Structure-preserving quality improvement of cone beam CT images using contrastive learning. (May 2023)

• Record Type: Journal Article
• Authors: Kang, Se-Ryong; Shin, Woncheol; Yang, Su; Kim, Jo-Eun; Huh, Kyung-Hoe; Lee, Sam-Sun; Heo, Min-Suk; Yi, Won-Jin
• Is Part Of: Computers in biology and medicine. Volume 158(2023)
• Publication Date: 2023-05
• Subjects: Contrastive learning -- Structure-preserving -- Cone-beam CT images -- Deep learning; Medicine -- Data processing -- Periodicals; Biology -- Data processing -- Periodicals; 610.285
• DOI: 10.1016/j.compbiomed.2023.106803
• Languages: English
• ISSNs: 0010-4825

Abstract

Abstract: Cone-beam CT (CBCT) is widely used in dental clinics but exhibits limitations in assessing soft tissue pathology because of its lack of contrast resolution and low Hounsfield Units (HU) quantification accuracy. We aimed to increase the image quality and HU accuracy of CBCTs while preserving anatomical structures. We generated CT-like images from CBCT images using a patchwise contrastive learning-based GAN model. Our model was trained on unpaired CT and CBCT datasets with the novel combination of losses and the feature extractor pretrained on our training dataset. We evaluated the quality of the images generated by our model in terms of Fréchet inception distance (FID), peak signal-to-noise ratio (PSNR), mean absolute error (MAE), and root mean square error (RMSE). Additionally, the structure preservation performance was assessed by the structure score. As a result, the generated CT-like images by our model were significantly superior to those generated by various baseline models in terms of FID, PSNR, MAE, RMSE, and structure score. Therefore, we demonstrated that our model provided the complementary benefits of preserving the anatomical structures of the input CBCT images and improving the image quality to be similar to those of CT images. Highlights: For the first time, we applied contrastive learning for unpaired image translation to the quality improvement of CBCT images. The proposed loss combination improved CBCT image quality while preserving anatomical structures. Our method demonstrated improved performance in terms of FID, PSNR, MAE, and RMSE. Our method could accurately quantify the HU of CBCT images.