Automated White Matter Lesion Segmentation in Alzheimer's Disease Using SegFormer3D

• Grant number: 25/07033-9
• Support Opportunities: Scholarships abroad - Research Internship - Scientific Initiation
• Start date: July 29, 2025
• End date: November 28, 2025
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Ricardo José Ferrari
• Grantee: Yasmin Victoria Oliveira
• Supervisor: Roger Tam
• Host Institution: Centro de Ciências Exatas e de Tecnologia (CCET). Universidade Federal de São Carlos (UFSCAR). São Carlos , SP, Brazil
• Institution abroad: University of British Columbia, Vancouver (UBC), Canada
• Associated to the scholarship: 23/15916-2 - Automating the Koedam Parietal Atrophy Scale in Alzheimer's Disease Using Attributes from T1-weighted Magnetic Resonance Imaging and Clustering Techniques, BP.IC

Abstract

Abstract Alzheimer's Disease (AD) is the leading cause of dementia and is marked by progressive neurodegeneration and cognitive decline. In addition to hallmark features such as cortical atrophy and amyloid-beta deposition, white matter lesions (WMLs) have emerged as important imaging biomarkers that reflect a complex interplay between cerebrovascular damage and neurodegenerative mechanisms. WMLs are not only prevalent in individuals with AD but also strongly associated with normal and pathological brain aging, often preceding overt cognitive symptoms. Their presence and burden have been linked to accelerated cognitive decline, reduced brain connectivity, and increased risk of progression from mild cognitive impairment to dementia. Despite their clinical relevance, the accurate segmentation of WMLs in magnetic resonance imaging (MRI) remains a significant challenge. WMLs exhibit substantial heterogeneity in appearance-varying in size, shape, intensity, and anatomical location-and are further confounded by inter-subject variability, acquisition artifacts, and differences in scanner protocols, particularly in large-scale or multi-center studies. These factors make traditional and even some modern segmentation methods prone to inconsistency and limited generalizability. To address these limitations, this project proposes a fully automated segmentation pipeline for WMLs using Segformer3D, a lightweight yet powerful transformer-based architecture optimized for 3D volumetric neuroimaging data. The model is designed to deliver accurate segmentation while maintaining computational efficiency, making it suitable for deployment in both research and clinical environments. Furthermore, the segmented WMLs may be integrated with established neuroimaging markers-such as the Koedam Score for parietal atrophy, Medial Temporal Atrophy (MTA), and Global Cortical Atrophy (GCA)-to enrich the multidimensional characterization of AD pathology and brain aging. To better understand model behavior across varying lesion burdens, segmentation performance will also be stratified by lesion size, acknowledging the unique difficulties posed by smaller and more subtle lesions. By enabling robust and scalable WMLs quantification, this work aims to enhance diagnostic precision and facilitate longitudinal tracking of neurodegeneration, ultimately contributing to a more nuanced understanding of the interplay between vascular burden, aging, and AD. (AU)