Spring wheat (Triticum aestivum cv Paragon) was grown as single plants in compost (3.5L Levington F2) in the greenhouse (day temperature set to 20 °C and night 15 °C, 14 h day length) until the sixth leaf stage and then split into 2 equal groups, one watered to 80% field capacity (FC) (high water—HW) and one to 40% FC (low water—LW). Plants were imaged and gravimetrically watered daily using a LemnaTec Scanlyser system until the primary tiller was at Growth Stage (GS) 39–41 (Zadoks scale) which approximates to meiosis (primary tiller was tagged). Plants were then subjected to different temperature regimes as follows: 25/20, 30/25 and 35/30 °C (day/night set air temperature respectively) for 4 days, and then returned to ambient conditions within the glasshouse to complete their life cycle and ripen. At harvest the primary tiller was weighed and retained for scanning. Other parameters including the total dry weight of the plant and ears and height of the primary tiller were also collected.

For each treatment, twelve representative, fully dried primary spikes were selected for scanning and placed in plastic holders (34 mm diameter, 70 mm height). The majority of the spikes were too tall to fit in the holders so they were cut into two pieces and each scanned separately. Pieces of thermoplastic starch were used to eliminate sample movement while scanning. Sample preparation and loading into the scanner takes around 30 min per 12 samples and after this time there is no more user input. The twelve holders were loaded into the sample changing carousel of a μCT100 scanner (Scanco Medical, Switzerland). This scanner has a cone beam X-ray source with power ranging from 20 to 100 kVp (pre-set and calibrated for 45, 55, 70, 90 kVp) and a detector consisting of 3072 × 400 elements (48 µm pitch) and a maximum resolution of 1.25 µm. The samples can be positioned at different distances from the X-ray source greatly improving resolution while keeping scanning time to a minimum. Spikes were scanned with the X-ray power set at 45 kVp and 200 µA with an integration time of 200 ms. Each spike was ~ 1000 slices (51 slices per stack), 125 projections/180° were taken and a binning of 6 was used. Output images were produced with a 0.2 megapixel (512 × 512) resolution (68.8 µm/pixel) in a proprietary ISQ file type format (Scanco Medical, Switzerland).

The 3D volume was reconstructed from the projections (raw data, including flat field correction data) using proprietary software supplied with the Scanco μCT100 scanner. After 3D volume generation, the developed processing pipeline makes use of standard computing hardware. A DELL XPS desktop computer with an Intel (i7 6700k) 64 bit CPU, 64 GB of memory and an NVIDIA GPU (GTX 1080) was used.

A computer vision and analysis protocol was developed using a combination of the MATLAB [26] image processing toolbox and Python [27] (Additional file 1). Visualisation of processed images, stored as TIFF files, used ImageJ’s 3D viewing plugin [28] and TomViz [29]. Post-processing of results used the Scientific Python collection of software (SciPy). A full list of additional software packages can be found in supplemental table (Additional file 2: Table S1). All reconstructed 3D volumes and segmented images can be accessed at https://www.aber.ac.uk/en/research/data-catalogue/a11df174-d73d-4443-a7fd-ab5b7039df79/ [30].

A high-throughput automated pipeline was developed with the goal of reducing human input and time. Reconstructed 3D volumes were retrieved from the µCT scanner and MATLAB scripting then performed feature extraction. All the source code as well as user instructions are available from https://github.com/NPPC-UK/microCT_grain_analyser. Analysis of the resulting data is performed using Scientific Python libraries.

Python scripts were used to automatically find data files and match them with information about their scanning parameters, correctly label and then compile data into tables (data frames) based on treatments, prior to analysis. Data was analysed using a collection of Scientific Python packages [27] and statistical analysis was performed using one-way ANOVA with significance calculated at p < 0.05.

Computer vision approaches have been previously used to extract quantitative grain characteristics from μCT images but these tend to require high resolution images and long scan times [25]; this makes μCT expensive, hard to scale up to population size samples and technically difficult to apply to new species. In order to overcome these issues we developed a high-throughput, automated method using relatively low resolution images acquired from a bench top scanner that is easily applicable to species with diverse spike morphologies and grain sizes.

The initial test population consisted of naturally ripened dry wheat spikes (cv. Paragon) harvested from plants which had been subjected to different defined watering and temperature regimes. Dry spikes were harvested and 12 spikes per treatment were scanned at a resolution of 0.2 megapixel (512 × 512 × ~ 1000; 68.8 µm/pixel). This resolution was chosen to allow fast scan times and increase throughput while still retaining sufficient image information necessary for accurate data acquisition and analysis. Scanning time for each spike was around 40 min. We found that performing the scans at higher resolution (1024 × 1024 × ~ 2000; 34.4 µm/pixel) increased scanning time to 60 min and this did not translate into an increase in the quality of the data output for the analysed grain traits (Additional file 3: Table S2). Therefore, we chose a resolution of 512 × 512 × ~ 1000; 68.8 µm/pixel for routine scans. Higher resolutions (2048 × 2048 × ~ 4000; 17.2 µm/pixel) increased the scanning time to 3.3 h and produced a 32 Gb that was too large for routine use, but these may be useful for measurement of tissue related traits (such as thickness of the bran layer or embryo size) as discussed below.

Our aim was to develop a pipeline that could automatically identify and measure different grain parameters from these reconstructed volumes. The measured parameters included spike height, grain number, grain height, width and depth, grain volume and surface area (Fig. 1).

A schematic representation of the pipeline used from scanning to data analysis is presented in Fig. 2. The pipeline is divided in three sections: μCT scanning and 3D volume reconstruction done by the Scanco software (Scanco medical, Switzerland); segmentation and 3D processing; and feature extraction and analysis both developed using MATLAB and Python. This pipeline is readily applicable to other plant species with varied spike and grain morphologies, and scanned at different resolutions, simply by adjusting the structured element size, the resolution and the minimum size the as detailed in Additional file 1 (setup.m). As default we set the structured element size at 5 and the minimum object size at 1000, parameters that perform well for most of the species tested. For species with very small seeds the minimum size object can be reduced.

After 3D volume reconstruction, the files were exported to a data storage server. The first computational task performed was converting these data into a malleable, workable format. For this, a file-reader in MATLAB (available from [31]) was developed to generate image stacks. The 3D reconstructed volume as collected from the μCT scanner has a dual peak distribution of grey values and the use of this information has been essential in constructing an effective method for removing all non-plant material from an image [25]. To further segment the plant material of interest, we developed an adaptive thresholding method that enabled both removal of non-plant material and segmentation of grain and non-grain data (Fig. 3a, b). This developed method relied on taking a cumulative sum of grey values across all slices in a single scan and computing a minimum value for plant material. Additional file 4: Fig. S1 illustrates that material with a density value within the pink shaded area is of interest for this method.

Post thresholding morphological operations were used to perform further sanitization on the segmented images. A disk shaped structuring element (SE) [32] was used to erode the image, and then we applied a median filter before dilating. This allowed for the removal the majority of the non-target plant material and artefacts of the scanning process and obtain a segmented image (Fig. 3c, d). From this image a minimum grain size parameter is used to filter out any remaining noise, thus an image is produced containing exclusively grain material. Finding and removing the largest cluster of connected pixels throughout the 3D image allowed elimination of the outer tube used for scanning (Fig. 3e). To precisely preserve the shape of the grains along with the grey levels and ensure that there was no data loss, this segmented black and white image was used to mask the original one (Fig. 3f). After this step, the cleaned image can be used for trait extraction.

After segmentation and separation of all fused objects, each isolated grain was orientated by calculating the major axis of the 3D shape and was fitted to an ellipsoid in order to calculate circularity (Fig. 5c–e). The length was calculated by measuring the size of the major axis while width and depth were found by examining a cross section of each grain and taking measurements of major and minor 2D axis respectively. Volume is the complete connected-pixel count for each given grain. Surface area was calculated by adapting previously successful methods [25]. The calculation of positional data for each grain required registration at a post-processing stage, due to splitting spikes for scanning (see below).

Once compiled, grain parameters in pixels were converted to metric units by the following equation (mm=pixel∗conversion1000) and this data was then exported as a CSV formatted file on a per-spike basis (Additional file 5: Table S3).

Due to size limitations of the available scanner some of the samples were scanned in two portions. The positional data in the Z axis was inverted before in silico spike rejoining was performed Z=maxZ-Zi, by incrementing top-portion scans globally by the size of the bottom scan (Zit=Zit+lengthZb). This could be easily generalised to 3 or more portions for longer and larger structures.

The beginning and end of a spike was calculated by locating the lowest and highest rachis points respectively in the Z axis. Detection of the rachis is done by relaxing the thresholding algorithm by 20% to detect less dense plant material.

In order to evaluate the accuracy of the software in determining grain number and volume, different approaches were taken. For grain number, three separate counts were done manually on reconstructed 3D volumes. One verifier used a counting technique that involved examining each individual Z slice while the other two examined the data set in a 3D image viewer. This provided a grain count on a per spike basis that could be directly compared to the one obtained from the computer vision approach. As can be seen in Additional file 6: Fig. S2a, the software was able to accurately identify and count grains.

To confirm that no data was lost during the segmentation process, several randomly selected processed image files were compared with their original counterparts, with contour maps drawn and manually examined. This showed that there was no data loss during the segmentation process (Additional file 6: Fig. S2b).

These ground truthing results clearly demonstrate the high degree of accuracy achieved with the developed method.

We also found a strong correlation (R ² = 0.75) between the total volume of all the grains in a spike as measured by our method and manually acquired spike weight for all the spikes (Additional file 6: Fig. S2c).

After establishing the robustness and accuracy of the data provided by the segmentation method, we then asked whether it could provide insight into the combined effect of water regime and temperature stress on grain characteristics.

μCT scanning confirms that grain development is differentially affected by water and heat and, in addition, that the developmental position along the spike modulates these effects. By using 3D reconstructions of whole spikes, the spatial distribution of grains along the spike is maintained and this can be overlaid with the traits measured for each grain (Fig. 6). This allows for a visual representation of how grain traits change along the spike and also how the different stresses affect those traits. A high degree of resolution was achieved in the y and z axis (Fig. 6a); this is illustrated by detection of occasional secondary spikelets that protrudes in the y-axis (Fig. 6a circled region). Generally, the middle region of the spike (in the z axis) contains more grains than the top and bottom and grains at the top of the spike are smaller (Fig. 6a, b). The effect of temperature on spike height and grain number is clearly visible with spikes grown at 35 °C showing reduced height and reduced number of grains (Fig. 6a). The major reduction in grain number occurs in the middle of the spike with the top and bottom regions being less affected (Fig. 6b). Increased temperature also leads to a reduction in grain volume at the top of the spike but this is only observed in the plants grown in a high water regime (Fig. 6b). On a per spike basis, we found an inverse correlation between average grain volume and grain number with the temperature stressed spikes harbouring fewer but bigger grains (Fig. 6c).

Temperature had an effect on spike height with spikes being shorter in both water regimes as the temperature increased while water supply alone did not have an effect on spike height (Fig. 7a). Temperature also had a dramatic effect on grain number with temperature increases significantly reducing seed number per spike in both watering regimes (Fig. 7b). Surprisingly, we observed that the lower water regime resulted in an increased grain count at a given temperature compared to the high water although this effect was lost at the highest temperature (Fig. 7b).

Grain volume was also strongly influenced by growth conditions, shown by a general increase of individual grain volume with increased temperature (Fig. 7c). As noted for Fig. 6c, the observed grain volume increase is associated with a decrease in grain number for a given treatment (Fig. 7b, c). The observed increase in volume is a result of a general increase in grain length, width and depth (Fig. 7d–f).

There is a scarcity of organ-level imaging approaches that lend themselves to rapid quantitative measurements suitable for in-depth physiological or genetic dissection and modelling. Light and electron microscopy are widely used but they provide limited information and tend to be labour-intensive to produce [38]. Other techniques using conventional cameras that rotate around the subject can also generate accurate 3D model but do not provide information about the internal structure of plant material [39, 40]. As the organs of interest are often embedded within other tissues, the techniques described above require the manual removal of the surrounding tissue. This can be time-consuming and spatial/developmental information is easily lost. Methods have been designed to allow automated removal of grains from the spike while retaining positional information, but these are highly specialized and expensive instruments [41].

These limitations can be largely overcome by μCT. μCT has traditionally been used, to great effect, in medical imaging, and its applications in plant science have increased over the last few years [17–25]. Methodologies developed in the medical field have been applied to wider biological studies, for example techniques used to model regions of the human heart [42, 43] have more recently been used to examine seed anatomy [18] and methods used to study metamorphosis in insects [44], modified to track root development in soil and non-destructive floral staging [19, 45].

One of the critical advantages of μCT imaging is that positional information of organs and tissues is preserved and can be analysed. This is extremely important when looking at changes throughout development and variation in grain traits within each spikelet or along the length of a spike is a good example. Imaging of internal tissues and organs without dissection is also possible, although this will require scanning at higher resolutions. Thus non-destructive imaging of the bran layer and the embryo, both of which are economically important traits, could be further developed and scaled for breeding and quality control applications. Finally, detailed study of specific 3D grain parameters such as circularity, surface area and crease volume which are agronomically relevant are also made possible by this method.

Underlying the increased use of μCT in plant biology has been the development of more affordable small, and even benchtop, μCT scanners with sample loading carousels more suitable for larger sample numbers. However, their use necessitates a number of trade-offs between sample number, size and data quality. For example, the loading carousel imposes physical limitations on the size of individual samples and we had to divide many spikes. To re-integrate measurements taken from separate portions of the same spike, we identified conjoining points along the rachis of each spike and rejoining images was introduced as an additional processing step. Further issues can emerge from the use of a fixed X-ray beam which rotates the subject to obtain a 360° image. This gives opportunity for movement during scanning resulting in minor image distortion. To limit movement, scanning material was held in place using thermoplastic starch that, although visible in the scan, can easily be removed by the application of morphological filters during image processing. The time required to produce and reconstruct high resolution scans represents, perhaps, the most serious bottleneck for routine grain analyses. For a wheat spike this can take several hours using typical hardware. To address this, the scans were performed at the lower resolution of 0.2 megapixels (512 × 512) rather than much higher resolutions used in previous studies, for example 5 megapixels (2048 × 2048) and larger is often used [19]. This also reduced the output file size on average by a factor of 16. The trade-off for this lower resolution was the decrease in spatial accuracy resulting in the incorrect joining of juxtaposed objects; this was rectified during the segmentation process.

During our initial attempts to analyse the data produced through μCT we discovered that there was a lack of software that could handle the volume of the data and implement modern computer vision algorithms easily and was well-suited to high-throughput automation. VGStudio Max, a commercially licensed software package, and BoneJ, a free and open source software package, are often used in biological and medical science for image analysis and visualisation [19, 46, 47]. However they require human interaction on a per image basis. While this level of interaction is justifiable for high value subjects (i.e. in a medical context), the scale required for crop biology demands minimal intervention.

This prompted us to design and create a new computer vision based methodology. Our aim was to develop an entirely adaptable system which we could build upon in future, and robust enough to work with a multitude of grain shapes and sizes. The MATLAB [26] scientific programming language and environment provided a widely available professional platform that has closely related open-source alternatives (Octave [48]) that can be used to implement our method, albeit with reduced functionality (some of the watershedding techniques are not yet implemented in Octave).

As a proof of principle, the developed methodology was used to study the effect of temperature and water regime on spike development and grain traits on a population of wheat plants. We found that temperature differentially affects grain formation along the spike with the middle of the spike being more sensitive to the stresses. Recent studies have shown that there are two discrete developmental stages where the spike is more sensitive to temperature: early booting when meiosis is occurring and anthesis [8–10]. Floret development along the spike is asynchronous [12] it is thus tempting to speculate that the florets in the middle were at a temperature-sensitive stage when the stress was applied. In agreement with previous reports [9] we also found an inverse relationship between grain number and grain volume across treatments. While high temperature and high water regime caused a decrease in the number of grains per spike, the average volume of grains increased, partially compensating for grain loss. It should be noted that the low water plants were slightly ahead in terms of spike development when the heat stress was applied and this could explain why in these plants’ temperature has a less detrimental effect on grain number per spike. Despite suggestions that grain height, width and depth are affected by independent sets of genes [49], our data indicates that the response of these traits to different growth conditions are highly correlated. It will be informative to extend these studies to diversity and mapping populations to explore how changes in spike architecture and grain traits in response to multiple stresses are genetically controlled.

Finally, to demonstrate the wider applicability of the method, we examined different species (foxtail millet, oats, darnel ryegrass and ryegrass) that illustrate a diversity of inflorescence and grain morphologies, from the dispersed open panicle structure of oats to the very congested structure of millet which has numerous small grains packed together (Fig. 8). In all cases, simply by adjusting two parameters (the structuring element size, and minimum grain size) our method identified the grains and performed grain feature extraction accurately (Additional file 7: Table S4).

Grain uniformity is of economic value in many cereals and is an active breeding target. Grain size changes not only along the spike but also within each spikelet. Relating the position of an organ in physical space to its position in developmental space is a wider challenge, well-illustrated by the contrast between oats and millet but also applicable to other grasses. Besides the grain morphometric traits, the use of μCT may also provide a handle on more difficult to measure traits like crease volume and thickness of the bran layer. Both of these traits are commercially important and determine grading of grains for the milling industry, but are extremely difficult to measure. Embryo size in the seed is also thought to be important in determining seedling growth and final biomass of the plant, but again this is difficult to measure in a non-destructive way [50].

The challenge now is to develop more advanced computing methods that are able to detect and measure these highly complex and variable traits. Recent developments in computer vision methods and machine learning modelling should prove to be very useful for this purpose.

The authors acknowledge Catherine Howarth (IBERS, Aberystwyth University), John Harper and Danny Thorogood (IBERS, Aberystwyth University) and Yuanhuai Han (Shanxi Agricultural University) for providing respectively oat, ryegrass and foxtail millet samples for scanning, Sam Nicholls for help with data analysis and, Roger Boyle for critically reading the manuscript. We thank the BBSRC for funding (BB/M009459/1 and BB/M018407/1 - FACCE ERA-NET/MODCARBOSTRESS)

The authors declare that they have no competing interests.

The datasets used during the current study are available from https://www.aber.ac.uk/en/research/data-catalogue/a11df174-d73d-4443-a7fd-ab5b7039df79/.

Not applicable.

Not applicable.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

The datasets used during the current study are available from https://www.aber.ac.uk/en/research/data-catalogue/a11df174-d73d-4443-a7fd-ab5b7039df79/.