Dermatology’s Next AI Chapter

I remember when the landmark Esteva et al. paper was published in Nature back in 2017. It was one of the first major studies to show that deep learning could be applied to medical imaging, successfully classifying skin cancer at a level comparable to board-certified dermatologists. Now, a new paper in Nature Medicine on the PanDerm foundation model illustrates the next chapter in this story, showing the evolution from that initial classification work to a more holistic, clinically-integrated approach.

💡 Multimodal Power in Practice
Instead of relying on a single image type, PanDerm was trained on over 2 million images from 11 clinical institutions across four distinct modalities. It learns to synthesize data just as a clinician does, using clinical photos for wide context, dermoscopic images for structural detail, total-body photography for patient-level patterns, and dermatopathology slides for ground-truth tissue analysis.

🏆 Performance on Real-World Clinical Tasks
The model’s performance was validated across a comprehensive suite of 28 different clinical benchmarks. These tests were designed to mirror complex, real-world tasks, such as predicting the risk of metastasis from a dermoscopic image, monitoring a lesion for subtle changes over time, and assessing a patient's overall photodamage from total-body photos. The model consistently achieved state-of-the-art results, often with just 10% of the labeled data others require.

🤝 Tangible Clinical Impact
The three reader studies provide concrete evidence of PanDerm's value. When assisting clinicians, it improved skin cancer diagnostic accuracy by 11%. For non-specialists evaluating 128 different skin conditions, it boosted diagnostic accuracy by 16.5%. Most impressively, in a longitudinal analysis, it identified early-stage melanomas 10.2% more accurately than clinicians by detecting suspicious changes in sequential images over time.

🏥 A Clearer Path to the Clinic
This paper provides a strong foundation for clinical translation through its extensive validation. The path forward now involves addressing the limitations the authors rightly identify, such as expanding the model's scope to more of the 1,000+ known dermatological conditions and conducting more comprehensive fairness assessments. Beyond that, the gold standard for adoption is a series of prospective RCTs, each designed to validate a specific clinical application—like early cancer detection or risk stratification—in a real-world setting.

PanDerm serves as an excellent blueprint for developing powerful, multimodal foundation models in other medical specialties. It effectively bridges the gap between AI research and practical clinical support by integrating diverse data sources to tackle a wide array of real-world clinical challenges.

Paper: https://www.nature.com/articles/s41591-025-03747-y