DUNEScan

DUNEScan estimates prediction uncertainty and visualizes model evidence for skin cancer classification using convolutional neural networks (CNNs).

Key Features:

• Uncertainty Estimation: Computes mean and variance estimates from multiple CNN models to quantify prediction confidence for skin lesion classifications.
• Model Comparison: Compares outputs from various CNN models on the same input image to evaluate consistency and relative performance.
• Grad-CAM Visualization: Applies Gradient-weighted Class Activation Mapping (Grad-CAM) to highlight image regions that most influence CNN predictions.
• UMAP Visualization: Uses Uniform Manifold Approximation and Projection (UMAP) to project high-dimensional features into a lower-dimensional space to visualize the classification manifold and assess similarity to lesions in the ISIC database.

Scientific Applications:

• Dermatological research: Evaluate and refine CNN architectures for skin cancer detection by quantifying uncertainty and comparing model outputs.
• Clinical decision support: Provide prediction confidence measures and visual explanations to inform diagnostic interpretation of skin lesions.
• Dataset similarity assessment: Assess how closely individual images resemble known lesions from the ISIC database to inform model reliability on specific cases.

Methodology:

Computes mean and variance of predictions across multiple CNNs, generates Grad-CAM saliency maps, and projects high-dimensional model features with UMAP to compare cases against the ISIC database.