Accelerating the Deep Learning based Brachial Plexus Nerve Trunks Recognition in Ultrasound Images using Edge Computing

Abstract

The brachial plexus block is an ultrasound image-guided regional anesthesia technique. It involves injecting local anesthetics into or around the brachial plexus to induce numbness in the upper limb, facilitating painless surgeries. Due to the difficulty in achieving clear visualizations of nerves in ultrasound images, this procedure is typically administered by highly trained anesthetists. Therefore, the limited availability of such skilled professionals constrain the widespread performance of these anesthesia procedures. Recent advancements in deep learning-based image segmentation methods, such as U-Net architecture, can potentially solve the segmentation of nerves in ultrasound images. This paper proposes an improved deep learning-based approach to segment brachial plexus nerve trunks in ultrasound images. We improved U-Net architecture by adding a residual block and a naive inception block, achieving a dice coefficient of 0.66. However, the computational complexity of deep learning methods presents a challenge to achieve real-time segmentation using edge computing. As a solution, we deployed the model in Xilinx KV260 multi-processor system on chip (MPSOC) to accelerate the image segmentation.