Author: George Agrotis, Marios Myronakis, Dimitrios Samaras, Kyriaki Theodorou, Ioannis Tsougos, Vassilios Tzortzis, Maria Vakalopoulou, Alexandros Vamvakas, Aikaterini Vassiou, Marianna Vlychou 👨🔬
Affiliation: Medical Physics Department, Medical School, University of Thessaly, Department of Radiology, University of Thessaly, Netherland Cancer Institute, Department of Urology, University of Thessaly, CentraleSupelec, University Paris-Saclay 🌍
Purpose: Prostate cancer (PCa) diagnosis remains challenging due to discrepancies in Gleason Scoring (GS) and risks of overdiagnosis and underdiagnosis. Multiparametric MRI (mpMRI), including Apparent Diffusion Coefficient (ADC) maps, offers potential for non-invasive evaluation but struggles to reliably differentiate clinically significant (csPCa) from clinically insignificant (cinsPCa) cancer. Radiomics, an emerging AI-driven approach, extracts high-dimensional imaging features to improve diagnostic accuracy. This study investigates the integration of radiomics features and ADC ratio with machine learning (ML) classifiers and the application of ComBat harmonization to address multicenter variability.
Methods: A total of 207 PCa lesions from a private multicenter dataset and the PROSTATEx dataset were analyzed. First and higher order radiomics features (n=1246) were extracted from original and filter-applied ADC maps following IBSI guidelines, and ADC ratio was calculated as the mean ADC value of tumor lesions relative to normal tissue. ComBat harmonization mitigated scanner variability. Feature selection methods, including Recursive Feature Elimination (RFE), were applied, and 15 ML classifiers were evaluated. Model performance was assessed via metrics such as Area Under the Precision-Recall Curve (AUC-PR), F1 score, and Balanced Accuracy using stratified cross-validation and external testing. Permutation testing was performed to statistically compare model performance, with a significance threshold of p<0.05.
Results: The radiomics-ADC ratio model combining RFE and Random Forest (RF) with ComBat harmonization achieved superior performance (AUC-PR: 0.92±0.04, F1: 0.86±0.04) compared to the radiomics-only (AUC-PR: 0.91±0.06, F1: 0.84±0.03) and ADC-only models (AUC-PR: 0.71±0.13, F1: 0.58±0.15). ComBat harmonization improved generalizability, and wavelet-transformed radiomic features consistently enhanced model accuracy.
Conclusion: This study demonstrates that integrating radiomics features with ADC ratio and applying ComBat harmonization improves the accuracy and robustness of PCa diagnosis in multicenter settings. These findings highlight the potential of combining imaging biomarkers with harmonization strategies for non-invasive, clinically relevant diagnostics, warranting further validation in larger cohorts.