AI for Medical Imaging & Diagnosis

This research develops advanced deep learning and explainable artificial intelligence (XAI) frameworks for medical image analysis, with applications in thoracic radiography and mammographic diagnosis. The work focuses on improving both predictive accuracy and clinical interpretability by integrating anatomically guided preprocessing techniques with state-of-the-art deep learning architectures, including convolutional neural networks and transformer-based models.

In chest X-ray analysis, a novel multi-channel representation is introduced, combining spatial and frequency-domain features alongside lung-region isolation to enhance COVID-19 detection. In mammographic imaging, fine-tuned vision transformer models, particularly the SWIN transformer, are investigated for high-precision classification of benign and malignant cases.

A central contribution of this research is the incorporation of explainability methods, such as Grad-CAM, to ensure that model decisions are transparent, clinically relevant, and aligned with radiological knowledge, thereby supporting trustworthy AI deployment in healthcare.

This research uses artificial intelligence to help doctors analyse medical images like chest X-rays and breast scans more accurately. It not only improves how well diseases such as COVID-19 and breast cancer are detected, but also shows doctors why the AI made a decision. This helps build trust and makes the technology safer and more useful in real medical settings.

A novel explainable deep learning framework for chest X-ray analysis achieved 95.3% accuracy, AUC of 0.983, and strong correlation performance, demonstrating reliable COVID-19 detection.
Multi-channel feature representation (combining spatial, texture, and frequency-domain information) significantly improves diagnostic performance over standard single-input approaches.
Explainability techniques (e.g., Grad-CAM) confirm that model predictions focus on clinically relevant lung regions, improving interpretability and trust.
Fine-tuned transformer-based models, particularly SWIN Transformer, achieved up to 99.9% accuracy in breast cancer classification, outperforming conventional CNN architectures such as ResNet and VGG.
Deep learning models demonstrate strong potential to reduce diagnostic errors and support early detection in both respiratory and oncological imaging.
The integration of accuracy + interpretability is critical for real-world clinical adoption of AI-driven diagnostic systems.