Conditional Diffusion Model for Abdominal CT Image Synthesis

Deep neural networks have revolutionized medical image analysis, with Generative Adversarial Networks (GANs) being prominent in synthesizing realistic images. However, GANs face challenges like mode collapse. As an alternative, Denoising Diffusion Probabilistic Models (DDPM) have emerged, showing promise in generating high-fidelity images, though their application to medical imaging is still nascent. In this study, we propose using DDPM to generate realistic computed tomography (CT) images guided by segmentation labels to ensure anatomical precision. We performed data preprocessing and downsampling on CT images from the TotalSegmentator dataset, trained our DDPM-based model, and subsequently applied super-resolution techniques to produce high-resolution images. Evaluation metrics such as Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR) confirm the effectiveness of our approach in generating CT images that meet clinical standards while maintaining visual fidelity. We demonstrate that the generated images can augment training datasets for deep learning-based medical image segmentation models, thereby enhancing segmentation accuracy. Additionally, these synthetic images hold significant potential for applications in medical education.