Learning from 100K Leaf Epidermal Segments

Hongyuan Zhang, Güney Tombak, Shuwei Ji, Ender Konukoglub and Diana Santelia
Department of Environmental Systems Science, ETH Zürich, Universitätstrasse 16, 8092 Zürich, Switzerland

Comprehensively examining leaf epidermal segments, such as pavement cells and stomatal complexes, plays a crucial role in advancing our understanding of plant ecophysiology and ontogeny. While machine learning models have indeed revolutionized the detection of these segments from microscope images, more complex segmentation tasks still rely on drudgery manual tracing. To tackle this challenge, we developed a semi-automatic pipeline that annotated 126K epidermal segments from 486 plant species in fine-grain detail, leveraging vision foundation models. Moreover, acknowledging the variability in source datasets in terms of image modalities and preparation methods, we have trained the first generalized model suitable for direct application or fine-tuning across various research scenarios. Our model achieved human-level accuracy in tasks such as stomatal density, size, and pavement cell tracing, and it additionally unlocks the potential for novel biological insights. Beyond serving as a mere tool, we envision this work as a rallying call for collective action toward advancing epidermal phenomics.