Author: Ali Ajdari, Thomas R. Bortfeld, Christopher Bridge, Gregory Buti, Marcela Giovenco, Fredrik Lofman, Gregory C. Sharp, Helen A Shih, Tugba Yilmaz 👨🔬
Affiliation: Massachusetts General Hospital, RaySearch Laboratories, Department Of Radiation Oncology, Massachusetts General Hospital (MGH), Massachusetts General Hospital & Harvard Medical School, Massachusetts General Hospital and Harvard Medical School 🌍
Purpose: Defining radiation target volumes with accurate integration of the neuroanatomy is one of the major difficulties in designing glioma treatments. We developed a deep learning network for normal brain anatomy that learns neural connections between brain substructures that may serve as tumor infiltration pathways. The predictions are applied to an automatic workflow that defines the gross tumor volume (GTV) to clinical target volume (CTV) expansion.
Methods: Two radiation oncologists delineated six brain substructures (hemispheres, brainstem, cerebellum, chiasm, optic nerves, ventricles, and midline) for 42 glioma patients. Expert knowledge of white matter tracts was incorporated into the delineations by overlapping structures at the location of the neuronal connections (e.g. cerebellum and brainstem fiber crossing at the cerebellar peduncles). NnU-Net was used for model training (34 patients for training, 8 for testing) with CT image as input and segmentation map as output. Anatomy-informed CTVs were generated for all patients by expanding manually-contoured GTVs by 15 mm using a shortest-path solver that constrains the expansion so that it cannot leak outside the brain, using predicted brain substructures as no-flux boundary conditions. Ground-truth CTVs were defined as a 15 mm expansion of the GTV adjusted for manually-contoured brain substructures.
Results: Mean (±std) Dice Similarity Coefficient (DSC) of the test set for the hemispheres, brainstem, cerebellum, chiasm, optic nerves, and ventricles were (98.1±0.6)%, (91.4±1.8)%, (95.0±2.4)% (64.9±25.9)%, (72.9±6.7)%, and (88.9±4.3)%. Mean 95% Hausdorff distance (HD95) for those structures were, in mm, 2.0±0.5, 4.1±3.3, 2.3±1.0, 3.2±1.6, 7.6±13.1, and 2.9±2.0. The automatic CTVs generated showed excellent surface similarity to the ground truth CTVs with mean (std) Surface DSC with 2mm tolerance and HD95 scores of (93.0±3.6)% and (2.6±1.1)mm, respectively.
Conclusion: We have successfully developed an automated workflow for radiation target volume definition that incorporates the prior knowledge of neuronal connections that serve as pathways for tumor infiltration.