Click here to


Are you sure ?

Yes, do it No, cancel

A Conditional Generative Adversarial Deep Neural Network for Automatic Segmentation of Head and Neck Structures

A Santhanam*, J Wang , B Stiehl , R Chin , M Cao , D Low , UCLA School of Medicine, Los Angeles, CA


(Wednesday, 7/17/2019) 10:00 AM - 10:30 AM

Room: Exhibit Hall | Forum 2

Purpose: Segmentation of head and neck anatomy in a planning kVCT is a critical step in the radiotherapy treatment planning. Real-time automatic and accurate segmentation of Organs at Risk (OARs) is a critical step for reducing the contouring time needed and enabling a more efficient treatment planning workflow.

Methods: We developed a machine learning platform that enables an automatic CT segmentation of all the OARs in the head and neck region. A conditional generative adversarial framework (cGAN) for machine learning was employed. The cGAN consisted of two deep neural networks where one focused on generating realistic segmentation results while the other focused on quantifying and improving the accuracy. A set of 45 previously contoured (manually) patient head and neck CT datasets were employed for the training purposes. 27 of them were used for the training, 6 were used for testing, and 12 for validation. An in-house GPU cluster was employed for the training and inferencing process. Each patient had approximately 30 contoured structures, for each of which a dedicated cGAN training was performed. The cGANs were interconnected to eliminate contour overlaps and false positive scenarios for small structures.

Results: An in-house GPU cluster consisting of 48 GPUs was employed for inferencing the segmentation. The OARs were inferenced for the test and validation datasets. The computation time for the segmentation of all OARs were up to 45 seconds. Our accuracy analysis showed that for large structures, an accuracy of greater than 93% (Dice coefficient) was observed. For small structures such as the parotid glands, an accuracy of ~77% was observed. The reduction in the accuracy was observed to correlate with the manual contour artifacts.

Conclusion: The cGAN framework demonstrated qualitatively and quantitatively good performance on segmentation of OARs.




Not Applicable / None Entered.

Contact Email