Article: Brain age estimation using cross entropy loss and sparse autoencoder features: NeuroAgeNet, a deep graph neural network Journal: International Journal of Biomedical Engineering and Technology (IJBET) 2025 Vol.49 No.4 pp.313 - 338 Abstract: MRI brain age evaluations may detect age-related neurological disorders early. Convolutional structures provide global structural information but overlook local neurological changes in standard DL models, restricting their applicability in different populations and high-dimensional neuroimaging. In NeuroAgeNet, DGNN and functional/anatomical brain interregional correlations predict brain age. SAs efficiently reduce high-dimensional input to compact feature representations, improving generality and eliminating redundancy. The regression-based aim solves class imbalance in discrete age groups and evaluates age-related patterns using a modified cross-entropy loss function. Multi-head self-attention and SAG Pool help DGNN layers filter node-level brain connection graphs and capture hierarchical dependencies. Experimental T1- and T2-weighted multi-site structural MRI using IXI dataset. NeuroAgeNet obtained 1.903 ± 0.054 MAE, 2.041 ± 0.067 RMSE, and 0.989 ± 0.003 PCC in 10 runs. Compared to baseline CNN and GNN techniques, paired t-tests significantly increased (p < 0.01). A graph-based model and sparse representation learning predict brain age better. Inderscience Publishers - linking academia, business and industry through research

Title: Brain age estimation using cross entropy loss and sparse autoencoder features: NeuroAgeNet, a deep graph neural network

Authors: L.K. Soumya Kumari; R. Sundarrajan

Addresses: Department of Computer Science and Engineering, School of Computing, Kalasalingam Academy of Research and Education, Anand Nagar, Krishnankoil, Tamil Nadu, 626126, India ' Department of Information Technology, School of Computing, Kalasalingam Academy of Research and Education, Anand Nagar, Krishnankoil, Tamil Nadu, 626126, India

Abstract: MRI brain age evaluations may detect age-related neurological disorders early. Convolutional structures provide global structural information but overlook local neurological changes in standard DL models, restricting their applicability in different populations and high-dimensional neuroimaging. In NeuroAgeNet, DGNN and functional/anatomical brain interregional correlations predict brain age. SAs efficiently reduce high-dimensional input to compact feature representations, improving generality and eliminating redundancy. The regression-based aim solves class imbalance in discrete age groups and evaluates age-related patterns using a modified cross-entropy loss function. Multi-head self-attention and SAG Pool help DGNN layers filter node-level brain connection graphs and capture hierarchical dependencies. Experimental T1- and T2-weighted multi-site structural MRI using IXI dataset. NeuroAgeNet obtained 1.903 ± 0.054 MAE, 2.041 ± 0.067 RMSE, and 0.989 ± 0.003 PCC in 10 runs. Compared to baseline CNN and GNN techniques, paired t-tests significantly increased (p < 0.01). A graph-based model and sparse representation learning predict brain age better.

Keywords: brain age estimation; deep graph neural networks; sparse autoencoder; cross-entropy loss; feature extraction; MRI analysis; neuroimaging.

DOI: 10.1504/IJBET.2025.150728

International Journal of Biomedical Engineering and Technology, 2025 Vol.49 No.4, pp.313 - 338

Received: 20 Mar 2025
Accepted: 01 Aug 2025
Published online: 22 Dec 2025 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: Brain age estimation using cross entropy loss and sparse autoencoder features: NeuroAgeNet, a deep graph neural network

Keep up-to-date