Differentiating hemorrhage from tumor vasculature in SWI-MRI images using Swin UNETR: Applications in Glioma Grading

Motivation: Glioma grading using intra-tumoral-susceptibility-signal (ITSS) from SWI is affected by the presence of hemorrhage within the tumor region. Separation of ITSS vasculature (IV) from hemorrhage is challenging. Goal(s): To develop a deep learning-based method for segmenting IV from SWI. Approach: Swin UNETR model was trained for separation between IV and hemorrhage components of ITSS tissues on SWI-MRI. Data of 214 Glioma patients was used. The potential of IV in glioma grading was also evaluated. Results: Swin UNETR provided a dice score of 0.910±0.060 for IV segmentation. The predicted and the ground truth masks showed similar glioma grading accuracies (P<0.001). Impact: This study presents a method for automatically segmenting IV from SWI images using Swin UNETR without multi-echo SWI or R2* maps, enhancing efficiency in resource-limited settings, showing high accuracy in classifying gliomas. It addresses the subjectivity inherent in manual methods.