Dermatology's artificial intelligence story has moved from narrow convolutional neural networks (CNNs) trained on single tasks to foundation models — large, general-purpose systems that promise broader clinical reach, according to a new Medscape analysis titled "From Convolutional Neural Networks to Foundation Models: A New Era for Artificial Intelligence in Dermatology."

The framing matters for clinicians who have followed AI in skin disease over the past decade. The first generation of dermatology AI was built almost entirely on CNNs: deep-learning architectures trained to do one thing well, most often image classification for skin-cancer detection from dermoscopic or clinical photographs.

What distinguished the CNN era?

CNN-based systems dominated the published literature on dermatology AI. Their typical design pattern was consistent:

• A curated dataset of labeled lesion images
• Supervised training on a single diagnostic task, such as melanoma versus nevus classification
• Performance benchmarks reported against dermatologists or board-certified readers

Those systems demonstrated that machines could match or approach expert-level accuracy on narrow visual tasks. But each model did one job. A CNN tuned to distinguish melanocytic lesions could not evaluate inflammatory disease, triage rashes, or generate clinical documentation without being retrained from scratch.

That limitation shaped how the technology could actually enter practice. A single-task classifier maps poorly onto a specialty where a single consultation may span skin-cancer screening, inflammatory disease, and cosmetic concerns in the same visit.

What do foundation models change?

Foundation models, the class of large-scale AI systems that includes today's multimodal and large language models, are trained on vast, heterogeneous datasets rather than task-specific image collections. The Medscape analysis positions them as the beginning of a new era for the field.

The practical implications for dermatology follow from the architecture:

• Generalization. One pretrained model can be adapted — fine-tuned or prompted — to many dermatologic tasks instead of one.
• Multimodality. Beyond images, these systems can work with clinical text, history, and structured data, opening pathways to decision support that goes beyond visual pattern recognition.
• Reduced data hunger per task. Adapting a foundation model to a new application can require far less labeled dermatologic data than training a CNN from zero.

For a specialty that has struggled with dataset bias — collections skewed toward lighter skin phototypes and academic referral centers — the adaptability of foundation models raises the possibility of broader, more flexible deployment. Whether that promise translates into equitable performance across skin tones and care settings remains a validation question, not a settled result.

Why the transition is more than an architecture swap

The shift from CNNs to foundation models also changes the questions clinicians should ask of the evidence. CNN-era studies could report a single metric — sensitivity and specificity for one classification task — and the clinical claim was bounded accordingly.

Foundation models invite broader claims: triage support, documentation, patient communication, differential generation. Each of those use cases carries its own risk profile, its own failure modes, and its own regulatory pathway. A model that drafts a reasonable note is not the same problem as a model that decides which pigmented lesion warrants biopsy.

The Medscape piece arrives at a moment when dermatology is arguably the visual specialty most exposed to AI-driven change. Skin disease is diagnosed primarily by inspection; imaging is cheap and ubiquitous; and the volume of dermatologic photographs — clinical, dermoscopic, and patient-generated — grows continuously. Those conditions made dermatology a proving ground for CNNs, and they now make it an early test case for what general-purpose models can and cannot do responsibly.

What comes next

The open questions the new era must answer are familiar in shape but larger in scope: external validation in real-world cohorts, performance across the full range of skin phototypes, integration into clinical workflows rather than parallel demonstrations, and clarity on liability when a foundation model contributes to a missed melanoma or an inappropriate treatment course.

The trajectory Medscape outlines suggests the field's next chapter will be defined less by whether AI can read a lesion image — CNNs largely settled that — and more by how safely general-purpose models can be woven into the complex, high-stakes decisions that make up actual dermatologic practice.