Unsupervised Causal Prototypical Networks for De-biased Interpretable Dermoscopy Diagnosis
New AI model isolates skin disease features from confounding artifacts, boosting diagnostic trust and accuracy.
A research team from Zhejiang University and other institutions has published a breakthrough paper on arXiv titled "Unsupervised Causal Prototypical Networks for De-biased Interpretable Dermoscopy Diagnosis." The work introduces CausalProto, a novel AI architecture designed to solve a critical flaw in current medical imaging models: their tendency to learn spurious correlations from biased clinical data. For example, a model might incorrectly associate a specific brand of dermatoscope or a patient's skin tone with a disease, creating misleading visual evidence that erodes clinician trust. CausalProto reframes the problem within a Structural Causal Model (SCM) to purify the chain of visual evidence.
The technical core of CausalProto is an encoder trained with an Information Bottleneck constraint to perform strict, unsupervised orthogonal disentanglement. This forces the model to separate true pathological features (like lesion morphology) from environmental confounders (like imaging artifacts or demographic markers) into independent prototypical spaces. It then uses the learned dictionary of spurious features to perform a backdoor adjustment via do-calculus, effectively marginalizing out the environmental noise through efficient expectation pooling. The result is a system that, as demonstrated on multiple dermoscopy datasets, achieves superior diagnostic accuracy compared to standard black-box models while providing transparent, high-purity visual prototypes for its decisions—eliminating the traditional compromise between interpretability and performance.
- Uses causal inference (do-calculus) to remove bias from skin lesion image data, preventing models from learning spurious correlations.
- Achieves unsupervised disentanglement of disease features from confounders via an Information Bottleneck-constrained encoder, requiring no labeled bias data.
- Outperforms standard black-box models in accuracy while providing transparent visual prototypes, avoiding the typical interpretability-performance trade-off.
Why It Matters
Paves the way for trustworthy AI diagnostics by providing clear, unbiased visual evidence doctors can actually use, moving beyond opaque 'black box' predictions.