Debiasing strategies for skin lesion analysis encompass methods aimed at preventing deep neural networks from exploiting spurious correlations between visual artefacts and diagnostic labels in dermoscopy images. Bissoto et al. (2020) provided the most systematic analysis of this problem, identifying 7 artefact types and evaluating the state-of-the-art debiasing method Learning Not To Learn (LNTL).
The 7 Artefact Types
- Dark corners (vignetting) — shadows at image edges from the dermoscope
- Hair — patient hair overlying the lesion
- Gel borders — edges of contact gel used in dermoscopy
- Gel bubbles — air bubbles in the contact gel
- Rulers — measurement rulers placed next to lesions
- Ink markings/staining — surgical markings or dye on the skin
- Patches — adhesive patches applied to the skin
Debiasing Approaches
Data-level approaches:
- Normalized background: Replace background pixels with pixel-average training image to eliminate background information (Bissoto et al., 2020)
- Trap sets: Construct train/test splits with amplified and reversed artefact-label correlations to test bias reliance
- Data augmentation: Add or remove artefacts during training to reduce their predictive value
Model-level approaches:
- Learning Not To Learn (LNTL) (Kim et al., 2019): Feature extractor with reversed gradients from bias classification heads — found insufficient by Bissoto et al. for entangled artefacts
- Domain confusion loss (Alvi et al., 2018): Classifier per known bias domain with confusion loss to discourage bias encoding
- Spatial specificity (Saab et al., 2022): Use finer spatial annotations to exclude background artefacts from training signal
- segmentation-for-classification (Hooper et al., 2023): Train segmentation networks that inherently focus on pathology regions
- Certified unlearning (Towards Certified Shortcut Unlearning in Medical Imaging): Formally “forget” spurious associations with provable guarantees
Key Insights
- Individual artefact-label correlations are weak, but models exploit cumulative weak correlations
- Networks detect artefacts with 80-98% AUC even on heavily occluded images, indicating deep feature-level encoding
- Background removal alone is insufficient because artefacts overlap with foreground (hair, ink on lesion)
- Gradient-reversal approaches (LNTL) fail because artefact features are entangled with diagnostic features in learned representations
- The most promising direction is combining spatial constraints (segmentation) with formal guarantees (certified unlearning)