| Title |
Enhancing Chest X-ray Classification withMulti-class Token Hybrid Transformers |
| Authors |
이창민(Changmin Lee) ; 신호경(Hokyung Shin) ; 남우정(Woo-jeoung Nam) |
| DOI |
https://doi.org/10.5573/ieie.2024.61.11.69 |
| Keywords |
Chest X-ray; Vision transformer; Multi-label classification; Deep learning |
| Abstract |
Chest X-rays are commonly used medical imaging techniques for diagnosing and detecting lung abnormalities. Computer-aided diagnosis (CAD) systems based on deep learning, especially convolutional neural networks (CNNs), have proven valuable in assisting doctors with these diagnoses. In this study, we propose a novel model, the Multi-Class Token CheXFormer (MCTCheXFormer), which integrates CNNs and transformers to enhance chest disease classification performance. This model leverages both convolutional operations and self-attention mechanisms to effectively capture local and global features in chest X-rays. MCTCheXFormer combines a transformer with CheXNet?a CNN model known for its strong performance in chest X-ray classification. Here, the transformer's class token is extended from a single-class token to a multi-class token, enabling it to learn and differentiate class-specific features and improve multi-label classification. Additionally, we introduce an Iterative Class Token Weighting (ICW) technique, which applies the confidence scores generated by CheXNet as attention weights to the multi-class tokens, further enhancing classification performance. The model’s transformer is based on the Pyramid Vision Transformer (PVT) and consists of four stages. A token refinement module is added to ensure that the dimensions of the patch tokens generated at each stage align with those of the multi-class tokens. We evaluated the proposed MCTCheXFormer on the ChestX-ray14 dataset, which provides class labels for 14 chest diseases, comparing it with existing CNN- and transformer-based models. MCTCheXFormer outperformed the other models, highlighting the potential of transformers for accurate and efficient chest disease diagnosis. |