Image based Lung Disease Detection: Comparing Swin Transformers and ConvNets

Each year, chest diseases prematurely claim 4 million lives. They are very difficult to detect even for experienced radiologists. Thus, there is a need to develop artificial intelligence-based detection systems to aid in the diagnosis of chest diseases. The paper proposes a Swin Transformer based approach to classify images into one or more out of 14 chest diseases. The model is trained using a large number chest X-ray images from the Chest X-ray14 dataset. The proposed architecture uses multiple projection heads for improved results. The proposed model is compared to existing solutions based on Convolutional Neural Networks available for the problem and produces an AUC of 0.801 with state-of-the-art performances in four pathologies.