PixCell: A generative foundation model for digital histopathology images.

Area of Science:

• Computational pathology
• Digital pathology
• Artificial intelligence in medicine

Background:

• Histology slide digitization generates large datasets for cancer research.
• Generative models offer solutions for data scarcity, privacy, and synthetic tasks like virtual staining.
• Existing models like contrastive self-supervised and vision-language models mine pathology data for representations.

Purpose of the Study:

• Introduce PixCell, the first diffusion-based generative foundation model for histopathology.
• Demonstrate PixCell's capability in generating diverse, high-quality synthetic histopathology images.
• Showcase PixCell's utility for data augmentation, privacy-preserving sharing, and inferring molecular data.

Main Methods:

• Trained PixCell on the PanCan-30M dataset (69,184 H&E-stained whole slide images).
• Employed a progressive training strategy and self-supervision-based conditioning for scalable training without annotated data.
• Utilized mask-guided generation for targeted data augmentation and H&E staining to infer molecular marker results (e.g., IHC staining).

Main Results:

• PixCell generates diverse, high-quality histopathology images across multiple cancer types.
• Synthetic images effectively replaced real data for training self-supervised discriminative models.
• Mask-guided PixCell improved downstream performance in cell segmentation tasks.
• PixCell successfully inferred IHC staining from H&E images, demonstrating molecular marker prediction.

Conclusions:

• PixCell represents a significant advancement in generative AI for histopathology.
• The model addresses key challenges in computational pathology, including data scarcity and privacy.
• Public release of PixCell models aims to accelerate research in digital and computational pathology.