Time-efficient Diffusion Models for Semantic Segmentation of Choroid Plexus from brain MRI

Since their introduction, diffusion models have achieved significant success in various applications, demonstrating superiority over other generative models. This work explores the feasibility of applying diffusion models to the segmentation of the choroid plexus, a crucial brain structure for the diagnosis and study of various pathologies. Using magnetic resonance imaging (MRI) and manual segmentations performed by expert radiologists on healthy subjects and patients with relapsing-remitting multiple sclerosis (RRMS), several models were trained to generate accurate and consistent segmentations from the input images. In addition to the direct training of Denoising Diffusion Probabilistic Models, advanced deep learning techniques were implemented to enhance performance and make the application of diffusion models in the medical field more feasible. Specifically, transfer learning and self-supervised learning techniques were adopted to assess whether efficient pre-training could reduce the number of images required for model training, given the limited availability of freely usable medical data. Additionally, patch-based techniques were employed to segment complete three-dimensional volumes, aiming to reduce the time needed to generate segmented volumes, which is a crucial factor in the medical context. Our results indicate that, while diffusion models were not initially designed for segmentation tasks, they can be successfully adapted for this purpose. However, further research is necessary to achieve state-of-the-art performance and fully leverage deep learning techniques such as data augmentation, transfer learning, and self-supervised learning. Nevertheless, our findings provide valuable insights for future research, guiding efforts toward improved performance and generalization. In conclusion, among various generative models, diffusion models present a promising alternative to traditional methodologies used in medical segmentation, with potential improvements achievable through the combination with other deep learning techniques and the ability to provide quick results even for three-dimensional segmentations.