Sub-voxel Susceptibility Mapping and Machine Learning to Detect Brain Iron Deposition and Its Cognitive Relevance in Beta-Thalassemia.
Authors
Affiliations (2)
Affiliations (2)
- From the Department of Radiology (M.Y., Y.H., C.Z., G.C., C.T., Y.L., J.L., R.K., J.L., P. P.), The First Affiliated Hospital of Guangxi Medical University, Nanning, China; NHC Key Laboratory of Thalassemia Medicine (C.T., P.P.), Nanning, China; Binzhou Medical University Hospital (M.Y.), Binzhou, China; and MR Research Collaboration Team (H.Z.), Siemens Healthineers Ltd., Shenzhen, China.
- From the Department of Radiology (M.Y., Y.H., C.Z., G.C., C.T., Y.L., J.L., R.K., J.L., P. P.), The First Affiliated Hospital of Guangxi Medical University, Nanning, China; NHC Key Laboratory of Thalassemia Medicine (C.T., P.P.), Nanning, China; Binzhou Medical University Hospital (M.Y.), Binzhou, China; and MR Research Collaboration Team (H.Z.), Siemens Healthineers Ltd., Shenzhen, China. [email protected].
Abstract
Brain iron dysregulation is increasingly recognized as a critical contributor to neurocognitive impairment in patients with beta-thalassemia major (β-TM). However, in-vivo characterization of region-specific iron accumulation and its relationship with cognitive function remains limited. Fifty β-TM patients and fifty age- and sex-matched healthy controls underwent 3T multi-echo gradient-echo MRI. Sub-voxel chi-separation decomposed magnetic susceptibility into paramagnetic (iron-related) and diamagnetic components. Regional paramagnetic susceptibility was extracted from anatomically defined regions of interest (ROIs). Group differences were assessed using FDR-corrected comparisons (<i>q</i> < 0.05). Partial Spearman correlations evaluated associations between regional susceptibility and Montreal Cognitive Assessment (MoCA) scores, controlling for age, sex, and education. Three classifiers-support vector machine (SVM), random forest (RF), and eXtreme Gradient Boosting (XGBoost)-were trained using ROI-based features. In each Monte Carlo cross-validation iteration, the data were divided into stratified 80/20 training/testing subsets. Preprocessing, feature selection, and hyperparameter optimization were performed using the training data only, and the held-out test subset was used exclusively for final evaluation. SHapley Additive exPlanations (SHAP) were used for model interpretability. β-TM patients showed significantly higher paramagnetic susceptibility in the hippocampus, insula, and anterior cingulate cortex (<i>q</i> < 0.05). Among classifiers, SVM with an RBF kernel demonstrated the highest performance (mean AUC = 0.919 ± 0.054), outperforming RF and XGBoost. SHAP analysis identified hippocampal and insular susceptibility as key features, with higher susceptibility linked to lower MoCA scores. ROI-based chi-separation detected iron-related changes in β-TM, and exploratory machine-learning analysis highlighted regions associated with cognitive vulnerability, supporting the potential value of susceptibility-based imaging features for studying neurocognitive risk patterns in this population.