Background Type 2 diabetes mellitus (T2DM) is associated with a higher risk of mild cognitive impairment (MCI), but conventional structural MRI (sMRI) metrics lack sensitivity for detecting early hippocampal alterations. Radiomic analysis of hippocampal subfields may capture diabetes-related microstructural abnormalities. Objective To develop and interpret a structural MRI-based model incorporating hippocampal subfield radiomic features for identifying MCI in T2DM patients. Methods This retrospective study included 149 T2DM patients. Based on Montreal Cognitive Assessment (MoCA) and Mini-Mental State Examination (MMSE) scores, participants were classified into T2DM-MCI and T2DM-NC groups. A total of 2232 radiomic features were extracted from 24 hippocampal subfields segmented by FreeSurfer. Maximum relevance minimum redundancy (mRMR) and least absolute shrinkage and selection operator (LASSO) were used to construct the hippocampal radiomics score (Hip-Radscore). Additional models were built using gray matter volume (sMRI-Radscore), subfield volumes (SFV-Score), and clinical variables. Model performance was evaluated by bootstrap resampling, ROC analysis, and SHapley Additive exPlanations (SHAP) interpretation. Results The Hip-Radscore model achieved an AUC of 0.76 (95% CI: 0.68–0.84), outperforming the gray matter volume (0.62) and SFV-Score (0.63) models. The combined model integrating imaging and clinical variables reached the highest accuracy (AUC = 0.87, 95% CI: 0.81–0.93), significantly superior to the clinical-only model (AUC = 0.70, p < 0.05). The Hip-Radscore correlated positively with diabetes duration (ρ = 0.47) and HbA1c (ρ = 0.56), and negatively with MMSE (ρ = −0.63) and MoCA (ρ = −0.65) (all p < 0.001). Conclusions Hippocampal subfield radiomics significantly improves the detection of MCI in T2DM and reflects diabetes-specific structural alterations linked to metabolic status and cognitive decline.
Wang et al. (Tue,) studied this question.