A Lightweight 2D Projection CNN Framework for MRI-Based Brain Age Prediction

Brain age is an important biomarker that quantifies age-related structural changes in the human brain, with the potential for early disease diagnosis and monitoring of healthy aging. However, conventional three-dimensional (3D) convolutional neural networks require substantial computational resources to achieve high accuracy. In this study, we propose a computationally efficient deep learning model based on two-dimensional (2D) projections that balances efficiency and accuracy. We integrated eight publicly available datasets comprising 7,649 healthy participants aged 5-89 years, using T1-weighted magnetic resonance imaging. In addition to the gray matter probability maps, we incorporated full-brain structural information and multiple projection statistics. These statistics include the mean, standard deviation, median, and maximum to capture the comprehensive morphological features. The proposed architecture comprises only three convolutional blocks with 414,785 parameters and an 86% reduction compared to the similarly high-performance simple fully convolutional network (SFCN). To mitigate systematic prediction biases, we implemented age-distribution-weighted training. The experimental results indicated that single-plane models achieved a mean absolute error (MAE) of approximately 2.7-2.8 years, whereas a three-plane ensemble reduced the error to 2.50 years. After bias correction, the error was 2.54 years, effectively mitigating age-related bias while maintaining accuracy. However, prediction reliability in the middle-aged subgroup remains limited owing to data scarcity, with MAE reaching 7.48 years in the 40-49 age range. The model outperformed the existing 2D projection methods with extremely low computational complexity, requiring only approximately 1.5 h of training. This training time is nearly two orders of magnitude faster than that of the 3D approaches. Furthermore, gradient-weighted class activation mapping visualizations confirmed biological plausibility, highlighting aging-related regions such as the ventricles, cortex, and hippocampus.