Related Experiment Video
Updated: Jun 29, 2025
06:08
Quantitative Visualization and Detection of Skin Cancer Using Dynamic Thermal Imaging
Published on: May 5, 2011
16.8K
From data to diagnosis: skin cancer image datasets for artificial intelligence
David Wen1,2, Andrew Soltan3,4,5, Emanuele Trucco6
1Department of Dermatology, Oxford University Hospitals NHS Foundation Trust, Oxford, UK.
Clinical and Experimental Dermatology
|March 29, 2024
Summary
Artificial intelligence (AI) for skin cancer diagnosis needs transparent, diverse datasets. Addressing dataset shifts and biases is crucial for equitable AI performance across all patient populations.
Area of Science:
- Dermatology
- Medical Imaging
- Artificial Intelligence
Background:
- AI, particularly deep learning, shows promise for skin cancer diagnosis.
- Development of AI models relies heavily on large, high-quality digital image datasets.
- Ensuring AI generalizability requires understanding dataset characteristics.
Purpose of the Study:
- To review datasets used in AI for skin cancer diagnosis.
- To highlight the importance of dataset transparency for AI evaluation.
- To discuss challenges and strategies for improving AI datasets.
Main Methods:
- Literature review of AI datasets for skin cancer diagnosis.
- Analysis of dataset curation challenges, including shifts and biases.
- Examination of strategies like federated learning and model-level bias mitigation.
Main Results:
- Dataset shifts due to demographics and methodologies impact AI performance.
- Limited representation of rare cancers and minority groups skews AI results.
- Lack of transparency hinders evaluation of AI generalizability.
Conclusions:
- Improving dataset transparency, representation, and interoperability is key.
- Federated and generative methods may enhance dataset diversity and privacy.
- Addressing dataset biases is essential for equitable AI in skin cancer diagnosis.
