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

Duway Nicolas Lesmes-Leon, Andreas Dengel, Sheraz AhmedbioRxiv 2023 · DOI

Abstract

Cell microscopy is the primary tool for studying microorganisms, yet limitations in sampling and technique often result in a shortage of high-quality data. Generative Adversarial Networks (GANs) offer a deep learning solution by generating synthetic samples that mimic real data distributions. This systematic review analyzes 32 studies to identify trends, popular architectures like StyleGAN, and common loss functions such as Vanilla and Wasserstein. The findings reveal a significant lack of consensus regarding performance metrics and reproducibility, highlighting the need for standardized practices in biological image augmentation.

Introduction

Deep learning has revolutionized computer vision, but its effectiveness in cell microscopy is often hampered by a "data bottleneck." Preparing biological samples is expensive, time-consuming, and prone to human error. Furthermore, physical constraints like phototoxicity can limit the acquisition of live-cell data. GANs address this by learning the underlying probability distribution of real cell images to generate authentic-looking synthetic samples. This review organizes the current research landscape to identify the most effective architectures and practices for this emerging field.

Methodology

This study followed the PRISMA 2020 guidelines to ensure a rigorous systematic review. The authors screened 321 publications across five major databases, ultimately selecting 32 studies focused on GAN-based image augmentation in cell microscopy. Data collection involved extracting 15 features per study, including architecture types, loss functions, and performance metrics.

Systematic search strategy across major scientific databases.

Technical Analysis of GAN Components

Architecture Evolution

There is a clear shift from basic models like DCGAN toward sophisticated architectures such as StyleGAN2 and Progressive Growing GANs (PGGAN). These newer models are better equipped to handle the high-resolution textures and fine structural details essential for scientific accuracy in microscopy.

Loss Functions and Stability

The choice of loss function is critical for training stability. While "Vanilla" cross-entropy loss remains common, the most successful recent studies favor Wasserstein loss. This approach significantly reduces "mode collapse"—a common failure mode where the GAN generates repetitive images—and provides a more reliable training signal.

Comparison of GAN loss functions, their mathematical foundations, and attributes.

Key Insights and Results

  • The Metric Mess: The review identified a staggering lack of consensus on evaluation. Over 20 different metrics were used across the 32 studies. While the Fréchet Inception Distance (FID) is gaining traction, many researchers still rely on qualitative "eye tests" or downstream task performance.
  • Dominant Modalities: Optical microscopy (59.38%) and Fluorescence microscopy (28.13%) are the primary targets for GAN augmentation.
  • Reproducibility Crisis: Only 28% of the reviewed studies shared their code, and 43% treated GAN augmentation as a secondary task, often omitting the rigorous ablation studies necessary to validate their configurations.

Conclusion and Future Directions

While GANs show immense promise for alleviating data shortages in biology, several hurdles remain. Future research must prioritize biological realism over mere visual quality, ensuring that synthetic images are scientifically valid representations of life rather than just "plausible noise." The authors call for a unified reporting framework and closer collaboration between deep learning engineers and biologists to establish "gold standards" for synthetic data in the life sciences.

Summary updated

Abstract

Abstract Cell microscopy is the main tool that allows researchers to study microorganisms and plays a key role in observing and understanding the morphology, interactions, and development of microorganisms. However, there exist limitations in both the techniques and the samples that impair the amount of available data to study. Generative adversarial networks (GANs) are a deep learning alternative to alleviate the data availability limitation by generating nonexistent samples that resemble the probability distribution of the real data. The aim of this systematic review is to find trends, common practices, popular datasets, and analyze the impact of GANs in image augmentation of cell microscopy images. We used ScienceDirect, IEEE Xplore, PubMed, bioRxiv, and arXiv to select English research articles that employed GANs to generate any kind of cell microscopy images independently of the main objective of the study. We conducted the data collection using 15 selected features from each study, which allowed us to analyze the results from different perspectives using tables and histograms. 32 studies met the legibility criteria, where 18 had image augmentation as the main task. Moreover, we retrieved 21 publicly available datasets. The results showed a lack of consensus with performance metrics, baselines, and datasets. Additionally, we evidenced the relevance of popular architectures such as StyleGAN and losses including Vanilla and Wasserstein adversarial loss. This systematic review presents the most popular configurations to perform image augmentation. It also highlights the importance of design good practices and gold standards to guarantee comparability and reproducibility. This review implemented the ROBIS tool to assess the risk of bias, and it was not registered in PROSPERO.
BibTeX
@misc{lesmesleon2023generative,
  author = {Duway Nicolas Lesmes-Leon and Andreas Dengel and Sheraz Ahmed},
  title = {{Generative adversarial networks in cell microscopy for image augmentation. A systematic review}},
  year = {2023},
  doi = {10.1101/2023.08.25.554841},
  howpublished = {bioRxiv},
}
Figure