An Improved EfficientNet Model and its Applications in Pneumonia Image Classification

ABSTRACT

Since the outburst of COVID-19, the medical system has been facing great challenges due to the explosive growth in detection and treatment needs within a short period. To improve the working efficiency of doctors, an improved EfficientNet model of Convolutional Neural Network (CNN) was proposed and applied for the diagnosis of pneumonia cases and the classification of relevant images in the present study. First, the acquired images of pneumonia cases were divided to determine the zones with target features, and image size was limited to improve the training speed of the network. Meanwhile, reinforcement learning was performed to the input dataset to further improve the training effect of the model. Second, the preprocessed images were inputted into the improved EfficientNet-B4 model. The depth and width of the model, as well as the resolution of the input images, were determined by optimizing the combination coefficient. On this basis, the model was scaled, and then its ability in extracting the features of deep-layer images was strengthened by introducing the attention mechanism. Third, the learning rate was adjusted by using the Adaptive Momentum (ADAM), and the training efficiency of the model was accelerated. Finally, the test set was inputted into the trained model. Results demonstrate that the improved model could detect 98% of patients with pneumonia and 97% of patients without pneumonia. The accuracy rate, precision rate, and sensitivity of the model were generally improved. Lastly, the training and test results of VGGNet, SqueezeNet-Elus, SqueezeNet-Relu, and the improved EfficientNet-B4 models were compared and evaluated. The improved EfficientNet-B4 model achieved the highest comprehensive accuracy rate, reaching 92.95%. The proposed method provides some references to the application of the CNN model in image classification and recognition. © 2022 School of Science, IHU. All Rights Reserved.