李晓峰 李东 王妍玮

摘 要：針对目前存在的图像弱边缘检测算法容易忽略边缘计算中阈值的选取问题，且并未对数据进行聚类分析，导致检测效果不佳的问题，提出基于全卷积神经网络的复杂医疗图像弱边缘检测算法。首先，采用Mean-shift（均值偏移）对复杂医疗图像滤波处理，并对滤波后的图像进行灰度像素增强;其次，使用可自适应调节的动态阈值方法对图像边缘点和内部候选点进行判定，获取边缘计算结果;最后，建立全卷积神经网络模型，并对模型进行训练，在模型中输入计算得到的图像边缘数据，采用量子遗传聚类方法构建聚类目标函数并求解，完成对输入数据的聚类，以聚类结果为基础，生成边缘信息概率图，对图像的弱边缘概率进行计算，完成弱边缘检测。实验结果表明，选取阈值的适应度较强，图像边缘计算耗时少，对于不同的邻域量子漫步空间数据集，所提算法均具有较好的聚类效果，且所提算法的弱边缘检测准确率最高可达到86%，检测效果好。

关键词：

全卷积神经网络;医疗图像;弱边缘;阈值;量子遗传聚类方法;检测

DOI：10.15938/j.jhust.2021.03.010

中图分类号： TP391

文献标志码： A

文章编号： 1007-2683（2021）03-0065-09

Weak Edge Detection Algorithm for Medical

Images Based on Full Convolution Neural Network

LI Xiao-feng1， LI Dong2， WANG Yan-wei 3

（1.Department of Information Engineering， Heilongjiang International University， Harbin 150025， China;

2.School of Computer Science and Technology， Harbin Institute of Technology， Harbin 150001， China;

3.Department of Mechanical Engineering， Purdue University， West Lafayette， Indianan IN47906， US）

Abstract：Aiming at the existing image weak edge detection algorithm， it is easy to ignore the selection of threshold in edge calculation， and there is no clustering analysis of the data， which leads to the problem that the detection effect is not good. A weak edge detection algorithm for complex medical images based on full convolution neural network is proposed. Firstly， Mean-shift is used to filter the complex medical image， and the gray pixel of the filtered image is enhanced. Secondly， the adaptive dynamic threshold method is used to determine the edge points and internal candidate points of the image， and the edge calculation results are obtained. Lastly， the full convolution neural network model is established， and the model is trained， and the calculated image edges are inputted into the model. The objective function of clustering is constructed and solved by quantum genetic clustering method， and the clustering of input data is completed. Based on the clustering results， the edge information probability map is generated， the weak edge probability of the image is calculated， and the weak edge detection is completed. The experimental results show that the fitness of selecting threshold is strong and the calculation of image edge is less time-consuming. For different neighborhood quantum walking spatial data sets， the proposed algorithms have good clustering effect， and the accuracy of weak edge detection of the proposed algorithm can reach 86%， and the detection effect is good.

Keywords：full-convolution neural network; medical image; weak edge; threshold; quantum genetic clustering method; detection

0 引 言

医疗图像是指医疗设备所获取的图像，随着医疗图像在临床医学中的广泛应用，医疗图像成为临床诊断与治疗的重要参考依据[1]。对医疗图像进行实际扫描处理时，通常形成的图像边缘过渡区较宽，易成为弱边缘，影响图像质量[2-3]。边缘是图像最基本的特征……