A Diagnostic Pipeline for Multi-Class Classification of Chest X-Ray Images Using Knowledge Distillation and Semi-Supervised Segmentation

Artificial Intelligence (AI)-powered medical diagnosis systems can play a vital role in healthcare by enabling early detection of COVID-19, pneumonia, and tuberculosis. While chest X-ray diagnostics continue to be an essential tool in clinical assessment, timely identification is crucial as it facilitates prompt medical intervention. In this research, we present a novel AI-driven diagnostic system that combines three operational components: applying knowledge distillation methods with semi-supervised segmentation procedures and multi-class classification features to achieve better accuracy rates and operational effectiveness. Our research employed 20,000 four-class chest x-ray images for evaluation following preprocessing operations involving downsampling and augmented techniques. Using semi-supervised segmentation, the UNet model achieved a Dice score of 98% and an IoU of 97%, effectively isolating lung regions to create a refined dataset. We then trained both subsets of the dataset - segmented and original - on CNN models such as InceptionV3,Vgg16, ResNet-50, DenseNet-121, and their Attention enhanced variants, followed by ensembled CNNs and CLIP-ViT-L/14. Of these, the ViT model demonstrated the highest score through training on segmented data at 97% accuracy, establishing it as the most suitable teacher model for knowledge distillation. This led to CLIP-KDViT, which distilled knowledge into the MobileNetV2 student model, achieving 99% accuracy on segmented images and classifying 3,000 unseen test images with 97.2% accuracy and a 2.8% misclassification rate. The optimized pipeline was then integrated into our web application, Respire Check, which incorporates Grad-CAM for explainability, demonstrating the full functionality of our AI-driven X-ray diagnosis system.