Conditional Effect of Retrieval-Augmented Generation in Dermatologic Imaging: Dependence on Input Format
Year: 2026
Type: article
Abstract: Introduction: Dermatologic ultrasonography is an commonly utilized, operator-dependent imaging modality in clinical practice, yet the role of multimodal artificial intelligence in this domain remains insufficiently defined, particularly in the absence of structured clinical input. Objectives: To evaluate whether retrieval-based knowledge integration and structured radiologist reporting improve the diagnostic performance of a vision-capable large language model in dermatologic ultrasound for increasing their utility in clinical practice. Methods: Dermatologic ultrasound cases with histopathological reference standards were assessed under four clinically relevant conditions: image-only without retrieval, image-only with retrieval, image plus radiologic report without retrieval, and image plus radiologic report with retrieval. A vision-capable large language model was used without fine-tuning. Outputs were evaluated for most likely diagnosis accuracy and differential diagnosis performance. Results: Image-only performance was limited and did not significantly improve with retrieval augmentation. In contrast, the incorporation of structured radiologic reports resulted in a marked improvement in diagnostic accuracy and differential diagnosis generation. The addition of retrieval further enhanced performance, with the highest accuracy observed in the image-plus-report with retrieval condition, indicating a synergistic effect between structured input and knowledge augmentation. Conclusions: The effectiveness of artificial intelligence in dermatologic imaging, is primarily determined by input structure rather than retrieval alone. Retrieval-based methods provide meaningful benefit only when coupled with expert-derived semantic representations, supporting a clinically integrated, collaborative model of artificial intelligence use rather than standalone image interpretation.
Institutions: Ankara 29 Mayis State Hospital, Ankara Bilkent City Hospital
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: diamond
OpenAlex ID: W7214916110