Methodology Research Of Deep Learning-based Segmentation Of Median Nerve On Ultrasound Images

The median nerve,one of the major branches of the brachial plexus,innervates most of the muscles in the flexion of the forearm,as well as most of the muscles in the radial half of the inner hand and the radial skin sensation of the palm.When the median nerve was injured in the arm,all branches were affected,and the forearm could not be pronated.Compression of the median nerve as it passes through the carpal tunnel can lead to carpal tunnel syndrome,which affects general health.Clinical diagnosis can be made according to the morphological characteristics of median nerve in ultrasound images,such as nerve compression,neuroinflammatory lesions and nerve swelling.However,the speckle noise and low resolution of ultrasound imaging reduce the image quality,as well as the boundary between median nerve and peripheral tissue is blurred,all of which make it difficult to define the nerve contour.In addition,the ultrasound images of median nerve currently obtained are mostly in two-dimension,thus median nerve can’t be visualized in three-dimension.Doctors make diagnosis according to the characteristics of median nerve in continuous images collected by moving the probe,such as the position,the gray and texture information of the median nerve.It relies heavily on the doctor’s experience and the ability of three-dimensional(3D)imagination.Therefore,VGG16-UNet model based on deep learning is proposed in this paper for automatic segmentation of median nerve ultrasound images.The model is based on U-Net and VGG16,in which encoder is VGG16 and the decoder is constructed by referring to the upsampling path of U-Net.The information of the same level in the encoder and decoder is connected by skip connections.The experimental results show that the evaluation metrics’ values of VGG16-UNet are all higher than those of U-Net,indicating that the segmentation results of VGG16-UNet are superior to the segmentation results of U-Net.To further improve the performance of the model,identity mapping operations are added to encoder and decoder respectively to construct residual blocks and then reuse the extracted features.Additionally,the spatial attention block is introduced to connect the features from deep and shallow layers of the model,so that the model can pay more attention to the representative feature information.UNet models(R-UNet,A-UNet and AR-UNet)and VGG16-UNet models(R-VGG16UNet,A-VGG16-UNet and AR-VGG16-UNet)are constructed by combing with residual block or/and attention block based on U-Net and VGG16-UNet.The comparative experimental results show that VGG16-UNet models are superior to the corresponding U-Net models,which proves the effectiveness of the proposed VGG16UNet network in improving segmentation performance,and the role of the proposed residual module and attention module in optimizing the models.It is interesting that AR-UNet combing with both attention block and residual block has the best performance among its variants,while AR-VGG16-UNet is slightly inferior to AVGG16-UNet with only attention block embedded,which indicates that the robustness of the proposed VGG16-UNet is not strong enough and there is a certain space for improvement.In addition,3D reconstruction and Visualization of two-dimensional median nerve image sequences are performed using Visualization Toolkits(VTK).In order to better display the shape and travel direction of median nerve in the visual image,the binary segmentation results of median nerve image sequence produced by A-VGG16-UNet are blended with the original image sequence using different weights,and then the blending results are used as inputs to reconstruct 3D image of median nerve by using volume rendering method based on light projection in VTK.The morphology and travel direction of the nerve can be more intuitively displayed from different perspectives,and it can be zoomed and flipped as needed,providing information for the assessment of nerve swelling or compression,nerve continuity,and clinical diagnosis.