Accurate Thoracic Disease Classification via Ensemble Networks

Convolutional Neural Networks (CNNs) are being adapted to various computer-aided diagnosis applications, including recognizing thoracic diseases. To improve classification performance, recent solutions alter the structure of existing networks or require additional prior information for training. In this paper, we propose an ensemble network to accurately recognize thoracic diseases, without additional layers nor extra input data for training. To this end, we perform an exhaustive set of experiments involving various single CNN models for thoracic disease classification. We then form an ensemble from the most accurate classifier based on these experiments. Our approach achieves state-of-the-art average AUROC score of 79.32% on the ChestX-ray14 dataset, which is 2.08% higher than the previous best result. Additionally, we also attain the highest AUROC for 12 of the 14 classes. Our code and trained models are publicly available at https://github.com/arvention/ChestXRay14-Classification-PyTorch.