Radiomics Feature Profiling of Brodmann Regions in Structural MRI: A Machine Learning Study of Intensive Verbal Memorisation (Huffaz vs Controls)

Memorisation-based cognitive training has been hypothesized to relate to experience-dependent brain plasticity; however, quantitative evidence at the regional level remains limited. We hypothesized that radiomics descriptors extracted from Brodmann-area volume-of-interest (VOI) regions in pre-processed structural MRI would contain sufficient information to discriminate Quran memorizers (Huffaz) from non-memorizers (controls), and we evaluated this hypothesis using a fully nested validation framework. T1-weighted MRI volumes were pre-processed using a voxel-based morphometry pipeline, and VOIs were defined using Brodmann-area masks. Using PyRadiomics, first-order and texture features were extracted per VOI and combined into a feature matrix for classification. Models were evaluated using repeated nested cross-validation (outer 5-fold × 10 repeats; inner 5-fold for tuning), with ROC-AUC as the primary metric. Random Forest achieved the strongest discrimination (AUC = 0.6704 ± 0.1792), followed by Logistic Regression (AUC = 0.5948 ± 0.2153), while SVM with an RBF kernel underperformed (AUC = 0.4356 ± 0.1927). One-sided testing against chance (AUC = 0.5) indicated above-chance performance for Random Forest and Logistic Regression, but not for SVM-RBF. These results should be interpreted as exploratory because the cohort is small (n = 47) and no independent external validation cohort was available. Practically, the observed effect sizes suggest that VOI-based radiomics may capture detectable group-associated imaging signatures under the current preprocessing and VOI assumptions, motivating validation on larger cohorts, sensitivity analysis (e.g., discretization/normalization settings), and assessment of probability calibration.