An empirical study for the early detection of Mpox from skin lesion images using pretrained CNN models leveraging XAI technique

AI can help detect monkeypox early, but isn't perfect: The study found that pre-trained AI models achieved 95% accuracy on binary classification (mpox vs. non-mpox), suggesting AI could be a useful diagnostic tool, though the remaining 5% error rate means human oversight remains important.

Explainability matters for medical AI: The researchers used Grad-CAM, an explainability technique, to show doctors which parts of skin lesion images the AI was analyzing when making decisions. This transparency is crucial for building trust in AI-assisted medical diagnosis and helping clinicians verify the AI's reasoning.

Different AI models work better for different tasks: InceptionV3 performed best for simple binary classification (95% accuracy), while MobileNetV2 was superior for multi-class classification (93% accuracy), demonstrating that one-size-fits-all AI solutions don't work in medicine—models must be carefully matched to specific diagnostic needs.

Overfitting remains a practical challenge: Despite high accuracy scores, the study found that some models showed overfitting (performing better on training data than real-world validation data), highlighting that reported accuracy numbers can be misleading and real-world performance testing is essential before deploying medical AI.

Transfer learning reduces data requirements: By using pre-trained models adapted for mpox detection rather than building from scratch, researchers could work with limited medical imaging datasets—an important consideration for rare diseases where large training datasets are difficult to obtain.