Generating and evaluating synthetic data in digital pathology through diffusion models.

Area of Science:

• Digital Pathology
• Computational Biology
• Artificial Intelligence

Background:

• Synthetic data offers solutions for data augmentation, scarcity, and privacy in computational pathology.
• Careful planning and evaluation are crucial to avoid clinically irrelevant artifacts in synthetic data.

Purpose of the Study:

• To introduce a comprehensive pipeline for generating and evaluating synthetic pathology data using diffusion models.
• To implement a multifaceted evaluation strategy with integrated explainability for synthetic medical data.

Main Methods:

• Utilized a diffusion model for synthetic pathology data generation.
• Employed an ensemble-like evaluation approach: data similarity metrics, deep learning model usability with explainable AI, and histopathological realism assessment by pathologists.
• Demonstrated the pipeline on the GTEx dataset, generating tiles from 650 Whole Slide Images across five tissues.

Main Results:

• The proposed evaluation pipeline provides complementary information, indicating the necessity of each assessment step for data quality.
• The pipeline successfully generated reliable synthetic pathology data, yielding promising results on the GTEx dataset.
• The approach addresses key aspects of synthetic data use in the medical domain, including clinical relevance and usability.

Conclusions:

• The developed workflow offers a comprehensive solution for generative AI in digital pathology.
• This pipeline can aid the digital pathology community in transitioning towards digitalization and data-driven modeling.
• Rigorous, multi-faceted evaluation is essential for the reliable application of synthetic data in medical imaging.