High resolution histopathology image generation and segmentation through adversarial training.

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence

Background:

• Deep learning models are crucial for computer-aided diagnosis in histopathology image analysis.
• Limited fully annotated datasets hinder the performance of these models.
• Existing Generative Adversarial Network (GAN) applications are mainly for lower-resolution medical images and classification tasks.

Purpose of the Study:

• To address the challenge of limited annotated data in histopathology image segmentation.
• To propose a novel multi-scale conditional GAN for generating high-resolution, large-scale histopathology images and their corresponding semantic masks.
• To improve the performance of semantic segmentation models using GAN-generated data.

Main Methods:

• Development of a multi-scale conditional Generative Adversarial Network (GAN) architecture.
• The GAN pyramid is designed for generating and segmenting images at various scales.
• Utilizing semantic masks to guide the generative process for realistic image synthesis.

Main Results:

• The proposed GAN successfully synthesizes visually realistic histopathology images with accurate semantic masks.
• The generated images and masks significantly improve semantic segmentation performance.
• Performance gains are particularly notable in semi-supervised learning scenarios.

Conclusions:

• Multi-scale conditional GANs are effective for data augmentation in histopathology image analysis.
• GAN-generated data can overcome limitations posed by small annotated datasets.
• This approach holds significant potential for advancing computer-aided diagnosis in digital pathology.