Research and Development of the Convolutional Neural Network for Pneumonia Recognition in Radiographs

In this paper, we present the results and analysis of using the convolutional neural network (CNN) for pneumonia recognition in radiographs. The network architecture was developed, and the author's software implementation of the CNN model using TensorFlow / Keras framework is proposed. The classes in the dataset are imbalanced, and we used a combination of approaches to prevent overfitting problem such as increasing the cost of classification mistakes on the minority class, increasing the amount of data (data augmentation), exclusion of a certain percentage of neurons during training (dropout) and others. The model gave us the next results: accuracy is 0.91, precision is 0.9, recall is 0.97, roc-auc is 0.96. The obtained results allow making a conclusion about the effectiveness of using the developed CNN model for pneumonia recognizing.