rosettR: a tool for screening seedling areas and growth rate
rosettR is a phenotyping protocol for tracking the growth of Arabidopsis seedlings over time (Fig. 1). The prerequisites for using rosettR are a fixed digital camera, a growth chamber, tissue culture plates, medium, a computer, the freely available software R [25], and the seeds of the genotypes to test.
Images of the plates with the growing seedlings are taken at regular intervals, and the rosette areas are estimated from the images, requiring minimal user interaction. Template reports are generated to support sowing, quality control, and area and growth rate comparisons for each experiment. The protocol is compatible with a wide range of stress treatments applied at any desired time point during the experiment. Treatments such as different light regimes or temperatures can be applied by simply shifting the plates to the desired condition. Other treatments can be applied from germination by supplementing the medium with sugars, sorbitol, salt, among others. Another option is to place a membrane on the solid media and sow directly on top of the membrane. The seeds/seedlings and the membrane can be transferred to new plates with the desired supplement at a later stage of the experiment.
Starting a new experiment
A new experiment can be started by loading the package in R and providing information regarding the genotypes, treatments, time-points at which pictures will be taken, the number of repetition blocks, and the genotype to use as reference. The package provides high-level functions to create a directory tree where images are to be placed and reports that define the randomized block design to facilitate sowing and the placement of the plates in the growth chamber. Images are then taken at the pre-defined time-points and saved in the corresponding directory. Once all images have been taken, they can be analyzed to compute areas and relative growth rates.
The tissue culture dishes we used are 150 × 25 mm and have a grid that divides the plate in 32 squares. A single seed is placed in each square, so each plate can have a total of 32 seedlings from different genotypes. Plants from different genotypes are sown on the same plate in alternating combinations to account for differences between plates.
Half-strength Murashige and Skoog media with 1% glucose is poured in the plates in sterile conditions, and the sterile seeds placed on the solid medium with a Vacuumseed or sterile toothpick/pipette in the corresponding square. The plates are then sealed with Urgopore tape, wrapped in groups of 10 with transparent foil, and placed at 4 °C in darkness for three nights for seed stratification. Plates are placed on the shelf in a growth chamber, and the plants allowed to grow at a temperature of 20–22 °C and 150 μmol m−2 s−1 light intensity.
Images should have a uniform bright background to avoid any shadows and allow for accurate detection of the seedlings. We recommend using a white backlit imaging table with homogeneous background, avoiding any formation of shades, and the camera mounted firmly right above the plates for the whole duration of the experiment. Preferably, the imaging is done inside the growth chamber to avoid temperature differences that may cause condensation on the lid of the plate or affect experiment treatments.
Image analysis
All images are expected to be taken with the same zoom factor which is manually defined by indicating two points on an image at a set distance in millimetres using the calibrateScale function. After that, the remaining estimation and recording of rosette areas for each seedling is achieved by a fully automated workflow depicted schematically in Fig. 2.
Step 1: Displacement correction Seedlings are expected to be sown in a pre-defined grid of configurable dimensions. The first step of the image analysis is to make sure that the grid is centered and at a right angle to the image edges. The exact x- and y-coordinate of the plate centers are determined using a Nelder and Mead [29] optimization algorithm that maximizes the fraction of dark pixels (plate) to light pixels (background) within the circle. Plate rotation is compensated by applying step-wise rotation of the image at a given interval (e.g. −5°:5°), and by interpolation choosing the angle that minimizes the common standard deviation of a multi-component normal distribution with means at the centres of each square in the grid. Conceptually, this can be thought of as moving along a forest with trees planted in rows, and then stopping when all trees align and one can see the other side of the forest. In order to test these correction steps, we performed a small simulation study applying known dislocation of the plate as well as small rotation to 100 images. The algorithm could accurately recover the displacement in horizontal (x) and vertical (y) direction (Fig. 3). Detecting the rotation was less precise but still of sufficient accuracy to correctly identify the grid in the corrected image.
Step 2: Thresholding The default behavior of rosettR is to convert images to greyscale by keeping only the blue channel as healthy plant material absorb blue light thereby appearing dark. For seedlings with strong discoloration, the weighting of the three channels (red, green, blue) can be adjusted by the user. A threshold for segmenting the plate in foreground (seedling) and background (plate) is determined by fitting a mixture model of two normal distributions truncated at 0 and 1 to the plate region of the image, yielding a tuned threshold per image. In our experience, this works well since the background is bright and homogeneous resulting in a distinct class of bright pixels, whereas all darker pixels can be assumed to be the seedling.
Step 3: Sorting image features Image features are identified using the bwlabel function in EBImage [26] and each feature is sorted to the square it occupies the most. The area of each plant is then finally estimated as the sum of all image features allocated to each corresponding square using the computeFeatures.shape function. This sorting procedure implies that leaves that are detected as detached from the rosette but still mostly in the right square, or still attached to the seedling but predominantly in the wrong square, will still be classified to the right seedling even if it reaches into the neighboring square.
Step 4: Quality control Finally, a quality control image is generated for each plate where the outline of the plate and each square is indicated [see example in Fig. 4 (Step 4)]. Features for the same plant are colored with the same color and squares where plants had been found to be merged between squares are outlined red (not shown in the figure). Plants residing in such ambiguous squares are marked in the final data sheet and can be ignored during comparisons of plant areas and relative growth rates.
Once image analysis has finished, a template report can be compiled that shows all quality control images in a convenient overview, as well as growth curves and boxplots highlighting plates with ambiguous squares or outliers (Fig. 4).
Data analysis and visualization
The main output of the image analysis is a spreadsheet with estimated areas and relative growth rates for each plant. With this data, it is straight-forward to perform statistical analysis to compare genotype and treatment effects as needed given the exact context of the experiment. For comparison between examined genotypes and a reference of choice, and as an example in general, the compare areas template report can be generated non-interactively after successfully completed image analysis. For estimation of effect sizes and significance, we use a 2-way ANOVA (genotypes and treatments) and the multiple comparisons framework described in [30]. An example of an area comparison report can be seen at [31].