Medical diagnosis of lung nodules by fusing CT image feature extraction and improved SVM algorithm
10.19745/j.1003-8868.2023242
- VernacularTitle:基于CT图像特征提取及改进SVM算法的肺结节诊断方法研究
- Author:
Jing-Jing LIU
1
;
Liang SHAO
Author Information
1. 上海市嘉定区安亭医院,上海 201805
- Keywords:
CT image;
feature extraction;
pulmonary nodule;
support vector machine;
particle swarm optimization algorithm;
simulated annealing;
deep learning
- From:
Chinese Medical Equipment Journal
2023;44(12):7-13
- CountryChina
- Language:Chinese
-
Abstract:
Objective To propose a lung nodule diagnosis method based on CT image feature extraction and improved support vector machine(SVM)algorithm to enhance the accuracy and efficiency of automatic identification of lung nodules.Methods A cascade feature extraction method combining deep learning-based feature extraction and traditional manual extraction was used for CT image feature extraction,and the extracted features were input into an improved SVM algorithm to complete automated identification of lung nodules,using a multiple kernel learning support vector machine(MKL-SVM)algorithm and a particle swarm optimization(PSO)algorithm that integrated simulated annealing(SA)algorithm for parameter optimizing.The performance of cascade features was tested by comparing traditional feature extraction,deep learning-based feature extraction and cascade feature extraction.Comparison tests were performed using single kernel functions(RBF kernel,Sigmoid kernel and polynomial kernel functions)to validate the performance of the MKL-SVM algorithm.Tests were carried out using SVM functions with Sigmoid kernel to compare the fitness curves of the PSO algorithm and the PSO-SA algorithm for optimization to validate the effectiveness of the PSO-SA algorithm.Comparison analyses were conducted with the existing computer aided diagnosis(CAD)models of lung under the same dataset to verify the diagnostic efficacy of the proposed model of cascade features combined with improved MKL-SVM(cascade features with improved MKL-SVM,CF with MKL-SVM).Results The performance test results showed that cascade feature extraction had the F value with a mean value of 0.934 1,a maximum value of 0.957 3,a minimum value of 0.919 5 and a median value of 0.939 7,which behaved better in accuracy than manual feature extraction and deep learning-based feature extraction.The kernel function comparison test results indicated that the MKL-SVM algorithm had the best diagnostic performance with the mean value of F value of 0.924 3,the maximum value of 0.935 0 and the AUC value of 0.987 3.The Sigmoid kernel comparison test results found that PSO-SA al-gorithm had the best fitness value of 0.943 7,which gained advantages over the PSO algorithm.The model comparison test revealed that compared with the lung CAD model,the CF+MKL-SVM model had advantages in generalization ability,AUC value(0.9845),the values of all the indexes(all higher than 0.9),specificity and precision.Conclusion The proposed method can be used for automatic recognition of lung cancer and enhances the accuracy for detecting lung cancer.
