• Gradient magnitude-based features were derived from the entorhinal cortex on T1-weighted MRI. • The proposed feature captured structural heterogeneity associated with Alzheimer’s disease progression. • Inclusion of the proposed feature consistently improved classification across three machine learning models. • SHAP analysis revealed that the proposed feature interacts differently with other biomarkers across disease stages. Gradient magnitude-based features were derived from the entorhinal cortex on T1-weighted MRI. The proposed feature captured structural heterogeneity associated with Alzheimer’s disease progression. Inclusion of the proposed feature consistently improved classification across three machine learning models. SHAP analysis revealed that the proposed feature interacts differently with other biomarkers across disease stages. Volumetric analysis of brain structures is widely used to detect pathological changes in Alzheimer’s disease (AD), but it has limited sensitivity to subtle tissue alterations. This study proposes a gradient magnitude-based feature derived from the normalized gradient magnitude (nGM) of the entorhinal cortex (EC) as a potential biomarker for AD detection. A total of 1,422 T1-weighted MRI scans from the Alzheimer’s Disease Neuroimaging Initiative were used. nGM features were extracted from the EC using Sobel filtering and beta distribution fitting. Two classification tasks were conducted to distinguish cognitively normal (CN) individuals from those with AD and mild cognitive impairment (MCI). Model performance, with and without the nGM feature, was evaluated using XGBoost, Random Forest, and Multi-Layer Perceptron (MLP) with repeated 5-fold cross-validation. Shapley Additive Explanations (SHAP) analysis was used to interpret feature contributions. The nGM feature demonstrated disease-stage-dependent distribution shifts and low correlation with volumetric features, indicating its complementary diagnostic value. Inclusion of the nGM feature consistently improved Accuracy across all models. In the CN vs. AD task, XGBoost improved from 0.828 to 0.845 (q < 0.001), Random Forest from 0.831 to 0.8849 (q < 0.040), and MLP from 0.813 to 0.834 (q < 0.001). In the CN vs. MCI task, improvements were observed from 0.681 to 0.695 (XGBoost, q < 0.001), 0.691 to 0.710 (Random forest, q < 0.001), and 0.677 to 0.698 (MLP, p = 0.003). These results highlight the value of EC-derived nGM texture features as a complementary modality to conventional volumetric measures for AD classification.
Lim et al. (Sun,) studied this question.