1. Introduction

Deep learning models, as one of the branches of machine learning, are widely used in the field of computer vision technology, and the field of image recognition, represented by medical image analysis, is also growing. Artificial Intelligence has far-reaching impact on labour force, economic dynamics, and ethical issues, in which the study of classification of X-ray images of neocoronary pneumonia, X-ray images of healthy individuals and X-ray images of community-acquired pneumonia has attracted much attention. Neocoronavirus pneumonia is an infectious disease caused by a novel coronavirus, and its imaging manifestations in the lungs are characterised by certain features on X-ray images. By classifying and analysing X-ray images of neocoronavirus pneumonia, it can help doctors diagnose whether patients are infected with neocoronavirus more quickly and accurately, so as to take timely and effective treatment measures. As a comparison group, X-ray images of healthy people play an important role in the classification of X-ray images of patients with neocoronavirus pneumonia. Through in-depth analysis and comparison of CT images of healthy people groups, the characteristic imaging manifestations of neocoronary pneumonia can be better distinguished and the accuracy of classification and diagnosis can be improved. Community-acquired pneumonia is an infectious disease of the lungs acquired in a community setting. Compared with neocoronal pneumonia, community-acquired pneumonia has different features on X-ray images. Therefore, differentiating community-acquired pneumonia from other types of lung diseases is also an important diagnostic problem in the field of medical imaging.

Deep learning algorithms play a vital role in these image classification problems. Deep learning algorithms are able to automatically learn features from a large amount of well-labelled training data and build efficient models for image classification. For example, convolutional neural network (CNN) is a deep learning algorithm commonly used for medical image classification tasks, which has strong adaptability and accuracy in processing medical images.

In addition, deep learning algorithms can be combined with techniques such as migration learning and data augmentation to further improve the generalisation ability and robustness of classification models. By continuously optimising the algorithm structure and adjusting the hyperparameters, the deep learning model can be made to achieve more accurate and reliable image classification between different categories such as new coronary pneumonia, healthy people and community-acquired pneumonia.

It is of great significance to use deep learning algorithms to classify different types of X-ray images such as new crown pneumonia, healthy people and community-acquired pneumonia in the field of medical imaging. In this paper, AlexNet, Vgg, GoogleNet and MobileNet models are introduced for comparative analysis to compare the effectiveness and accuracy of various deep learning models in pneumonia image classification. With the continuous progress of technology and the accumulation of data, we believe that deep learning algorithms will provide more effective auxiliary information for doctors in the future, and help make greater breakthroughs in clinical diagnosis.

2. Data set sources and data analysis

The dataset used in this paper is from the Kaggle open-​source dataset (https:​/​/​www.​kaggle.​com/​datasets/​pranavraikokte/​covid19-​image-​dataset?rvi=​1), which currently has 213 citations, and it contains about 137 X-​ray images of COVID-​19 images, in addition to 317 X-​ray images containing community-​acquired pneumonia and 118 normal chest X-​ray images.​The X-​ray images of COVID-​19 are shown in Fig.​ 1, the X-​ray images of community-​acquired pneumonia are shown in Fig.​ 2, and the normal chest X-​ray images are shown in Fig.​ 3.​

Figure 1. COVID-19.

Figure 2. Community-acquired pneumonia.

Figure 3. Normal chest.

3. Method

3.1. AlexNet

AlexNet is a deep convolutional neural network model proposed by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton in 2012 and is widely regarded as one of the major milestones in the field of deep learning. It achieved great success in the ImageNet image recognition challenge at that time and pushed deep learning technology to new heights. The model structure of AlexNet is shown in Fig. 4.

Figure 4. The model structure of AlexNet.

The structure of AlexNet consists of five convolutional layers and three fully connected layers. The first layer is a convolutional kernel of size 11x11, followed by a ReLU activation function and a maximum pooling layer. Subsequent convolutional layers follow a similar structure, extracting more abstract and complex features by continuously decreasing the size of the convolutional kernel and increasing the number of feature mappings. The fully connected layer is then responsible for classifying the features extracted by the convolutional layers. AlexNet introduces the Dropout technique, which prevents overfitting by randomly dropping some of the neurons during the training process, thus improving the generalisation of the model. In addition, AlexNet uses Data Enhancement technique to perform operations such as panning, flipping, and scaling on images in the training set in order to increase diversity and thus improve the model performance.

3.2. Vgg

VGG is a deep convolutional neural network model proposed by the Computer Vision Group of the University of Oxford in 2014, whose main feature is the use of smaller convolutional kernels and deeper network structure. The VGG model consists of 16 (VGG16) or 19 (VGG19) convolutional layers, which all use 3x3 sized convolutional kernels and convolutional operations with a step size of 1. Meanwhile every two convolutional layers are connected to a pooling layer for downsampling. The model structure of Vgg is shown in Fig. 5.

Figure 5. The model structure of Vgg.

The VGG model extracts features at different abstraction levels in an image by superimposing multiple convolutional layers, enabling the network to better understand the image content. Due to the use of smaller convolutional kernels and deeper network structure, the VGG model is able to reduce the number of parameters while maintaining a certain level of accuracy, which is conducive to training deeper neural networks.

During training, the VGG model usually uses optimisation methods such as cross-entropy loss function and stochastic gradient descent to continuously adjust the network parameters to improve the classification accuracy.

3.3. GoogleNetAlgorithm Optimisation

GoogLeNet is a deep convolutional neural network model proposed by the Google team in 2014, whose main feature is the use of the Inception module to extract features at different scales. Compared with the traditional approach of simply stacking convolutional layers, the Inception module is able to use different sized convolutional kernels at the same time to capture information at different scales in the image, thus improving the network's ability to understand the content of the image. The model structure diagram of GoogLeNet is shown in Fig. 6.

Figure 6. The model structure diagram of GoogLeNet.

GoogLeNet contains multiple Inception modules, each Inception module internally consists of multiple parallel convolutional layers including 1x1, 3x3, 5x5 sized convolutional kernels as well as a maximal pooling layer, which enriches the representation of features learned by the network by means of feature extraction at different scales and with different sensory fields.

3.4. MobileNet

MobileNet is a lightweight deep convolutional neural network model proposed by Google team in 2017, aiming to achieve efficient image recognition in resource-constrained environments such as mobile devices. MobileNet mainly uses deep separable convolution to reduce the number of parameters and computational complexity. The model principle of MobileNet is shown in Fig. 7.

Figure 7. The model principle of MobileNet.

Deep separable convolution consists of two steps: deep convolution and point-by-point convolution. The deep convolution stage applies a separate convolution kernel to each input channel, while the point-by-point convolution stage uses a 1x1 convolution kernel to fuse the features of different channels together. This separation allows the network to drastically reduce the number of parameters while maintaining good performance. MobileNet also introduces a width multiplier and a resolution multiplier to further compress the model. The width multiplier controls the number of channels in each layer of the network to reduce the complexity of the model, while the resolution multiplier adjusts the resolution of the input image to further reduce the amount of computation.

4. Result

In this paper, experiments are conducted using a standalone device with NVIDIA GeForce RTX 3090 graphics card for GPU, 32 GB of RAM, Intel Core i9 processor, and PyTorch, a deep learning framework, with Python version 3.9.

For the experimental parameter settings in this paper, Epoch is 50 rounds, Loss function is Cross Entropy Loss Function, Batch size is 32, Optimiser is Adam Optimiser, Initial Learning Rate is 0.001, Learning Rate Decay Strategy is used, Regularisation is L2 Regularisation Term, Regularisation Coefficient is 0.0001. The image size is uniformly adjusted to 224x224 pixels, and the data enhancement includes random horizontal flip, random rotation.

With the above experimental parameter settings, we will compare and analyse the effect of different models on the task of pneumonia X-ray image classification under the same conditions in order to evaluate their performance on this task. The loss variation curves of each deep learning algorithm are output, as shown in Fig. 8.

Figure 8. The loss variation curves of each deep learning algorithm.

As can be seen from Fig. 8, in terms of the convergence rate of the loss, GoogleNet reaches convergence first (3rd epoch), and the other three models gradually converge, and finally the loss of AlexNet is 0.426, the loss of Vgg is 0.566, the loss of GoogleNet is 0.936, and the loss of MobileNet is 0.626. The change curve of the validation set accuracy is shown in Fig. 9.

Figure 9. The change curve of the validation set accuracy.

The weights of the best trained round of the four models were selected for testing and the classification accuracy results were obtained: the Accuracy of AlexNet was 0.889, the Accuracy of Vgg was 92.6%, the Accuracy of GoogleNet was 85.2%, and the Accuracy of MobileNet was 96.3%. MobileNet showed the best prediction on the test set.

5. Conclusion

In this study, we introduced four deep learning models, AlexNet, Vgg, GoogleNet and MobileNet, to classify pneumonia images and compare their effectiveness and accuracy. In terms of the convergence speed of the loss function, GoogleNet performs well and reaches the convergence state in only the 3rd epoch, while the other three models gradually stabilise. The final loss values show 0.426 for AlexNet, 0.566 for Vgg, 0.936 for GoogleNet, and 0.626 for MobileNet. Further, after selecting the best performing round of training weights among the four models for testing, the classification accuracy results were obtained; AlexNet achieved 88.9% accuracy, Vgg 92.6%, GoogleNet 85.2%, and MobileNet up to 96.3% accuracy. This indicates that MobileNet shows the most superior prediction on the test set with more powerful classification ability. Therefore, the comprehensive experimental results show that MobileNet performs well in the pneumonia image classification task, not only in terms of faster loss function convergence, but also far ahead of other models in terms of accuracy. Therefore, in practical applications, MobileNet can be considered as the preferred model for the pneumonia image classification task to obtain better classification performance and prediction results.