Fully automated segmentation of brain and scalp blood vessels on multi-parametric magnetic resonance imaging using multi-view cascaded networks

Background and objective: Neurosurgical navigation is a critical element of brain surgery, and accurate segmentation of brain and scalp blood vessels is crucial for surgical planning and treatment. However, conventional methods for segmenting blood vessels based on statistical or thresholding techniques have limitations. In recent years, deep learning-based methods have emerged as a promising solution for blood vessel segmentation, but the segmentation of small blood vessels and scalp blood vessels remains challenging. This study aimed to explore a solution to overcoming the challenges.

Methods: This study proposes a multi-view cascaded deep learning network (MVPCNet) that combines multiple refinements, including multi-view learning, multi-parameter input, and a multi-view ensemble module. We evaluated the proposed method on a dataset of 155 patients, which included annotations for brain and scalp blood vessels. Five-fold cross-validation was conducted on the dataset to assess the performance of the network.

Results: Ablation experiments showed that the proposed refinements in MVPCNet significantly improved the segmentation of small blood and low-contrast vessel performance, which segmented scalp blood vessels from the original images, increasing the Dice and the 95 % Hausdorf distance (HD), from 0.865 to 0.922 and from 1.28 mm to 0.47 mm, respectively, compared to the baseline model.

Conclusions: The proposed method in this study provided a fully automated and accurate segmentation of brain and scalp blood vessels, which is essential for neurosurgical navigation and has potential for clinical applications.