Author: John Byun, Juan J Cardona, Steven D Chang, Cynthia Fu-Yu Chuang, Xuejun Gu, Yusuke Hori, Hao Jiang, Fred Lam, Lianli Liu, Weiguo Lu, David Park, Erqi Pollom, Elham Rahimy, Deyaaldeen Abu Reesh, Scott Soltys, Gregory Szalkowski, Lei Wang 👨🔬
Affiliation: Department of Radiation Oncology, Stanford University, Department of Neurosurgery, Stanford School of Medicine, Department of Neurosurgery, Stanford University, Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, Department of Radiation Oncology, Stanford University School of Medicine 🌍
Purpose:
Intraventricular tumors pose significant challenges in neurosurgery due to their complex location. Therefore, brain SRS could be a better treatment option. At our institution, some patients treated with CyberKnife stereotactic radiosurgery (SRS) for intraventricular tumors developed hydrocephalus, necessitating shunt procedures. We hypothesize that radiation effects on the choroid plexus or ventricular ependymal cells may contribute to this complication. To investigate, in this preliminary study, we retrospectively analyzed dosimetric data from 11 patients (13 treatment courses) and explored the use of deep learning (DL)-based auto-segmentation to accurately determine ventricular wall doses, aiming to guide future planning and mitigate post-SRS hydrocephalus.
Methods:
We employed NeuralRad-Brain, an AI platform developed in-house and validated for CyberKnife SRS. The DL-based auto-segmentation algorithm, built on the nn-UNet framework, was trained using paired MRI-CT images and publicly available datasets from The Cancer Imaging Archive. Ventricle contours were validated on 10 brain SRS patients and confirmed by neurosurgeons for accuracy. For 13 intraventricular treatments, ventricles were auto-segmented, expanded by 2 mm, and analyzed for ventricular wall doses.
Results:
The average mean dose to intraventricular lesions was 23.28 Gy (range: 16.8–30.5 Gy), with an average lesion volume of 1.68 cc (range: 0.2–2.6 cc). The average maximum dose to the ventricular wall was 25.54 Gy (range: 18.2–35.4 Gy). The DL-based auto-segmentation successfully delineated ventricles and 2 mm-thick ventricular walls, enabling precise dosimetric analysis in this preliminary study.
Conclusion:
We demonstrated the accuracy and clinical utility of DL-based ventricle auto-segmentation using the NeuralRad-Brain platform. This tool facilitates the study of ventricular wall doses and supports the development of future dosing guidelines to reduce the risk of post-SRS hydrocephalus in intraventricular tumor treatments.