DermL2V: Unifying Heterogeneous Dermatology Texts Via LLM-Driven Encoding.

Dermatology diagnosis requires text encoding models that map unstructured descriptions to clinical semantic representations. Using public dermatology corpora to train such models remains highly challenging due to the lack of a specialized dataset offering unified semantic triplets as well as the difficulty of adaptively weighting clinically informative tokens and enhancing discrimination between clinically different but lexically similar descriptions. In this paper, we first construct DermSCL, a dermatology-specific dataset that organizes heterogeneous dermatology texts into 140,278 structured anchor-positive-negative semantic triplets and provides a unified supervision signal for text representation learning. Pretrained on DermSCL, we propose DermL2V, an LLM-based dermatology text encoder that improves representations by adaptively aggregating clinically relevant token information within a text and distinguishing semantic differences across different texts. First, we introduce a novel Clinical-Aware Relevance Aggregation module, where multiple relevance heads learn complementary token-level relevance patterns under dermatology-specific contrastive supervision to strengthen the representation. Second, we design a Boundary-Aware Semantic Mixup Strategy, which constructs intra- and inter-triplet near-boundary supervisory signals to enhance discrimination between clinically distinct yet lexically similar descriptions. Together, these two components synergistically improve the semantic robustness and discriminative capacity of the learned representations for dermatology. Furthermore, we evaluate DermL2V on important clinical tasks, including dermatology knowledge retrieval, embedding space analysis, concept classification and cross-modal retrieval. Across four out-of-domain dermatology retrieval datasets, DermL2V achieves the best average NDCG@10 and Recall@10 scores while delivering the best performance on the other downstream tasks. These results show DermL2V's potential as a dermatology-specific semantic encoder, providing reliable representations for retrieval-centered clinical applications. The project source code is publicly available here.