Building a Clinically Relevant and Technically Robust Synthetic Histopathology Dataset for Breast and Gastric Cancer

Area of Science:

• Digital pathology
• Artificial intelligence
• Medical image synthesis

Background:

• Developing generalizable AI in digital pathology is hindered by data scarcity, inter-institutional variability, and privacy concerns.
• Existing generative models struggle with fine-grained morphology and class-specific diversity in pathology images.
• Current diffusion models lack sufficient class-conditional synthesis for heterogeneous, multi-institutional pathology data.

Purpose of the Study:

• To introduce the Efficient Pathology Diffusion Pipeline (EPDP), a class-conditional diffusion framework for generating subtype-specific synthetic histopathology images.
• To enable the training and validation of diagnostic AI models using high-fidelity, clinically relevant synthetic datasets.
• To lower barriers in developing robust AI diagnostic tools and support standardized evaluation in digital pathology.

Main Methods:

• EPDP integrates a customized denoising U-Net with nuclear detail preservation, learnable class embeddings, and CycleGAN-based stain normalization.
• Reference-guided alignment ensures stain-invariant visual consistency across institutions.
• Multi-institutional H&E whole-slide images were curated and reviewed to create subtype-labeled training and evaluation sets.

Main Results:

• EPDP achieved lower real-synthetic FID scores than real-real FID (11.0% for breast, 8.6% for gastric).
• Subtype 1-LPIPS and SSIM gaps were minimal (≤12% and ≤0.3%, respectively), indicating high image fidelity.
• AI classifiers trained solely on synthetic data performed comparably to those trained on real data (within ~1% F1 score).
• Pathologists could not reliably distinguish synthetic from real images in a visual Turing test (accuracies 50-56%).

Conclusions:

• EPDP effectively generates high-fidelity, subtype-specific synthetic histopathology images.
• The framework preserves fine-grained nuclear morphology and class-specific diversity while ensuring stain invariance.
• EPDP facilitates the development of robust AI diagnostic tools and standardized evaluation frameworks for digital pathology.