Usage of AI in neuroimaging to detect asymptomatic Alzheimer's disease

Abstract

Background: Alzheimer's disease (AD) is the leading cause of dementia worldwide. In 2019, it was estimated that approximately 57 million people had dementia, with projections that this number will triple by 2050. The main clinical challenge is the late diagnosis of these conditions, often occurring when cognitive decline is already quite evident. However, neuroimaging changes such as hippocampal atrophy and cortical thinning tend to occur several years before the onset of clinical symptoms. Therefore, machine learning (ML) and deep learning (DL) methods applied to magnetic resonance imaging (MRI) are increasingly being studied as a strategy for the pre-symptomatic detection of dementia. Objective: Analyze the effectiveness of artificial intelligence-based ML technologies involving MRI for the early detection of AD, assessing the effectiveness of the systems and their practical application. Methods: The search was conducted in the PubMed database using terms like "machine learning" and "Alzheimer's" over the last 10 years. The inclusion criteria required the full text to be available in English and be based on clinical studies or systematic reviews and meta-analysis. Works that did not meet these criteria were directly excluded. Finally, the relevance of each article was evaluated for inclusion. Results: Despite studies acknowledging methodological limitations related to the low number of participants and a certain risk of bias, it was still possible to obtain good performance from ML in MRI for subclinical detection of AD. In this sense, the different techniques included analysis of cortical thickness, hippocampal volume, and brain activity patterns on MRI, in addition to finer detections from the entorhinal cortex, precuneus, lateral temporal lobe, and prefrontal cortex. Among the studies, the approaches achieved robust diagnostic performance, with minimal variations in sensitivity and specificity estimates, approximately 92.5% and 95.7%, respectively. Among the most researched models, Convolutional Neural Networks (CNNs) stand out especially in identifying spatial relationships in volumetric MRI images. A 2024 systematic review reinforced the robustness of this approach for multiple stages of AD. In recent years, Vision Transformers (ViTs) have emerged as a complementary model to CNNs. ViTs utilize self-awareness mechanisms to better identify global dependencies between distant cortical regions. This is advantageous in the early detection of AD, since its anatomical changes are distributed. Hybrid ViT-CNN models reported an accuracy of 92.14%, pooled sensitivity of 93.27%, and specificity of 89.95% in differentiating between AD, mild cognitive impairment, and healthy controls. Conclusion: The advancement of DL and ML techniques highlights a promising future for the clinical application of these tools. However, to consolidate the implementation of these technologies in medical practice, better methodological standardization is essential, as well as the performance of prospective multicenter studies.