DEPAS: De-novo Pathology Semantic Masks using a Generative Model

Area of Science:

• Digital pathology
• Artificial intelligence
• Computational biology

Background:

• AI in digital pathology offers automation but faces challenges with biased datasets due to tissue variability and labeling needs.
• Synthetic histological images are a promising solution for debiasing datasets, requiring photorealistic generation and controllable cellular features.

Purpose of the Study:

• To introduce DEPAS (De-novo Pathology Semantic Masks), a scalable generative model for creating high-quality semantic masks of tissue structures.
• To demonstrate the utility of DEPAS-generated masks in producing photorealistic synthetic histology images with controllable cellular features for AI training.

Main Methods:

• Developed DEPAS, a scalable generative model for de-novo semantic mask generation.
• Utilized image translation models to convert semantic masks into photorealistic histology images.
• Generated multi-label semantic masks to control cellular feature distribution in synthetic images.

Main Results:

• DEPAS successfully generated high-resolution, state-of-the-art semantic masks for skin, prostate, and lung tissues.
• Synthetic histology images of cancer were produced using DEPAS masks and image translation, showcasing realism across different staining techniques.
• On-demand cellular features were generated in synthetic histology images by leveraging multi-label semantic masks.

Conclusions:

• DEPAS provides a scalable and effective solution for generating controlled synthetic histological images.
• This approach addresses the limitations of real-world datasets, paving the way for more generalizable AI algorithms in digital pathology.
• The ability to control semantic information in synthetic images enhances their utility for AI model development and validation.