AI in Clinical Dermatology? New Model Supports Diagnosis

Integration of Imaging Modalities

PanDerm is one of the first so-called foundation models—large, general-purpose AI systems trained on diverse datasets and designed to handle a broad range of clinical tasks. Unlike traditional models tailored to a single task (e.g., melanoma detection), PanDerm is built to flexibly address the diagnostic needs of clinical dermatology.

It utilizes four types of imaging inputs: clinical photographs, dermatoscopic images, histopathology slides, and full-body photos. As Associate Professor Zongyuan Ge from Monash University’s Faculty of IT explained, this integration of modalities offers a major advantage over earlier systems, which typically relied on just one type of data. “PanDerm was designed as a support tool for clinicians. It helps them interpret complex data and make informed decisions with greater confidence,” he added.

Higher Accuracy, Wider Range of Tasks

The model was trained on more than two million images from 11 international institutions and tested across a wide spectrum of clinical tasks—from skin cancer screening and recurrence/metastasis prediction, to skin type classification, mole counting, lesion segmentation, and longitudinal tracking. In each task, PanDerm achieved best-in-class performance, often using only 5–10% of the annotated data typically needed for training similar models.

Clinical evaluations showed that PanDerm improved the diagnostic accuracy of skin cancer by up to 11% for dermatologists and up to 16.5% for non-specialist physicians. It also helped detect high-risk lesions earlier than conventional examinations would allow.

Broad Clinical Applications

In a clinical setting, PanDerm functions as a decision-support system: it analyzes images commonly captured by physicians and generates probabilistic diagnostic suggestions. This capability is particularly useful for non-specialists, aiding in the interpretation of visual findings, tracking subtle changes over time, and assessing individual patient risk levels.

The model was developed through collaboration among experts from Australia, Europe, and Asia. According to Professor Harald Kittler of the Medical University of Vienna, this international diversity ensures PanDerm’s applicability across various healthcare systems.

One of PanDerm’s key strengths is efficiency: it maintains high performance even when trained with relatively small datasets, making it especially valuable in resource-limited regions.

Further Validation Needed

Despite its impressive research results, PanDerm will require additional validation before routine clinical deployment. The research team plans to implement standardized evaluation protocols across diverse demographic groups and clinical environments to ensure consistent and equitable performance.

“The multimodal approach allows us to approximate how an experienced dermatologist thinks—by synthesizing different visual inputs when making a diagnosis,” said Siyuan Yan, the study’s lead author from Monash University.

Editorial Team, Medscope.pro

Source:

Yan S., Yu Z., Primiero C. et al. A multimodal vision foundation model for clinical dermatology. Nat Med 2025 Jun 6, doi: 10.1038/s41591-025-03747-y.