EDB-Diff: a EdgeDevice based diffusion network for brain tumor image segmentation

Manually segmenting brain tumor images is time-consuming and not conducive to timely treatment for patients. In recent years, significant progress has been made in the research and development of automated brain tumor image segmentation. CNN-based methods still face instability due to the non-uniqueness of data labels, while diffusion model-based methods have shown significant improvement in stability but are known to have a considerable computational burden. We propose EDB-Diff, a method that has been lightweighted and features a feature separation module based on the analysis of the properties of multi-sequence MRI brain images. Additionally, we have incorporated a broad modality attention mechanism into the denoising network, which enhances the network’s sensitivity to specific features of each sequence without compromising its ability to integrate common features. We evaluated our method on the BraTS2023 dataset, achieving a 60.66% reduction in the number of parameters and a 72.87% increase in inference speed on edge computing devices, while maintaining comparable Dice scores and exhibiting better HD95 stability.