A straightforward framework to harmonize computational pathology

Area of Science:

• Computational pathology
• Digital pathology
• Medical informatics

Background:

• Inconsistent terminology in computational pathology datasets conflates biological units, lab preparations, and digital data.
• This ambiguity complicates data interpretation, limits cross-study comparability, and impacts claims on dataset scale and clinical relevance.
• Existing foundation models show significant variability in reporting disease representation, necessitating a standardized framework.

Purpose of the Study:

• To propose a harmonization framework for computational pathology dataset descriptions.
• To align dataset reporting with the DICOM hierarchical information model.
• To enable transparent characterization, improve reproducibility, and facilitate clinical translation of AI in pathology.

Main Methods:

• Review of widely cited foundation models in computational pathology.
• Identification of variability in disease representation reporting.
• Development of a hierarchical framework distinguishing clinical, lab, and digital domains.

Main Results:

• Substantial variability observed in how disease representation is reported across foundation models.
• A proposed framework that aligns with DICOM, distinguishing key data domains.
• Recommendations for explicit reporting across hierarchical levels.

Conclusions:

• Adoption of the proposed framework enables transparent dataset characterization in computational pathology.
• Standardized terminology improves reproducibility and facilitates regulatory and clinical translation.
• The framework offers an immediately implementable pathway towards harmonized reporting in AI-driven pathology research.