3763 - Synthesis of Contrast-Enhanced CT Images via Multi-Task CycleGAN Integrating Vision Transformer

Purpose/Objective(s): Contrast-enhanced CT (CECT) is pivotal in the diagnosis of tumors and the formulation of radiation therapy plans. However, iodinated contrast agents (ICAs) carry inherent safety risks, including renal impairment and severe allergic reactions in susceptible patients, as well as an additional radiation dose. This study aims to train a multi-task generation model, Multitask-VIT-CycleGAN, which incorporates Vision Transformer to generate CECT images from non-contrast enhanced CT (NECT) images, reducing the use of ICAs.

Materials/Methods: A total of 196 pairs of chest positioning NECT/CECT data from 2020 to 2022 were retrospectively collected, among which 176 cases were used for training and 20 cases for testing. We build a multi-task segmentation and generation model, Multitask-VIT-CycleGAN, which incorporates the Vision Transformer in the generator structure. The segmentation labels include artificially reviewed organs and tissues that exhibit significant contrast enhancement on CECT images, such as the aorta and heart. The quantitatively metrics: normalized mean absolute error of tissues (MAE_tissue), structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) were used to assess the quality of synthetic CECT images. In addition, we set up two sets of control experiments, plain CycleGAN and VIT-CycleGANwithout multi-task.

Results: As detailed in Table 1, compared to the plain CycleGAN, The synthetic CECT images generated by our model exhibited a higher similarity to real CECT images (MAE_tissue: 0.015, SSIM: 0.885, PSNR: 36.35 dB, P < 0.01). Incorporating the segmentation multi-task could further enhance the image generation quality within the delineated label region (MAE_tissue: 0.019 to 0.017, SSIM: 0.969 to 0.972, PSNR: 40.98 to 41.51 dB, P < 0.05 except for the P-value of PSNR, which was 0.08), which performed a center crop of size [256, 256] on the label in the XY direction and retained images within the label range.

Conclusion: The preliminary experimental results show that our proposed model significantly improves the quality of synthetic CECT images and devotes greater attention to the generation details of contrast-enhanced tissues and organs, holding potential for clinical application to reduce the use of ICAs. Subsequently, we intend to further evaluate the feasibility of using synthetic CECT as a substitute for real CECT in the target volume delineation of lung cancer and esophageal cancer, and combine specific disease types to supplement training data and specifically optimize the model.