Computer-assisted lung diseases detection from pediatric chest radiography using long short-term memory networks

Abstract

Pneumonia is one of the most common lung diseases for children. Computer-aided diagnosis (CAD) can be utilized to achieve automated lung disease detection using artificial intelligence techniques. This paper proposes a hybrid deep model of pretrained convolutional neural network (CNN) and long short-term memory (LSTM) networks to identify bacterial pneumonia versus viral pneumonia diseases in pediatric chest radiography. Extensive experiments have been conducted in two main scenarios, namely, pneumonia versus normal cases and bacterial pneumonia versus viral pneumonia infections. Evaluation of the results showed that the proposed deep classifiers achieved an accuracy and area under curve (AUC) of 98.6% and 99.9%, respectively, for normal versus pneumonia scenario. Meanwhile, the obtained accuracy and AUC are 92.3% and 94.5%, respectively, for bacterial pneumonia versus viral pneumonia classification. Compared to deep models in previous studies, our hybrid classifiers of CNN and LSTM networks have relatively better performance to assist radiologists during the scanning procedure of the child's lungs.