Expert Insights: Artificial Intelligence and Dermatology

During the third day of Spring Dermatology Week 2024, Roxana Daneshjou, MD, PhD, an artificial intelligence (AI) researcher and board-certified dermatologist, delivered a comprehensive talk on the intersection of AI and dermatology during her session, “AI in Dermatology: Pros and Pitfalls.”

She began by outlining the learning objectives, which included reviewing computer vision models, understanding biases, exploring pitfalls, and discussing large language models.

Computer vision models, she explained, allow computers to analyze visual inputs and make predictions. These models primarily use deep learning, particularly supervised learning, where computers learn from labeled examples. Dr Daneshjou emphasized the importance of data in training these models, highlighting the need for large datasets and prospective testing for accurate evaluation.

One major concern in AI dermatology is biases, especially in datasets. Dr Daneshjou discussed the underrepresentation of diverse skin tones in datasets, which can lead to biased outputs in AI algorithms. She stressed the importance of diverse datasets for training models to ensure fair and accurate results.

To address the lack of diversity in dermatology datasets, Dr Daneshjou and her team developed the Diverse Dermatology Images dataset. They found that algorithms trained on diverse images performed better across different skin tones compared to those trained on predominantly white skin images.

Despite potential biases, Dr Daneshjou highlighted the promise of computer vision in dermatology. She discussed how AI algorithms could improve clinical workflows by enhancing the quality of telemedicine images and aiding in melanoma diagnosis.

Transitioning to large language models, Dr Daneshjou explained that these models learn from vast amounts of text data. Fine-tuning and prompt-based learning further refine these models to generate human-like language.

She cautioned against blindly trusting outputs from large language models due to their tendency for "hallucinations,” or incorrect responses. Dr Daneshjou emphasized the need for careful evaluation and validation of these models, especially in medical settings.

In conclusion, Dr Daneshjou emphasized the importance of trials and ongoing research to validate the utility and safety of AI technologies in dermatology. Despite challenges, she remains optimistic about the potential for AI-human partnerships to enhance clinical care and improve patient outcomes.

For more meeting coverage, visit the Spring Dermatology Week 2024 newsroom.

Maximize your experience by adding exclusive sessions available only on demand via the Premium Plus package.

Reference
Daneshjou R. AI in dermatology: pros and pitfalls. Presented at: Dermatology Week; May 8–11, 2024; Virtual.