Zone-specific logistic regression models improve classification of prostate cancer on multi-parametric MRI

Area of Science:

• Radiology
• Urologic Oncology
• Medical Imaging

Background:

• Multiparametric MRI (mp-MRI) is crucial for prostate cancer detection and localization.
• Prostate cancer exhibits distinct patterns in the peripheral zone (PZ) and transition zone (TZ).
• Zone-specific logistic regression (LR) models may improve diagnostic accuracy.

Purpose of the Study:

• To evaluate the interchangeability of mp-MRI logistic regression (LR) models between the peripheral zone (PZ) and transition zone (TZ) for prostate cancer classification.
• To determine if zone-specific models can be reliably applied across different prostate zones.

Main Methods:

• Development and validation of uni/multi-variate mp-MRI LR models for PZ and TZ cancer classification using transperineal-template-prostate-mapping biopsy data.
• Assessment of inter-zonal performance by applying TZ models to the PZ cohort and vice-versa.
• Receiver operating characteristic area-under-curve (ROC-AUC) analysis to compare model performance.

Main Results:

• Univariate models showed T2 signal normalization (T2nSI) in TZ (ROC-AUC=0.77) and normalized early contrast-enhanced T1 signal (DCE-nSI) in PZ (ROC-AUC=0.79) had the best performance.
• Bivariate/trivariate modeling did not significantly improve classification performance.
• PZ models with DCE-nSI and TZ models with maximum enhancement performed poorly when applied to the other zone.

Conclusions:

• Logistic regression models relying solely on dynamic contrast-enhanced MRI (DCE-MRI) parameters are not interchangeable between prostate PZ and TZ.
• Models based on T2-weighted imaging and/or apparent diffusion coefficient (ADC) demonstrate greater robustness for inter-zonal application.
• Contrast enhancement parameters show significant differences between benign PZ and TZ, impacting cross-zonal model performance.