TReND: Transformer derived features and regularized NMF for neonatal functional network delineation

Motivation: Neonatal brain functional organization, characterized by extensive immature networks, remains poorly understood. Few methods were established to accurately parcellate neonate functional networks. Goal(s): We aim to develop novel DL approach to extract reliable features, combined with regularized clustering algorithm for robust neonatal functional network delineation. Approach: We developed a novel computational framework, including transformer-based autoencoder to extract feature from BOLD signals, coupled with regularized NMF clustering algorithm to parcellate simulated and real-world neonatal fMRI. Results: TReND surpasses competing feature extraction techniques like PCA, UMAP, and TD, and outperforms clustering methods: K-PCA, ICA, and NMF, demonstrating high stability and robustness in neonate brain parcellation. Impact: We established TReND, a novel and robust framework, for neonatal functional network delineation. TReND-derived neonatal functional networks could serve as a neonatal functional atlas for perinatal populations in health and disease.