Fig. 6. Gradient-weighted class activation mapping (GradCAM) analysis for different types of artifacts with model “attention” heatmaps (yellow and red representing areas in patch used by model for classification): one representative patch with from a tumor class (Dataset 3) is shown.

Dark spots, fingerprints, synthetic threads and squamous epithelia in general do not generate new features and act by obfuscating patch content and existing features. During focus deterioration, even after misclassification (starting from focus level 4) to false negative benign category, roughly same regions are used for classification. On contrary, elastic deformations produce visible feature instability and volatility of patch regions used for classification implying potential of elastic deformations to generate new features.