上海科技大学知识管理系统

Accurate brain tissue segmentation is a vital prerequisite for charting infant brain development and for diagnosing early brain disorders. However, due to inherently ongoing myelination and maturation, the intensity distributions of gray matter (GM) and white matter (WM) on T1-weighted (T1w) data undergo substantial variations in intensity from neonatal to 24 months. Especially at the ages around 6 months, the intensity distributions of GM and WM are highly overlapped. These physiological phenomena pose great challenges for automatic infant brain tissue segmentation, even for expert radiologists. To address these issues, in this study, we present a unified infant brain tissue segmentation (UinTSeg) framework to accurately segment brain tissues of infants aged 0-24 months using a single model. UinTSeg comprises two stages: 1) boundary extraction and 2) tissue segmentation. In the first stage, to alleviate the difficulty of tissue segmentation caused by variations in intensity, we extract the intensity-invariant tissue boundaries from T1w data driven by edge maps extracted from the Sobel filter. In the second stage, the Sobel edge maps and extracted boundaries of GM, WM, and cerebrospinal fluid (CSF) are utilized as intensity-invariant anatomy information to ensure unified and accurate tissue segmentation in infants age period of 0-24 months. Both stages are built upon an attentionbased surrounding-aware segmentation network (ASNet), which exploits the contextual information from multi-scale patches to improve the segmentation performance. Extensive experiments on the baby connectome project dataset demonstrate the superiority of our proposed framework over five state-of-the-art methods.