Privileged Lesion-Context Relational Distillation for Mask-Free Skin Lesion Classification

Automated skin lesion classification often benefits from segmentation masks, which help models focus on the lesion rather than background artifacts. However, requiring these masks or an auxiliary segmentation model during clinical deployment increases computational complexity and limits practical utility. This paper asks: can we use lesion masks as privileged information during training to guide a student model, such that the student achieves high diagnostic accuracy using only raw dermoscopic images at inference?

The authors propose Privileged Lesion-Context Relational Distillation (PLCRD). The framework employs a dual-branch 'privileged teacher' that processes both the original image and a mask-guided lesion view to learn diagnostic representations. An 'image-only student' is then trained to mimic the teacher's knowledge through three mechanisms:

1. Logit Distillation: Transferring diagnostic distributions to capture inter-class uncertainty.
2. Attention Distillation: Guiding the student to focus on the same spatial regions as the teacher.
3. Relational Distillation: Decomposing features into lesion-focused and contextual embeddings, then transferring their relational organization (e.g., inter-lesion similarity and lesion-context affinity) rather than raw feature vectors.

Because the relational distillation focuses on the organization of embeddings rather than direct feature matching, the teacher and student can have different, heterogeneous architectures.

PLCRD was evaluated on the HAM10000 dataset and externally validated on ISIC 2018. The model achieved a macro-F1 score of 0.773 on HAM10000 and 0.732 on ISIC 2018. The results demonstrate that the student successfully internalizes mask-informed diagnostic structures, allowing it to perform effectively without needing masks at deployment. This approach provides a more practical and interpretable alternative to standard mask-dependent classification pipelines.