Segmentation of Skeleton Ultrasound Images Based on MMA-SUISNet

Abstract

When exploring the use of ultrasound to provide real-time, radiation-free 3D imaging for fracture surgery, A skeleton ultrasound image segmentation network based on the fusion of mixed multiple attention (MMA-SUISNet) was proposed to solve the problems of excessive noise, small skeleton features, and difficult boundary division in the ultrasound image. The model uses the Squeeze Exception (SE) module to complete the encoding function, constructs cross-layer connections, and improves the ability to identify small targets; By adding Convolutional Block Attention Module (CBAM) to the encoder, the model can adaptively adjust the weights of channels and positions to better extract features and reduce the impact of noise; By adding Attention Gates (AG) to the decoder, features are adaptively emphasized and transmitted, allowing the network to focus on skeleton boundary information. For the collected skeleton ultrasound images, this paper shows through segmentation, ablation, and generalization experiments that the proposed model has improved Dice, IoU, and F1-Score indicators by 13.87%, 10.01%, and 13.80% compared to the original U-Net model, respectively.