Anat-SFSeg: Anatomically-guided superficial fiber segmentation with point-cloud deep learning

• Published in: Medical image analysis Vol. 95; p. 103165
• Main Authors: Zhang, Di, Zong, Fangrong, Zhang, Qichen, Yue, Yunhui, Zhang, Fan, Zhao, Kun, Wang, Dawei, Wang, Pan, Zhang, Xi, Liu, Yong
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 01.07.2024
• Subjects: Anatomy; Deep learning; Diffusion MRI; Point cloud; Superficial white matter; Tractography; Deep learning; Diffusion MRI; Point cloud; Anatomy; Superficial white matter; Tractography
• ISSN: 1361-8415; 1361-8423; 1361-8423
• DOI: 10.1016/j.media.2024.103165

•We propose a novel deep learning framework Anat-SFSeg for superficial white matter fiber segmentation guided by grey matter anatomy, which performs great accuracy.•Anat-SFSeg contains a descriptor FiberAnatMap to represent both individual-level and group-level anatomical features for each streamline.•Two new metrics FARP and ARFC are proposed. They are used to quantify the proportion of fibers in anatomical brain regions and the average fiber number in each cluster, enabling the comparison of segmentation methods and assessment of inter-subject differences respectively.•Applications on Alzheimer's disease and mild cognitive impairment reveal that diffusion metrics, along with our novel metric ARFC show disorder severity associated alterations, and they are considered as potential neuroimaging biomarkers. Diffusion magnetic resonance imaging (dMRI) tractography is a critical technique to map the brain's structural connectivity. Accurate segmentation of white matter, particularly the superficial white matter (SWM), is essential for neuroscience and clinical research. However, it is challenging to segment SWM due to the short adjacent gyri connection in a U-shaped pattern. In this work, we propose an Anatomically-guided Superficial Fiber Segmentation (Anat-SFSeg) framework to improve the performance on SWM segmentation. The framework consists of a unique fiber anatomical descriptor (named FiberAnatMap) and a deep learning network based on point-cloud data. The spatial coordinates of fibers represented as point clouds, as well as the anatomical features at both the individual and group levels, are fed into a neural network. The network is trained on Human Connectome Project (HCP) datasets and tested on the subjects with a range of cognitive impairment levels. One new metric named fiber anatomical region proportion (FARP), quantifies the ratio of fibers in the defined brain regions and enables the comparison with other methods. Another metric named anatomical region fiber count (ARFC), represents the average fiber number in each cluster for the assessment of inter-subject differences. The experimental results demonstrate that Anat-SFSeg achieves the highest accuracy on HCP datasets and exhibits great generalization on clinical datasets. Diffusion tensor metrics and ARFC show disorder severity associated alterations in patients with Alzheimer's disease (AD) and mild cognitive impairments (MCI). Correlations with cognitive grades show that these metrics are potential neuroimaging biomarkers for AD. Furthermore, Anat-SFSeg could be utilized to explore other neurodegenerative, neurodevelopmental or psychiatric disorders. [Display omitted]