COSTA: A Multi-Center TOF-MRA Dataset and a Style Self-Consistency Network for Cerebrovascular Segmentation.

Area of Science:

• Medical Imaging
• Neuroscience
• Computer Vision

Background:

• Time-of-flight magnetic resonance angiography (TOF-MRA) is a non-invasive, radiation-free method for cerebrovascular imaging.
• Significant challenges exist in accurate cerebrovascular segmentation due to imaging artifacts, inter-site/vendor heterogeneity, and limited annotated data.

Purpose of the Study:

• To develop a robust cerebrovascular segmentation method for TOF-MRA data.
• To address the challenges of feature granularity and image style heterogeneity in TOF-MRA segmentation.
• To create and release the largest and most diverse TOF-MRA dataset (COSTA) for research.

Main Methods:

• Construction of the COSTA dataset, comprising manually annotated TOF-MRA volumes from 8 imaging centers.
• Proposal of the CESAR network featuring a coarse-to-fine architecture for iterative segmentation refinement.
• Implementation of an automatic feature selection module for fusing global and local cerebrovascular features.
• Introduction of a style self-consistency loss to standardize diverse TOF-MRA image styles.

Main Results:

• The CESAR network demonstrated superior performance in cerebrovascular segmentation compared to state-of-the-art methods on the COSTA dataset.
• The proposed methods effectively tackled feature granularity and image style heterogeneity.
• The COSTA dataset and CESAR source code were made publicly available to foster research.

Conclusions:

• The CESAR network offers an effective solution for accurate and robust cerebrovascular segmentation in TOF-MRA.
• The COSTA dataset provides a valuable resource for advancing research in TOF-MRA segmentation.
• The study highlights the importance of addressing data heterogeneity for reliable medical image analysis.