CoFiDA-M: Concept-Aware Feature Modulation for Cross-Domain Adaptation with Image-Only Inference

Researchers introduced CoFiDA-M, a cross-domain skin cancer screening method that uses MONET concept probabilities during training.

A teacher model modulates visual features with concept-aware FiLM-style editing, then distills the improved representation into an image-only student.

The student keeps inference practical by using only images at test time, while benefiting from privileged concept information during training.

CoFiDA-M outperforms prior approaches on multi-dataset benchmarks, improving melanoma and broader skin cancer screening across domains.