Shortcut learning occurs when deep learning models exploit spurious statistical correlations in training data rather than learning the true causal features relevant to the task. In medical imaging, this manifests when models learn to associate background artefacts (e.g., surgical ink markings, rulers, gel bubbles, dark corners, chest tubes) with diagnostic labels, because these artefacts are correlated with certain pathologies in the training distribution.
The problem is particularly acute in segmentation-classification pipelines where coarse segmentation masks (e.g., bounding boxes) include background regions alongside the pathology of interest. The model can then learn to use background features as predictive shortcuts, since the coarse mask cannot distinguish pathology from co-occurring artefacts.
Shortcut learning has been formally modelled using a generative data model where the spurious feature mask S and the true pathology mask Y are causally linked through the global disease label y and artefact presence s, drawn from a joint distribution P_{y,s}. Under this model, training with coarse masks Y^(r) allows spurious overlap, enabling the model to learn S as a predictive shortcut for class k.
Key Details
- Saab et al. (2022) showed that increasing spatial specificity (from image-level labels to bounding boxes to segmentation masks) reduces reliance on shortcuts
- Towards Certified Shortcut Unlearning in Medical Imaging proved that certified unlearning provides formal guarantees on shortcut reduction via the global spurious mutual information metric
- In the ISIC melanoma detection benchmark, shortcuts include: dark corners, hair, gel bubbles, gel borders, rulers, ink markings, and patches
- The disjoint support assumption (spurious features do not physically overlap with the lesion) enables formal analysis without the conditional independence assumption required by prior work
- Bissoto et al. (2020) introduced controlled spurious correlation benchmarks for skin lesion analysis