Click here to

Session: Deep Learning Image Processing and Segmentation [Return to Session]

An Enhancement Method That Improves the Auto-Segmentation Performance On Abdominal Region Cone-Beam CT

J Duan1, G Huang2, X Feng2,Q Chen1*, (1) University of Kentucky, Lexington, KY, (2) Carina Medical LLC, Lexington, KY


SU-H400-IePD-F6-3 (Sunday, 7/10/2022) 4:00 PM - 4:30 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 6

Purpose: Cone-beam computed tomography (CBCT) imaging quality can be improved with deep learning. However, how much would this improvement benefits segmentation hasn’t been fully investigated. This study developed a CBCT enhancement model based on cycleGAN+U-NET neural network and evaluated its benefits on auto-segmentation.

Methods: 40 abdominal cases from a TCIA public dataset with one planning CT and two CBCT in each were used. Training-test split is 30:10. A combined model of CycleGAN and UNet was used as the network structure, including a UNet as the generator and an image classifier as the discriminator. Training losses included a masked l1-loss, GAN loss, and cycle loss. Due to the respiratory motion and anatomy changes, the pixels with over 400 HU difference between CBCT and CT were ignored in calculating l1-loss. A patch-based approach was used to tackle the memory issue. A commercial auto-segmentation software was applied to original CBCT and enhanced CBCT separately. The structures segmented were compared with manual reference contour and geometric metrics including Dice similarity coefficient (DSC), 95 percentile Harsdorf distance (HD95), and surface dice similarity coefficient (SDSC) were computed.

Results: Overall geometric metrics have been improved with CBCT enhancement. Average SDSC3mm was significantly improved by 0.13, 0.16, and 0.15 for Kindey_L, Kidney_R, and Liver respectively. The DSC was statistically improved by 0.04 for Liver. HD95 decrease of on average of 8.25mm for stomach, although not statistically significant. CBCT enhancement also improves the soft tissue HU consistency (i.e., more visible soft tissue on the same HU window level).

Conclusion: Overall, the proposed enhancement model is effective in improving the auto-segmentation performance on CBCT at the abdominal region.

Funding Support, Disclosures, and Conflict of Interest: NIH:R44CA254844 NIH:75N91020C00048 Varian Research grant:75N91020C00048 Xue Feng and Gaofeng Huang are employees of Carina Medical LLC. Quan Chen is shareholder of Carina Medical LLC


Not Applicable / None Entered.


IM/TH- image Segmentation: General (Most aspects)

Contact Email