Few-shot learning in diffusion models for generating cerebral aneurysm geometries

The study of brain vessel pathologies is critical for the advancement of neurovascular medicine, yet researchers often face significant hurdles due to the scarcity of imaging data for certain uncommon types of aneurysms. Generative deep learning models have been proposed to address the lack of high-quality labeled medical images - however, the shortage of data also presents a unique challenge in training generative models. To address this issue, our work explores the efficacy of training latent diffusion models (LDMs) with fewshot learning, enabling the generation of detailed vessel segmentations from as few as five images per class. By incorporating set-based vision transformers for class embeddings and leveraging signed distance functions (SDFs) as a novel form of conditioning, our method reduces the need for extensive datasets for training. Comparative studies with established generative models, including variational autoencoders (VAEs) and generative adversarial networks (GANs), highlight the robustness of our approach. Our model not only successfully generates high-quality segmentations of brain vessels with aneurysms but also significantly outperforms the standard generative models.