基于改进UNet模型的眼球超声图像分割算法研究

Research on Segmentation Algorithm of Eye Ultrasound Image Based on Improved UNet Model

摘要: 在医学图像分割领域，提高分割性能一直是一个具有挑战性的任务。超声图像具有边缘模糊、噪声污染等缺点，为了解决眼球超声图像分割结果不理想这一难题，本文提出了一种基于UNet的改进分割算法。首先，本文采用了残差网络(ResNet)结合UNet，有效地解决了模型退化的问题，进一步提高了模型的精度和泛化能力；其次，在主干特征提取部分引入高效多尺度注意力(EMA)机制，以增强分割模型的特征表示能力；最后，通过RAVIR数据集进行泛化性实验，证明了所提出模型的泛化能力。实验结果显示，改进的UNet算法在超声眼球图像数据集上获得的MIoU和Dice的值分别达到了82.5%和82.3%，相比UNet模型分别提升了1.1%和1.4%，具有更好的医学图像分割效果。

Abstract: In the field of medical image segmentation, improving segmentation performance has been a challenging task. Ultrasound images have disadvantages such as blurred edges and noise pollution; in order to solve the complex problem of unsatisfactory segmentation results of eye ultrasound images, this paper proposes an improved segmentation algorithm based on UNet. Firstly, the residual network (ResNet) combined with UNet is used in this paper to effectively solve the problem of model degradation and further improve the accuracy and generalization ability of the model; secondly, the efficient multi-scale attention (EMA) mechanism is introduced in the central feature extraction part to enhance the feature representation ability of the segmentation model; finally, generalizability experiments are carried out with the RAVIR dataset, which proves the generalization ability of the proposed model’s generalization ability. The experimental results show that the improved UNet algorithm achieves 82.5% and 82.3% values of MIoU and Dice on the ultrasound eye image dataset, which are 1.1% and 1.4% higher than the UNet model, and it has better medical image segmentation results.