Mitigating hallucinations in synthesized clinical texts to improve multimodal deep learning for dermatology

Despite recent advancements in the development of foundation models and multimodal (MM) architectures in dermatology, their translation to clinical practice remains limited by the scarcity of large-scale multimodal (MM) datasets, as most publicly available resources are small, unimodal, and lack expressive clinical text. This paper investigates strategies to synthesize and exploit clinical notes paired with dermatological images to effectively train a MM architecture, focusing on solutions to limit the inherently hallucinated contents introduced by Large Language Models (LLMs) and to identify conditions under which synthetic clinical notes can be reliably leveraged.

The paper proposes a MM architecture trained on real dermatological images paired with LLM-synthesized clinical notes. We systematically evaluate different note generation strategies, including metadata-guided prompting, alignment of image representations with specific keywords, sentence-level filtering of clinical notes, network architectural designs. Experiments involve 16,000 image-note couples collected from six public datasets for model training and over 37,000 images from fifteen public datasets as external data for generalization assessment. Performance is assessed on cross-modal retrieval and zero-shot learning tasks to quantify robustness and generalization.

Results show that metadata inclusion into the prompts reduces the hallucinations within LLM outputs, providing more reliable notes. The resulting MM model trained with these notes show superior performance on multiple downstream tasks.

Synthesized clinical notes can be paired with real dermatology images under specific conditions, providing a valuable resource to develop foundation models that can help reduce the dermatologists' workload.