An Improved Method for Pneumonia Image Classification Based on CoAtNet

The most effective way to identify pneumonia is through the use of Chest X-Ray (CXR), but accurately interpreting these images can pose a formidable challenge for even seasoned professionals. This difficulty can potentially result in patients receiving incorrect treatment. To address this issue, a novel neural network called CoAtNet is adopted in this study to aid doctors in achieving more precise pneumonia diagnoses. This network structure combines the Convolutional Neural Networks and Vision Transformer. It better generalization obtained from convolution blocks and higher model capacity by applying Transformer mechanism. Totally, 5216 photos from Kaggle are used as training set and 624 photos are used as test set. At the beginning of each epoch, 30% photos in training set are chosen to be inputted into validation set randomly. By applying data augmentation to the training set, the model is less likely to be overfitting and has better generalization. When compiling the model, Adam is chosen to be the optimizer and Categorical Cross Entropy is adopted to be the loss function. After 10 epochs’ training, both the validation set and training set converge sufficiently. After running the model on the test set, the model shows great performance with a high accuracy of 86.70%, high precision of 83.44% and 90.22% F1-Score.