XAI-Enhanced Hybrid CNN–Transformer Framework For Multi-Class Lung Disease Classification
TL;DR
The results suggest that the hybrid CNN-Transformer model provides a strong level of diagnostic accuracy and meaningfully understood visual rationale so it can serve as an excellent decision support mechanism for hospitals and radiologists in their daily operations.
Abstract
Chest radiograph images have become a critical research area for applying deep learning in radiological interpretation for the classification of pulmonary diseases. But, to achieve both high accuracy and good interpretability continues to be a major hurdle for many researchers. In this research, we offer a hybrid architecture that incorporates CNNs and Transformer techniques for classifying different respiratory diseases using chest radiograph images. The CNN component provides a mechanism to capture many of the fine, local details found in an image, while the Transformer provides a self-attentive mechanism to capture the overall context of an X-ray image. In addition, a range of approaches exist to improve overall performance of the CNN and Transformer architecture, including structured preprocessing, data augmentation and class balancing. All of these techniques will improve model learning performance and help to effectively manage class imbalance when dealing with imbalanced datasets. To make our model more transparent to users and clinically useful, we employed explainability methods like Grad-CAM and Attention Visualizations to provide users with evidence of the specific area in an X-ray where the model is basing its prediction, thereby providing a greater amount of trust on the part of radiologists in interpreting the model's output. Based on our findings from testing the 6 Classes Chest Xray dataset, the proposed system proved to achieve a very impressive final testing accuracy of 94.42%. It classifies tuberculosis and healthy patients particularly well, with precision, recall, and F1-scores of 0.99 and 0.97, respectively, but provides good performance across the other disease types too. Furthermore, confidence analysis of predicted labels exhibited that when there was an accurate prediction, the assigned probability score was usually much higher than the assigned probability score for an incorrect prediction. Thus, these results suggest that the hybrid CNN-Transformer model provides a strong level of diagnostic accuracy and meaningfully understood visual rationale so it can serve as an excellent decision support mechanism for hospitals and radiologists in their daily operations.