Brain Age Prediction Using a Lightweight Convolutional Neural Network

• Main Author: Eltashani, Fatma
• Publication Date: 2025
• Other Authors: Parreno-Centeno, Mario, Cole, James H., Paulo Papa, Joao [UNESP], Costen, Fumie
• Format: Article
• Language: eng
• Source: Repositório Institucional da UNESP
• Download full: http://dx.doi.org/10.1109/ACCESS.2025.3526520; https://hdl.handle.net/11449/305560
• topic: Biological age estimation; brain ageing; brain imaging; convolutional neural network; deep learning; magnetic resonance imaging

Much interest has recently been drawn to brain age prediction due to the significant development in machine learning and image processing techniques. Studies based on brain magnetic resonance images showed a strong relationship between the brain ageing process and accelerated brain atrophy, suggesting using brain age prediction models for early diagnosis of neurodegenerative disorders, such as Parkinson's, Schizophrenia, and Alzheimer's disease. However, data availability, acquisition protocols diversity and models' computational complexity remain limiting factors for clinical adoption. This study proposes a low-complexity convolutional neural network (CNN) model that tackles these challenges, focusing on three main aspects: performance accuracy, computational complexity, and adaptability to new, external datasets. We developed a brain-age prediction system using a minimally preprocessed T1-weighted MRI images with a multi-site dataset of healthy individuals covering the whole human lifespan (2251 subjects, age range 6-90 years). We proposed a lighter version of the Simple Fully Convolutional Network (SFCN) that contain only 1.2 million parameters. Computational load was further reduced by cropping the brain images. Finally, we employed transfer learning approach to achieve domain adaptation to external, unseen sites. We demonstrated that leveraging the cropped brain images reduced the computational time for training by 50%, maintaining a comparable accuracy to using the entire brain. The model achieved a Mean Absolute Error (MAE) of 3.557 for the full brain and 4.139 for the cropped images with a Pearson correlation r = 0.988 $ between the full and cropped brain predictions when evaluated on the same test set. Domain adaptation of our model to new external data showed a significant improvement in the prediction performance, reducing MAE from 7.219 to 4.750 for full brain images and from 12.107 to 5.770 for the cropped images. This study is the first to demonstrate comparable prediction accuracy using only a small segment of a 3D full brain MRI scan. Our results show that it is feasible to build lightweight CNN models trained on small-scale, heterogeneous datasets and fine-tuned to new external clinical data, making significant steps toward practical clinical application.