Study on the Depth of Convolutional Neural Network for Pneumonia Detection

Abstract

Pneumonia is a dangerous but common disease to human. Due to the similar symptom between cold and pneumonia, it is essential to have an effective method to diagnose pneumonia. In this paper, a prevalent concept called computer vision was introduced to help detect pneumonia based on chest X-ray images. More specifically, a Convolutional Neural Network (CNN) was proposed to classify the images into normal class or pneumonia class. The image dataset was preprocessed by unifying, normalizing, shuffling and augmentation. To finally achieve a model with relatively reasonable and accurate prediction, four CNN models were built and compared. To keep them comparable and have certain similarity, a small architecture with one convolutional layer, one max pooling layer and one batch normalization layer was defined and piled up repeatedly. The channel numbers in the convolutional layer were increased by proportion. These four models were eventually evaluated on their accuracies both on training set and testing set and F1 scores on two categories. Experimental results indicated that the model with four convolutional layers had the best performance which achieved an accuracy of 0.8606 on test set. Also, the model had F1 scores of 0.86 on normal class and 0.88 on pneumonia class.