Background

With the advent of an aging society, the number of patients with Alzheimer's Disease (AD), a neurodegenerative brain disorder, is rapidly increasing. Early diagnosis, which identifies the disease before cognitive abilities are lost, plays a crucial role in preserving the patient's quality of life. Currently, clinical diagnosis relies on Positron Emission Tomography (PET) or cerebrospinal fluid analysis. However, these methods are limited by high costs, restrictions related to the administration of radioactive substances, and the discomfort associated with lumbar punctures, which reduces patient accessibility. As an alternative, non-invasive Magnetic Resonance Imaging (MRI) analysis has emerged. However, it is not easy to visually identify subtle hippocampal atrophy and other early signs of Mild Cognitive Impairment (MCI). Existing Artificial Intelligence (AI)-based models also face limitations due to the enormous computational resources required to process large amounts of three-dimensional brain image data. There is a need for a high-performance algorithm that is lightweight enough to be used immediately in clinical settings and reduces the rate of misdiagnosis.

Key Findings

In a paper published in the international journal Nature, the researchers introduced AlzheiNN, a discriminative algorithm based on Convolutional Neural Networks (CNNs). This model focuses on brain regions closely associated with AD, such as the hippocampus and temporal lobe. It was trained using a large-scale clinical dataset, the Alzheimer's Disease Neuroimaging Initiative (ADNI). In classification tests, AlzheiNN accurately distinguished between cognitively normal (CN) individuals and AD patients with an accuracy of 96.8%. It also achieved an accuracy of 88.5% in distinguishing MCI patients from CN individuals, a task that is challenging for early detection. The key diagnostic indicators of sensitivity (95.2%) and specificity (97.1%) provide evidence of a significant reduction in misdiagnosis rates. This was achieved by incorporating a Spatial Attention mechanism, which differs from existing methods that analyze the entire brain image indiscriminately. Thanks to maximized computational efficiency, the time required for diagnosis is only 4.8 seconds per patient. In the analysis of an external dataset (Open Access Series of Imaging Studies, OASIS) acquired in a different environment, the accuracy decreased by less than 2%. This demonstrates its stability, as it is not affected by specific equipment or imaging conditions.

Significance and Prospects

AlzheiNN has laid the foundation for popularizing early screening for AD using MRI equipment, even in small and medium-sized hospitals with limited access to expensive equipment or advanced diagnostics. The goal is to identify patients in the early stages and proactively secure the optimal treatment window. However, the academic community points out that the precision should be verified in the future for diverse patient populations worldwide, including those with different ethnicities and age groups. Combining multi-modal information, such as genomic information or cognitive behavioral questionnaire data, is expected to further improve its screening capabilities. The establishment of legal standards for AI-based diagnosis in medical settings and the acquisition of regulatory approvals are also issues that need to be addressed. The research team explained that they plan to collaborate with various medical institutions to conduct safety evaluations based on actual patient clinical data.