MSCA: A Universal Plug-and-Play Multi-Scale Co-Attention Framework for Brain Age Estimation from sMRI

Structural Magnetic Resonance Imaging(sMRI) data has been used to estimate brain age. The gap between the estimated age and chronological age is a potential biomarker in neurodegenerative disorder. Developing a highly accurate brain age estimation model is crucial for clinical diagnosis. The challenging issue is how to integrate multi-scale three-dimensional features. We developed a universal plug-and-play Multi-Scale Co-Attention framework (MSCA) inspired by the co-attention mechanism. This method contains a multi-scale splitting module, a feature extraction module, and a feature fusion module. The multi-scale splitting module divides sMRI data into multiple three-dimensional local patches and rescales the whole data to a global patch. The feature extraction module extracts the features from the patches by convolutional neural networks and Tran-Resnet as backbone networks. Finally, the feature fusion module fuses the multi-scale features to estimate brain age. Among three different backbone networks, our method demonstrated a lower mean absolute error (MAE) in age estimation after training on a public dataset comprising 10,934 individuals aged 6-90. Additionally, we employed a block occlusion technique to visualize important brain regions used to make the estimation.