Comparing deep learning CNN method with traditional MRI-based hippocampal segmentation and volumetry for early Alzheimer’s disease diagnosis across diverse populations

Abstract

The advent of artificial intelligence (AI) driven software has impacted numerous aspects of medicine, leading to automated algorithms that assist in performing feature extraction, making measurements on diagnostic imaging, and aiding in diagnosing disorders. In neuroimaging, AI-based convoluted neural networks (CNN) facilitate the automated segmentation of the hippocampal volume observed on MRI diagnostic imaging, thereby aiding in the diagnosis of Alzheimer’s disease (AD). Traditional voxel-based morphometry (VBM) used for measuring hippocampal volume can be time-laborious and sensitive to pre-processing errors, thus CNN-based algorithms can minimize the time and reduce human errors. We utilized HippoDeep, an open-source CNN-based algorithm, to compare the MRI-derived hippocampal volumes from a Caucasian population dataset with a Southeast Asian AD and cognitively healthy control (HC) population dataset. ROC analysis demonstrated enhanced diagnostic performance using HippoDeep, yielding AUCs of 0.918 (left hippocampus) and 0.882 (right hippocampus), in contrast to VBM’s 0.788 and 0.741, respectively. We determined the cut-off thresholds for hippocampal volume to further improve the HippoDeep-driven classification method. CNN-based method outperformed traditional semiautomated method in segmentation accuracy (p < 0.001) with non-significant interpopulation differences. Moreover, HippoDeep-derived hippocampal volumes exhibited stronger correlations with MMSE scores, with smaller volumes being associated with lower cognitive performance (r = 0.63 vs. r = 0.42). HippoDeep offers accurate, reproducible, and generalizable hippocampal segmentation, supporting its potential as a clinical tool for early AD diagnosis across diverse populations.