一种基于改进ResU-Net的角膜神经分割算法

角膜神经图像的自动分割对于糖尿病神经病变等疾病的诊断与筛查至关重要。针对由于角膜神经图像存在对比度低且包含非神经结构而造成分割效率较低的问题,在ResU-Net结构基础上引入多尺度残差、注意力机制、多尺度图像输入与多层损失函数输出模块,提出一种基于注意力机制的角膜神经分割算法。多尺度残差模块通过在残差模块中加入多尺度表征信息以提高卷积层提取多尺度特征的能力,而注意力机制模块在双重注意力作用下,利用网络对编码器与解码器中的目标特征进行权重优化,使得在增强图像目标区域特征的同时抑制背景及噪声区域,并采用多尺度图像输入与多层函数输出模块以监督网络中每一层的特征学习。实验结果表明,与主流分割算法相比,该算法的分割效果更优,且曲线下面积与敏感度分别可达到0.990和0.880。

A Corneal Nerve Segmentation Algorithm Based on Improved ResU-Net

The automatic segmentation of corneal nerve images is crucial to the diagnosis and screening of several diseases such as diabetic neuropathy,but it suffers from the low segmentation efficiency caused by the low contrast of corneal nerve images and the existence of non-neural structures.To address the problem,this paper proposes a novel corneal nerve segmentation algorithm based on attention mechanism,which introduces multi-scale residual module,attention mechanism module,multi-scale image input module,and multi-layer loss function output module into the ResU-Net structure.The multiscale residual module is used to add multi-scale representation information into the residual module to improve the multi-scale feature extraction ability of the convolutional layer.The attention mechanism module consisting of channel and spatial attentions uses the network to optimize the weight of the target features in the encoder and decoder,so as to enhance the features of the target area as well as suppress the features of background and noise area.Furthermore,the multi-scale image input and multi-layer function output modules are added to supervise the feature learning of each network layer.Experimental results show that the proposed method outperforms the existing mainstream segmentation algorithms with its Area Under Curve(AUC)reaching 0.990 and its Sencificity(Sen)reaching 0.880.