Classification of pulmonary lesions based on CNN and chest X-ray images. Issue 2 (June 2021)

• Record Type: Journal Article
• Authors: Xue, Wen
• Is Part Of: Journal of physics. Volume 1952:Issue 2(2021)
• Publication Date: 2021-06
• Subjects: Pulmonary lesions -- Chest imaging -- Predict -- AUC scores; Physics -- Congresses; 530.5
• DOI: 10.1088/1742-6596/1952/2/022025
• Languages: English
• ISSNs: 1742-6588

Abstract

Abstract: This article mainly introduces the convolutional neural network (CNN) and uses CNN to realize the processing and classification prediction of chest X-ray images (CXR), to determine whether the lung has lesions, and finally the final AUC score of 0.85556 through CNN. In order to further improve the accuracy, after referring to many documents and considering the actual situation, I chose to perform principal component analysis (PCA) in the image preprocessing part, replace the random initial sample with the principal component initial sample, and replace the random initial kernel with the principal component initial kernel. To avoid staying in the local optimum when stochastic gradient descent finds the optimal kernel. The PCA + CNN model predicted an AUC score of 0.89333, an increase of 0.03777.