Learning Generative Models of Tissue Organization with Supervised GANs.

Area of Science:

• Computational Biology
• Microscopy Imaging
• Artificial Intelligence

Background:

• Understanding cellular and tissue spatial organization is crucial for biological research.
• Accurate generative models are needed to represent complex biological structures.
• Electron microscopy (EM) provides high-resolution images essential for this analysis.

Purpose of the Study:

• To develop a novel generative model for synthesizing realistic electron microscope (EM) images.
• To accurately model the spatial organization of cell membranes and mitochondria within EM images.
• To propose a supervised, two-stage Generative Adversarial Network (GAN) approach for image synthesis.

Main Methods:

• A two-stage Generative Adversarial Network (GAN) procedure was implemented.
• Stage 1: Synthesized label images from noise inputs.
• Stage 2: Used synthesized labels to supervise the generation of realistic EM images, creating label-image pairs.

Main Results:

• Generated synthetic EM images were validated using shape features, global statistics, and segmentation accuracy.
• User studies confirmed the realism and accuracy of the generated images.
• Enforcing a reconstruction loss on intermediate labels unified the stages into an end-to-end framework, improving results.

Conclusions:

• The proposed two-stage GAN framework successfully generates realistic EM images with accurate spatial organization.
• The method provides a powerful tool for modeling complex cellular structures.
• Further improvements were achieved by integrating a reconstruction loss for an end-to-end generative model.