基于卷积神经网络与特征选择的医疗图像误差预测算法

Error Prediction Algorithm of Medical Image Based on Convolution Neural Network and Feature Selection

针对传统医疗图像误差预测算法无法很好的选择图像特征,存在图像误差预测值与实际值拟合度低、预测耗时长等问题,提出基于卷积神经网络与特征选择的医疗图像误差预测算法.首先,选取5种集成规则构建自适应多分类器,对医疗图像区域进行分类;其次,训练卷积神经网络,利用训练完成的神经网络提取不同类别医疗图像区域特征,以此为基础计算区域距离,寻找出相似度最小的区域,完成图像可疑区域定位;再次,融合多评价标准生成特征子集,从中搜索得到最优特征子集,完成可疑区域图像特征选择;最后,以选择得到的特征区域像素点作为训练样本,建立预测样本与训练样本之间的多元线性回归矩阵,实现误差预测.实验结果表明,所提算法的集成规则适应度较高,分类性能好,区域距离计算准确率高达95%左右,特征选择的AUC值(Area Under Curve)高,且预测结果拟合度和预测耗时均优于传统算法.

In order to address the problem that traditional medical image error prediction algorithm can not select image features well,there are some problems such as low fitting degree of image error prediction value,low actual value and long prediction time,a medical image error prediction algorithm based on convolution neural network and feature selection was proposed.Firstly,five integrated rules were selected to construct adaptive multi-classifiers to classify medical image regions.Secondly,the training convolution neural network was used to extract different types of medical image area features by using the training neural network.Then,multiple evaluation criteria were combined to generate special features.The optimal feature subset was searched to complete the feature selection of suspicious region image.Finally,the multiple linear regression matrix between the prediction sample and the training sample was established to realize the error prediction by taking the pixel points of the feature region as the training sample.The experimental results show that the proposed algorithm has high fitness of integration rules and good classification performance,the accuracy of region distance calculation is about 95%,the AUC value of feature selection is high,and the fitting degree and prediction time of the prediction results are better than those of the traditional algorithm.