StomaGAN, developed by Drs. Jonathon and Alexandra Gibbs, is a specialised AI model designed to generate highly realistic images—so convincing that even the model itself struggles to distinguish them from real ones. These artificially generated images play a crucial role in training deep learning models, helping them learn to recognise and analyse specific features in real images. StomaGAN is particularly focused on stomata, the tiny openings on plant leaves that regulate gas exchange.
Crop yields largely depend on two main factors: how efficiently plants can perform photosynthesis and the availability of water. At the heart of this process are stomata, tiny openings on plant surfaces that play a crucial role in gas exchange. Each stoma is surrounded by guard cells, which control when the stomata open and close. During photosynthesis, plants take in carbon dioxide (CO₂) through these openings, which is essential for producing energy. At the same time, water vapor escapes from the leaves in a process called transpiration. This delicate balance between CO₂ intake and water loss directly influences plant growth and, ultimately, crop yields.
By studying stomatal properties such as density, distribution, size, and the rate of opening and closing, also known as the rhythm, we can enhance photosynthesis and improve water use efficiency. This knowledge can lead to increased crop yields, especially in challenging conditions like drought, contributing to better agricultural outcomes and food security.
In the last seven years, significant progress has been made in the development of deep learning tools for detecting and annotating stomata, providing a quick and effective method for automating stomatal analysis. Annotation involves adding metadata to images that details the identification of the boundaries of the stomata and their morphological characteristics such as size and density. However, the accuracy and precision of stomatal identification and characterization almost entirely depends on the quality of the underlying dataset. Deep learning learns what it is told to, and if the dataset provided is poorly annotated, the deep learning model will also produce poorly annotated images.
Creating this training dataset is time-consuming and tedious. Researchers capture microscopy images of stomata using fixed leaf samples, epidermal peels, or imprints. Software tools like ImageJ, LabelImg, and SegmentAnything are then used to annotate these images. Although these tools can automate stomatal detection, they often struggle with identification when there is insufficient contrast. Furthermore, users are still required to manually annotate stomatal morphology, adding to the overall time investment.
The use of high-quality artificial images provides a solution to this issue.
Dr. Jonathon Gibbs and Dr. Alexandra Gibbs at the University of Nottingham developed a new tool for generating artificial images of stomata using deep learning techniques. The tool, StomaGAN, is based on Generative Adversarial Networks (GANs), a type of neural network used to generate realistic images.
GANs consist of two key parts: a generator and a discriminator, which work against each other in a process called an adversarial game.
The generator is like an artist trying to paint realistic pictures. It starts by producing random images and gradually learns to improve by identifying important details such as edges, shapes, and textures from real images. Its goal is to create images that look as real as possible.
The discriminator is like an art critic. It examines images and decides whether they are real (from an actual dataset) or fake (created by the generator). Initially, the generator’s images may not be very convincing, but over time, it improves by learning from the discriminator’s feedback.
This back-and-forth competition between the generator and the discriminator helps the generator get better at creating realistic images. The more they train, the harder it becomes for the discriminator to tell real images apart from fake ones. Eventually, the generator can create images that are so realistic they can fool even a human observer.
To create StomaGAN, the authors first had to manually annotate 559 microscopy images of nail varnish-based surface impressions. Next, the images underwent pre-processing. During this step, individual stoma were identified, extracted, resized, and rotated for consistent orientation.
Following pre-processing, StomaGAN was trained by the authors to balance data generation and discrimination. The artificial images created by StomaGAN were then augmented to enhance the size and variety of the dataset by resizing and/or flipping the images. StomaGAN generated 10,000 artificial images from the 559 real images used to train it.
To evaluate the usefulness of the artificial dataset, the authors compared the performance of an existing detection tool in identifying stomata in real images when trained using only real images or trained using a combination of artificial and real images.
The detection tool trained solely on real data detected stomata with 94.7% accuracy, whereas the model incorporating artificial data was 99.7% accurate, misclassifying only one out of 5,000 stomata present in the images. This demonstrates that StomaGAN can effectively generate high-quality synthetic datasets that enable reliable stomatal detection. This approach reduces the need for extensive manual data collection and simplifies complex morphological assessments, making the process more efficient and accessible for researchers.
StomaGAN can be used to facilitate large-scale data collection and analysis, improving our understanding of photosynthesis and water use efficiency, as well as how plants respond to environmental changes.
READ THE ARTICLE:
Jonathon A Gibbs, Alexandra J Gibbs, StomaGAN: Improving image-based analysis of stomata through Generative Adversarial Networks, in silico Plants, 2025;, diaf002, https://doi.org/10.1093/insilicoplants/diaf002
The dataset(s) supporting the conclusions of this article are openly available on the Stomata Hub, https://www.stomatahub.com/ and in the GitHub repository https://github.com/DrJonoG/StomaGAN.
The post Synthetic Stomata: Leveraging Artificial Imagery for Improved Stomatal Analysis appeared first on Botany One.
