Plant materials and sample preparation
A panel of 161 unique J. regia accessions from worldwide was analysed. All the accessions are maintained at the Prunus and Juglans Genetic Resources Center and located in the Fruit Experimental Unit of INRAE in Toulenne (latitude 44°34’37.442’’N – longitude 0°16’51.48’’W), near Bordeaux, France (Additional file 1). The panel choice was made thanks to a previous work based on genetic diversity and phenotypic variation results [51].
In-shell walnut sampling was performed during harvest season in September 2018 and walnuts were dried following classical French industrial recommendations, for 2 days at 25 °C using a food dryer, and then stored until analyses in a cold room set to 2 °C. For each accession, a selection of 50 walnuts was performed based on their sanitary state and sent to the GEVES laboratory (Beaucouzé, France). All the samples were stored in an environmentally controlled room at 10 °C and 47.75% (± 3.4) relative humidity until use. During the preparation, the walnuts were embedded in a floral foam sample holder (9 cm length × 8 cm width × 21 cm height) to keep the samples from any abrupt or slight movement during the scanning process in order to avoid producing distorted images. The floral foam was chosen based on preliminary trials on different low-density materials in order to observe the level of attenuation of the X-rays passing through these materials.
The walnuts were scanned in batches, knowing that, the sample size was not fixed due to the huge variation between the walnuts in size and the limited scanning scope of the detector. The sample size ranged from 5 up to 16 walnuts per scan.
Image acquisition and reconstruction
Scans were obtained at constant electron acceleration energy of 120 kV, an electric current of 300 µA and a rotation speed of 4.99 degrees/s resulting in a scan duration of 14 m34s. A total of 2,164 images (or radiographs) in a .tif format were used for reconstructing each 3D image using North Star Imaging© reconstruction software EFX-CT (version 1.9.5.12) where the resulting 3D images were exported in a .nsihdr format with a resolution of 992 × 992 × 2991 voxels (voxel size of 0.1 × 0.1 × 0.1 mm).
After 3D reconstruction, a multi-stage workflow was applied to all CT images in order to eventually achieve a quantitative study. This workflow consists of three key steps as illustrated in Fig. 1d-f: preprocessing steps, walnuts individualization, and morphological traits extraction and quantification.
(i) The preprocessing: it begins with automatically loading each image in our image collection I (1,x,y,z),...,I (m,x,y,z)sequentially (where is the number of images) and then denoising them in order to eliminate the noise and artifacts introduced by the X-ray system during image acquisition. Subsequently, all the voxels which represent the sample holder are eliminated and only the voxels which represent the walnuts are preserved by discarding all the voxels below a certain threshold τ resulting in a binary mask M(x,y,z) according to the Eq. (1).
The individualization: the task of extracting the features of the walnuts and quantification is challenging especially if the walnuts are touching. To overcome this difficulty, all the walnuts in the images were separated and individualized (Fig.
2). The individualization step consists of multiple sub-steps such labelling, masking, convex hull estimation and exporting. Labelling is based on voxel connectivity in the whole 3D volume in order to determine the regions of interest which represent the walnuts by assigning identical values to all the voxels that belong to an individual walnut. Each walnut in the image was assigned a unique value starting from 1 to
consecutively where n represents the total number of walnuts in the image. Then, labelling was followed by generating a set of masks using the determined regions of interest then finally, given the original input image and the generated set of masks
K (x,y,z) as shown in the Eq. (
2), each walnut was exported in a separate sub-image
i(x,y,z) in a .nsihdr format after estimating the convex hull of each walnut.
(iii) The morphological traits extraction and quantification: at this step, the principal components of the walnuts whose morphological features were segmented using multi-level thresholding and watershed algorithm which is a transformation that treats the image like a topographic map [52]. As a consequence, in our case, each main part of each walnut was segmented and was given a unique label as shown in Fig. 1f, considering the kernel, the shell, and the empty space between the kernel and the shell. The optimum threshold τ and greyscale ranges of the principal parts of the walnuts were estimated experimentally based on the analysis of the histogram that corresponds to the distribution of the intensities of the images. An additional movie file shows this in more detail (Additional file 2).
A fully automated in-house image processing pipeline was developed using the Thermo Scientific Avizo© software V9.0.0 built-in functions, the MATLAB© version 7.7.0 R2008b image processing toolbox [53] from The MathWorks©, Inc. (Massachusetts, USA), the TCL scripting language and Spyder Python IDE. To use this pipeline, walnuts that have no damage on the shell are required.
Measurements of a total of 14 morphological and shape descriptors were obtained: the nut length, the nut face diameter, the nut profile diameter, the nut volume, the nut shape VA3D, the nut Feret shape 3D (defined by D/d where d is the minimum Feret diameter and D is the maximum Feret diameter in the orthogonal direction, so 90° from the minimum Feret diameter; the maximum Feret diameter is the maximum diameter of an object as if it were freely rotating in three dimensions using a caliper [54]), the nut surface area, the shell volume, shell thickness, the kernel volume, the kernel filling ratio, and the empty space volume (Table 1).
Table 1
Walnut morphological traits measured by the workflow.
Morphological trait | Symbol | Description | Unit |
Nut | | | |
Nut Length | L | The largest length of the nut from the base to the end | mm |
Nut Face Diameter | F | The largest longitudinal section of the nut through suture | mm |
Nut Profile Diameter | P | The largest longitudinal section of the nut perpendicular to suture | mm |
Nut Volume | Vn | Total volume of the nut, Vn = Vs + Vk + Ve | mm3 |
Nut Shape VA3D | S1 | Shape factor of the nut | - |
Nut Feret Shape 3D | S2 | Feret shape factor of the nut | - |
Nut Surface Area | A | Surface area of the nut | mm2 |
Nut Sphericity | Ψ | Index of nut roundness | - |
Shell | | | |
Shell Volume | Vs | Volume of the shell | mm3 |
Shell Thickness | T | Thickness of the shell | mm |
Shell Rugosity | Ω | Index of shell surface roughness | - |
Kernel | | | |
Kernel Volume | Vk | Volume of the kernel | mm3 |
Kernel Filling Ratio | R | Ratio of the kernel volume Vk to the total volume of the nut Vn | % |
Empty Space | | | |
Empty Space Volume | Ve | Volume of the empty space | mm3 |
In addition, two supplementary traits, the nut sphericity (close to roundness) (3) and the shell rugosity (or surface roughness) (4) indexes, were measured using a non-standard calculation since they are not supported by the Avizo© software:
where Vn is the nut volume and A is its surface area. The sphericity of a sphere is 1 and any object which is not a sphere will have sphericity less than 1. The nut shape VA3D is defined by .
Experiments were run on a workstation equipped with an Intel® Xeon® dual-core processor running at 3 GHz using 64 MB of RAM and running Windows® version 10. The dataset was assessed using R software [55] with the package “tidyverse” [56]. Pearson correlation matrices were performed using the package “corrplot” [57] and Principal Component Analysis (PCA) using the package “FactoMineR” [58].