Click here to

Session: Deep Learning for Image-guided Therapy [Return to Session]

Deep-Learning Method for Segmenting Head and Neck Tumours in KV Images Acquired During Radiotherapy

M Gardner1*, A Mylonas1, M Mueller1, Y Ben Bouchta1, J Sykes2, P Keall1, D Nguyen7, (1) University of Sydney, Sydney, NSW, AU (2) Blacktown Hospital, Blacktown, AU,(3) University of Technology Sydney, Ultimo, NSW


WE-C1030-IePD-F2-1 (Wednesday, 7/13/2022) 10:30 AM - 11:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 2

Purpose: This study investigated the feasibility of automatic segmentation of head and neck (H&N) tumors in kV images using simulated projections. Success will lead to the ability to track tumor motion during H&N radiotherapy with the goal of removing the need for the thermoplastic mask for radiotherapy.

Methods: Patient data was derived from CT scans and contoured Gross Tumor Volume (GTV) volumes of 15 patients from the HNSCC database from the Cancer Imaging Archive. A patient-specific conditional Generative Adversarial Network (cGAN) was trained to detect and segment the tumor location in each Digitally Reconstructed Radiograph (DRR). To create the training data, the planning CT was deformed 10 different ways using a novel method for creating synthetic realistic H&N deformations. From each of these 11 CT volumes and their corresponding GTV volumes, 3600 H&N DRRs were generated (0.1° projection angle difference), with a total of 39600 training DRRs used. To create the testing data, the planning CT and contoured GTV were again deformed using a deformation of a different magnitude and were used to create 360 testing DRRs (1° projection angle difference). The estimated tumor location from the cGAN was compared with the ground truth for each patient by measuring the centroid error, the DICE similarity coefficient (DSC) and the Mean Surface Distance (MSD).

Results: For all 15 patients the magnitude of the mean ± standard deviation cGAN segmentation centroid error was 1.1 ± 0.6 mm. The mean DSC and MSD values were 0.90 ± 0.03 and 1.7 ± 0.6mm respectively.

Conclusion: These results demonstrate the ability to detect and segment H&N GTVs in DRRs. A novel method for creating H&N deformations was also introduced. Further development of this technology will pave the way for the removal of the need for thermoplastic masks for treating head and neck cancers during radiotherapy.

Funding Support, Disclosures, and Conflict of Interest: This project was funded by Cancer Australia, funded by the Australian Government


Segmentation, Patient Movement, Target Localization


Not Applicable / None Entered.

Contact Email