to 22.7% fewer biopsies.
• Fairness analyses indicated generalizability across categories, though performance varied slightly across centers and with PI-RADS scores.

Conclusions:

• A multimodal model integrating bpMRI radiomics offers a temporally generalizable and superior predictor of csPCa compared to PI-RADS.
• This approach has the potential to significantly improve prostate cancer diagnostics and reduce invasive procedures.
• Further validation and implementation could enhance clinical decision-making in prostate cancer management.