Synth-to-Segment: MRI Brain Tumor Segmentation with Diffusion Transformers and Attention U-Net

Mudassar Ali, Tong Wu, Maryam Mansoor, Haoji Hu

Abstract: Accurately segmenting brain tumors from MRI images is very important for effective treatment planning and mostly suffers from a scarcity of annotated datasets and complex tumor morphology. In this paper, the authors propose a new approach by generating high-quality synthetic MRI images using the BraTS 2020 dataset with a Modality-Conditional Diffusion Transformer that has Multi-Scale Attention and a Hierarchical Latent Space model. These synthetic images are combined with real data for training an attention-enhanced U-Net for segmentation. Our approach focuses on critical multi-scale features, therefore improving the accuracy and robustness of segmentation. Experimental results show significant improvements over the state-of-the-art and offer a very accurate and robust solution to brain tumor segmentation, particularly in data-scarce scenarios.

External IDs:dblp:conf/micad2/AliWMH24