Subcategorisation of Data for AI Models in Healthcare: A Case Study in Mammography

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Science

Background:

• Reliable artificial intelligence (AI) models require accurate data subcategorization for effective training and validation.
• Clinical and technical factors significantly influence mammography image appearance and interpretation, impacting AI decision support tools.
• Existing mammography datasets often lack comprehensive metadata regarding crucial subcategorization features.

Purpose of the Study:

• To demonstrate the complexity of factors affecting mammography image interpretation and AI model performance.
• To examine the metadata availability in current mammography image datasets.
• To advocate for improved data practices in AI development for medical applications.

Main Methods:

• Utilized mammography as a case study to explore data subcategorization challenges.
• Analyzed screened population characteristics (age, ethnicity) and image acquisition factors (system brand, exposure, processing).
• Reviewed existing studies and datasets, summarizing available metadata.

Main Results:

• Identified numerous clinical and technical features impacting AI model output in mammography.
• Highlighted variability in data equitability across screened populations and image acquisition parameters.
• Demonstrated the heterogeneity of image data and metadata in available mammography datasets.

Conclusions:

• Recommend training and evaluating AI models with data subcategorized by clinical and technical features for increased equitability and coverage of image heterogeneities.
• Advocate for clear definition of AI model validity subcategories when comprehensive data is unavailable.
• Emphasize that these data practices can enhance AI reliability, clinical efficiency, and diagnostic accuracy.