李立 董现玲 刘会玲
摘 要:随着卷积神经网络在医疗图像领域的成功应用,进一步推动了医学影像设备性能的提升。通过利用卷积神经网络方法,对甲状腺影像进行系统的分析,有效预测患者的病情发展态势,切实保障诊断效率的提高,整体看具有较为显著的临床应用价值。鉴于此,该文通过分析卷积神经网络的基本结构、工作机制、基本特点,探讨基于卷积神经网络的甲状腺影像识别方法,并为甲状腺影像诊断提出几点思考,以期有效推动卷积神经网络在甲状腺影像诊断中的应用改革进程。
关键词:甲状腺疾病 医学影像 卷积神经网络 神经网络模型
Abstract: With the successful application of convolutional neural network in the field of medical imaging, the performance of medical imaging equipment is further improved. Through the use of convolution neural network method, systematic analysis of thyroid imaging can effectively predict the development trend of the patient's condition and effectively ensure the improvement of diagnosis efficiency, which has a more significant clinical application value. In view of this, this paper analyzes the basic structure, working mechanism and basic characteristics of convolutional neural network, discusses the thyroid image recognition method based on convolutional neural network, and puts forward some thoughts for thyroid imaging diagnosis in order to effectively promote the application and reform process of convolutional neural network in thyroid imaging diagnosis.
Key Words: Thyroid disease; Medical imaging; Convolutional neural network; Neural network model
1 卷积神经网络的基本原理
1.1 基本结构
卷积神经网络是由输入层、卷积层、池化层、全连接层、输出层等组成的,其组织结构比较固定。
其中,该方法的运用需要明确以下事项:卷积层的每一层均包含着各种可学习的参数,同时相当于一个滤波器,便于获取相关物体的边缘与颜色特征;池化层一般在卷积层的后面,具有减少网络训练参数、降低输出结果维度、对输入做采样与降维等作用;激活函数位于卷积层与池化层之间,同时也是解决神经网络中非线性问题的关键,Tanh、Sigmoid、ReLUd等是较为常用的激活函数。
1.2 工作机制
一是卷积神经网络的反向传播算法。其本质与BP神经网络算法基本一致,主要区别在于卷积神经网络的反向求导需要明确参数连接的各个神经元。
二是梯度下降法。朝着目标函数梯度反向上更新模型参数,从而实现最小化目标函数,即该方法机制的基本思想。……