The new tool, described in the open-access journal Plant Methods, is built on a vision transformer architecture—a class of neural network models that have revolutionized computer vision by processing images as sequences of small patches, allowing them to capture both local textures and long-range structural relationships within an image. In the context of plant anatomy, that means the model can look at a stained cross-section of a rice root and reliably distinguish the epidermis, the cortical tissue, and the voids of the aerenchyma lacunae carved into the cortex, even when lighting, staining quality, magnification, and root developmental stage vary dramatically between samples. This generalization ability is precisely what previous image-analysis pipelines lacked; classical thresholding or parameter-tuned segmentation approaches tended to work well only under the specific conditions for which they were calibrated, then faltered when confronted with material from a different laboratory or growth environment.

Training such a model requires annotated data, and here the team invested heavily in both quantity and diversity. The researchers assembled a dataset of 1,760 annotated images of rice root cross-sections collected across multiple countries, growth stages, cultivation systems, and experimental contexts. Critically, the annotations followed a collaboratively defined protocol, meaning that experts from different institutions agreed in advance on the criteria for labeling cortical tissues and lacunae. This kind of standardization matters enormously in biological image analysis, where inconsistent labeling is one of the largest sources of noise in training data. By anchoring the annotations to a shared protocol and spanning such a broad range of imaging conditions, the team gave the transformer model the variety it needed to learn features that hold up in the real world rather than in a single, idealized microscope setting.

The performance numbers reported in the study are striking. On independent test data, the model achieved mean intersection-over-union scores—a standard metric of segmentation accuracy that measures the overlap between predicted and true pixel regions—exceeding 0.92 for both cortical tissues and lacunae. Translating segmentation into a biologically meaningful quantity, the pipeline computes the lacuna-to-cortex ratio, essentially the fraction of root cortical space devoted to air channels, and this measure showed strong agreement with manual expert quantification, with a coefficient of determination of 0.98. In plain terms, when the model and a trained human each measured the same roots, their answers tracked each other almost perfectly. Beyond raw accuracy, an independent expert review panel found that the model’s predictions were at least as internally consistent as manual annotations from different human experts, and that the algorithm actively reduced the large inconsistencies that plague manual labeling efforts—particularly for subtle or ambiguous lacunar structures.

Why does this anatomical trait deserve such technological firepower? Aerenchyma lacunae sit at the intersection of several major questions in rice biology and climate science. Internally, they form the ventilation system of the plant, transporting oxygen from aerial tissues down to the submerged and buried roots, which is essential for survival in anaerobic soils. They also influence hydraulic conductivity, meaning the degree of aerenchyma formation can limit how easily water moves through the root—potentially an important trait for drought tolerance in rice varieties grown under water-saving or rainfed systems. On the climate side, the same air channels that help the root breathe can serve as conduits for methane produced in flooded soils to escape to the atmosphere, so the extent of lacuna formation may directly modulate how much methane a rice field emits. A trait that touches drought adaptation, gas transport, and greenhouse gas flux is exactly the kind of target that climate-smart breeding programs want to measure—and measure cheaply and consistently.

That is where scalability becomes the central theme. Conventional manual phenotyping of aerenchyma requires sectioning roots, mounting and staining samples, capturing images, and then tracing tissue boundaries by hand, which can take an expert substantial time per cross-section and introduces inter-observer variability at every step. For a breeding program screening hundreds of genotypes across multiple water regimes and locations, the workload becomes prohibitive, and anatomical traits like aerenchyma are typically left out of selection decisions even when they are known to matter. The new pipeline, released as open-source software with an interactive online demonstrator and an accompanying online test dataset, allows any laboratory to upload cross-section images and obtain quantitative lacuna measurements without parameter tuning or coding expertise. The open release is designed explicitly to support testing and reproducibility, addressing a recurring criticism of machine learning tools in biology that are described in papers but never usable by others.

To demonstrate that the tool is more than a benchmark, the researchers applied it across six independent experimental use cases, spanning differences in genotype, water regime, environment, and developmental stage. The pipeline reproducibly detected known patterns of aerenchyma variation—for instance, differences in lacuna formation between genotypes and shifts associated with contrasting water availability—suggesting that it captures biologically real signal rather than imaging artifacts. This kind of cross-context validation is what distinguishes a genuinely transferable phenotyping tool from a model that merely memorizes its training set. The authors argue that the approach now makes it feasible to integrate aerenchyma quantification into breeding pipelines, physiological studies, and climate-focused crop improvement programs, where anatomical measurements have historically been too expensive to deploy at population scale.

The research emerged from a collaboration between the AGAP Institut, a joint research unit of INRAE and CIRAD in Montpellier, France, and the International Center for Tropical Agriculture (CIAT) in Cali, Colombia, with additional involvement from Icesi University. The work was funded through the project “Enhancing breeding strategies to reduce methane emissions in rice cultivation,” supported by the Global Methane Hub, and made use of high-performance computing resources from GENCI in France as well as the MRI imaging facility, part of the national France BioImaging infrastructure. The corresponding authors are Maria Camila Rebolledo and Romain Fernandez, who led a multidisciplinary team combining plant physiology, computational imaging, and breeding expertise—reflecting the reality that modern agricultural challenges increasingly demand tools that bridge biology and artificial intelligence.

The broader significance of the work lies in what it signals for the future of plant phenotyping. Root anatomy remains one of the least-explored dimensions of crop improvement, largely because it is hidden underground and technically demanding to measure, in contrast to canopy traits that satellite and drone platforms can now capture routinely. Vision transformers, with their ability to learn robust, generalizable visual representations from diverse training data, are rapidly becoming the workhorses of this anatomical revolution, enabling automated quantification of internal structures that once required a trained anatomist’s eye for every single sample. If aerenchyma can now be scored across thousands of genotypes with minimal human effort, breeders can begin selecting for optimal lacuna formation—balancing oxygen transport and drought resilience against methane emissions—in ways that were previously unthinkable.

For rice, a crop cultivated in some of the world’s most climate-vulnerable regions and simultaneously responsible for a meaningful share of global anthropogenic methane emissions, the timing of such a tool is opportune. Climate-smart agriculture will require cultivars that maintain productivity under erratic water supplies while minimizing environmental footprint, and internal root anatomy is emerging as a key lever in that optimization. By transforming a labor-intensive, subjective measurement into a fast, consistent, and open computational pipeline, the research team has effectively removed one of the last excuses for ignoring root anatomy in large-scale breeding. The pipeline’s open-source release means that laboratories across the rice-growing world—from advanced research centers to national programs—can adopt and adapt it, potentially accelerating the development of rice varieties that feed billions while treading more lightly on the climate.

As with any machine learning system, the authors note that continued community contributions of annotated images from new environments and genotypes will further strengthen the model’s robustness, and the availability of the online demonstrator and test dataset lowers the barrier for exactly that kind of engagement. The study stands as a concrete example of how transformer-based deep learning, careful collaborative annotation, and open science practices can combine to crack a long-standing measurement bottleneck—one cross-section, and now one click, at a time.