Improving Cone-Beam CT Auto-Segmentation Accuracy Using Cycle GANs for Domain Adaptation
K Shah1*, J Shackleford1, N Kandasamy1, G Sharp2, (1) Drexel University, Philadelphia, PA, (2)Massachusetts General Hospital, Boston, MA
Presentations
SU-F-201-2 (Sunday, 7/10/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Room 201
Purpose: Cone-beam CT (CBCT) images are used in adaptive radiation therapy to re-optimize treatments for changing anatomy, which requires segmentation of the target and organs at risk (OARs). OAR segmentation is both challenging and time-consuming, due to the poor soft tissue contrast of the CBCT image. Auto-segmentation using deep learning techniques requires extensive amount of labeled data, which is hard to collect for CBCT images. This work evaluates the application of domain adaptation using a cycle GAN architecture to map CT images to CBCT images, allowing segmented CT images to be used for training data.
Methods: The approach was evaluated using a publicly available pelvic reference dataset. The dataset consists of CT and CBCT images of 58 patients. 12 patients with an intact prostate were selected for this study. The prostate was manually segmented on the CT and CBCT images by a medical expert. A cycle GAN architecture was trained to map the CT images to CBCT images using the remaining 46 patients. The synthetic images generated from cycle GAN (sCBCT) along with the CBCT images were used to train a U-Net architecture to automatically segment the prostate from the CBCT images using 10 patients. The two patients were used for independent validation of the network.
Results: Accuracy of the U-Net was quantified using the Dice Similarity Coefficient (DSC), with and without the use of the cycle GAN generated sCBCT images. Results demonstrate much improved segmentation accuracy with the sCBCT images, with the average DSC of the prostate increased from 0.74 to 0.82.
Conclusion: Domain adaptation using cycle GAN architecture was found to increase the U-Net segmentation accuracy by increasing the availability of labeled training data. This allows for a significant improvement in terms of auto-segmentation accuracy.
Funding Support, Disclosures, and Conflict of Interest: This material is based upon work supported by the National Science Foundation under Grant Nos. 1553436, 1642345, and 1642380 and the National Institutes of Health under NCI R01CA229178.
Keywords
Cone-beam CT, Segmentation, Computer Vision
Taxonomy
IM- Cone Beam CT: Machine learning, computer vision
Contact Email