Retrospective Analysis of Imaging Data Using Artificial Intelligence Methods for Diagnostic Assessment of Amyloid PET in Alzheimer's Disease

• Grant number: 24/22139-5
• Support Opportunities: Scholarships in Brazil - Master
• Start date: March 01, 2025
• End date: February 28, 2027
• Field of knowledge: Health Sciences - Medicine
• Principal Investigator: Marcelo Tatit Sapienza
• Grantee: Gabriel Dias Dadalto
• Host Institution: Faculdade de Medicina (FM). Universidade de São Paulo (USP). São Paulo , SP, Brazil
• Associated research grant: 21/11905-0 - Center of Science, Technology and Development for innovation in Medicine and Health: inLab.iNova, AP.CCD

Abstract

Alzheimer's disease (AD) is the most frequent cause of dementia. The presence of betaamyloid (A²) plaques characteristic of AD can be detected in vivo by positron emission tomography combined with computed tomography (PET/CT) with amyloid tracers, usually analyzed by visual criteria with occasional assistance from semiquantitative techniques. This retrospective study aims to employ artificial intelligence (AI) methods in the analysis of images from 108 patients previouslysubmitted to PET/CT with [11C]PIB for amyloid deposition detection. Examination data already analyzed and classified by specialist doctors will undergo pre-processing for data control and fairness assessment. Subsequently, the imageswill undergo a semiautomatic segmentation process of the brain regions of interest to generate training data for AI segmentation models. With the prepared data, training of the classification networks EfficientNet, Inception (GoogLeNet),ViT, and ResNeXt will begin, using only preclassified images as binary trainingdata (positive or negative for AD). Subsequently, training of the segmentation architectures 3D UNet, DeepMedic, UNETR++, and Attention VNet will be conducted, employing segmented data from regions of interest as training inputs.Precision and efficiency results of each model will be obtained, allowing comparisons with previous analyses performed by specialists and among the applied models. All training processes of segmentation and classification networks willhave their flow and results monitored by techniques and frameworks aimed at fairness application in AI models such as conditional statistical parity tests, odds equalizations, AI Fairness 360, and Fairlearn. In this way, efforts will be made to identify and mitigate possible biases of AI models and contribute to supportingspecialists in faster and safer diagnoses.