Main Session
Sep
29
PQA 05 - Physics
Presenter(s)
Darren Yang, BS - Stony Brook School of Medicine, Stony Brook, NY
D. L. Yang1,2, D. G. Hsu3, C. Gui4,5, B. S. Imber5, J. N. Stember1, and A. Holodny1; 1Department of Radiology, Memorial Sloan Kettering Cancer Center, New York, NY, 2Renaissance School of Medicine at Stony Brook University, Stony Brook, NY, 3Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, 4Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Baltimore, MD, 5Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY
Purpose/Objective(s):
Clinical target volume (CTV) delineation of glioblastomas (GBM) requires knowledge of anatomical barriers limiting infiltrative tumor spread across the parenchyma. The CTV is typically defined as an isometric expansion of the gross tumor volume (GTV), followed by manual reduction around barriers to spread. However, this process is time consuming and subject to variability, leading to unintended radiation of regions not readily accessible to microscopic spread. Automated segmentation methods can detect barriers of interest but are limited in performance with distortions due to tumor resection cavities or post-treatment effects. We investigated the accuracy of automatic segmentation tools to delineate anatomical barriers for radiation planning.Materials/Methods:
We obtained 5,559 volumetric T1 post-contrast MR scans from patients without brain tumor diagnosis at our institution and 1,064 volumetric T1 MR scans from the Human Connectome Project for model pretraining. We performed Fastsurfer reconstruction to obtain masks for the barriers of interest (falx cerebri, tentorium cerebellum, and sylvian fissure). We trained an ensemble 3D U-Net model using the nnU-Net framework to segment the barriers. We then obtained 422 volumetric T1 post-contrast MR scans with known GBM diagnosis and distortions and manually corrected barriers segmented by the model for fine-tuning. We evaluated the model’s segmentation quality with Dice similarity coefficient (DSC) and Hausdorff distance (HD) using a held-out test set of 85 manually segmented post-contrast MR scans with distortions. We also performed a case study of 6 patients to generate barrier-informed CTVs using the fast marching method initiated on the GTV. We compared these volumes to CTVs generated based on a 15 mm isometric expansion of the GTV alone.Results:
The model was able to predict the falx cerebri, tentorium cerebelli, and sylvian fissure. The median DSC for these structures in our test set was 0.95 (IQR 0.94–0.96), 0.93 (IQR 0.90–0.94), and 0.85 (IQR 0.80–0.88) respectively, and the median HD was 5 mm (IQR 3–8), 6 mm (IQR 4–7), and 8 mm (IQR 6–14) respectively. Case studies demonstrated a 12% (range 8–17) reduction in planning target volume using the barrier-informed technique compared to isometric expansion. We observed the greatest reduction with expansion across the falx cerebri, although there was significant reduction across all barriers of interest (p = 0.002).Conclusion:
Anatomic barriers can be delineated by automatic segmentation models in the setting of GBM radiotherapy. This method can be applied to other structures based on knowledge of microscopic patterns of spread, setting the groundwork for automated, barrier-informed delineation of CTVs. Ultimately, this methodology may not only improve efficiency and reproducibility of target delineation but also minimize adverse effects of radiation therapy for patients with GBM.