Geração de tomografia cardíaca com contraste sintético a partir de imagens de tomografia cardíacas sem contraste utilizando modelos de difusão

Abstract

This dissertation proposes the use of a deep learning-based adversarial diffusion model to address the translation of contrast-free computed tomography (CT) images of the heart into synthetic images compatible with images acquired with contrast injection. The work investigates challenges in medical image translation by combining concepts from generative adversarial networks (GANs) and diffusion models. The results were evaluated through several statistics such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Frechet Inception Distance (FID), and Root Mean Square Error (RMSE) to demonstrate the model’s performance in generating images with synthetic contrast while preserving quality and visual similarity. The proposed model achieved the following best results: PSNR = 32.85, SSIM = 0.766, and FID = 42.348. CyTran obtained the best RMSE of 0.14, while the model had the worst of 0.2. A comparison of the results obtained with CyTran, Cycle-GAN, and Pix2Pix networks is also presented. Although the obtained results are promising, the analysis of RMSE indicates that there are still challenges to be overcome, highlighting the need for continuous improvements. The intersection of GANs and diffusion models promises future advancements, significantly contributing to clinical practice.