Vessel Segmentation in χ-separation using Deep Learning

Motivation: A recently proposed region-growing algorithm-based vessel segmentation method is sensitive to hyperparameters, leading to user-dependent results that can be inconsistent and inconvenient. Goal(s): We develop a deep learning method that generates high-quality vessel masks of χ-separation maps without hyperparameter tuning. Approach: Vessel masks generated by a region-growing algorithm and manual correction for errors were used as training labels. A deep neural network was trained which takes χ-separation maps as input and generates vessel masks. This model was evaluated on 3T and 7T data with different resolutions. Results: The model successfully produced high-quality vessel masks for both 3T and 7T χ-separation maps. Impact: The proposed deep neural network can produce high-quality vessel masks, eliminating the need for hyperparameter tuning in the original region-growing method. The result may improve the accuracy and efficiency of χ-separation map analysis, supporting precise susceptibility studies.