Structure preserving adversarial generation of labeled training samples for single-cell segmentation.

Area of Science:

• * Computational biology
• * Image analysis
• * Microscopy

Background:

• * Accurate instance segmentation of microscopy data is crucial for analyzing complex tissue structures.
• * Traditional data augmentation methods may not fully capture the variability in biological samples.
• * Generative adversarial networks (GANs) offer a powerful approach for synthetic data generation.

Purpose of the Study:

• * To introduce a novel generative data augmentation strategy for improving instance segmentation of microscopy data.
• * To enhance the accuracy of segmenting complex tissue structures using synthetic data.
• * To evaluate the effectiveness of GAN-based augmentation against traditional methods.

Main Methods:

• * Development of a pipeline using regular and conditional generative adversarial networks (GANs) for image-to-image translation.
• * Generation of synthetic microscopy images and corresponding instance masks simulating object distribution, shape, and appearance.
• * Training of instance segmentation networks (e.g., StarDist, Cellpose) with the generated synthetic data.

Main Results:

• * Demonstrated improved accuracy in downstream instance segmentation tasks on two single-cell-resolution tissue datasets.
• * Showcased superior performance compared to training strategies using raw data or basic augmentations.
• * Quantified mask quality using Fréchet Inception Distances, indicating synthesized masks are closer to ground truth than those from traditional simulation methods.

Conclusions:

• * The proposed generative data augmentation strategy significantly enhances instance segmentation accuracy for microscopy data.
• * GAN-based synthetic data generation provides a robust alternative to traditional methods for complex biological imaging.
• * This approach holds promise for advancing quantitative analysis in cell biology and tissue imaging.