Open Access

ARTICLE

A Pneumonia Recognition Model Based on Multiscale Attention Improved EfficientNetV2

Zhigao Zeng¹, Jun Liu¹, Bing Zheng², Shengqiu Yi¹, Xinpan Yuan¹, Qiang Liu^1,*

1 School of Computer Science, Hunan University of Technology, Zhuzhou, 412007, China
2 Hubei Central China Technology Development of Electric Power Co., Ltd., Wuhan, 430000, China

* Corresponding Author: Qiang Liu. Email:

Computers, Materials & Continua 2025, 84(1), 513-536. https://doi.org/10.32604/cmc.2025.063257

Received 09 January 2025; Accepted 03 April 2025; Issue published 09 June 2025

Abstract

To solve the problems of complex lesion region morphology, blurred edges, and limited hardware resources for deploying the recognition model in pneumonia image recognition, an improved EfficientNetV2 pneumonia recognition model based on multiscale attention is proposed. First, the number of main module stacks of the model is reduced to avoid overfitting, while the dilated convolution is introduced in the first convolutional layer to expand the receptive field of the model; second, a redesigned improved mobile inverted bottleneck convolution (IMBConv) module is proposed, in which GSConv is introduced to enhance the model’s attention to inter-channel information, and a SimAM module is introduced to reduce the number of model parameters while guaranteeing the model’s recognition performance; finally, an improved multi-scale efficient local attention (MELA) module is proposed to ensure the model’s recognition ability for pneumonia images with complex lesion regions. The experimental results show that the improved model has a computational complexity of 1.96 GFLOPs, which is reduced by 32% relative to the baseline model, and the number of model parameters is also reduced, and achieves an accuracy of 86.67% on the triple classification task of the public dataset Chest X-ray, representing an improvement of 2.74% compared to the baseline model. The recognition accuracies of ResNet50, Inception-V4, and Swin Transformer V2 on this dataset are 84.36%, 85.98%, and 83.42%, respectively, and their computational complexities and model parameter counts are all higher than those of the proposed model. This indicates that the proposed model has very high feasibility for deployment in edge computing or mobile healthcare systems. In addition, the improved model achieved the highest accuracy of 90.98% on the four-classification public dataset compared to other models, indicating that the model has better recognition accuracy and generalization ability for pneumonia image recognition.

---

Keywords

---