Abstract

Info

Title:: Identification and multi-class classification of pediatric pneumonia based on DenseNet

Author(s):: YANG Chuanjie1; MA Zhiqing1; ZHAO Wenhua1; ZHAO Shuang1; 2; 1. School of Medical Information Engineering, Shandong University of Traditional Chinese Medicine, Jinan 250000, China 2. Experimental Center of Shandong University of Traditional Chinese Medicine, Jinan 250000, China

Keywords:: Keywords: image classification medical image processing pediatric pneumonia space and channel reconstruction convolution spatial group-wise enhance

Abstract:: Abstract: Given the challenges in imaging-based diagnosis of pediatric pneumonia, this study presents an improved diagnostic model for pediatric pneumonia based on the DenseNet algorithm, aiming to enhance diagnostic accuracy. With DenseNet as the basic model, the spatial and channel reconstruction convolution is integrated to compress the convolutional neural network by leveraging the spatial and channel redundancy in features, thus improving inference efficiency. A spatial group-wise enhance module is embedded into the network, which utilizes attention masks to scale and adjust feature vectors at different positions, thereby strengthening the robustness of each feature group in the spatial dimension. Meanwhile, the convolutional and pooling kernels in the early stages of the network are increased to improve the models expressive capability. Furthermore, the introduction of A2-Nets dual attention network significantly improves image recognition performance through efficient feature aggregation and propagation mechanisms. Experimental results demonstrate that the proposed method achieves remarkable performances, with an accuracy of 97.8% for the binary classification of normal and pneumonia, 82.3% for the binary classification of bacterial and viral pneumonia, and 83.1% for the 3-class classification of normal, bacterial pneumonia, and viral pneumonia.