CT cardiac image and label generation by a 3D latent diffusion model

Computed Tomography (CT) imaging continues to evolve rapidly. This technique provides detailed views of internal structures, aiding in disease detection, radiation therapy planning, tumor monitoring, and emergency diagnostics. CT cardiac imaging is particularly challenging due to the heart's constant motion, requiring rapid snapshots to avoid blurred images. Variability in heart rates, contrast timing, radiation exposure, and artifacts from pacemakers or calcifications add to the complexity. Organ segmentation in medical imaging, especially CT cardiac image segmentation, is crucial for identifying and classifying heart structures. This segmentation supports applications like 3D visualization, arrhythmia detection, and surgical planning. Challenges include the heart's complex anatomy, size variability, and dynamic movements. Latent Diffusion Models (LDMs) in Deep Learning generate new image data through a stochastic diffusion process in latent space, offering GPU memory efficiency and high-quality image generation. LDMs are used in computer vision and medical imaging for tasks like image synthesis and restoration. Recent advancements include 3D image generation for CT imaging, enhancing segmentation performance. This paper introduces a 3D latent diffusion model for CT cardiac imaging that generates both images and labels, improving training and validation data. A dual-step training scheme is proposed, pre-training with diffusion-generated data and fine-tuning with ground truth data. Extensive experiments demonstrate the model's effectiveness in enhancing cardiac chamber image segmentation.