The genotypes used in this study were chosen from a previous common bean population screened by F. brasiliense to represent diversity in disease severity (Wang et al., 2018). From the Cal96 × MLB49‐89A population, 13 diverse recombinant inbred lines and the two parental lines were included, for a total of 15 lines used. Two bean seeds were soaked in 10% bleach sterilization solution for 20 min and then washed five times in sterile deionized water. Seedlings were germinated and grown in CYG germination pouches (Mega International, Roseville, Minnesota, USA), then placed in a BioChambers BigFoot series growth chamber (BioChambers, Winnipeg, Manitoba, Canada) with 250 mE light intensity, 70% humidity, and 25/20°C day/night temperature on a 14‐hour day cycle. After seven days, one seed was treated with sterile deionized water (Mock), while the second was treated with F. brasiliense inoculum (FRR). Seedlings were grown for an additional seven days before data collection. This experiment was repeated five times.

Isolates of F. brasiliense (F_14‐42) were grown using aseptic techniques on plates of potato dextrose agar medium for approximately 25 days at 21–24°C. Inoculum was prepared at a 1 × 10⁵ concentration by scraping the plate, rinsing with 50 mL of sterile deionized water, filtering through a 100‐nm mesh, and using a hemocytometer to count macroconidia. Using small spray bottles, seedling roots were sprayed evenly with either deionized water or inoculum.

Seven days post‐inoculation, plants were harvested and roots were scored for disease severity as described previously by two trained researchers (van Schoonhoven and Pastor‐Corrales, 1987; Wang et al., 2018; Haus et al., 2020, 2021). The roots were scored on a scale from 1–9, where “1” indicates no disease symptoms or discoloration and “9” indicates severe discoloration and extreme necrosis. Disease scores were averaged between scorers. Root images were collected using an Epson Perfection V550 photo scanner (Epson, Tokyo, Japan) to preserve quality, lighting, and consistency. Roots were separated from the stem where the stem color shifted from white to purple or green. The roots were dried in a drying oven for three days and then weighed for biomass.

To collect root image measurements, nine root images were carefully traced in Fiji (version 2.6.0; ImageJ version 1.53q; Schindelin et al., 2012) using the freehand tool, and non‐root portions of the image were deleted. Manual curation was averaged across two researchers, and each image took between 6–24 hours to complete. The rectangle tool was used to encapsulate the entire root system (excluding the stem), and the height and width of the resulting box was obtained by using the Measure capability. The area was measured by counting the total pixels remaining after manually removing the background using the Analyze Particles feature.

PlantCV has been used in a variety of studies for phenotyping both dicots and monocots (Gehan et al., 2017; Enders et al., 2019; Acosta‐Gamboa et al., 2020; Kumar et al., 2020). We employed PlantCV as the backbone of the RootDS preprocessing pipeline by using the naive Bayes classifier function to distinguish and classify pixels based on red‐green‐blue (RGB) values (Gehan et al., 2017). More documentation can be found at https://plantcv.readthedocs.io/en/latest/naive_bayes_classifier/, and a tutorial is available at https://plantcv.readthedocs.io/en/v3.11.0/machine_learning_tutorial/.

Prior to running RootDS, a probability density function (PDF) classifier file was created using Fiji (version 2.1.0) to train the Python program. RGB color values were sampled in Fiji (version 2.1.0) using the Pixel Inspector tool for four categories: Background, Shoot, Non‐Diseased Root, and Diseased Root. The PDF file was created manually using a subset comprising approximately 10% of the images from the original image set. Each category can have differing levels of complexity, such that some categories may require only 20 RGB data points for accurate classification, whereas more complex categories may require close to 100 RGB values. A category with too few RGB values will not be precise enough, resulting in miscategorized pixels, but a category with too many RGB values can oversample, causing over‐prioritization of that category. Each category in this data set contained a sample of approximately 100 different RGB values. Once defined, the PDF file is used by the naive Bayes machine learning algorithm to classify pixels into one of the four categories, and each pixel is binned during the RootDS processing steps.

Parameters were assigned prior to running RootDS, indicating file locations for the PDF file and input and output destinations (Figure 1). The pipeline batch‐processes the image set in a recursive loop by first creating the necessary output folders and printing the original image for future reference. Each pixel in the original image is classified using a naive Bayes algorithm. Four binary masks are created based on the four categories created in the PDF classifier file, and separate binary images for both the diseased and non‐diseased root portion are saved (Figure 2). A compilation of the binary images of the root system (stem, diseased roots, and non‐diseased roots) is also saved. Merging the masks into a pseudo‐colored root image allows for visualization of how each part of the root system was classified.

In addition to the segmented and merged images, the percentage of the diseased root system and the ratio of root system pixel count to total image pixel count is compiled into a CSV file for each image. Root disease severity was calculated by removing all pixels binned as stem and background, then dividing the sum of pixels classified as diseased by the total root pixel count to find root disease severity expressed as percentage of diseased pixels in the total root system. To calculate root area, the total root pixel count is divided by the total pixel count of the image to get the ratio of root area to total image area. This method is equally as accurate as compiling the pixel count of the root system but allows for multiple images of multiple different resolutions to be processed in the same data set. The code used in this study can be found in GitHub (https://github.com/HausMJ/RootDS_PythonCode).

Rhizovision Explorer (version 2.0.3; Seethepalli et al., 2021): Root scans were analyzed using the “Whole root” analysis mode, inverted using the inversion tool, and thresholded at a value of 170. The region‐of‐interest (ROI) tool was used to remove any stem from both images. This was done by using the ROI box to capture the entire root system from the uppermost root to the lowermost root, as well as including the entire width of the system. The images were analyzed and the following measurements in pixels were recorded: “Depth (px)” for length, “Maximum Width (px)” for width, and “Network Area (px2)” for area.

DIRT (Das et al., 2015): None of the original root scans gave a successful output from DIRT. To overcome this, root scans were thresholded in ImageJ by splitting into their RGB channel and selecting the red channel. These images were uploaded into a single root collection, and each marked collection was analyzed at the following settings: Masking threshold of 10.00; scale marker of 0.00 (which provided pixel measurements); has root crown; requires segmentation; 0 excised root.

RootDS binary images were run through both Rhizovision and DIRT using the same settings mentioned above.

All statistics were performed in R (version 4.0.4; R Core Team, 2022). Histograms were plotted for both manually generated disease scores as well as computer‐generated disease percentage. Disease percentage was correlated with manual disease score, and root weight was correlated with root pixel percentage using Pearson correlation (cor.test function in R). Graphs were made using ggplot2 (Wickham, 2016).

The percent differences with manually calculated measurements were compared between original images with no preprocessing and images that underwent RootDS preprocessing. Differences were averaged for each combination of processing and software used. Standard error of the mean was calculated for each combination.

We developed a Python‐based (version 3.8.6) software package for use in disease severity phenotyping and in the generation of binary root system images for processing in other RSML software (Figure 1). Examples of the output images made from each input image can be found in Figure 2. From the original image, two binary masks are made, one for the diseased and one for the non‐diseased root tissue. Additionally, two merged images are created: one pseudo‐colored image to visually assess tissue classification and another binary image for input into RSML software programs.

To evaluate the accuracy of the pipeline results, disease percentage and root area were compared to manually calculated measurements (manual disease scores and root weight). The percentage of pixels comprising the diseased root mask relative to pixels comprising the whole root system was used as a quantitative representation of disease severity. The results showed that computer‐generated disease scores provide greater capture of variation than manual disease scores (Figure 3). When comparing equivalent histogram bins between the data sets, computer‐generated disease percentages have broader variation for both FRR‐treated and Mock‐treated roots (Figure 3A) compared to manually annotated scores, which are only semi‐quantitative and so are not continuously distributed.

Computer‐generated disease assessment shows a strong correlation with manual hand‐scoring methods (R ² = 0.9154, P = 2.2 × 10⁻¹⁶), indicating a high accuracy rate when using RootDS (Figure 4A). Outliers were attributed to human factors, for example, uneven lighting throughout the image or the presence of debris that resulted in misidentified pixels. The ratio of root system pixels (disease and non‐diseased) to total pixels in the image denotes the area of the root system relative to the area of the total image. Computer‐generated whole root system pixel percentage had a strong correlation with root weight (R ² = 0.8962, P = 2.2 × 10⁻¹⁶), indicating a high accuracy in distinguishing plant tissue (Figure 4B).

To demonstrate the pipeline's potential as a processing step prior to use of RSML software, both manually calculated and merged binary images were processed through two available RSML software programs, Rhizovision Explorer (Seethepalli et al., 2021) and DIRT (Das et al., 2015). The depth, width, and area were measured for each image in both the manually traced and RootDS‐processed form (Figure 5). The RootDS processing pipeline led to results that closely aligned with manually calculated values and produced data that were similarly accurate to manual measurements using Fiji (Figure 5). Rhizovision Explorer was more accurate than DIRT for depth and width in either processing pipeline but performed similarly for area.

The available code and a subset of the images are available on GitHub (https://github.com/HausMJ/RootDS_PythonCode). The full data set is available upon request (email hausmira@msu.edu) and will be made available in the same GitHub repository after a one‐year embargo period after article publication.