A methodology for developing dermatological datasets: lessons from retrospective data collection for AI-based...

Area of Science:

• Dermatology
• Medical Informatics
• Artificial Intelligence

Background:

• The advancement of artificial intelligence in dermatology necessitates standardized, high-quality datasets.
• Current dermatological datasets lack uniformity in development, hindering AI integration.
• There is a critical need for multimodal data, defined metadata standards, and inclusive population representation.

Purpose of the Study:

• To propose a practical methodology for creating structured dermatological datasets for skin tumor classification.
• To address the scarcity of health data by incorporating multimodal information and diverse populations.
• To establish a reproducible and scalable framework for dataset creation.

Main Methods:

• A four-stage methodology was developed: IRB approval and clinical information analysis, data recording and structuring, clinical data and image processing, and quality assessment.
• The methodology integrates clinical images and patient metadata from electronic health records.
• Experience from building datasets for Chilean and Mexican populations informed the methodology.

Main Results:

• The proposed methodology facilitates the creation of well-organized, reproducible dermatological datasets.
• Practical guidance is provided for common challenges, including image metadata and technical validation.
• The framework supports interdisciplinary collaboration between dermatologists and computer scientists.

Conclusions:

• A reproducible, scalable, and interdisciplinary framework for dermatological dataset creation is presented.
• The methodology is particularly beneficial for countries beginning dataset development.
• Recommendations and common pitfalls are highlighted to aid future dataset initiatives.