Diffusion Posterior Sampling for High-Resolution 3D Lung CT Generation

The development of reliable AI-based diagnostic tools is constrained by the scarcity of high-quality and fully annotated medical imaging data, due to costly acquisition procedures and privacy restrictions. Although generative AI has the potential to address this gap, 3D generation remains computationally demanding while 2D lacks the spatial context required for many applications. Accordingly, we propose a hierarchical pipeline that combines low-resolution 3D CT generation with slice-wise super-resolution using an adaptation of the Posterior Sampling with Latent Diffusion (PSLD) algorithm, enabling a computationally efficient two-stage synthesis process. Preliminary results indicate the successful reconstruction of high-resolution CT volumes, with fine detailsand overall slice consistency, while preserving the variability of the generated samples.