StomaGAN: Improving image-based analysis of stomata through Generative Adversarial Networks

Gibbs, Jonathon; Gibbs, Alexandra

Abstract

Stomata regulate gas exchange between plants and the atmosphere, but analysing their morphology is challenging due to anatomical variability and artifacts during image acquisition. Deep learning (DL) can address these challenges but often requires large and diverse datasets, which are costly and error prone to produce. Generative adversarial networks (GANs) offer a solution by generating artificial data via unsupervised learning. However, GANs often suffer from problems including mode collapse, vanishing gradients, and network failure, particularly with small datasets. Here, we present StomaGAN, a deep convolutional GAN (DCGAN) with tailored modifications to address common GAN issues. We collected 559 stomatal impressions of field, or faba bean (Vicia faba) consisting of ~3,000 stoma, 80% of which were used to train StomaGAN. Evaluation metrics, including generator and discriminator loss progression and a mean Fréchet Inception Distance (FID) score of 61.4 across eight experimental runs confirms successful training. To validate StomaGAN, we generated artificial images to train a deep convolutional neural network (DCNN) based on the DeepLabV3 framework for stomata detection from real, unseen images. The DCNN achieved a mean Interception over Union (IoU) of 0.95 on artificial training images and a 0.91 on real, unseen, images across varying magnifications. Our results demonstrate that StomaGAN effectively generates high-quality synthetic datasets, enabling reliable stomatal detection and enhancing phenotypic analysis. This approach reduces the need for extensive manual data collection and simplifies complex morphological assessments.

Citation

Gibbs, J., & Gibbs, A. (2025). StomaGAN: Improving image-based analysis of stomata through Generative Adversarial Networks. in silico Plants, https://doi.org/10.1093/insilicoplants/diaf002

• Journal Article Type: Article
• Acceptance Date: Feb 17, 2025
• Online Publication Date: Feb 24, 2025
• Publication Date: Feb 24, 2025
• Deposit Date: Feb 24, 2025
• Publicly Available Date: Feb 25, 2025
• Journal: in silico Plants
• Print ISSN: 2517-5025
• Electronic ISSN: 2517-5025
• Publisher: Oxford University Press
• Peer Reviewed: Peer Reviewed
• DOI: https://doi.org/10.1093/insilicoplants/diaf002
• Public URL: https://nottingham-repository.worktribe.com/output/45848947
• Publisher URL: https://academic.oup.com/insilicoplants/advance-article/doi/10.1093/insilicoplants/diaf002/8037836