Conditions for X-ray CT scanning
A scanning time of 10 min for each plant sample was determined as the X-ray CT condition for application to high-throughput crop RSA phenotyping. Because additional time was required for machine operation, X-ray generator start-up, and saving CT images, 15 min was the actual elapsed time needed per single sample. Thus, 32 samples could be processed in an eight-hour day, and 160 individuals, which is sufficient to perform genetic analysis, could be scanned in a week. To observe RSA development in rice continuously, until the roots reached the pot wall, pot diameter and depth were set as 20 cm and 25 cm, respectively, based on the maximum size that could be scanned by the detector of the CT scanner used in this study.
To obtain CT volumes showing clear root shapes, we determined the best soil substrate for CT scanning. We used an upland rice cultivar Kinandang Patong (KP) as the test sample, as upland rice usually has thicker roots than lowland rice [34], making it relatively easy to isolate the root segments from the CT volumes. To select a soil substrate suitable for rice root CT scanning, we examined the CT images of KP grown in five different soil substrates: calcined clay, volcanic ash soil, andosol, alluvial soil, and sand (Additional file 1). Among these five types, calcined clay gave the clearest root imagery in the CT process (Additional file 2). Based on this result, we decided to use calcined clay in our rice root CT scanning experiments.
As the tube voltage and current used in X-ray CT scanning affect image quality, we worked to identify the combination that exhibited the highest root–to–soil contrast. KP was grown in a growth chamber in pots filled with calcined clay, for five weeks, and was subjected to CT scanning—with representative CT images shown in Fig. 1. The inside of the pot was invisible in the 3-D reconstructed volumes (Fig. 1a). In the horizontal and the vertical slices (Figs. 1b and 1c, respectively), the roots are visible as dark pixels. The pixels of lower values were colored in black, which indicated that there was lower X-ray absorbance in the rice roots compared to that of calcined clay. To evaluate the influence of tube voltage and current on CT image quality, we scanned pots using tube voltages of 125, 150, 175, 200, and 225 kV, and tube currents of 100, 200, 300, 400, and 500 μA. The scaled-up CT slices created using these combinations are shown in Fig. 2, where it can be seen that higher voltage and current produced the highest contrast images. This was supported by the fact that the contrast-to-noise ratio (CNR) increased with higher tube voltage and current, attaining its highest value at 225 kV and 500 μA, respectively. These results conclusively indicated that a 225 kV voltage and 500 μA current were the best combination for rice root scanning, under our conditions.
Image processing
To visualize rice root segments automatically, we developed an image processing algorithm. The process flow involved the following three steps: (1) a 3-D median filter process to increase the root-to-soil contrast; (2) an edge detection process to isolate root segments; and (3) slice stacking to construct 3-D volumes.
The first step was to increase the root-to-soil contrast by reducing noise in the CT image. Noise is caused by mineral particles, voids in the soil, and short scanning time, so to reduce the noise level, we applied a 3-D median filter to the CT volume. Fig. 3a shows horizontal slices with five different kernel sizes—one, three, five, seven, and nine—where an image processed with the kernel size of one is equivalent to a non-filtered image. We calculated the CNR for each condition, and found it to be the highest for the 3-D median filter of kernel size seven—while the image created using the kernel size of nine was the most blurred. According to this result, we decided to use a 3-D median filter with a kernel size of seven.
The second step was to isolate root segments. CT slices that were processed using the 3-D median filter were inverted, as the CT image had soil voxels with higher value intensity and root voxels with lower value intensity. Root segments were isolated using an edge detection process that subtracted the blurred slices from their corresponding non-blurred counterparts, down to zero soil value intensity. Edge detection results achieved using various kernel sizes, from 1–57, are shown in Fig. 3b. Because an image blurred using kernel size one is the same as a non-filtered image, edge detection with kernel size one resulted in all-zero images. In the images processed using a kernel size > five, (Fig. 3b), signals were observed at the positions where the root could be seen in Fig. 3a. The root segments indicated by white arrowheads in Fig. 3b were not visible when kernel size five was used but visible when kernel size over 9 was used, and their size increased as the kernel size increased. Some roots, such as the root segments indicated by yellow arrowheads in Fig. 3b, were showed combined, when large kernel sizes, such as 57, were used, and the boundaries became ambiguous. Based on our observations, we decided to use kernel size 21 for edge detection in this study, as we found that when it was used, all the root signals indicated in Fig. 3a were both visible and unambiguous.
The third step was to construct a 3-D volume. All slices processed using a 3-D median filter and edge detection were stacked, while, to eliminate the influence of pot walls on RSA development, the inside regions of the CT slices were cropped. A horizontal projection and 3-D animation of the 3-D rendered volume are shown in Fig. 3c and Additional file 3, respectively. The RSA in the soil was successfully visualized, although non-root segments could also be seen in the image and the movie. Because root segmentation depends on contrast difference, all the voids in the soil were visualized. Non-root segments at the bottom were caused by the soil collapsing, while segments at the top were cracks, caused by plant growth and incompletely packed soil close to the ground surface. Small particles appearing everywhere were voids or water gradients in the soil. These small, non-root segments could be removed using thresholding and size opening processes (Fig. 3d and Additional file 4), but we did not remove them in this study as there was a risk of inadvertently erasing the small root segments of young seedlings at the same time.
We implemented the algorithm using python script (Additional file 5), and measured processing time using different hardware (Table 1). All the hardware we tested took less than 8 min for image processing, with the elapsed time dependent on the speed of the central processing unit. The fastest processing time, 2 min, was achieved using Intel® Xeon® E5-2650 v4. Because python is an interpreter language, batch operations were easily achieved.
Influence of X-ray dose on rice growth
X-rays affect plant growth [26], which is a problem for 4-D RSA phenotyping using X-ray CT systems, as repeated CT scanning increases the cumulative X-ray dose. Given that tube voltage and current were constant, we identified two ways to reduce X-ray doses.
The first way was to shorten the scanning time, applying the assumption that X-ray dose is proportional to scanning time. Fig. 4a shows horizontal projections of the image-processed CT volume of a KP at eight weeks after sowing, scanned under eight conditions and with its scanning time varied from 33 s to 10 min. We observed similar RSA in all conditions, despite the degraded image quality at faster scanning conditions. These results indicated that faster scanning was acceptable for the thick roots encountered at later growth stages, but would not be appropriate to use for the thinner roots associated with early growth stages.
The second way was by using metal filters to reduce the proportion of low-energy X-rays. Scaled-up horizontal slices of unprocessed and processed CT volumes, with and without 0.5, 1.0, and 2.0 mm copper (Cu) filters, under 10 min scanning conditions, are shown in Fig. 4b. The noise level of the unprocessed CT slices increased with increased filter thickness, while the quality of the processed CT slices obtained when using Cu filters were very similar to those achieved when no filters were used. Generally, thinner metal filters have been widely used in evaluating X-ray dose effects on plant growth—for example, 0.2 mm Cu filters have been used in rice [26], and 0.5 mm Cu filters have been used in wheat (Triticum aestivum), faba bean (Vicia faba), and barley (Hordeum vulgare) experiments [35, 36]. Although X-ray dose is determined by several factors, including tube voltage, tube current, and source–rotation axis distance, we considered that a 1.0 mm Cu filter would be sufficient to reduce the X-ray dose under the conditions applied in our testing.
We estimated X-ray doses using Rad Pro Dose Calculator (http://www.radprocalculator.com/). When a 225 kV tube voltage and a 500 μA current were used, the X-ray dose of the material placed 900 mm from the X-ray source, with a 0.5 mm Cu filter in place, the X-ray dose on the material was estimated as 0.55 Gy / hr. With a scanning time of 10 min, and using a 1.0 mm Cu filter, the dose to rice plants was estimated to be < 0.09 Gy per scan. Because 0.09 Gy was sufficiently < 33 Gy, which has been reported as the threshold affecting plant growth [26, 37], we considered that sequential X-ray scanning did not pose a problem for rice growth. It has been reported that scanning rice daily, for nine days with a dose of 1.4 Gy—that is, administering a total dose of 12.6 Gy—did not negatively impact rice growth [26]. If 12.6 Gy was taken as the upper limit for X-ray CT exposure, a simple arithmetic calculation indicated that 140 scanning procedures were permissible under our scanning conditions, before the rice plants would be potentially harmed.
To evaluate the influence of sequential X-ray doses on plant growth, KP was cultivated for two weeks, and subjected to daily CT scanning for seven days. The results indicated no apparent differences between shoot and root shape of mock- and X-ray-treated plants, at 21 days after sowing (DAS, Fig. 5a–b). We collected the shoot and root samples, and measured plant height, total root length of crown and lateral roots, shoot dry weight, or root dry weight (Figs. 5c–f). As a result, there were no significant differences between mock and X-ray treated plants. These results indicated that, under our scanning conditions, X-ray doses did not pose any potential impact on rice plant growth, and so, based on these results, we determined that, in our 4-D RSA phenotyping process flow, a 10 min scanning condition was appropriate for observing RSA from its early to late development stages.
4-D visualization of RSA development
To evaluate the developed process flow, we cultivated KP for one week, and then subjected it to daily CT scanning for three weeks. Horizontal projections of the image-processed CT volumes, and a 3-D movie, are provided in Additional files 6 and 7, respectively. From seven to 13 DAS, the root shape in the image was hazy, although daily root growth could be observed. From 14 to 20 DAS, the root shape became bolder, root length increased rapidly, and, by DAS 20, many root tips had developed beyond the imaging scope. From 21 to 27 DAS, the root shapes became increasingly bolder, while the general RSA shape remained unchanged. Overall, this test indicated that our process flow could be used for monitoring RSA development.
Verification of root fragments in the processed CT volumes
We applied a wired basket method to verify the lengths of various diameter roots detected using our process flow. Wire baskets keep the RSA in situ when the basket is unearthed and the soil is removed, ensuring that we could collect root samples that corresponded to the root fragments observed in the X-ray CT volumes. We cultivated rice plants for 21 d, and then unearthed the baskets. Because we had observed that 21-DAS KP had many roots of different thicknesses (Additional files 6 and 7), 21-DAS rice was suitable for investigating the detection limit for root diameters. To exclude the influence of root distribution in the soil, we used three genotypes that had different RSAs [38]—KP (thick and deep-root type), IR64 (thin and shallow-root type), and Dro1-NIL (thin and intermediate-root type). Image-processed CT volume vertical projections, images shot from directly above the basket, and the vertical projection in which roots were classified by their depth and observability as different color and line types, can be seen in Additional file 8.
We found that, when compared with the camera images, there were 68 detected and 12 non-detected roots in the image-processed CT volume. We then collected root segments and compared the root diameters of the detected and non-detected roots and found that many roots with a diameter of < 0.3 mm were not visualized, irrespective of the RSA type (Fig. 6a). As the root diameter threshold used when distinguishing the radicle and the crown roots from lateral roots was known to be approximately 0.2 mm [39], our results showed that we could not detect 15% of the roots of the 21-DAS rice plants. To visualize all radicle and crown roots, we used a smaller diameter pot (16 cm) and adjusted the source–detector and source–rotation axis distances to 800 mm and 407 mm, respectively. We repeated the wired basket assay test (Additional file 9) and found 82 detectable roots and zero non-detectable roots (Fig. 6b). The diameter of the detected roots was > 0.2 mm. In this case, we could detect almost all radicle and crown roots, with the X-ray dose per scan estimated to be 0.44 Gy. This meant that, as 12.6 Gy had been established as the safe upper limit for X-ray CT exposure, a simple arithmetic calculation indicated that plants would be unaffected by up to 28 scanning procedures. These results showed that the root detection limit could be influenced by adjusting pot diameter, source–detector distance, and source–rotation axis distance.