Platform for the radiomics analysis of brain regions: The case of Alzheimer's disease and metabolic imaging

• High accuracy and Interpretability: this study presents a novel radiomics platform (RAB-PET) achieving exceptional accuracy (AUC of 0.976) in distinguishing AD from healthy controls using FDG-PET images. Additionally, RAB-PET identifies crucial brain regions and features associated with AD, enhancing interpretability and potential clinical utility. • Automated region and feature Selection: unlike conventional methods relying on predefined regions, RAB-PET employs an automated approach to identify the most relevant brain regions and features for AD diagnosis. This eliminates potential bias and streamlines the analysis process. • Superior performance compared to existing Methods: the proposed RAB-PET platform demonstrates significant improvement in accuracy compared to previous studies, as shown in Table 1 . This highlights the platform's potential for advancing AD diagnosis and management. • Potential for early AD Detection: identifying critical regions like the entorhinal cortex, implicated in early AD stages, suggests RAB-PET's potential for earlier and more effective intervention strategies. • Broad applicability beyond AD: the study emphasizes the platform's adaptability, suggesting its potential application in detecting other neurological diseases and predicting various AD characteristics, paving the way for broader clinical use. This study introduces a PET-based platform for brain radiomics analysis. We automatically identify key brain regions and features associated with Alzheimer's disease (AD), enabling more accurate diagnosis and staging compared to using predefined regions. To create an integrated platform that covers all the phases of radiomics, we obtained FDG-PET images of 549 individuals from the ADNI database. We used FastSurfer to segment the brain into 95 regions. We then obtained 120 features for each of the 95 ROIs. We employed eight feature selection methods to select and analyze the features. We finally utilized nine different classifiers on the 20 most significant features extracted. For all three predictions AD vs. cognitively normal (CN), AD vs. mild cognitive impairments (MCI), and CN vs. MCI the Random Forest (RF) classifier with LASSO demonstrated the highest accuracy with an AUC of 0.976 for AD vs CN, AUC=0.917 for AD vs MCI, and AUC=0.877 for MCI vs CN. This is the highest performance that we encountered compared to the studies in the literature. Three subregions hippocampus, entorhinal, and amygdala could then be identified as critical. A brain radiomics platform can enable an efficient, standardized, and optimally accurate AD and MCI diagnosis from FDG PET images by using an automated pipeline. The three regions identified as having the highest discriminating power confirm the findings of previous clinical research results on AD. While the focus was AD in this study, the platform can potentially be used to address other brain conditions.