Abstract
Brain age prediction based on anatomical MRI scans, as an essentially new measure in neuroimaging and aging research, provides a crucial marker for the early diagnosis of neurodegenerative diseases, cognitive health appraisal, and biological age prediction. Conventional machine learning models rely on handcrafted features, which can result in low accuracy and generalizability because they fail to capture the complex spatial, contextual, and structural information inherent in MRI images. While deep learning methods like CNNs and Transformers enhance feature extraction, they fail to adequately capture the brain's structural connectivity patterns, leading to more significant prediction errors and lower reliability. To address these limitations, this study introduces NeuroAgeFusionNet: A hybrid deep learning framework leveraging CNNs, Transformers, and Graph Neural Networks (GNNs) to improve brain age estimation. The proposed framework uses a feature fusion mechanism with a hybrid modeling approach that optimizes spatial, contextual, and structural features for more comprehensive feature representation. Moreover, an uncertainty quantification module is built into the model to make predictions more robust by safeguarding them against unreliable estimates. On the UK Biobank dataset, our model achieves state-of-the-art performance with an MAE of 2.30, a Pearson correlation of 0.97, and an R(2) score of 0.96, significantly surpassing conventional approaches. A high-level abstract of the brain age estimation framework, which shows excellent potential for accuracy, low variance, and intelligible characteristics. Such advancements position NeuroAgeFusionNet as a valuable tool for clinical neuroscience applications that facilitate improved brain aging monitoring and early neurodegenerative disease detection.