Background

Singapore’s aging population is expected to double by 2030, contributing to a 22% rise in neurodegenerative diseases from 2013 to 2017 (1, 2). Since aging is a primary risk factor, early detection is crucial for effective intervention (3, 4). Conditions like schizophrenia also accelerate brain aging, highlighting the need for timely diagnosis (5–6).

Magnetic resonance imaging (MRI) offers detailed, non-invasive brain imaging but often relies on qualitative assessments, leading to variability and low interobserver agreement (7). Quantitative tools are needed to improve the detection of subtle brain volume changes and enhance diagnostic accuracy.

Brain volumetric analysis is challenged by structural variations influenced by age, gender, ethnicity, and geography. Existing brain templates, such as the Talairach and MNI templates, are based on Western populations and lack applicability in diverse regions like Singapore (8-9). The absence of a normative brain aging template for the Singaporean population limits diagnostic precision, underscoring the need for a population-specific brain atlas.

Advances in deep learning (DL) and machine learning (ML) automate brain volumetric analysis, improving accuracy and enabling early detection of neurodegenerative diseases. AI-driven tools can track disease progression, monitor treatment efficacy, and identify biomarkers for healthy aging (10-12).

Methods

Study Design and ParticipantsThis retrospective study analyzed 1,826 cognitively healthy Singaporean Chinese individuals (838 males, 988 females), aged 1 to 100 years, using 3D T1-weighted MRI scans from Sengkang General Hospital and the GUSTO cohort (2018–2023). Only participants with normal MRI scans and no significant brain pathologies were included.

Inclusion and Exclusion CriteriaIncluded participants had no history of neurological diseases, cognitive impairments, or conditions affecting brain structure. Exclusion criteria covered ischemic stroke, traumatic brain injury, metabolic disorders (e.g., diabetes), neurodegenerative diseases (e.g., Alzheimer’s, Parkinson’s), cranial implants, prior cranial surgery, and MRI artifacts or incomplete brain coverage.

MRI AcquisitionMRIs were acquired using a 3D T1-weighted MPRAGE sequence with 1 mm³ voxel resolution on a single scanner at SKH for consistency. Scans were aligned to the AC-PC line and independently reviewed by three board-certified radiologists to ensure quality.

Data PreprocessingDICOM files were converted to BIDS-compliant NIfTI format. Preprocessing steps included bias field correction, skull stripping, and registration to MNI space. Automated segmentation of grey matter (GM), white matter (WM), and cerebrospinal fluid (CSF) was performed using FreeSurfer.

Brain Segmentation and VolumetryAssemblyNet, a deep learning model based on U-Net architecture, was used for brain segmentation and volumetry. Segmented regions were normalized to MNI space using ANTs, generating 20 age- and gender-specific brain templates for the Singaporean Chinese population. These templates facilitated the creation of a population-specific brain volumetry database, calculating absolute volumes, intracranial-normalized percentages, and left-right asymmetry.

Statistical AnalysisGM and WM volumes were analyzed across age groups, normalized to intracranial volume. Multivariate regression models assessed age-related changes, while Student’s t-tests, Wilcoxon rank-sum tests, and Kruskal-Wallis tests examined gender differences. Statistical significance was set at p < 0.05, with analyses conducted in SPSS.

AI Model Development

A convolutional neural network (CNN) based AI model was developed to predict brain age from MRI-derived volumetric and structural features. The model was validated using 10-fold cross-validation, with performance assessed by accuracy, precision, recall, and ROC-AUC.

Ethical Approval

Ethical approval was granted by the Institutional Review Board (IRB) at Sengkang General Hospital. All MRI data were anonymized to protect patient confidentiality, adhering to institutional and regulatory guidelines for managing sensitive medical information.

Technical Implementation

The AI framework for brain segmentation and volumetric analysis was developed using a scalable architecture. The front end was built with React, and Flask was used for back-end processing. PostgreSQL managed the database for efficient storage and retrieval of volumetric data. The entire system was containerized with Docker to ensure scalability and reproducibility across various computational environments.