(Sunday, 7/29/2018) 4:00 PM - 4:55 PM
Room: Karl Dean Ballroom C
Purpose: To develop and validate a recursive ensemble neural network based on 3D UNet for automatic organs-at-risks (OARs) segmentation in MR images for brain cancer radiotherapy.
Methods: Brain OARs often include eyes, brainstem, optical nerves and chiasm. We used a recursive framework to segment these OARs in two levels: (level I) the eyes and brainstem, which have regular shapes and high contrast, were segmented first; (level II) the optical nerves and chiasm, which have irregular shapes, were segmented with constraints from level-I segmentation. An ensemble network utilizing global and local features was used for each OARsegmentation. First, a global 3D UNet was constructed and trained on the entire images for OARs localization and global image feature maps (IFMs) extraction. Then, organ-specific 3D UNet were trained for organ segmentation and extract organ-specific IFMs. The final ensemble network employs a convolution layer to combine global and organ-specific IFMs to obtain refined organ segmentation. This segmenation algorithm was evaluated on 80 brain patients’ T1-weighted MR images. 60 images were used for training, and the rest 20 images for testing. The segmentation accuracy is quantitatively evaluated by the Dice’s similarity coefficient (DSC).
Results: The segmented contours match well with the physician-delineated contours, even for the optical nerves and chiasm. The ensemble segmentation achieves the results with the mean DSC of 94.9(±1.6)%, 94.9(±1.3)%, 90.3(±3.2)%, 79.6(±9.0)%, 76.9(±11.4)%, and 68.3(±11.4)% for right and left eyes, brain term, right and left optical nerves and chiasm, respectively.
Conclusion: We developed and validated a recursive ensemble neural network for brain OARs segmentation. The auto-segmentation results demonstrated the accuracy and effectiveness of the recursive ensemble segmentation approach, rendering it a potential tool for automatic OARs segmentation in brain cancer radiotherapy.
Keywords
Taxonomy
Contact Email
