Abstract:BACKGROUND AND OBJECTIVE: Amyloid plaques are one of the physical hallmarks of Alzheimer's disease. The objective of this study is to predict amyloid positivity non-invasively from FDG-PET images using a radiomics approach.METHODS: We obtained FDG-PET images of 301 individuals from various groups, including control normal (CN), mild cognitive impairment (MCI), and Alzheimer's Disease (AD), from the ADNI database. Following the utilization of the CSF Aβ1-42 (192) and Standardized Uptake Value Ratio (SUVR) (1.11) thresholds derived from Florbetapir scans, the subjects were categorized into two categories: those with a positive amyloid status (n = 185) and those with a negative amyloid status (n = 116). The process of segmenting the entire brain into 95 classes using the DKT-atlas was utilized. Following that, we obtained 120 characteristics for each of the 95 regions of interest (ROIs). We employed eight feature selection methods to analyze the features. Additionally, we utilized eight different classifiers on the 20 most significant features extracted from each feature selection method. Finally, in order to improve interpretability, we selected the most important features and ROIs.RESULT: We found that the GNB classifier and the LASSO feature selection method had the best performance with an average accuracy of (AUC=0.924) while using 18 features on 15 ROIs. We were then able to reduce the model to three regions (Hippocampus, inferior parietal, and isthmus cingulate) and three gray-level based features (AUC=0.853).CONCLUSION: The FDG-PET images which serve to study metabolic activity can be used to predict amyloid positivity without the use of invasive methods or another PET tracer and study. The proposed method has superior prediction accuracy with respect to similar studies reported in the literature using other imaging modalities. Only three brain regions had a high impact on amyloid positivity results.

Predicting amyloid positivity from FDG-PET images using radiomics: A parsimonious model