An improved X-ray image diagnosis method for COVID-19 pneumonia on a lightweight neural network embedded device

With the continuous variation and evolution of COVID-19, Deep learning-based X-Ray image classification models of COVID-19 have been emerged a lot and achieved high accuracy rates. However, the current mainstream neural network algorithm generally has too many parameters, complex models and high computational complexity. It is usually deployed on high-performance computing servers, which is not conducive to embedded deployment and real-time computing. In order to better enable the neural network to play its role on embedded devices without reducing the classification accuracy of the model, this paper proposes MECNet (MobileExprCovidNet) by redesigning the network structure and transforming the attention mechanism with cutting-edge knowledge, which ensures the accuracy rate compared to MobileNetV3 has been greatly improved. Compared with networks with larger number of relative parameters and higher computational complexity, MECNet does not have significant disadvantages in terms of accuracy and recall. MECNet is more suitable for implementation on various embedded architecture compared to the original MobileNetV3 network structure. Experimentally, the proposed network improves 4%,5%, and 5% in accuracy, precision, and recall compared to the base model MobileNetV3 for the publicly accessible COVID-19 Radiography Database dataset image class. Outperformed many metrics on embedded device experiments GhostNet, ResNet50, MobileNetV3. The accuracy, precision, and recall rates are 92%, 90%, and 91%, while only 73MFLOPs are required, and the number of parameters is only 0.85M, which meets our requirement for deployment on embedded machines. The Chest X-ray dataset is used to further verify the generalization ability of the network, and Score-CAM is used to verify its effectiveness. The experiments demonstrate that the MECNet proposed in this paper has a good classification effect for COVID pneumonia X-rays. Finally, the model was deployed to an embedded device, and the processing speed was measured to reach 183.3ms, and the accuracy rate reached 89.7%, which directly proves the usability and efficiency of the model on the embedded device.