Enhancement Guidance Network for Classification of Pneumonia in Chest X-Rays

Authors:
Open Access
Publisher website
Google Scholar
Abstract
In this paper, we propose a deep learning-based lung disease classification scheme SGC-Net, which does not rely on the work experience of doctors and realizes accurate and rapid classification of pneumonia. SGC-Net is composed of partition network and classification network, the output of the segmentation network guides the classification network to focus on the lung region, it solves the problem of classification errors caused by background interference in the traditional pneumonia classification network and the loss of global information when using lung CXR alone. In this paper, we use global CXR and lung CXR to reconstruct a dataset suitable for classification network training. Meanwhile, the KL divergence consistency loss function is also introduced into the classification network, which makes the classification result of the global CXR obtained from the classification network consistent with its own lung CXR as much as possible, so that the lung CXR can guide the global CXR and achieve lung location enhancement guidance. In addition, this paper presents CXR-cam, a visualization method based on Grad-cam for classification networks, which is used to study and analyze the classification ability of classification networks and to automatically label lesion regions. conclusion: The binary (pneumonia, Normal) mode constructed using SGC-Net reached ACC score of 0.959 and AUROC score of 0.965 on the test dataset of ChestXRay2017,the Mean-IoU score of the segmented network reached 0.943, and compared with the traditional classification network, the scheme proposed in this paper achieved an advanced level of classification. In summary, this classification scheme has strong classification ability and good robustness, which can be used to construct an accurate and rapid auxiliary system disease of the lung, which solves the problems such as relying on the working experience of doctors when classifying pneumonia, and also realize the precise annotation of the location of the lesions, which provides a new idea for the classification of pneumonia.