H Wang*, T Qu, K Bernstein, D Barbee, D Kondziolka, NYU Langone Medical Center, New York, NY

• 
  H Wang

PO-GePV-M-203 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Segmentation of vestibular schwannomas (VS) for treatment planning and long-term follow-up assessment is labor intensive. This study aims to develop a deep learning (DL) method to automatically segment VS from MRI acquired in radiosurgery for VS.

Methods: This study enrolled 450 patients who have received Gamma Knife Radiosurgery for VS. Treatment planning T1-weighted isotropic MRI and corresponding tumor contours were used for model development. The DL model was a 3D convolutional neural network (U-Net) built on ResNet blocks. Spatial attenuation module was integrated in each decoder layer and explicitly supervised to focus on small target volumes. The model was implemented using MONAI framework and trained with data augmentation. Performance of the model for VS segmentation was assessed against the GTVs manually contoured for treatment planning.

Results: The dataset was split into 300 patients for training, 50 for validation and 100 for testing. Measured on the test data, the automated method achieved average Dice Similarity Coefficient (DSC) 0.89 ± 0.14, and 95% Hausdorff distance of 1.78 ± 1.87 mm.

Conclusion: The DL method has shown high accuracy in segmentation of VS on MRI. The method will greatly facilitate volumetric measurements of VS longitudinal assessment and aid in management of patients with VS.

Keywords

Segmentation, MRI

Taxonomy

IM- MRI : Machine learning, computer vision

Contact Email