Abstract
PURPOSE: This study sought to develop an nnTransUNet model for synthesizing virtual contrast-enhanced CT from non-contrast CT, and to evaluate its feasibility for target volume delineation in cervical cancer radiotherapy by comparison with conventional non-contrast and contrast-enhanced CT. METHODS: A total of 210 patients with cervical cancer who underwent noncontrast CT and contrast-enhanced CT scans before and after intravenous administration of iodine contrast agent were selected. The "nnTransUNet" network architecture was used to convert noncontrast CT images into virtual contrast-enhanced CT images. Noncontrast, enhanced and virtual contrast-enhanced CT images were designed for cervical cancer radiotherapy, and their image similarity measurements, supervisor image quality evaluations, CT value distributions and dosimetric evaluations were compared. RESULTS: The virtual contrast-enhanced CT images achieved scores of 0.958, 48.332, 0.997, and 0.976 in terms of mean squared error (MSE), peak signal-to-noise ratio (PSNR), universal quality index (UQI), and structural similarity index (SSIM), respectively. The Dice similarity coefficient (DSC) and Hausdorff distance (HD) for tumor delineation were 0.95 and<7.1 mm. In the subjective evaluation, the virtual contrast-enhanced CT images achieved a score of 4 points in terms of artefacts, noise, image structural integrity and image distortion, which was consistent with the scores for contrast-enhanced CT images. In terms of anatomical structure clarity, the score was slightly lower than that of the contrast-enhanced CT image (3.7 points vs. 4 points). The CT values of the virtual contrast-enhanced CT images were close to those of the contrast-enhanced CT images, and the CT values of the blood vessels and bone marrow were much greater than those of the noncontrast CT images. Compared with that of contrast-enhanced CT, the dose matching between virtual contrast-enhanced CT and noncontrast CT images was closer, and the relative dose difference in the target area was less than 2%. No significant difference in the organs at risk (OARs) dose distribution between the virtual contrast-enhanced CT images and noncontrast CT images. CONCLUSIONS: We developed a deep learning model based on the nnTransUNet architecture for generating virtual contrast-enhanced CT images from non-contrast CT scans, and validated its feasibility in terms of image quality assessment, radiotherapy dose calculation, and target volume delineation for cervical cancer.