Author: Evan Calabrese, Edward Robert Criscuolo, Deshan Yang 👨🔬
Affiliation: Duke University, Department of Radiation Oncology, Duke University 🌍
Purpose: Glioblastoma (GBM) is the most common and aggressive form of brain cancer. Deformable image registration (DIR) is a powerful tool to compute anatomical changes in longitudinal MRI scans, which can detect recurrence and assess treatment response. However, large deformations in postoperative GBM MRIs make DIR difficult. Blood vessel tracking can be used for DIR verification and guidance, but ground truth data for vessel segmentation models is lacking. Therefore, we developed a novel procedure for efficiently simulating biologically accurate blood vessels on T1-weighted post-contrast (T1-C) MRIs to train a segmentation model. This model will allow vessel tracking and DIR verification between GBM scans.
Methods: Thirty-seven T1-C and T1 MRI images from GBM patients were processed. Blood vessels were simulated in between the gyri of each case. Artificial T1-C images were created by rewriting the intensity of the T1 image corresponding to the simulated vessels. After noise texture addition and intensity adjustment, the real T1 image was subtracted from the artificial T1-C image to get a simulated subtraction image. This and the ground truth vessel simulation mask were used to train an nnUNet model to segment the vessels embedded in the brain. A separate nnUNet model was trained on manually annotated data to segment the dural venous sinuses.
Results: On the test data, preliminary results of small vessel model had a dice score of 0.78, while the venous sinus model had a dice score of 0.80. The simulated vessel data was empirically realistic and the trained model had high sensitivity.
Conclusion: With a limited training dataset, we achieved high vessel segmentation accuracy. To our knowledge, this is the first deep model to use T1-C subtraction images to segment blood vessels in brain MRIs, increasing accuracy while remaining highly generalizable. This will support highly accurate and verifiable DIR for GBM MRIs.