KGT: Knowledge-guided graph transformer for neurodegenerative disease diagnosis and brain age prediction with MRI

Deep learning methods have significantly advanced the analysis of brain imaging data for various downstream tasks such as disease diagnosis and age prediction. However, most existing methods train deep models on large amounts of imaging data, neglecting prior domain knowledge about brain structure and disease. To address this limitation, we propose KGT, a knowledge-guided graph transformer network that integrates medical domain knowledge with brain images to learn more relevant features and complex associations from regions of interest (ROIs), achieving superior performance in multiple tasks including diagnosing neurodegenerative diseases and predicting brain age. First, a convolutional autoencoder is built to extract ROI features from brain images. Then, we construct a brain ROI-oriented knowledge graph from public medical datasets, followed by a fine-tuned text encoder to generate knowledge embeddings. Next, we build a hybrid brain graph by integration of image features, spatial proximity and knowledge embeddings. Finally, a graph transformer is used to learn feature interaction and fusion from the whole brain ROI graph for disease diagnosis and age prediction. Our method is evaluated on structural MRI (sMRI) data from the Alzheimer's Disease Neuroimaging Initiative (ADNI), the Parkinson's Progression Markers Initiative (PPMI), and the UK Biobank (UKB). Experimental results demonstrate that KGT improves both feature representation and connectivity of brain ROIs, achieving superior performance in neurodegenerative disease diagnosis and brain age prediction.