Subregional Biomarkers in FDG PET for Alzheimer's Diagnosis and Staging: An Interpretable and Explainable model

Abstract

ObjectiveTo investigate the radiomics features of the hippocampus and the amygdala subregions in FDG-PET images that can best differentiate Mild Cognitive Impairment (MCI), Alzheimers Disease (AD), and healthy patients. MethodsBaseline FDG-PET data from 555 participants in the ADNI dataset were analyzed, comprising 189 cognitively normal (CN) individuals, 201 with MCI, and 165 with AD. The hippocampus and amygdala were segmented based on the DKT-Atlas, with additional subdivisions guided by probabilistic atlases from Freesurfer. Then radiomic features (n=120) were extracted from 38 hippocampal subregions and 18 amygdala nuclei using PyRadiomics. Various feature selection techniques, including ANOVA, PCA, Chi-square, and LASSO, were applied alongside nine machine learning classifiers. ResultsThe Multi-Layer Perceptron (MLP) model combined with LASSO demonstrated excellent classification performance: ROC AUC of 0.957 for CN vs. AD, ROC AUC of 0.867 for MCI vs. AD, and ROC AUC of 0.782 for CN vs. MCI. Key regions, including the accessory basal nucleus, presubiculum head, and CA4 head, were identified as critical biomarkers. Features including GLRLM (Long Run Emphasis) and Small Dependence Emphasis (GLDM) showed strong diagnostic potential, reflecting subtle metabolic and microstructural changes often preceding anatomical alterations. ConclusionSpecific hippocampal and amygdala subregions and their four radiomic features were found to have a significant role in the early diagnosis of AD, its staging, and its severity assessment by capturing subtle shifts in metabolic patterns. Furthermore, these features offer potential insights into the diseases underlying mechanisms and model interpretability.