Pretraining a dermoscopic vision foundation model via an enhanced self-distillation framework

Dermoscopy is widely used for diagnosing skin tumors, yet automated analysis remains constrained by scarce high-quality annotations and task-specific supervised models. Existing weakly supervised or unsupervised approaches mainly rely on language supervision, teacher–student distillation, or self-distillation. Language supervision requires high-quality text annotations, whereas teacher–student methods depend on external models that may lack dermoscopy-specific knowledge. Conventional self-distillation avoids these requirements but does not explicitly address lesion-centered spatial distributions, unreliable supervision from ambiguous views, or local structural consistency. To address these limitations, we present ROSE, a visual foundation model pretrained on more than 110,000 dermoscopic images. ROSE integrates Gaussian resized cropping to concentrate global and local views on lesion-relevant regions, entropy-weighted distillation to reduce noisy supervision from semantically ambiguous views, and dense contrastive regularization to preserve local feature consistency without compromising global representations. We evaluate ROSE against state-of-the-art methods on multiple public datasets for classification, few-shot learning, segmentation, and skin-tone-stratified analysis. ROSE consistently achieves competitive or superior performance. On HAM10000, it obtains a weighted F1-score of 0.859 and an AUROC of 0.973 for classification, as well as a Dice similarity coefficient of 0.886 and a Jaccard index of 0.833 for segmentation. In a reader study, ROSE achieves a balanced accuracy of 0.809, providing diagnostic performance competitive with human experts while offering substantially greater inference efficiency. Ablation studies confirm the complementary contributions of all components. These findings demonstrate that ROSE learns robust, transferable representations across diverse dermoscopic tasks and has potential for clinical translation.