Characterizing Stomatal and Epidermal Traits Using Peels, Clearing, and AI-Based Image Analysis

Summary

This protocol presents a comparative method to assess stomatal and epidermal traits in two distinct species using epidermal peels, clearing techniques, and AI-assisted image analysis. The goal is to enable high-throughput, reproducible quantification of cellular structures underlying stomatal function for agricultural applications and broad plant research.

Abstract

Accurate characterization of stomatal morphology and surrounding epidermal traits is essential for understanding stomatal function. A comparative protocol is presented using two complementary techniques-epidermal peels and leaf clearing-to assess stomatal and epidermal cell size, shape, and density in two ecologically distinct dicot species, Fraxinus excelsior and Taraxacum officinale. The protocol includes detailed steps for sample preparation, imaging, and trait quantification using cost- and labor-effective manual measurements (tissue clearing and peeling), and AI-assisted image analysis with the potential to facilitate high throughput applications. Three segmentation tools-a YOLOv8-based stomatal detection model, a generalist deep learning tool for cell segmentation, and a pixel-based classifier-were evaluated for their accuracy and suitability across methods. Significant differences in stomatal and epidermal traits are consistently observed between peeling and clearing techniques in both manual and AI analyses. AI tools enabled reproducible quantification, with YOLOv8-based stomatal detection model (StoManager1) performing best for stomatal traits and a generalist deep learning tool for cell segmentation (Cellpose) for epidermal segmentation. AI-based trait detection was highly consistent in T. officinale but limited in F. excelsior, results still diverged from manual measurements, illustrating how species-specific anatomy and image quality impact method suitability. These findings suggest that while AI-assisted analysis provides a useful framework for stomatal and epidermal trait quantification, expert validation and pilot testing remain essential before broader or high-throughput application.

Introduction

Stomata are microscopic pores in the epidermis that play a vital role in regulating plant gas exchange and transpiration. Thus, understanding the morphological basis of stomatal functioning is key to improving crop performance¹. While most studies of stomatal structure and function focus solely on stomatal morphology, a full understanding of stomatal responsiveness also requires characterization of the size and shape of surrounding epidermal cells. Accurate quantification of these traits is fundamental to understanding the structure-function relationships underlying gas exchange and crop improvement.

As crop improvement increasingly relies on advanced gene-editing approaches, it is critical to understand the key morphological traits- such as stomata- that influence photosynthesis, intrinsic water-use efficiency, and carbon gain. This need is especially pressing under ongoing environmental change, as the demand for resilient crops that maintain both photosynthetic capacity and yield grows in the face of more frequent droughts and heatwaves^2,3. Consequently, there is a strong need for standardized imaging and analysis protocols. Stomatal peeling with varnish/nail polish is the simplest method for obtaining leaf surface impressions to measure stomatal and epidermal size, shape, and density^4,5,6. Its procedure is fast, non-destructive, allowing a rapid processing of samples. Tissue clearing method is an alternative approach allowing simultaneous characterization of vein network and potentially more comprehensive characterization of leaf hydraulics.

Chemical clearing softens and renders tissues optically transparent, allowing in situ imaging of stomata and epidermal cells without mechanical disruption. Clearing is broadly applicable across a range of tissue types, including dried or delicate specimens, and can be used not only for stomatal imaging but also for visualization of venation networks, fluorescence imaging, and internal anatomical structures^{7,8,9,10,11,12}, thus allowing stomatal traits to be analyzed alongside other functional traits. Various clearing methods exist, involving organic solvents, aqueous solutions, hyperhydration, or hydrogel embedding^13,14,15. These differ in duration, imaging compatibility, and reagent availability. The protocol used here is simple, cost-effective, and requires no specialized skills. Both techniques are inexpensive and rapid compared to more labor-intensive methods such as light microscopy or scanning electron microscopy, which require the chemical fixation, embedding, and sectioning of samples.

Traditional methods for analyzing stomatal and epidermal characteristics often rely on manual segmentation of microscopy images, which is labor-intensive, time-consuming, and subject to observer bias, which limits throughput and reduces the efficiency of the analysis pipeline¹⁶. Furthermore, anatomical features such as trichomes and irregular cell walls present additional challenges to consistent trait quantification. While automated image analysis tools offer great promise for high-throughput data extraction, they have been optimized for specific plant tissues and are not readily transferable across leaf types and preparation methods. For example, StoManager1 (Supplementary Figure 1) provides high accuracy in stomatal detection using a YOLOv8-based object detection model but is specifically trained for stomatal traits. Cellpose (Supplementary Figure 2), is a large, pre-trained, generalist segmentation model that performs well when the input data resembles its pre-trained models (e.g., cyto, cyto2), based originally on animal cells and organelles. Ilastik (Supplementary Figure 3) is a pixel-based classification tool relying on user-defined annotations. Both Cellpose and Ilastik are user-friendly, but each may require retraining or fine-tuning to accommodate variations in image quality, background artifacts, and tissue-specific features.

Furthermore, stomatal and epidermal morphology show high variability across species¹⁷. Traits such as stomatal density and size are especially variable, yet they are typically inversely related and together determine species differences in stomatal kinetics. Larger stomata can partly compensate for lower density, but they generally exhibit slower kinetics⁵, which may be disadvantageous under increasingly frequent heatwaves, where rapid stomatal responses are critical to maintaining carbon assimilation and preventing excessive water loss. Altogether, this variability complicates the measurement of stomatal traits and may compromise the validity of the selected protocol and the robustness of AI-based results.

This protocol integrates two classical preparation techniques-epidermal peels and chemical clearings-with modern artificial intelligence (AI) and machine learning (ML) tools for trait segmentation and measurement. Using two common light-demanding dicot species with contrasting ecological strategies (F. excelsior and T. officinale), we examined variability in stomatal and epidermal traits to assess how three methods compare and where potential inaccuracies may arise. The performance of three AI tools: StoManager1 (a YOLOv8-based deep learning model for stomatal detection), Cellpose (a generalist deep learning segmentation tool), and ilastik (a pixel-based classifier for epidermal tissues) were assessed. This integrative approach enables reproducible, high-throughput quantification of stomatal and epidermal traits, offering a robust and scalable framework for phenotypic analysis in both ecological and agricultural research. However, anatomical features such as dense trichomes, thick cuticles, or low cell-boundary contrast may reduce segmentation accuracy and should be considered when selecting between peeling and clearing approaches and when applying AI-based quantification.

To our knowledge, this is the first step-by-step protocol that directly compares epidermal peeling and chemical clearing approaches in combination with multiple AI-based segmentation tools for both stomatal and epidermal trait quantification. By uniting classical microscopy techniques with accessible ML workflows, the protocol reduces user bias, increases throughput, and facilitates reproducible phenotyping across species. This makes the approach broadly applicable for ecology, crop improvement, and climate-change research, and provides a foundation for ongoing tool expansion through model retraining or transfer learning.

Protocol

NOTE: No specific permits or approvals were required for the collection of T. officinale and F. excelsior leaves, as these non-protected species were sampled from publicly accessible garden areas. All samples' collection complied with institutional and local guidelines.

1. Plant material collection

1. Select healthy leaves for comparative analysis. For current experiment, healthy, sun-exposed mature leaves from two dicot species with contrasting stomatal distribution were selected: the hypo stomatous tree F. excelsior, which bears stomata only on the lower leaf surface, and the amphistomatous perennial forb T. officinale.
2. For each sample type/location take 3-5 replicate leaves.
3. Detach the leaf using clean scissors or a razor blade. Place the cut leaf immediately into water to prevent dehydration.
4. After collection, fully submerge petiole in water and store them at 4°C.
5. Proceed with protocol sections 2 or 3 within 48 h.

2. Preparation and viewing stomata impression peels

NOTE: All chemical waste generated during the clearing protocol was handled according to institutional and local safety regulations. Glutaraldehyde-containing solutions were collected separately and disposed of as hazardous chemical waste. Sodium hydroxide (NaOH) solutions were neutralized to pH 6-8 prior to disposal, where permitted. Diluted bleach and ethanol solutions were disposed of down the drain with copious amounts of cold water in places where allowed by institutional and municipal regulations. TBO solutions were collected and disposed of as chemical waste according to local guidelines. Users should consult their institutional safety office for approved disposal procedures.

1. If necessary, gently rinse the selected leaf surface with water to remove dust or debris. Immediately blot the leaf surface dry using filter paper or a soft paper towel to remove excess moisture. Do not air dry, as this may cause shrinkage of the tissue, leading to inaccurate size measurements of stomata and epidermal cells.
2. Label the microscope slide with sample details (e.g., species, leaf side, and date of preparation) using a permanent marker to ensure traceability of each sample throughout imaging, analysis, and data interpretation, and enables proper storing of sample for future use.
3. For species with non-glandular (hairy) trichomes, remove the trichomes from the leaf surface with a razor blade prior to peel preparation, as their presence can obscure cell boundaries and reduce image quality (Figure 1).
4. Using a small brush, apply a thin, even layer of transparent, non-extendible nail polish (avoid "stretch" or "peel-off" formulas) to the target area. Always select the central portion of the leaf for peel preparation, avoiding the tip and base.
5. Allow the nail polish to dry completely (typically 2-5 min. at room temperature). Prolonged exposure may cause tissue discoloration (e.g., red tint on softer leaves), which may be visible in the final peel, and interfere with downstream analyses requiring uniform coloration.
6. Once dry, apply a strip of transparent tape onto the coated area. Start at one edge of the nail polish, and gradually lay the tape down across the surface, gently pressing to ensure good adhesion and to avoid air bubbles or creases. Wear nitrile gloves to avoid fingerprints.
7. Carefully peel away the tape together with the nail polish film (epidermal impression).
8. Place the tape (sticky side down) onto a clean microscope slide, avoiding folds or overlaps.
9. Store prepared slides at room temperature, protected from dust and direct sunlight.
10. Examine the slide under a compound light microscope at 400x magnification. Save images as per the software guide, along with the relevant scale.

3. Preparing and viewing cleared sections

1. Cut leaf pieces of approximately 1 cm². Perform all steps wearing nitrile gloves and working under a fume hood.
2. Transfer the samples to a labelled vial, cover them with fixating solution (4% glutaric aldehyde in 0.1 M phosphate buffer, pH = 6.9) (CAUTION), and store at 4°C for 24 h (Table 1, for preparation of fixating solution).
3. Remove the fixating solution with a pipette and cover the samples with distilled water. Leave for 30 min. Repeat thrice (Figure 2A-C).
4. Remove the water and cover the samples with 1 M NaOH solution (dissolve 20 g NaOH in 0.5 L of distilled water).
5. During leaf clearing, the NaOH solution will turn green. Replace the NaOH solution once per day with fresh solution. Continue for 3 to 4 days, or until the samples turn pale yellow and the NaOH solution remains clear (Figure 2D-F).
6. Remove the NaOH solution and cover the samples with bleach (5% NaClO). Commercial household bleach with this concentration is suitable.
7. Incubate the samples in bleach for 1-2 min. The samples will turn white and translucent (Figure 2G-I).
8. Remove the previous solution and cover the samples with distilled water for 2 min. Dehydrate the samples through a graded distilled water-ethanol series as follows: 75:25, 50:50, 25:75, and 100% ethanol.
9. Remove the ethanol and cover the samples with a 3% alcoholic solution of toluidine blue O (TBO, dissolve 3 g TBO in 100 mL of 100% ethanol). Leave overnight.
10. Remove the staining solution and cover the samples with 100% ethanol for 2 min. Repeat this step using a graded distilled water-ethanol series as follows: 25:75, 50:50, 75:25, and finally distilled water (Figure 2J-L). Refer to Figure 3 for common mistakes and troubleshooting steps associated with the clearing process.
11. Keep the samples in distilled water at 4°C until viewing.
12. Transfer the samples onto a glass slide, add a drop of distilled water, and cover with a glass coverslip for microscopic viewing and imaging.
13. Examine the slide under a compound light microscope at 400x magnification. Save images as per the software guide, along with the relevant scale. See the Table of materials for further information.

4. Image analysis and measurements

1. Use Fiji²³ (ImageJ distribution, version 1.54p) for image processing and analysis.
2. Using the Line tool, draw a line over the scale bar, then select Analyze > Set Scale, modify the known distance to the length of the scale bar, and change the unit of length to µm.
3. Go to Analyze > Set Measurements. Ensure the Area and Feret's diameter options are checked.
4. Use the Polygon Selection Tool to outline the object of interest. Record the measurement by pressing the M key.
5. Measure the following stomatal and epidermal traits in Fiji and calculate for both cleared samples and stomatal peels (Figure 3).
6. Calculate the number of stomata (Ns).
   1. Determine the average stomatal area (As).
   2. Determine average stomatal width (Ws): using Feret's minimum diameter.
   3. Calculate stomatal density (SD), by the following formula:
      
      ; per unit Ia (the area of leaf selected for measurements, mm²).
   4. Determine the average epidermal cell area (Ae).
   5. Determine the epidermal cell density (ECD) using the following formula:
      

5. Statistics

1. Use ANOVA and Tukey's HSD post-hoc test to test differences between sample groups.
2. Conduct all statistical analyses in R²¹.
3. Analyze a total of 3 replicate leaves per species, and view 3 peels from each leaf.

6. AI assisted analysis

1. Select machine learning tools for segmentation.
   1. Select appropriate machine learning-based segmentation tools based on the target structure (stomata or epidermal cells), image quality, and analysis goals.
   2. Use a stomata-specific deep learning model for automated detection and measurement of stomatal traits and use generalist deep learning or pixel-based segmentation tools for epidermal cell segmentation.
   3. Refer to the Machine Learning Tools and Software section for specific software implementations (Supplementary file).
   4. Perform pilot testing on a subset of images prior to large-scale analysis to evaluate segmentation performance and identify potential sources of false positives.
2. Prepare the computing environment.
   1. Use a workstation equipped with an 8-12 core processor, at least 32 GB RAM, and an NVIDIA CUDA-enabled GPU with a minimum of 24 GB VRAM.
   2. Use a high-resolution display (preferably 4K) to facilitate annotation and visual inspection.
   3. Install all selected software tools according to the developers' instructions and record software versions to ensure reproducibility.
3. Prepare images for machine learning analysis.
   1. Import microscopy images in TIFF or PNG format, using 8 or 16 bit grayscale images when possible.
   2. Ensure that each image contains a visible scale bar or has been spatially calibrated prior to analysis.
   3. Select regions of interest with minimal debris, folds, or air bubbles, even illumination, and clearly defined cell boundaries.
   4. Exclude images with blurred edges, strong shadows, or low contrast, as these reduce segmentation accuracy.
4. Perform stomatal segmentation.
   1. Load images into the stomata-specific deep learning model (Supplementary file section 1).
   2. Configure detection parameters, including object size limits and confidence thresholds, according to species and image resolution.
   3. Run automated detection to segment stomatal complexes and extract output metrics such as stomatal number, area, and Feret's diameters.
   4. Export segmentation masks and measurement tables for downstream analysis. Visually inspect segmentation overlays and manually exclude incorrectly detected objects if required.
5. Perform epidermal cell segmentation.
   1. Load images into a generalist deep learning segmentation tool or a pixel-based classifier (Supplementary file section 2 - 4).
   2. Train or fine-tune the model using representative images with manually annotated examples of epidermal cells, cell boundaries, and background.
   3. Run segmentation to generate object masks and export the masks for post-processing in ImageJ/Fiji or equivalent image-analysis software.
   4. Apply size-based or quantile-based filtering to remove small false-positive objects and verify segmentation accuracy through visual inspection before extracting epidermal cell metrics such as cell area and cell density.
6. Document parameters and perform quality control.
   1. Record all software versions, parameter settings, filtering thresholds, and hardware specifications used during analysis.
   2. Retain representative segmentation overlays to document segmentation quality.
   3. Visually validate automated segmentation outputs prior to statistical analysis or data reporting.
7. Image preparation for all ML processing
   1. Sample imaging
      ​NOTE: Because ML-based segmentation relies on intensity gradients, edge continuity, and consistent object geometry rather than biological context, image pre-processing and region selection critically determine model performance and error structure.
      1. Capture high-resolution images of cleared or peeled leaf surfaces with a scale bar visible. For good machine-learning workflows, ensure that the measurable objects are sharp, continuous, and high-contrast throughout the image.
   2. Region selection and image quality
      ​NOTE: Following image acquisition in Section 6.7.1, images are visually inspected and, where image scale permits, artifact-free regions of interest are manually selected for downstream segmentation. Select regions of interest with minimal artifacts (e.g., debris, folds, air bubbles, uneven illumination). For stomatal models (e.g., StoManager1), prioritize clear pore visibility and a well-defined guard-cell silhouette. For epidermal models (e.g., Cellpose, ilastik), ensure stable intensity gradients along cell contours; discontinuous or low-contrast edges substantially reduce segmentation performance. Images with poorly defined boundaries should be excluded or reprocessed prior to segmentation.
   3. File format
      ​Save files in TIFF or PNG format (8- or 16-bit grayscale recommended for segmentation).
   4. Optional image enhancement
      1. Enhance contrast or subtract background in ImageJ/Fiji if boundaries are weak.
      2. Focus on edge geometry, as these can create false gradients interpreted as cell borders.
      3. Apply such adjustments consistently across all images within a dataset.
         NOTE: Particular attention should be paid to edge geometry, as shadows, halos, or uneven background gradients can create artificial intensity transitions that ML models may misinterpret as cell borders.
8. StoManager1 (v1.0.0 GUI version)
   NOTE: StoManager1 enables automated, high-throughput detection, measurement, and analysis of plant leaf stomata using convolutional neural networks (YOLOv8-seg-x). The tool quantifies key stomatal metrics (area, density, pore/guard cell ratio, aggregation indices, etc.) from various image types, reducing manual effort and bias, and facilitating large-scale ecological and physiological studies. A user-friendly Windows app is available, requiring no coding skills except for optional post-processing (Supplementary file Section 1)²³.
   1. Running StoManager1: install and configure
      1. Access via GitHub (https://github.com/JiaxinWang123/StoManager1) or the standalone Windows application ((Supplementary file section 1)(https://zenodo.org/doi/10.5281/zenodo.7686022)²³.
      2. Ensure compatible library versions (Python dependencies); conflicts may occur if other ML/image software is installed.
   2. Start StoManager1 via GUI or command-line interface.
   3. Choose your input directory containing leaf images.
   4. Model selection; Check the YOLOv8-seg-x model for segmentation-based, more detailed metrics. Keep it unchecked for the bounding box model (YOLOv3), fewer metrics.
   5. Parameter tuning; Adjust detection thresholds, image resolution, mask dilation, and object size limits as needed for your imaging modality/species. Please use default parameters, when some stomata objects remain undetected, mainly (1) decrease the detection confidence threshold to recover more true stomata while checking for false positive detection; and (2) increase the image resolution, as larger value can improve detection of small stomata at the cost of higher computational demand.
   6. Run segmentation:
      1. Click " Start Process" to detect, segment, and measure stomata and whole stomata.
      2. Click " Statistical Analysis" to process group-level statistics.
   7. Collect results. Output is saved to YOUR_OUTPUT_PATH/Predict-output/Output_csv/ (segmentation mode) or directly to YOUR_OUTPUT_PATH (bounding box mode).CSV files contain per-object and summary metrics; masks and result images are also generated.
   8. Launch batch processing.
   9. Upon completion, collect the output: segmentation masks, and per-image CSV result files. Results per image include: Stomatal number and stomatal area, Feret's diameter (max/min), guard cell metrics, aggregation/evenness/divergence indices and pore/guard cell ratios.
   10. Review outputs using Excel, R, or Python for further statistics and visualization.
   11. Quality control; Visually inspect masks and overlays. Remove outlier objects by size, shape, or manual curation if needed (see below).
       NOTE: Detailed protocols and post-processing steps for Cellpose and ilastik are provided in Supplementary file. For optimal performance with large datasets and complex images, we recommend a system with an 8-12 core processor, at least 32 GB RAM, and an NVIDIA CUDA-enabled GPU with a minimum of 24 GB VRAM. A large, high-resolution display (preferably 4K) is also strongly recommended for efficient annotation and visualization.

Results

Comparison of results between peeling and clearing

Both T. officinale and F. excelsior showed variability in stomatal and epidermal traits depending on whether samples were cleared or peeled (Figure 4). Cleared samples generally provided sharper cell boundaries and clearer visualization of guard cell shape and pore aperture, although overlapping tissue layers occasionally obscured smaller stomata and biased size estimates upward. In F. excelsior, average stomatal area and stomatal width were significantly higher in cleared samples than in peels, whereas in T. officinale no significant differences were observed. Stomatal density was significantly higher in peeled samples of F. excelsior, while no method effect was detected in T. officinale. Epidermal cell area did not differ between methods in either species, but epidermal cell density was slightly higher in peeled samples of F. excelsior.

Overall, method-dependent differences were trait- and species-specific, suggesting that peeling provides a reliable option for stomatal measurements, particularly when stomata vary in size (e.g., in developing leaves), while epidermal traits were largely unaffected (Figure 5).

AI results and how they resonate with peeling and clearing methods

The stomata-specific model StoManager1 detected stomata more reliably on peeled images (Figure 6), whereas performance declined markedly on cleared samples (Figure 7). To further evaluate the performance of StoManager1 across preparation methods, we compared stomatal trait detection in T. officinale and F. excelsior using both peels and cleared images (Figure 8). In T. officinale, the model reliably detected stomatal presence and density across the preparation methods. However, the resulting values differed significantly from manual measurements (Table 2 and Figures 5 and 8). Most stomatal traits differed significantly (p < 0.05) between peeled and cleared samples (Figure 8), indicating that AI detection pipelines are highly sensitive to preparation methods and require careful validation before large-scale application. By contrast, in F. excelsior, detection was limited to the abaxial surface due to the absence of adaxial stomata, and in cleared samples detection was inconsistent, with frequent false positives (Figure 8). These results highlight the strong influence of species-specific anatomical features and image quality on the effectiveness of AI-based analyses. Generalist ML tools like Cellpose (Figure 9A) and Ilastik (Figure 9B) segmented epidermal cells more effectively on peeled than on cleared images. However, both tools produced numerous false positives with both preparation techniques (peeling and clearing), especially under low-contrast conditions or when cell contours were blurred. In such cases, indistinct stomatal or epidermal contours were often over-segmented into multiple tiny objects misclassified as individual epidermal cells (see Figure 10). As a result, post-processing is essential, not just outlier removal but also filtering by upper quantiles, to retain reliably identified cells.

Figure 1: Stomatal peels of F. carica leaves before and after removal of trichomes using a razor blade. Large trichomes can leave impressions on the peel, forming air pockets that result in uneven surfaces and hinder consistent focusing (visible as large grey dots). F. carica was included as an outlier to extend the protocol to hairy leaves. (A) At 200x magnification, image before trichome removal is partly obstructed with the trichome impressions. (B) Quality of image is improved after trichome removal. (C) At a 400x magnification image before trichome removal is to most extent obstructed with trichome impressions. (D) Obstruction is minimal in image after trichome removal with a razor blade. Scale bars 100 µm (A, B) and 50 µm (C, D). Please click here to view a larger version of this figure.

Figure 2: Consecutive steps of the clearing protocol. Representative images of T. officinale, F. excelsior, and F. carica illustrate successive stages of the clearing protocol. (A-C) Fresh leaves. (D-F) leaves after clearing with NaOH. (G-I) leaves after bleaching with NaClO. (J-L) leaves after staining with toluidine blue O. This visual aid shows the expected sample appearance before proceeding to the next step and highlights protocol variability among species. F. carica was included as an outlier species to demonstrate protocol applicability across a wider range of leaf and venation morphologies. Please click here to view a larger version of this figure.

Figure 3: Example measurements of stomatal and epidermal traits. Measured traits are illustrated on the image as follows: As = stomatal area; Ls = stomatal length; Ws = stomatal width; Ae = epidermal cell area. Stomatal density and epidermal cell density were calculated within the marked area, as described in the protocol. Image at 400x magnification. Please click here to view a larger version of this figure.

Figure 4: Comparison of images obtained with clearing protocol and peeling protocol. Representative images of T. officinale and F. excelsior comparing clearing and stomatal peel images showing the visibility and variability of stomata and epidermis cell size and density. Scale bar 50 µm, and images are taken at 400x magnification. Please click here to view a larger version of this figure.

Figure 5: Differences in manual trait measurement results between clearing and peeling protocol. Box plots showing differences in stomatal and epidermal traits between clearing and epidermal peel methods in T. officinale and F. excelsior. (A) Average stomatal area (B) average stomatal width (C) stomatal density (D) average epidermal cell area, and (E) epidermal cell density. Asterisks indicate levels of statistical significance (P < 0.05; *P < 0.01; **P < 0.001). Statistical significance was assessed using one-way ANOVA followed by Tukey's HSD post-hoc test. Sample size per group: n = 3-5. Abbreviations: As = stomatal area; Ws = stomatal width; SD = stomatal density; Ae = epidermal cell area; ECD = epidermal cell density. Please click here to view a larger version of this figure.

Figure 6: Example of segmentation results using a stomatal detection model on peel sample. StoManager1 segmentation result for a peeled sample of T. officinale (lower side of a mature shade-exposed leaf), showing good detection accuracy. Please click here to view a larger version of this figure.

Figure 7: Example of segmentation result using a stomatal detection model on cleared sample. StoManager1 segmentation result for a cleared sample of T. officinale (lower side of a mature shade-exposed leaf), showing reduced detection accuracy compared to peeled samples. Please click here to view a larger version of this figure.

Figure 8: Differences in AI assisted trait measurement results between clearing and peeling protocol. Box plots showing AI-derived measurements of stomatal anatomical traits from images processed by StoManager1 in T. officinale and F. excelsior. Traits include stomatal area (As), stomatal width (Ws), stomatal length (Ls), and stomatal density (Ds). Units are as in Figure 5. Statistical significance was assessed using one-way ANOVA followed by Tukey's HSD post-hoc test. Asterisks indicate levels of statistical significance (P < 0.05; *P < 0.01; **P < 0.001). Sample size per group: n = 3-5. Abbreviations: As = stomatal area; Ws = stomatal width; Ls = stomatal length; Ds = stomatal density. Please click here to view a larger version of this figure.

Figure 9: Screenshots showing segmentation with generalist deep learning tool and pixel-based classifier. Peeled F. excelsior sample (abaxial leaf surface) segmentations with (A) generalist deep learning tool (Cellpose (cyto2)), and (B) pixel-based classifier (ilastik). Please click here to view a larger version of this figure.

Figure 10: Ilastik segmentation mask for a peeled sample of F. excelsior (abaxial side of a leaf), showing numerous false positives. Tiny dots are misclassified as distinct epidermal cells. During post-processing, manual (visual) inspection is required to set size-based thresholds to effectively filter out false positives. Please click here to view a larger version of this figure.

Solution	Preparation	Final volume
Base for buffer 1	7.12 g (Na2HPO4 *2H2O) in 200 mL of dist.H2O	200 mL
Base for buffer 2	7.12 g of Na2HPO4 *2H2O) in 200 mL of dist.H2O	200 mL
Phosphate buffer	Mix 36 mL of base 1, 14 mL of base 2 + 50 mL  of dist.H2O	100 mL
Fixative solution	Mix 9.25 mL of 5% glutaraldehyde with 12.5 mL of phosphate buffer and 28.25 mL of dist.H2O	50 mL

Table 1: Recipe for fixating solution²⁴. Composition of the solutions used in this study.

Species	Side	Method	Man (As)	AI (As)	Man (Ws)	AI (Ws)	Man (Ls)	AI (Ls)
Fraxinus	Abaxial	Peel	177.43	9.32	12.71	2.85	17.58	3.58
Taraxacum	Abaxial	Cleared	408.09	18.84	20.1	4.19	26.14	5.39
Taraxacum	Abaxial	Peel	262.54	13.93	15.18	3.47	22.2	4.68
Taraxacum	Adaxial	Cleared	414.48	14.91	20.11	3.44	26.09	5.71
Taraxacum	Adaxial	Peel	258.72	11.37	15.72	3.3	21.15	4.08

Table 2: Indicative comparisons on manual measurements and AI. ML (StoManager1) scaled results. Man and AI indicate, correspondingly, manual and AI (StoManager1) results. Units are as in Figure 5. Manual and AI measurements differ significantly. Some comparisons are not available for various reasons, including missing stomata (e.g., adaxial F. excelsior) or failed detection (e.g., stomata were more difficult to detect in cleared samples due to fuzzier cell contours).

Supplementary figure 1: Examples of common mistakes occurring during leaf clearing that affect the results. Images shown for troubleshooting purposes: (A) uneven staining due to incomplete rehydration of the sample during section 3.10; (B) mechanical damage resulting from mishandling of the chemically processed very fragile sample; (C) green hue in the center of the sample due to incomplete removal of chlorophyll during section 3.5. Please click here to download this figure.

Supplementary figure 2: Example of StoManager1 user interface. Screenshot of StoManager1 user interface showing the input setup and results table (left), and micrograph views with original images (top right) and stomatal detection overlay (bottom right, adaxial surface of a T. officinale peeled sample).Please click here to download this figure.

Supplementary figure 3: Example of Cellpose user interface. Screenshot of Cellpose segmentation in progress through the GUI with adjustable parameters (e.g., model selection - recommended pretrained cyto2 model or a custom-trained model, thresholds, and cell diameter calibration).Supervised manual annotation of epidermal cells highlights each cell in a unique color to guide model calibration. Overlay shows segmentation of a F. excelsior abaxial epidermal peel. Please click here to download this figure.

Supplementary figure 4: Ilastik segmentation process. Overlay illustrates pixel-based classification on a F. excelsior abaxial epidermal peel image. (A) Screenshot of ilastik segmentation in progress showing supervised training with manually labeled classes: epidermal cells (yellow), cell contours (blue), and non-epidermal background such as trichomes (red). (B) Screenshot showing only the epidermis class (yellow), illustrating how the classifier isolates individual epidermal cells. Note the presence of multiple false-positive "tiny" detections, particularly in regions of low contrast or uneven illumination. (C) Screenshot showing only the contour class (blue), illustrating how the classifier isolates cell boundaries from the surrounding epidermal tissue. Note that some contours are discontinuous and that multiple false-positive "tiny" detections occur, particularly in regions of low contrast or uneven illumination. (D) Screenshot showing the background class. Mask highlights stomatal complexes, glandular trichomes, and miscellaneous debris (red), while excluding epidermal cell bodies. Numerous small false-positive "tiny" detections are visible within epidermal regions, arising from grayscale artefacts and texture-based misclassification. Please click here to download this figure.

Supplementary figure 5: Cellpose training settings. Example of Cellpose training settings (GUI) for a cleared T. officinale sample (both leaf surfaces), showing default model parameters prior to adjustment. The pre-trained cyto2 model is selected, with a learning rate of 0.1 and 100 training epochs (default values), which determine the update speed of model weights and the number of training passes, respectively. These defaults are commonly modified (e.g., learning rate = 0.15; epochs = 12) to reduce training time and adapt the model to plant epidermal morphology; however, lower epoch numbers may reduce detection performance and increase false positives. Please click here to download this figure.

Supplementary figure 6: Cellpose "labels" mask. Applying Cellpose "labels" mask for a peeled F. excelsior sample (abaxial leaf surface), showing the result of supervised training in which each segmented object is assigned a unique identifier. Epidermal cells (marked e) dominate, while stomatal guard cell pairs (marked g) are also detected. False positives are evident where trichomes are misclassified as epidermal cells (upper left) or stomata (lower central), as well as false negatives where epidermal cells remain unlabelled across portions of the image. This mask type is used for downstream quantification and refinement of supervised models. Please click here to download this figure.

Supplementary figure 7: Quantification of segmented epidermal cells in ImageJ/Fiji. Screenshots showing segmented epidermal cell quantification following ilastik pixel classification for a peeled F. excelsior sample (abaxial leaf surface). (A) In the exported mask, each detected pixel cluster corresponds to an individual particle, enabling automated extraction of object-level metrics (e.g., area, perimeter, Feret's diameters). The Results window (bottom) displays raw measurements for thousands of objects; note the exceptionally high number of small particles arising from "tiny" dot noise and fragmentary contours, indicating the presence of numerous false positives. These artifacts require subsequent biologically informed filtering to isolate true epidermal cells prior to downstream statistical analysis. (B) Area-based filtering of ilastik-segmented epidermal objects in ImageJ/Fiji. Small fragments and noise are excluded by applying a minimum area threshold of 150-350 µm², reducing the number of detected particles by approximately an order of magnitude (tenfold) relative to the unfiltered mask in panel A. Although this filtering step substantially removes tiny false positives, imperfect classification persists: fragmented contours and residual non-epidermal elements remain and require additional refinement, such as circularity thresholds or biologically informed post-screening. Please click here to download this figure.

Supplementary figure 8: Post-processing of a Cellpose "labels" mask in ImageJ/Fiji. By using the Labels to ROIs plugin for a peeled F. excelsior sample (abaxial leaf surface), each segmented object in the mask is converted into a region of interest (ROI), enabling individual inspection and measurement of epidermal cells and stomatal guard cell complexes. ROIs are overlaid on the original micrograph (left), while the Results window (bottom right) displays cell-level metrics (e.g., area, Feret's diameters, perimeter). This step enables biologically informed curation, allowing users to verify whether each ROI corresponds to a true epidermal cell and to discard artefacts such as trichomes or mis-segmented stomata that persist after threshold filtering, thereby informing subsequent analysis steps. Please click here to download this figure.

Discussion

For consistent stomatal measurements, peels should be taken from the central lamina for the standardized results, while avoiding major veins, as stomatal density shows strong within-leaf variation, decreasing toward the leaf base, tip, and margins^25,26. Fresh leaves yield the most reliable peels, provided that non-shrinking, transparent nail polish and high-quality transparent tape are used to avoid distortion or artifacts. Delays beyond ~12 h reduce contour clarity through dehydration. Because peels are temporarily mounted and may trap air, early imaging is recommended, but quality remain acceptable also after one month.

Various clearing methods exist for specific tissue types, imaging goals, and safety considerations^13,14,15. Chemical clearing was performed using an aqueous NaOH-based protocol for its accessibility, low toxicity, and broad applicability across leaf types. Because tissue responses vary among species, testing and optimization is required. Sclerophyllous leaves may require ≥2 weeks, and residual green tissue can interfere staining and obscure anatomical features (Supplementary figure 1C). For faster clearing, NaOH concentration can be increased to 2.5 M or 5 M. Soft leaves become fragile after bleaching (Section 3.7; Supplementary figure 1B), and overexposure to sodium hypochlorite (NaClO) can degrade tissue or cause epidermal layer separation. While 30 s to 2 min of bleaching is sufficient for most soft leaves, tougher species may require up to 1 h²⁷. Although clearing is slower than peel preparation, requiring ~5 days and no specialized training it can be adjusted for diverse species. Cleared and stained samples can be stored in distilled water at 4°C, with permanent mounting recommended for long-term preservation^15,27.

Stomata specific StoManager1 reliably detected stomatal complexes across most image types (Supplementary figure 2), with detection results generally consistent with manual annotation for peel-based methods. However, stomatal area measurements differed systematically from manual measurements (Table 2), necessitating calibration based on a small manually measured subset (e.g., 10-20 stomata per species) prior to large-scale analysis. StoManager1 substantially increases throughput, with manual verification required in non-standard cases.

Simultaneous stomata-epidermis segmentation is challenging because optimal sharpness differs, and heterogeneity in morphology and membrane clarity introduces segmentation error. Same way training on large image set (e.g., adaxial/abaxial surfaces, peeled/cleared) lose specificity. No pre-trained Cellpose model exists for plant epidermis; the standard "cyto" and "cyto2" models, trained primarily on animal cells, underperform on epidermal images. Custom training is therefore required for reliability (Supplementary section 2). Even after training, over segmentation and false positives remain common, particularly in low-contrast or uncleared images (Supplementary figure 3).

In ilastik (Supplementary section 3, supplementary figure 4), object masks should be generated for the target cell classes (e.g., class 1 for epidermis), while retaining class masks for auditing and manual correction. Batch processing and subsequent ImageJ/Fiji workflows can save time and reduce manual error. Consistent filtering by area, standard deviation, or quantiles is essential to suppress noise and spurious detections. After filtering, median cell area and cell counts approach manual values. Benchmarking against manual measurements and alternative ML tools (Weka, RootPainter) remains essential for tuning parameters and assessing segmentation quality. ML batch processing is resource intensive. Stable training and large-scale batch segmentation require a NVIDIA CUDA-capable GPU and ≥32 GB RAM. On mid-range systems (e.g., 16 GB RAM with integrated graphics), supervised training typically requires ~5-10 min for <10 images and frequently crashed when multiple annotation classes or heterogeneous image types were used. CUDA increased processing speed by ~10-20x relative to CPU or integrated-GPU execution.

Differences in performance among Cellpose, ilastik, and StoManager1 arise from their distinct feature representations. Cellpose predicts vector-flow fields toward object centers and is sensitive to optical halos and abrupt refractive boundaries, which disrupt flow and generate clusters of false positives in cleared F. excelsior. ilastik, as a human-guided pixel classifier, oversegments high-relief pavement cell walls or halos unless explicitly labeled as background. StoManager1 detects stomatal complexes using object-level shape patterns (paired guard cells with a central pore) and performs well on peel preparations, including F. excelsior²⁸; however, halo-induced blurring affect pore visibility on clearing to produce false negatives rather than misclassification (Supplementary filesection 1 and 3). Accordingly, Cellpose and StoManager1 operate as automated segmenters, whereas ilastik functions as a human-guided classifier (Supplementary filesection 1-3)^28,.

Species-specific cells architecture modulates AI performance. In T. officinale, smooth, elliptic guard-cell margins match the StoManager1 training distribution, enabling detection on peeled samples (Supplementary figure 2), whereas generalist tools frequently confused guard and pavement cells. In F. excelsior, recessed stomata and high-relief pavement cells lead to false negatives on cleared images. These discrepancies reflect species- and preparation-dependent imaging geometry. Systematic differences between manual and AI-derived measurements arise from two sources: the lack of scale encoding from image by AI and preparation-induced optical artifacts. Mitigation strategies include visual triage of blurs and halos from tissue preparation, preferential use of peels under conditions described above, manual calibration with linear correction on a small subset of stomata (manual = a + b × AI). Additional strategies include taxon-specific augmentation and auxiliary class design for training (however, own trained are likely to remain lower than pre-trained models on large, diverse datasets), and post-segmentation upper quantiles filtering.

Because AI segmentation errors are systematic non-random, post-processing is essential. Image-quality diagnostics can follow a simple decision tree. Start by checking epidermal contrast and sharpness: if sharp, move to false positives; if blurred, expect small high-frequency artifacts and adjust filters. If false positives (e.g., dots, debris) are minimal, ±2-3 SD outlier filtering usually suffices. If small artifacts dominate, use quantile-based filtering. Then compare trait distributions (e.g., median Ae) to biological expectations or manual data. Refine thresholds as needed - keeping only upper 10-50% quantiles in artifact-rich datasets ensures meaningful outputs (Supplementaryfigures 7 and 8). Due to the high false-positive rates produced by epidermal ML tools, reliable stomatal density estimates could not be derived automatically, whereas StoManager1 provided precise stomatal measurements²⁸ when boundaries were clear (see Figures 6 and 7). For epidermal median Ae proved robust, being less sensitive to small false positives than the mean.

Method choice influenced trait detection. Peeling outperformed clearing for stomatal density by fully exposing stomata (Figure 3), whereas clearing often obscured smaller stomata in layered or pigmented tissues, despite sometimes better preserving stomatal shape²⁷. Among AI tools, StoManager1 consistently provided robust stomatal measurements with minimal curation, particularly on peel preparations²³, and peeling also improved epidermal measurements by increasing correctly identified cells and image coverage². Future gains are most likely to arise from large, taxonomically diverse pre-trained models for integration into automated imaging pipelines for ecophysiology and crop improvement.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Disinfectant liquid cleaner NaClO <5%	Domestos, GB	H000023244	Skin burns
Disodium hydrogen phosphate dihydrate (Na2HPO4. 2H2O)	PENTA, CZ	10028-24-7	NA
Ethanol, abs. 100% a.r.	Chem-Lab NV, BE	CL00.0505.1000	Highly inflammable
Glass vial 2 mL	VWR Life Science, US	548-0045
Glutaraldehyde 50% solution	VWR Life Science, US	23H2856331	Fatal if inhaled. Toxic if swallowed.  Wear protective gloves,  clothing,  glasses.
Microscope slides	Normax, PT	5470308A	NA
Nail care top coat fast dry	BK (INEZA), LT	LT-02244	NA
Nikon Eclipse E600 and Nikon DS0Fi1 5 MP	Nikon Corporation, JP	https://www.microscopyu.com/museum/eclipse-e600	NA
Pipette and pipette tips	Thermo Scientific, FI	KJ16047	NA
Single edge steele blades	Oxford instruments	51-1625-0182	NA
Sodium dihydrogen phosphate dihydrate (NaH2PO4 . 2H2O) pure	PENTA, CZ	13472-35-0	NA
Sodium hydroxide pellets (NaOH) pure	Merck KGaA, DE	1310-73-2	Severe burns.
Stationery clear tape	FOROFIS, LV	91231	NA
Toluidine blue, general purpose grade	ThermoFisher Scientific, GB	2045836	NA