DeiT-GAN: a segmentation-driven transformer framework for multi-class chest X-ray disease classification

Chest diseases pose a significant health challenges and requires earlier diagnosis. Chest X-ray (CXRs) images are globally utilized due to their accessibility and cost-effectiveness, but physical intervention is time-consuming and error-prone. Recent improvements in Artificial Intelligence (AI) have enabled faster and more accurate automated disease classification. To address the challenges of limited data and longer training time, this research employs three stage deep learning pipeline for chest disease classification:(1)Bounding box guided U-Net based segmentation to isolate disease affected region, reducing irrelevant background noise and focusing on the clinically significant area (2) Conditional Generative Adversarial Networks (cGAN) and conventional geometric transformation based augmentation technique to address the issue of limited training data and improve the model generalization, (3) attention-based classification mechanism by incorporating transformer-based models which effectively captures long range dependencies within the affected region. This method has been evaluated on two Datasets comprising 4800 images, which were generated from 880 images selected from the NIH Chest X-ray 14 data source across 8 classes through traditional and cGAN-based augmentation techniques. The U-Net + cGAN + DeiT achieves 93.23% accuracy, 93.53% precision, 93.23% recall, 93.28% F1-Score, and requires 98.08sec training time, illustrating its effectiveness in chest diseases.