Author: John Byun, Steven D Chang, Cynthia Fu-Yu Chuang, Xuejun Gu, Melanie Hayden Gephart, Yusuke Hori, Fred Lam, Gordon Li, Lianli Liu, Weiguo Lu, David Park, Erqi Pollom, Elham Rahimy, Deyaaldeen Abu Reesh, Scott Soltys, Gregory Szalkowski, Lei Wang, Xianghua Ye, Kangning Zhang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, Department of Neurosurgery, Stanford University, Department of Radiation Oncology, Stanford University, Department of Radiation Oncology, Stanford University School of Medicine 🌍
Purpose: Accurate and automated delineation of vestibular schwannoma (VS) volume is crucial for disease management, as both treatment approaches (stereotactic radiosurgery and invasive surgery) and monitoring approaches (observation and periodic assessment) require precise volume measurement. To address this need, we developed a deep learning (DL)-based VS auto-segmentation module and integrated it into our web-based NeuralRad-Brain for clinical use.
Methods: The DL-based VS auto-segmentation module was built on the nn-UNet framework and trained using paired MRI images and structure datasets of VS patients, publicly available on The Cancer Imaging Archive. We validated the trained VS auto-segmentation module on 90 intact VS patients who underwent radiosurgery treatment at our institution. To facilitate the validation process, we developed scripts to automatically handle the data import and segmentation output evaluation. The performance of VS auto-segmentation model was quantitatively evaluated by comparing model segmentation and clinicians’ manual contour with four metrics including DICE coefficient (DC), mean surface-to-surface distance (mSSD), 95 percentile Hausdorff distance 95 (HD95), and center-of-mass distance (COMD).
Results: Quantitative evaluations showed the model was able to accurately auto-delineate most VS lesions, with DC of 0.84±0.08, HD95 of 0.95±0.45 mm (~1-2 voxels on MR images), mSSD of 0.22±0.18 mm and COMD of 0.64±0.37 mm. Segmentation accuracy was lower for lesions with a small volume (<0.5 cc), extensive infiltration into the auditory canal, large cystic components, or multiple foci. The entire process, from DICOM import to final DICOM RT structure output, took approximately 2 minutes per case.
Conclusion: The high accuracy and efficiency of these auto-segmentation results suggest that the DL-based VS module in the NeuralRad-Brain platform can effectively assist clinical workflows and is feasible to clinically adoption for VS disease management whether for treatment or monitoring.