— The most common way to find out if someone has pneumonia, one of the most common causes of death around the world, especially in children under five, is through chest X-rays. But it is hard to spot because expert radiologists have to look for small patterns. Artificial intelligence (AI) provides a scalable solution for automating diagnosis via deep learning (DL) models. Though things are getting better, current methods still face two major problems. One is that they are using CNNs that only capture local features and may miss global features. The other is that they are using pre-trained models on natural image datasets such as ImageNet, which does not have the same context as medical imaging and hence is not as effective. To tackle these problems, we propose DAViT (Domain-Adapted Vision Transformer), a hybrid architecture of Vision Transformers (ViTs) and shallow CNNs with domain adaptation. The ViT extracts global features through self-attention, and the CNN extracts local features. We fine-tune the model on a large variety of chest X-ray data to narrow down the domain gaps. We evaluate DAViT on a real-world dataset of 5856 chest X-rays. The results show that DAViT achieves the best performance in pneumonia detection with 97% F1-score and 96% AUC. This outperforms twelve other methods that were used as a baseline. For classifying the types of pneumonia, DAViT achieves an F1 score of 81% and an AUC of 84% which is 25% to 74% better than other methods. We perform an ablation study to show that all three components, i.e., domain adaptation, ViT and CNN, are important, and together they bring an improvement of 21%. Finally, we run GradCAM on top of DAViT to generate heatmaps highlighting the most important regions for bacterial and viral pneumonia cases. These heatmaps can inform medical professionals to help them make decisions. The results suggest that DAViT could help physicians diagnose pneumonia by improving the accuracy and interpretability of the model. We release the training code and pre-trained models at https://github.com/awsmresearch/DAViT.
Mr.K.SATHISH et al. (Fri,) studied this question.