MRI-Based Surrogate Imaging Markers of Aggressiveness in Prostate Cancer: Development of a Machine Learning Model...

Area of Science:

• Radiology
• Oncology
• Machine Learning

Background:

• Prostate cancer (PCa) diagnosis relies on invasive methods.
• Accurate identification of clinically significant PCa (csPCa) is crucial for treatment decisions.
• Noninvasive biomarkers are needed to assess PCa aggressiveness.

Purpose of the Study:

• To develop and validate a noninvasive machine learning (ML) model for identifying csPCa (Gleason Score ≥ 7) in Hispanic men.
• To evaluate the model's performance using biparametric MRI (bpMRI) radiomic features and clinical data.
• To compare the ML model's accuracy against PI-RADS and prostate-specific antigen density (PSA-D).

Main Methods:

• Retrospective study of 86 Hispanic men with PCa undergoing prebiopsy 3T MRI.
• 2D segmentation of lesions in T2WI/ADC images by two observers.
• Multivariate ML model development using Recursive Feature Elimination (RFE) with radiomic and clinical features (PV, PSA-D).
• Classification of csPCa (GS ≥ 7) vs. non-csPCa (GS = 6).
• Stratified train/test (80%) and validation (20%) sets.

Main Results:

• Radiomic features from T2WI/ADC images are associated with Gleason Score (GS) in PCa patients.
• The best multivariate ML model integrated T2WI/ADC radiomic features with clinical data (PV, PSA-D).
• The model achieved a validation area under the curve (AUC) of 0.80 for csPCa detection, outperforming PI-RADS (AUC: 0.71) and PSA-D (AUC: 0.78).

Conclusions:

• The developed multivariate ML model accurately classifies csPCa, surpassing PI-RADS v2.1 and PSA-D.
• MRI-derived radiomics (T2WI/ADC) show potential as a robust biomarker for assessing PCa aggressiveness.
• This noninvasive approach offers a promising tool for PCa management in Hispanic populations.