Diffusion meets edge awareness: a unified framework for high-precision ultrasound nerve segmentation

Nerve segmentation from ultrasound images is still challenging due to speckle noise, poor contrast, and unclear anatomical edges. These limitations have an adverse effect on deep learning segmentation models working on raw ultrasound data.

In order to overcome these difficulties, a diffusion-guided edge-aware multi-channel segmentation approach that can take advantage of raw ultrasound images, diffusion-processed features, and edge information is proposed. Moreover, a new EdgeFusion-U-Net model is developed for integrating these complementary features. Experiments were performed on two benchmark datasets of ultrasound nerve segmentation and the proposed method was compared with other state-of-the-art techniques such as U-Net, Residual U-Net, UNet++, TransU-Net, nnU-Net, and Attention U-Net.

As evidenced by experiments, the proposed method shows better results than existing techniques. Our method reached Dice scores of 94.62% and 93.91% on two benchmark datasets, while baseline methods had about 83% and 84% Dice scores without applying diffusion denoising and edge fusion.

It can be seen that incorporating diffusion filter-based speckle noise removal and edge-based feature fusion techniques helps to achieve better results in terms of boundary detection. This technique not only helps in obtaining accurate information about the nerves but can also be used to support computer-assisted clinical analysis using ultrasounds.