Article Impact Level: HIGH Data Quality: STRONG Summary of Nature Medicine. https://doi.org/10.1038/s41591-025-03693-9 Dr. Siyuan Yan et al.
Points
- Researchers developed PanDerm, a multimodal AI model, by training it on over two million skin images from eleven institutions across four different real-world imaging modalities.
- The model achieved state-of-the-art performance across 28 diverse clinical benchmarks, often outperforming existing systems while using only ten percent of the usually required labeled data.
- PanDerm demonstrated a superior ability for early melanoma detection through longitudinal analysis, outperforming clinicians by 10.2 percent in identifying concerning changes in lesions over time.
- The tool improved clinicians’ skin cancer diagnostic accuracy by 11 percent in reader studies and enhanced non-dermatologist providers’ differential diagnostic skills by 16.5 percent across 128 conditions.
- This advanced system is designed as a clinical support tool that integrates into workflows to help doctors interpret complex imaging and make more confident diagnoses.
Summary
This technical summary outlines PanDerm, a multimodal dermatology foundation model designed to address the complex diagnostic needs of clinical practice. The model was developed through self-supervised pretraining on a large-scale dataset comprising over 2 million real-world images of skin diseases. This extensive dataset was aggregated from 11 clinical institutions and encompassed four different imaging modalities, providing a comprehensive basis for the model’s learning. The objective was to create a single, versatile tool that can synthesize information from multiple visual sources, mimicking the diagnostic process of a skilled dermatologist.
The model’s efficacy was rigorously tested across 28 diverse benchmarks that reflect real-world clinical tasks. These evaluations included skin cancer screening, risk stratification, differential diagnosis for common and rare skin conditions, lesion segmentation, longitudinal patient monitoring, and the prediction of metastasis and prognosis. PanDerm achieved state-of-the-art performance across all evaluated tasks. Notably, the model demonstrated high data efficiency, often outperforming existing specialized models when utilizing only 10% of the typically required labeled data for training.
Three reader studies were conducted to validate its clinical potential, yielding significant results. For early-stage melanoma detection through longitudinal analysis, PanDerm outperformed clinicians by 10.2%. The model improved clinicians’ skin cancer diagnostic accuracy by 11% when used as a support tool for interpreting dermoscopy images. Additionally, PanDerm enhanced the differential diagnostic accuracy of non-dermatologist healthcare providers by 16.5% across 128 different skin conditions presented in clinical photographs. These results highlight PanDerm’s potential to improve diagnostic accuracy and patient outcomes across various clinical settings.
Link to the article: https://www.nature.com/articles/s41591-025-03747-y
References Piccinno, G., Thompson, K. N., Manghi, P., Ghazi, A. R., Thomas, A. M., Blanco-Míguez, A., Asnicar, F., Mladenovic, K., Pinto, F., Armanini, F., Punčochář, M., Piperni, E., Heidrich, V., Fackelmann, G., Ferrero, G., Tarallo, S., Nguyen, L. H., Yan, Y., Keles, N. A., … Segata, N. (2025). Pooled analysis of 3,741 stool metagenomes from 18 cohorts for cross-stage and strain-level reproducible microbial biomarkers of colorectal cancer. Nature Medicine. https://doi.org/10.1038/s41591-025-03693-9
