Generative adversarial networks in cell microscopy for image augmentation. A systematic review

Title: Generative adversarial networks in cell microscopy for image augmentation. A systematic review

Authors: Duway Nicolas Lesmes-Leon, Andreas Dengel, Sheraz Ahmed

Published: Aug 28 2023

Link: https://www.biorxiv.org/content/10.1101/2023.08.25.554841v1

Summary (Generated by Microsoft Copilot):

Introduction:

• Cell microscopy is crucial for studying microorganisms but has limitations in data availability.
• Generative adversarial networks (GANs) can generate synthetic samples to augment data.

Challenges:

• Data limitations in cell microscopy due to intrinsic technique limitations and sample preparation challenges.
• Lack of consensus on performance metrics, baselines, and datasets.

Methods:

• Systematic review of 32 studies using GANs for cell microscopy image augmentation.
• Analysis of 21 publicly available datasets.

Novelties:

• Identification of popular architectures like StyleGAN and PathologyGAN.
• Highlighting the importance of design good practices and gold standards.

Results:

• 18 studies focused on image augmentation as the main task.
• Vanilla GAN loss was the most popular adversarial loss.

Performances:

• Inception score (IS) and Fréchet inception distance (FID) are common performance metrics.

Limitations:

• High risk of bias due to single-person data collection and lack of suitable bias assessment tools.

Discussion:

• Emphasis on the need for reproducibility and comparability in experimental designs.
• Challenges in evaluating GAN performance due to varied metrics and lack of robust evaluation methods.