Data-Efficient Vision Transformers for Multi-Label Disease Classification on Chest Radiographs

Abstract Radiographs are a versatile diagnostic tool for the detection and assessment of pathologies, treatment planning or navigation localization purposes in clinical interventions. However, their interpretation by radiologists can be tedious error-prone. Thus, wide variety deep learning methods have been proposed to support interpreting radiographs. Mostly, these approaches rely on convolutional neural networks (CNN) extract features from images. Especially multi-label classification pathologies chest radiographs (Chest X-Rays, CXR), CNNs proven well suited. On Contrary, Vision Transformers (ViTs) not applied this task despite high performance generic images interpretable local saliency maps which could add value ViTs do convolutions but patch-based self-attention contrast CNNs, no prior knowledge connectivity is present. While leads increased capacity, typically require an excessive amount training data represents hurdle medical domain as costs associated with collecting large sets. In work, we systematically compare different set sizes evaluate more data-efficient ViT variants (DeiT). Our results show that while between par small benefit ViTs, DeiTs outperform former if reasonably available training.