Answers are grounded in this paper and cite their sources.
Text-derived class prototypes offer a way to incorporate clinical descriptions into dermoscopic classification, but their contribution is difficult to assess when datasets overlap or text inputs encode sample labels. We evaluated a CLIP-inspired classifier using seven fixed class descriptions encoded by frozen Bio-ClinicalBERT, a trainable text projection, a Vision Transformer image encoder, and optional patient metadata. Four configurations crossed text-derived versus learned prototypes with metadata use. Contrastive weights were selected by validation performance: λ =1 for text-prototype models and λ =0 for learned-prototype models. Models were developed on lesion-grouped HAM10000 partitions and evaluated on 14,268 external ISIC 2019 images after removing 10,015 shared images and 621 precautionary exclusions. MetaBlock, BiomedCLIP zero-shot and linear-probe classifiers, and a metadata-only model were comparators. Trainable final configurations used three seeds; uncertainty was estimated by lesion-clustered bootstrap. The text-prototype configuration with metadata exceeded the learned-prototype configuration with metadata in external balanced multiclass accuracy (0.388 versus 0.364; Δ =+0.024 , 95% CI 0.005–0.043, Holm-adjusted $$p=0.032$$ ), but this comparison combines prototype source and contrastive weight. Adding metadata within the text-prototype family did not yield a statistically significant improvement ( Δ =-0.015 , 95% CI $$-0.043$$ to 0.014). The image-only text-prototype model had the highest observed external balanced accuracy (0.402); the BiomedCLIP linear probe achieved 0.381. Full-model balanced accuracy declined from approximately 0.63 internally to 0.39 externally, melanoma recall was 0.32, and maximum-softmax confidence did not reliably identify out-of-taxonomy squamous cell carcinomas. The evaluated text-prototype configurations performed better than their learned-prototype counterparts, but matched-objective comparisons are needed to isolate the contribution of prototype source. These results do not support clinical deployment.
No takes yet. Share an insight, caveat, or question.
NR et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: