Pneumonia Image Classification Based on ResNet Model and Its Optimization

Chest X-Ray (CRX) is the most common method to detect pneumonia, but analyzing the images precisely is a challenge even for an experienced doctor. This paper introduces a new convolutional neural network (CNN) named ResNet to help doctors achieve more accurate pneumonia diagnoses. This classifying model is built based on thousands of CRX images obtained from the Kaggle dataset, which respectively belong to healthy people and pneumonia patients. The built-in ImageDataGenerator of Keras is utilized to augment the dataset and avoid over-fitting. Five different optimizers models are also introduced, including the Adagrad, RMSprop, Adadelta, Adam and Nadam models. Moreover, their classification performance is quantitatively analyzed from the aspects of training set accuracy, loss and convergence rate, testing set accuracy, loss and confusion matrix. Finally, Adam is identified as the most suitable optimizer for the classification of medical pneumonia images in the ResNet neural network structure with the highest training set accuracy (96%), the fastest convergence rate and 0.934 F1-Score.