The Latest AI Developments in Dermatology

Among the medical specialties, dermatology has witnessed some of the most promising AI applications in recent years.

By John Halamka, M.D., Diercks President, Mayo Clinic Platform, Nneka Comfere, MD, Division Chair of Dermatopathology and Medical Director of Digital Health, Artificial Intelligence and Innovations (DHAI) in the Department of Dermatology at Mayo Clinic, and Paul Cerrato, MA, senior research analyst and communications specialist, Mayo Clinic Platform

Artificial intelligence has played an important role in dermatology for over a decade. In 2016, investigators analyzed clinical images, using a convolutional neural network (CNN) to detect melanoma. They concluded that the CNN was “superior in terms of diagnostic accuracy in comparison with the state-of-the-art methods” available at the time. Similarly, Andre Esteva and associates provided evidence to suggest that a CNN had value in diagnosing several skin cancers and found it was capable of classifying skin cancer “with a level of competence comparable to dermatologists.” They went on to state that using such neural networks have the potential to “extend the reach of dermatologists outside the clinic.” Despite weaknesses in the methodology used by Esteva et al, their work generated interest among clinicians and researchers worldwide.

Since these early investigations, several others have found that AI models have value in the diagnosis of a variety of skin disorders. Since the introduction of generative AI, the attention of clinicians and researchers has focused on foundation models and the subset of large language models. Recently, researchers from Australia and Austria developed a multimodal vision foundation for dermatology called PanDerm. Just as large language models like ChatGPT learn from reading millions of sentences, PanDerm learns by ‘seeing’ millions of medical images of skin conditions. Using self-supervised learning techniques, the model was trained on over 2 million images from 11 institutions. The team evaluated the model’s performance on a broad array of tasks related to several key clinical questions, including skin cancer detection, early melanoma detection, and differential diagnosis. Yan et al found that their model outperformed clinicians by more than 10% for detecting early-stage melanoma and improved differential diagnosis by more than 16% amongst nonspecialists.

Mayo Clinic has also invested time and resources into a variety of AI models to improve the diagnosis and treatment of skin disorders. The Department of Dermatology has a Digital Health, Artificial Intelligence and Innovations (DHAI) Program with a vision to use data and technology to deliver accessible, high quality and cost-effective dermatologic care. It consists of five pillars: Cutaneous Oncology and Dermatologic Surgery; Data Management; Digital Dermatopathology; Skin Imaging; and Teledermatology, within which multiple flagship projects led by clinicians in the practice are conducted.

In addition, a cross-enterprise team at Mayo Clinic, in partnership with Aignostics GmbH, a computational pathology company, was instrumental in the creation of Atlas, a pathology foundation model they developed in collaboration with Charité – Universitätsmedizin Berlin, a German University Hospital. The model used over 1 million whole slide images covering different types of tissue, stains (hematoxylin and eosin, special and immunohistochemical stains), across multiple magnifications and using different slide scanners, to achieve state-of-the-art performance in digital pathology for tasks like identifying diseases in tissue samples.

DHAI is leveraging the Atlas foundation model (FM) as the foundation for task-specific dermatopathology AI. Built on the Atlas foundation model, DHAI created an AI-powered dermatopathology triage system that automatically sorts dermatopathology slides into broad diagnostic categories (four classes), giving pathologists an early preview that speeds interpretation, streamlines workflow, and enhances diagnostic efficiency. Ongoing work includes expansion into more diagnostic classification tasks specific to dermatopathology.

PanDerm, the aforementioned multimodal vision foundation model, also has potential to transform dermatology. It presents an opportunity for DHAI to leverage existing Mayo Clinic Platform data to build a Mayo Clinic dermatology foundation model integrating whole slide images, clinical and dermoscopic images, and clinical information in the EHR. Such a model would unify image-based and clinical context, supporting comprehensive patient centered AI decision support from triage and differential diagnosis through longitudinal monitoring and prognostication.

Dermatology, like other early adopters of AI, has taught us that a human practicing with AI is a stronger diagnostician than a human without AI. AI can bring expert knowledge, analysis of features invisible to the human eye, and second opinion oversight to the examination of skin lesions. The cost effectiveness and competitive margin are key differentiators for medical practices that adopt AI.