Large Blood Vessel Segmentation using Swin Transformers for Improved Quantitative DCE-MRI of Gliomas

Motivation: Quantitative DCE-MRI parameters provide a noninvasive way to evaluate tumor microvasculature, crucial for tumor assessment. However, the presence of normal large-blood-vessels(LBVs) within tumor regions can interfere with accurate tumor evaluation. Goal(s): To develop a transformer based LBV segmentation method aimed to enhance tumor grading accuracy and treatment planning in glioma patients. Approach: To generate the ground truth LBV masks, k-means clustering was applied to combined DCE-MRI parameters (CBV and Slope-2 maps), with manual corrections as needed. The performance of Swin UNETR was compared against U-Net for LBV segmentation. Results: Swin UNETR model outperformed the standard U-Net, demonstrating superior performance in LBV segmentation. Impact: The proposed transformer based LBV segmentation algorithm can aid radiologists in achieving more objective and accurate tumor assessment, potentially overcoming limitations associated with manual or semi-automatic segmentation techniques.