Indiv-Brain: Individualized Brain Network Partition Learned from Individual fMRI Data using Deep Clustering with Vertex-level Attention

Individualized functional network partitioning is increasingly critical in elucidating individual differences in cognition, development, and behavior. The previous studies grouped brain regions that were pre-defined by common brain parcellations into individualized brain networks. The employment of common brain region parcellations ignores the individual variations in brain structure and reduces the spatial resolution of the results. Moreover, some studies trained models on a group of subjects to guide individual network partitioning. These methods largely depend on the sample size and encounter challenges in the case of limited subject data. In this paper, we propose Indiv-Brain which automatically partition brain vertices into different brain networks based on individual fMRI data without any prior brain parcellation. The Indiv-Brain consists of three sequential modules: stacked denoising autoencoder (SDAE) for mapping the raw fMRI data to a latent embedding space, masked vertex modeling (MVM) for learning attention-enhanced representations of brain vertices, and deep embedding clustering with spatial attention (DEC-A) for unsupervised clustering on the learned vertex representations. The experiments on Human Connectome Project (HCP) demonstrate that the accuracy of Indiv-Brain outperforms existing methods. We compared the model with methods like SDAE, DEC, IDEC, and k-means. For the results of 8 subjects, the accuracy of Indiv-Brain was consistently the highest, averaging 6.65 percentage points higher than the IDEC method. To the best of our knowledge, this is the first study to obtain individualized brain network partition based on individual fMRI data with deep learning models. Our study provides a novel insight into understanding individualized brain networks, especially suitable for special patients.