DR Concept Density Explorer

Select an image from the browser

◀ Collapse

Apply Attribution Mask: zero out the top-X% of attribution pixels (computed with Integrated Gradients) for the selected concept, then re-run the model to measure how the prediction changes. This helps estimate how important those pixels are for the model's DR decision.

Run the attribution-based counterfactual evaluation across many images. The batch report shows how masking high-attribution pixels affects predicted DR class and class probability. Important fields:

• Mean class shift: average change in predicted class (masked - original).
• Fraction decreased: proportion of images where the predicted class decreased after masking.
• Per-level mean shift: mean class shift broken down by original DR level.