Click here to

Session: AI in Imaging [Return to Session]

Hybrid Adversarial Network for Ultra-Quality Pulmonary Anatomy Imaging From Cone-Beam CT Images

J Zhu1, W Chen2, Y Huang1*,A Nicol3, Y Lam4, J Cai5, G Ren6, (1) Hong Kong Polytechnic University, Hong Kong (2) Chinese Academy of Sciences, Shenzhen Advanced Technology Academe, Shenzhen,CN(3) The Hong Kong Polytechnic University, Hong Kong(4) Hong Kong Polytechnic University, Hong Kong(5) Hong Kong Polytechnic University, Hong Kong (6) Hong Kong Polytechnic University, Hong Kong


TH-D-207-5 (Thursday, 7/14/2022) 11:00 AM - 12:00 PM [Eastern Time (GMT-4)]

Room 207

Purpose: Pulmonary cone-beam computed tomography (CBCT) suffers from severe scattering artifacts and low soft-tissue contrast, and conventional global enhancement techniques fail to extract the detailed textural information for further analysis. This study aims to develop a dual-task hybrid adversarial neural network (DAN), to reduce artifacts and enhance the textural details on lung CBCT images.

Methods: Thoracic CBCT from 100 patients with pulmonary diseases were retrospectively collected from Queen Mary Hospital. The pCTs were collected one week before RT with fixed positioning. For each patient, 3D rigid registration was applied from CBCT to pCT first, then left/right lung was segmented from registered CBCT and pCT pairs. The image pairs were then divided into 14,210 2D slices in the transverse plane (9947 for training;4263 for testing). These images were processed by a deep learning-based framework to translate CBCT to enhanced synthesized CT (sCT), based on the corresponding pCT. The framework was built in a dual-task hybrid form, consisting of a dual-stream pyramid network (DSPN), and a texture-aware enhancement (TAEN) network. The structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), and mean absolute error (MAE) between the pCT and the corresponding sCT were computed to assess the statistical enhancement performance. Global and local HU value distributions and difference maps were used to assess perceptual image enhancement performance.

Results: For statistical slice-wise agreement, the global evaluation shows a high SSIM (0.9217 ±0.0231), high PSNR (30.6429±2.1246), and low MAE (16.2780±0.0746) between the pCT images and sCT images. For visual enhancement evaluation, sCTs show clear reductions in artifacts and prominent improvements in image quality.

Conclusion: We developed a DL-DAN model to achieve ultra-quality pulmonary anatomy imaging from CBCT images for patients with lung diseases. This method holds great promise to provide improved detailed information for the CBCT based lung structure analysis.

Funding Support, Disclosures, and Conflict of Interest: Start-up Fund for RAPs under the Strategic Hiring Scheme (P0038378); Seed Fund for External Research Grant Applications (SF2122-GR)


Lung, Cone-beam CT, Modeling


IM/TH- Cone Beam CT: Machine learning, computer vision

Contact Email