École Nationale Supérieure de
Techniques Avancées

Abstract

State-of-the-art models such as Unet and nnUnet excel in semantic segmentation of medical images, but their training methods can occasionally yield structurally inconsistent results, featuring gaps and disconnected elements. In contrast, deep generative models have recently demonstrated substantial success in producing structurally coherent outputs for both natural and medical images. This work investigates the potential of diffusion models in generating anatomically coherent segmentations while maintaining competitiveness with nnUnet in segmentation tasks. Various diffusion model implementations are evaluated for non-conditional generation and segmentation tasks. The study presents and discusses both visual and quantitative results, which are promising, showcasing competitiveness with the nnUnet baseline in segmentation tasks and apparent improvements in mask structure. This work suggests a promising research path in semantic segmentation of medical images.