Ultrasound Nerve Segmentation of Brachial Plexus Based on Optimized ResU-Net

Abstract

The accurate ultrasound nerve segmentation has attracted wide attention, for it is beneficial to ensure the efficacy of regional anesthesia, reducing surgical injury, and speeding up the recovery of surgery. However, because of the characteristics of high noise and low contrast in ultrasonic images, it is difficult to achieve accurate neural ultrasound segmentation even with U-Net, which is one of the mainstream network in medical image segmentation and has achieved remarkable results in Computed Tomography (CT), Magnetic Resonance Imaging (MRI) and Optical Coherence Tomography (OCT). Addressing this problem, an optimized and effective ResU-Net variation to segment the ultrasound nerve of brachial plexus is proposed. In our proposed method, median filtering is first employed to reduce the speckle noise which is spatially correlated multiplicative noise inherited in ultrasound images. And then the Dense Atrous Convolution (DAC) and Residual Multi-kernel Pooling (RMP) modules are integrated into the ResU-Net architecture to reduce the loss of spatial information and improve the robustness of the segmentation with different scales, thus boosting the accuracy of segmentation. Our fully mechanism improves the segmentation performance in the public dataset NSD with the dice coefficient 0.7093, about 3% higher compared to that of the state-of-the-art models.