Author: Yongha Gi, Jinju Heo, Jinyoung Hong, Yunhui Jo, Yousun KO, HyeongJin Lim, Sang Yoon PARK, Myonggeun Yoon 👨🔬
Affiliation: Korea University, Institute of Global Health Technology (IGHT), Korea University, Republic of Korea 🌍
Purpose: To evaluate the effectiveness of the gradient magnitude (GM) feature of the entorhinal cortex, observed in T1 MR images, in dementia classification.
Methods: A total of 1,422 ADNI T1 MR dataset were segmented using the FastSurfer automated segmentation tool. Based on the segmentation results, seven volumetric features and GM features of the entorhinal cortex were extracted. Two dementia classification models were trained using the same XGBoost algorithm. Model 1 used only the seven volumetric features, while Model 2 incorporated the GM feature in addition to the seven volumetric features. Feature importance and Shapley Additive Explanation (SHAP) analyses were performed on Model 2.
Results: The accuracy, recall, precision, specificity, and F1 score of Model 1 were 0.8920, 0.8409, 0.9367, 0.9432, and 0.8862, respectively. For Model 2, the corresponding metrics were 0.9148, 0.8864, 0.9398, 0.9432, and 0.9123. In the feature importance analysis, the GM feature achieved an F-score of 161, ranking 3rd among the 8 features. Additionally, the average absolute SHAP value for the GM feature was 0.7403, ranking 4th among the 8 features.
Conclusion: This study demonstrates that the GM feature extracted from the entorhinal cortex enhances dementia classification performance in AI-based tasks, showing a contribution level comparable to that of major volumetric features