ThesisResearchAI/ML 8-12 months
Explainable AI for Medical Image Diagnosis
A study of explainability methods (Grad-CAM, SHAP, attention) for deep models on medical images, evaluating faithfulness and clinical usefulness.
PythonPyTorchGrad-CAMSHAPCaptum
Find GitHub repos & source code →How to build it — step by step
- 1Models: Train strong classifiers on a medical-imaging dataset (e.g. chest X-ray/skin lesion).
- 2Explanations: Generate saliency/attribution maps with multiple methods.
- 3Faithfulness: Quantify explanation faithfulness with perturbation and deletion metrics.
- 4Clinical study: Assess whether explanations help expert decision-making and trust.
Key features to implement
- ✓Multiple explanation methods
- ✓Faithfulness metrics
- ✓Comparative evaluation
- ✓Clinical usefulness study
- ✓Failure-case analysis
💡 Unique twist to stand out
Propose and validate a quantitative metric that correlates explanation quality with downstream diagnostic accuracy.
🎓 What you'll learn
Explainable AI, medical imaging, evaluation methodology, and human-AI interaction.