Využití kvantitativních markerů získaných z obrazů magnetické rezonance při diagnostice extrakapsulárního šíření

Abstract

Prostate cancer is one of the most common malignancies in men. A key factor in treatment decisionmaking is the presence of extracapsular extension (ECE); however, its assessment on MRI is predominantly qualitative and subject to interobserver variability. The aim of this study was to evaluate whether ECE can be predicted using radiomic features extracted from the peripheral and transition zones of the prostate on T2-weighted MR images. A cohort of 289 patients examined on a 3T Siemens MAGNETOM Prisma scanner was analyzed. After preprocessing (N4ITK bias field correction and Z-score normalization), 26 features were extracted from each zone, of which 11 independent features were selected for classification. The models were trained using cross-validation and optimization in two tasks (binary and three-class classification). In the binary task, the best performance was achieved by a Support Vector Machine (AUC = 0.6621; accuracy = 66.7%), while in the three-class task, the Random Forest model performed best (AUC = 0.6309; accuracy = 48.8%), with the ECE 1 class being the most challenging to classify. The most informative features were texture features from the peripheral zone, specifically glcm_SumSquares, glcm_MCC, glcm_InverseVariance. In conclusion, zonal radiomics contains some information related to ECE; however, its predictive performance remains limited and is insufficient for clinical application.