Secure Pulmonary Diagnosis using Transformer-Based Approach to X-Ray Classification with KL Divergence Optimization

Vatsala Anand¹Mohammed Shuaib²Irfanullah Khan³Mehran Ullah^4*Shadab Alam²

• ¹Akal University, Bathinda, India
• ²Jazan University, Jizan, Saudi Arabia
• ³King Fahd University of Petroleum & Minerals, Dhahran, Saudi Arabia
• ⁴University of the West of Scotland, Paisley, United Kingdom

Lung disease classification plays a significant part in the early discovery and diagnosis of respiratory conditions. This paper proposes a novel approach for lung disease classification utilizing two advanced deep learning models, MedViT and Swin Transformer. The models are evaluated on the Lung X-Ray Image Dataset that includes 10,425 X-ray images categorized into three classes: Normal with 3,750 images, Lung Opacity with 3,375 images, and Viral Pneumonia with 3,300 images. A series of data augmentation methods, including geometric and photometric augmentation, are combined to improve model performance and generalisation. The results illustrate that both MedViT and Swin Transformer accomplish promising classification accuracy. MedViT showing particular strength in medical image-specific feature learning due to its hybrid convolutional and transformer design. The impact of different loss functions is also examined, where Kullback–Leibler Divergence yields the highest accuracy and effectively handles class imbalance. The best-performing MedViT model achieves an accuracy of 98.6% with a test loss of 0.09, highlighting its potential for reliable clinical decision support. Whereas, the Swin Transformer model achieved 0.92 accuracy and 0.16 test loss.