AI-driven multi-class classification of X-ray images for disease detection

Cheang, Jia Zhi (2025) AI-driven multi-class classification of X-ray images for disease detection. Final Year Project, UTAR.

PDF Download (4087Kb)

Abstract

There are critical challenges in medical imaging diagnostics, including human error in radiological analysis, slow diagnostic processes, and radiologist shortages. This project developed an AI-driven multi-class classification system for chest X-ray disease detection. The research employed a DenseNet121 architecture with transfer learning to classify 13,482 chest X-ray images into five categories: COVID-19, pneumonia, tuberculosis, lung opacity, and normal cases. The methodology implemented a two-stage progressive unfreezing strategy, comprehensive data preprocessing with augmentation techniques, and class weight balancing to address dataset imbalances. Training utilized TensorFlow and Keras frameworks with GPU acceleration, incorporating early stopping and learning rate reduction callbacks for optimization. Gradient-weighted Class Activation Mapping (Grad-CAM) was integrated for AI interpretability, and a comprehensive Streamlit dashboard was developed featuring real-time processing capabilities. The DenseNet121 model achieved exceptional performance with 94.52% validation accuracy, 94.7% test accuracy, 95% precision and recall, and 0.99 AUC score across all disease categories. The system successfully demonstrated clinical-grade interpretability through visual attention mapping and deployed as a functional web application with hospital finder and weather health monitoring features. This research establishes a foundation for AI-assisted medical diagnosis, potentially improving healthcare accessibility and diagnostic reliability while maintaining transparency in AI decision-making processes for clinical deployment.

Actions (login required)