Skip to main content

Convolutional neural networks combined with conventional filtering to semantically segment plant roots in rapidly scanned X-ray computed tomography volumes with high noise levels

Abstract

Background

X-ray computed tomography (CT) is a powerful tool for measuring plant root growth in soil. However, a rapid scan with larger pots, which is required for throughput-prioritized crop breeding, results in high noise levels, low resolution, and blurred root segments in the CT volumes. Moreover, while plant root segmentation is essential for root quantification, detailed conditional studies on segmenting noisy root segments are scarce. The present study aimed to investigate the effects of scanning time and deep learning-based restoration of image quality on semantic segmentation of blurry rice (Oryza sativa) root segments in CT volumes.

Results

VoxResNet, a convolutional neural network-based voxel-wise residual network, was used as the segmentation model. The training efficiency of the model was compared using CT volumes obtained at scan times of 33, 66, 150, 300, and 600 s. The learning efficiencies of the samples were similar, except for scan times of 33 and 66 s. In addition, The noise levels of predicted volumes differd among scanning conditions, indicating that the noise level of a scan time ≥ 150 s does not affect the model training efficiency. Conventional filtering methods, such as median filtering and edge detection, increased the training efficiency by approximately 10% under any conditions. However, the training efficiency of 33 and 66 s-scanned samples remained relatively low. We concluded that scan time must be at least 150 s to not affect segmentation. Finally, we constructed a semantic segmentation model for 150 s-scanned CT volumes, for which the Dice loss reached 0.093. This model could not predict the lateral roots, which were not included in the training data. This limitation will be addressed by preparing appropriate training data.

Conclusions

A semantic segmentation model can be constructed even with rapidly scanned CT volumes with high noise levels. Given that scanning times ≥ 150 s did not affect the segmentation results, this technique holds promise for rapid and low-dose scanning. This study offers insights into images other than CT volumes with high noise levels that are challenging to determine when annotating.

Background

Plant root distribution in the soil is a crucial phenotype that determines nutrient and water availability. Generally, nutrient distribution in the soil is greatly influenced by fertilizer application, whereas water distribution is greatly influenced by the weather. Altering root distribution to gain access to soil resources is a practical strategy to increase crop yield. Changes in root distribution in the soil increase yield by enhancing nutrient uptake under irrigated paddy [1] and water uptake under upland conditions with drought [2] in rice (Oryza sativa). Therefore, root distribution that maximizes benefits strongly depends on the plant growth environment [3,4,5]. As the environment changes significantly during the growing period, the optimal root system also changes. For example, under top-dressing fertilization conditions, shallow, wider, and denser root systems are advantageous [6]. In contrast, deeper, narrower, and denser root systems are advantageous under drought conditions as deeper soils contain more water [7,8,9,10]. An ideal root system responds to the environment by increasing root density in soil areas with fertilizer and sufficient moisture. Therefore, these effects of changes in the root system must be evaluated in an integrated, time-series manner to improve crop root morphology.

Non-destructive measurements can help evaluate root growth dynamics over time. In particular, X-ray computed tomography (CT) and magnetic resonance imaging, which can measure changes in root distribution in three dimensions (3D), are powerful tools for assessing changes in root distribution in heterogeneous soil environments [11, 12], such as those used for local fertilization [13, 14], local soil compaction [15, 16], and in the field [17,18,19]. However, 3D volume data are large, and image analyses, such as extracting root segments from the 3D volume, calculating root traits, and analyzing time-series 3D data, are labor-intensive. Therefore, many methods have been developed for efficiently analyzing these 3D volumes. These include methods to isolate root segments by conventional filtering, such as thresholding and edge extraction using uniform soil [20, 21], region growing [22, 23], root tracking [24, 25], root filtering with shape-based filters [26, 27], and efficient analysis of time-series CT volumes [28,29,30]. However, the conditions under which these methods can be used limit their applicability. Therefore, a versatile analysis method independent of these conditions is required.

Semantic segmentation is a powerful method for extracting regions of interest (ROI) from the image data [31]. Generally, semantic segmentation produces image data in which ROIs are labeled as the foreground and the other regions as the background using convolutional neural networks (CNNs). Therefore, a model can be trained to extract only the ROI from an image with a complex background using appropriate training data. This advantage is also effective in extracting roots from heterozygous soil images. For example, conventional image processing algorithms, such as thresholding and edge detection, detect some backgrounds as false positives unless homozygous soil grains are used [20]. In this case, only complete root information must be extracted using root tracking or other methods [24, 25, 32]. Alternatively, semantic segmentation can be used independently of the background. Recently, semantic segmentation has been frequently used in plant root phenotyping. In 2D, semantic segmentation can extract roots from images captured from the side of a rhizobox, a container with transparent sides to observe roots, and with a mini-rhizotron, an instrument for acquiring an image of a surface by burying it in soil [33,34,35,36,37]. These images often include voids, bubbles, and light reflections on the transparent sides as the background. Moreover, digital images captured outdoors may exhibit changes in the color, brightness, and shadows of the soil. Therefore, incorporating these changes in the training data enables root extraction using semantic segmentation [38].

Semantic segmentation has been used for 3D volume and 2D images. However, 3D segmentation reconstructed from 2D segmentation processed slice-by-slice tends to overlook 3D connections; therefore, a model utilizing 3D images as input is superior [39]. Semantic segmentation in 3D is used to extract roots of various thicknesses [40]. A conventional filtering method, such as shape-based filtering [26, 27, 41], can be used if the root thickness variation is minimal, whereas alternative approaches are required for large variations [40]. This variation can be addressed by adding a spatial pyramid pooling layer [42] or multi-resolution encoder–decoder networks [43] to a CNN extracting multi-scale features. However, few studies have used 3D semantic segmentation for plant roots, likely owing to the labor-intensive creation of 3D label data and lengthy training of 3D segmentation models. Moreover, methods using artificially synthesized training data have been proposed [44]; however, issues, such as creating synthetic data close to actual root segmentation, hinder their application. Another issue with 3D segmentation is that few applications use rapidly scanned CT volumes with high noise levels. Specifically, in the context of throughput-prioritized crop breeding, the CT volume often exhibits blurring, and the noise level increases owing to the expansion of the scan range and shortening of the scan time [20].

We constructed a non-destructive measurement platform for crop root systems with X-ray CT, mainly for lowland and upland rice cultivars [20, 25, 28]. This platform uses two software, the root system architecture (RSA) visualization software RSAvis3D [20] and RSA vectorizing software RSAtrace3D [25], to measure root traits. RSAvis3D reduces noise levels through 3D median filtering and extracts root features through edge detection, enabling the rapid visualization of crop roots in the soil [20]. RSAtrace3D converts edge-detected images into vector data to extract information, such as root shape, connection, and distribution [25]. In addition, to conduct phenotypic analysis of 3D root distribution for variety development, large pots and rapid CT scanning are required [20]. However, as the pot size increases, the resolution of the CT volume decreases, and the CT volume becomes noisier as the scan time decreases. Therefore, the present study aimed to determine the feasibility of semantic segmentation with rapidly scanned CT volumes of large pots and ascertain the extent to which scan time can be reduced without affecting the results. To this end, we trained the segmentation model with CT volumes obtained at various scan times and compared the results.

Methods

Plant materials and cultivation conditions

We used the upland rice cultivar Kinandang Patong. After sterilizing the seed surface using a fungicide, the seeds were soaked in tap water. After 3 days, three seeds were sown in the center of the pots with 20 cm diameter and 25 cm depth, filled with a calcined clay (Profle® Greens Grade, PROFILE Products, Buffalo Grove, IL, USA) as a soil-like root growth substrate. These soils were saturated with 3 × Kimura B hydroponic solution containing 1094.4 µM (NH4)2SO4, 273 µM K2SO4, 1641 µM MgSO4, 549 µM KNO3, 1095 µM Ca(NO3)2, 546 µM KH2PO4, and 53.4 µM FeC6H5O7·H2O. The pH was adjusted to 5.5 using HCl and KOH.

The rice plants were cultivated in a greenhouse under natural light and with 400 W metal–halide lamp irradiation. The duration of the day light was 12 h. The temperature was maintained at 30 °C during the day and 26 °C at night, and humidity at 50%. Tap water was supplied to the bottom of each pot during cultivation.

We prepared two pots: one for model training and the other to evaluate the trained model. The rice plants were grown in different seasons and, therefore, exhibited slight differences in growth. One pot with shallower root systems was used for model training, and lateral roots could not be observed in the CT volume. The other pot with deeper root systems was used for model evaluation, and some lateral roots were observed in the CT volume.

X-ray CT scan and reconstruction

Four-week-old plants were scanned using an X-ray CT system (InspeXio SMX-225CT FPD HR; Shimadzu Corporation, Kyoto, Japan) with varying scanning times. Details of the scanning conditions are shown in Table 1. Conditions of scan times of 33, 66, 150, 300, 600, and 3600 s are referred to as S0033, S0066, S0150, S0300, S0600, and S3600, respectively. The pots were placed at the center of the CT scanner table. The distance between the CT scanner table center and the X-ray generator was 900 mm, and that between the detector and the generator was 1200 mm. The tube voltage and current were 225 kV and 500 μA, respectively. Cu filters (0.5 mm thickness) were used to harden the X-ray beam. The projection image (1024 × 1024 pixels) was captured using a detector (without binning: 3000 × 3000 pixels at 4.0 frames per second [fps]; with binning: and 1000 × 1000 pixels at 18.0 fps) and computed. The spatial resolution was 0.3 mm. Finally, 860 slices were reconstructed.

Table 1 CT scanning times and conditions

Rescaling intensity and cropping of CT volume

The intensity of the CT volume was rescaled from a 16-bit to an 8-bit grayscale. The histogram of pixel intensity exhibited a bimodal distribution under our experimental conditions, with peaks corresponding to air and soil voxels. As rescaling using minimum and maximum values eliminates differences in image intensity when converted to 8-bit, the intensities of the CT volumes were rescaled by setting the peak of the air at 0 and that of the soil at 128. Any values below 0 were replaced with 0, and values exceeding 255 were set to 255. Because values below 0 and over 255 were derived from the air, sand, or mineral particles contaminated in the soil substrates [28], no information was lost in root segmentation. To eliminate regions close to the pot wall, an area with an 9 cm radius from the center of the pot was cropped. This is because roots touching the pot wall changed their direction of elongation, making them less informative for evaluating a variety characteristics [20].

Image processing with RSAvis3D

An RSA visualization software RSAvis3D was used for extracting root features [20]. RSAvis3D reduced the noise level using 3D median filtering and extracted root features using edge detection. We used RSAvis3D version 1.5 (https://github.com/st707311g/RSAvis3D/, branch 1.5), in which RSAvis3D uses multiple graphics processing units (GPUs) to accelerate extracting roots.

Generating ground truth data

Ground truth (GT) data were generated using the RSA vectorization software RSAtrace3D [25] version 1.12 (https://github.com/st707311g/RSAtrace3D/, branch 1.12). RSAtrace3D is a GUI-based software that helps users vectorize monocotyledon RSA. The vector data can be converted into the GT volume.

Volume segmentation and assembling

The CT volume dimensions were 860 (H) × 599 (W) × 599 (D) voxels (Fig. S1a), which were too large for deep learning. Therefore, the CT volume was divided into sub-volumes of 64 (H) × 64 (W) × 64 (D) voxels dimensions (Fig. S1b). Regions at the edges of sub-volumes are less informative; therefore, the results for those regions are generally excluded [45]. Herein, we designated unique and overlapping regions in the sub-volumes (Fig. S1c). Each of the eight voxels at the edges of the sub-volumes was an overlapping region, and the remaining voxels were unique regions. After segmentation prediction, the overlapping regions were deleted, and the unique regions were assembled into the entire RSA segmentation.

Semantic segmentation model

VoxResNet [46], a deep voxel-wise residual network model typically used for semantic segmentation tasks, was utilized in this study. The model structure was similar to that described previously [46], except that the activation function at the end was changed from softmax to sigmoid, as the number of segmentation classes was only 1: roots. Model training and prediction were performed using PyTorch [47] version 2.0.1 in Python [48] version 3.10.4.

For model training, we used CT volumes S0033, S0066, S0150, S0300, and S0600. All volumes were completely aligned in 3D, as the sample remained fixed on the scan table during sequential scanning and only the imaging conditions were changed. GT was created using the S3600 sample; however, these datasets were not used for model testing. Before model training, the CT volumes (value range: 0–255) were rescaled to 0.0–1.0 by dividing by 255.0, and the GT volumes were converted to Boolean. Each volume was divided into 3042 sub-images. During training, random flips were performed along each axis. Dice loss [49] was used as the loss function to evaluate the spatial overlap between the segmentation results and GT [46]. Adam [50] was used as the optimizer. Given that the learning rate of 0.001 was unstable, a learning rate of 0.0001 was used. The batch size was set to 64, resulting in 48 iterations per epoch. The models were trained for 1500 or 3000 epochs. Training efficiency was determined based on the learning curve of Dice loss and the model weight after 1500 or 3000 epochs. Given that Dice loss exhibited a stable decrease, we judged that the lower the loss after 1500 or 3000 epochs of learning, the better the model performance.

For model testing, using the reconstructed predicted volume and corresponding GT volume, three metrics, namely precision, recall, and Dice were calculated. These metrics were calculated using the following formulas:

$$Precision=TruePositive/\left(TruePositive+FalsePositive\right)$$
(1)
$$Recall=TruePositive/\left(TruePositive+FalseNegative\right)$$
(2)
$$Dice=2\times Precision\times Recall/\left(Precision+Recall\right)$$
(3)

For model evaluation, additional CT volumes were used with the scanning conditions of S0033, S0066, S0150, S0300, and S0600, which were not used for model training. The predictions between scanning conditions were graphically compared.

Image quality restoration model

The VoxResNet model was trained using the “Noise2Clean” strategy, which involves converting an image with a short scan time to an image with a long scan time. The volumes acquired with short and long scan times were completely aligned in 3D. The last activation function, softmax, was removed because the purpose of this model was image quality restoration, not classification. The mean squared error (MSE) loss [51] was used as the loss function, and Adam [50] was used as the optimizer with a learning rate of 0.00001. The batch size was 64, and the number of iterations per epoch was 48. The model was trained for 500 epochs.

Volume rendering

Volume data were visualized using the VG Studio MAX 3.2 software (Volume Graphics, Heidelberg, Germany).

Machine specifications

For image processing and model training, a desktop personal computer (Xeon W-2295 CPU, 96 GB memory, and 3 × NVIDIA GeForce RTX 3090) was used. The operating system was the Ubuntu 20.04.6 LTS.

Results

Outline of root segmentation in rapidly scanned CT volumes with high noise levels

Large pot sizes and short scanning times resulted in high noise and low-resolution CT volumes. The root segments scanned from a pot with a 9 cm diameter were easily recognized (Fig. 1a). In contrast, those scanned from a pot with a 20 cm diameter, the maximum size determined by the detector used in this study, were hardly recognizable, with no evident contours (Fig. 1b). Root segmentation was hindered by low obscurity, necessitating developing GT volumes corresponding to these high-noise and low-resolution CT volumes to train the segmentation model. One approach involves scanning the same pot twice, with short and long scan times, and then creating a GT volume from the CT volume with long time scan. Another approach is to use conventional image processing filters to extract root features. Based on these filters, GT images can be created using tracing software [20, 25]. In this study, we implemented both approaches. We scanned the same pot several times with varying scanning times. Subsequently, we generated GT images from the CT volume scanned for a longer duration by tracing the root segments using RSAvis3D and RSAtrace3D software (Fig. 1c). RSAvis3D enriched root features using an edge detection algorithm [20], thereby enabling RSAtrace3D to produce an RSA vector to generate RSA segmentation.

Fig. 1
figure 1

Segmentation model for high-noise and low-resolution root segments in computed tomography (CT) volume. a Configuration of root segments in a CT slice. The pot diameter is 9 cm, and the scan time is 10 min. b High-noise and low-resolution root segments in a CT slice. The pot diameter is 20 cm, and the scan time is 10 min. The regions with root segments are circled. c Outline of obscured root segmentation

Scanning time and training efficiency

Using the paired dataset of CT and GT volumes, we trained the semantic segmentation models and measured the training efficiency of the segmentation models of S0033, S0066, S0150, S0300, S0600, and S3600. As CT image quality decreases in proportion to scan time when scan time is ≤ 600 s [20], we set six different scan times, adding a scan condition of 3600 s, which exhibits a lower noise level. The edge-detected volumes obtained using RSAvis3D are shown in Fig. 2a–f. Longer scan times tended to result in less noise. The GT volumes were created using the S3600 volume (Fig. 2g), which was paired with each edge-detected volume for model training (Fig. 2h).

Fig. 2
figure 2

Model training with samples with various scan times. Visualization via RSAvis3D of root distribution of a S0033, b S0066, c S0150, d S0300, e S0600, and f S3600. g Ground truth of root segmentation created from S3600. The shape of the bounding box is 25.8 (H) × 18.0 (W) × 18.0 (D) cm. h A diagram comparing training efficiency. (1) S3600 was used for segmentation to generate ground truth data. (2) With the ground truth, S0033, S0066, S0150, S0300, and S0600 were used to train the segmentation model. i Change in Dice loss for each sample during training

For model training, we used VoxResNet, a CNN model of a deep voxel-wise residual network which can improve model performance by increasing the CNN layers. VoxResNet consists of 25 layers and has been widely used for segmentation [46]. We compared the loss changes among datasets of S0033, S0066, S0150, S0300, and S0600 during VoxResNet training (Fig. 2i). Although Dice loss did not fully decrease after training for 1500 epochs, model performance was compared at 1500 epochs because the Dice loss decreased steadily under all conditions. The Dice losses of S0150, S0300, and S0600 were similar (Fig. 2i) at 1500 epochs, indicating that certain noise levels did not affect the training efficiency of the segmentation model. In contrast, the Dice losses of S0033 and S0066 were higher than those at other conditions (Fig. 2i). Notably, these two samples were scanned with binning (Table 1), which substantially affected the training efficiency of the segmentation model. The Dice loss of S0033 was slightly greater than that of S0066 (Fig. 2i). These results indicate that, to a certain degree, noise levels do not affect the training efficiency of VoxResNet.

Deep learning-based restoration of image qualities before model training

Assuming that the reduced segmentation efficiency in volumes with a shorter scan time was caused by image quality degradation, restoring image quality can help increase segmentation efficiency in volumes with a shorter scan time. To test this hypothesis, we trained VoxResNet as an image quality restoration model; the volumes of S0033 and S0066 were used as input data, and the volume of S3600 as the GT (Fig. 3a). After training 500 epochs, the MSE losses for S0033 and S0066 were approximately 0.006 (Fig. 3b). The noise levels of CT slices were apparently reduced (Fig. 3c–f), and the peak signal-to-noise ratio of S0033 and S0066 were 28.27 and 28.90 before restoration and 31.80 and 31.94 after restoration, respectively. These results indicate that the image qualities of S0033 and S0066 were restored. Using samples with and without image-quality restoration, we trained a segmentation model and compared the Dice loss changes of S0033 (Fig. 3g) and S0066 (Fig. 3h) over 1500 epochs. No differences were observed in the Dice losses between samples with and without image-quality restoration. These findings indicate that, at least in this case, deep learning-based image quality restoration is ineffective in increasing the training efficiency of the segmentation model.

Fig. 3
figure 3

Effects of image quality restoration on training efficiency of segmentation model. a A diagram of training of segmentation model combined with image quality restoration. (1) Restoration model was trained with S0033 and S0066 as input data and S3600 as the ground truth. (2) The ground truth of segmentation was generated from S3600. (3) Segmentation models of S0033 and S0066 were trained using the samples with image quality restoration. b Change in mean squared error (MSE) loss during training of image quality restoration models. Slices of S0033 (c) before and (d) after image quality restoration. Slices of S0066 (e) before and (f) after image quality restoration. The area enclosed by the red dashed line is enlarged. Changes in Dice losses of (g) S0033 and (h) S0066 during segmentation training with and without image quality restoration

Preprocessing before root segmentation

Assuming that prefiltering image processing, not based on deep learning, increases the training efficiency of the segmentation model, we trained two types of segmentation models. One used the CT volume, and the other used the preprocessed CT volume using RSAvis3D. Subsequently, we compared the Dice loss at 1500 epochs (Fig. 4a). The results revealed that preprocessing improved the Dice loss by approximately 10% (Table 2). No cases of reduced efficiency were observed, indicating that the feature extraction of RSAvis3D, which is not based on deep learning, effectively improved the training efficiency of the segmentation model.

Fig. 4
figure 4

A diagram of segmentation model training combined with RSAvis3D. (1) The ground truth of segmentation was created from S3600. (2) The CT volumes were filtered using RSAvis3D. (3) VoxResNet segmentation models were trained with S0033, S0066, S0150, S0300, S0600, and S3600

Table 2 Effect of preprocessing on Dice loss of the training dataset

To evaluate the effects of preprocessing, we calculated three model metrics, namely precision, recall, and the Dice values. Metrics were computed using the entire reconstructed 3D volumes, rather than segmented sub-blocks. Our findings revealed that preprocessing improved all metrics for almost all samples, with a particularly notable increase in precision (Table 3). This result suggested that preprocessing increased segmentation metrics by reducing false positives or increasing true positives. In addition, the precision of S0066 and S0600 without processing was lower than that of other samples, suggesting that preprocessing increased the stability of model training. Using the model trained with preprocessing and CT volumes used for model training, root segmentation was predicted (Fig. S2), with nearly all the root segments accurately identified.

Table 3 Effect of preprocessing on metrics of the training dataset

To validate the models, we additionally cultivated a rice plant and scanned its pots to prepare CT volumes that were not used for model training. Scanning conditions were S0033, S0066, S0150, S0300, and S0600. False positives (blue) were noted in all conditions at similar frequencies (Fig. 5), whereas false negatives (red) were more frequently observed in S0033. False positives were observed in the shoot area, near the ground surface, and on the surface of the true segments (white). This result is presumably attributed to limited shoot data in the training data, numerous voids due to incompletely packed soil particles near the ground surface, and incomplete GT data due to high-noise and low-resolution CT volumes. Almost all false negatives observed under all conditions appeared to be lateral roots, likely because the training data contained few lateral roots. In addition, we calculated the precision, recall, and Dice of the reconstructed prediction volumes and found that all metrics were lower than those of the training datasets, particularly in S0033 and S0066 (Table 4). These results indicate that more training datasets with scan times ≥ 150 s should be used to train the model.

Fig. 5
figure 5

Comparison of segmentation results with ground truth. Predicted segmentation of (a) S0033, (b) S0066, (c) S0150, (d) S0300, and (e) S0600. White: predicted segments overlapping the ground truth, blue: predicted segments not overlapping the ground truth, red: ground truth segments not predicted. The shape of the bounding box is 25.8 (H) × 18.0 (W) × 18.0 (D) cm

Table 4 The metrics of the validation dataset

High-precision models that can be developed based on our trials

To assess the accuracy of the root prediction models, we trained the models over a greater number of epochs, as the Dice loss did not decrease completely at 1500 epochs (Fig. 2i). In addition, we trained the model with CT and RSAvis3D-processed volumes, given that RSAvis3D may eliminate some features that appeared only in the original CT volumes. After 3000 epochs of training for S0150, the Dice loss was 0.093, a 17% improvement compared to the loss observed at 1500 epochs with only RSAvis3D-processed volumes. Using this model, we predicted root segments using S0150 for model evaluating (Fig. 6, Movies S1 and S2). The false-negative (red) segments were similar to those predicted by the model trained for 1500 epochs with only RSAvis3D-processed images; however, many false-positive segments disappeared (Fig. 6b). These results indicate that false positives can be reduced by increasing epochs or including original CT volumes or both and that false negatives cannot be reduced. In other words, roots with features absent in the training data were identified as false negatives in the prediction data.

Fig. 6
figure 6

High-precision segmentation model. a RSAvis3D-filtered volume of S0150. b Predicted segmentation. White: predicted segments overlapping the ground truth, blue: predicted segments not overlapping the ground truth, red: ground truth segments not predicted. The shape of the bounding box is 25.8 (H) × 18.0 (W) × 18.0 (D) cm

Discussion

Semantic segmentation is a powerful tool for extracting ROIs from images with complex backgrounds. In the present study, we demonstrated the feasibility of root segmentation in rapidly scanned CT volumes with high noise levels. The approach involved creating a pair of CT volumes with short and long scan times, generating GT volumes from the CT volumes with a long scan time, and training the model with the GT and CT volumes with a short scan time. In addition, using a pot with a 20 cm diameter, we demonstrated that reducing the scan time to 150 s had minimal effect on the segmentation results. However, some roots were not predicted by segmentation, mainly because these were lateral roots not included in the training data. This limitation can be addressed by expanding the training data. Currently, the demand for creating semantic segmentation models to isolate ROIs from high-noise and low-resolution images is high. In the medical field, semantic segmentation is used to isolate organs to accurately diagnose diseases [46, 52]; however, shortening the scanning time is necessary to minimize exposure [51, 53,54,55]. Furthermore, owing to limitations in equipment performance, only low-quality images are frequently available. We believe our findings will contribute to addressing these challenges.

In the present study, we observed no differences in the model training efficiency between S0150, S0300, and S0600, indicating that the semantic segmentation model can be used for high-throughput and low-dose root phenotyping. We required 1 min for RSAvis3D preprocessing and 5 days for model training. Thus, model training is the bottleneck of this method, and using fewer GPUs requires an extended training duration. Moreover, the model was trained at a constant learning rate, and introducing a learning rate scheduler could potentially shorten the training time. As the optimal learning rate depends on the soil type and plant species in the CT volumes, examining the learning rate when training the model under new conditions is necessary. Once the model training is complete, model prediction can be accomplished in only a few minutes, which is sufficiently rapid. This high-throughput technique can be used for genetic analysis requiring multiple samples. In addition, CT imaging is non-destructive and can measure changes over time; however, X-ray exposure must be considered when imaging the same individual repeatedly [20, 29]. Therefore, the segmentation model used in this study is advantageous, as shorter scan times allow for more frequent imaging of the same individual.

The minimum number of volumes required for the method presented in our study is two: one for model training and the other for model evaluating, indicating that only two GT data sets should be generated. Generally, segmentation model training requires less training data than other model training, such as classifiers. For example, in 2D, root prediction in the rhizobox uses 50 images [33], and in the trench profile, the surface with exposed roots in trenches dug near the plants utilizes 30 images [38]. In contrast, approximately 50,000 images are used to train a classification model for plant diseases [56]. This discrepancy arises because semantic segmentation can address a few classes, such as root segmentation, and it allows for the subdivision of one image into several images. In the trench profile, ten images are divided into 360 images for model training [38]. In the segmentation model in the present study, one CT volume was divided into 3042 sub-volumes. Therefore, if a new segmentation model of the X-ray CT volume is trained, at least one (or two for evaluation) dataset pairs of CT and GT volumes are required. Using more data sets for training will result in a more generalized model and mitigate potential biases in the data sets used for training, such as the lack of lateral root segmentation in this study.

In addition, image quality restoration before segmentation for increasing training efficiency did not increase training efficiency. This result may be because we used the same CNN structure in the restoration and segmentation models. Segmentation accuracy of degraded images, including reduced size, high noise levels, and blur, increases using image restoration based on a generative adversarial network [57]. Moreover, pretraining the decoder in the CNN structure is effective for few-shot semantic segmentation [58], indicating that pre-segmentation image quality restoration is effective. We postulated that a model with a different structure than the image quality restoration model used in this study may improve segmentation accuracy. Although image quality restoration did not affect the learning efficiency globally, the efficiency may have varied locally. However, as the objective of this study was to increase global learning efficiency reduced by shortened scan times, a more detailed analysis was not performed. Moreover, Dice losses during 1500 epochs of segmentation model training in S0150, S0300, and S0600 were similar despite different noise levels, indicating that this noise level does not affect the segmentation results using VoxResNet.

In the present study, the RSAvis3D preprocessing improved the model learning efficiency by approximately 10%. The main RSAvis3D algorithms are 3D median filtering and edge detection [20]. Some semantic segmentation models use median filtering [59, 60] and edge detection [61, 62] to increase segmentation accuracy, implying that 3D median filtering simplifies the complexity of the image texture, while edge detection focuses on the information of root features. In this study, the training efficiency of the segmentation model was robust at certain noise levels. These results suggest that edge detection plays a role in increasing training efficiency; however, we could not elucidate whether 3D median filtering or edge detection was more effective.

The model in this study was trained for rice roots. However, we need to consider a few conditions when applying it to other crops. First, the proportion of classes in the training data should be determined. Rice RSAs are considerably simple, mainly consisting of adventitious roots with lateral roots. In contrast, the RSAs of legume crops, such as soybean (Glycine max), are more complex, with seminal and lateral roots with nodules. Therefore, at least two classes, root and nodule, are required. Generally, the proportion of nodule to root voxels is considerably less than 1 (possibly zero for individuals without nodules), indicating imbalanced training data, which is a limitation for proper training [63]. GTs should be created from multiple CT volumes to ensure adequate training data for the nodules, and the extremely imbalanced training data set should be adjusted to address these issues. Furthermore, this study only created segmentations and did not evaluate traits. As the segmentation is just an image, computing the RSA traits by vectorization or other methods is necessary [64]. Moreover, given that segmentation is error-prone, segmentation-free image analysis techniques have been developed [65]. However, segmentation can also be used as a preprocessing step to extract ROIs from images with complex backgrounds. For example, segmentation can be performed with this method on a noisy image scanned at high speed and combined with another method that computes some features from the resulting root system image. This study contributes to advancing root research using image analysis and root phenomics.

Conclusions

In the present study, the semantic segmentation model was trained for root segmentation in high-noise and low-resolution CT volumes. The trained segmentation model can be used for root segmentation in 5 min, including scanning, image processing, and segmentation prediction. This study facilitates root system analysis using CT imaging and analysis of high-noise and low-resolution images, such as medical images created by low-dose scanning.

Availability of data and materials

The scripts for the training and prediction of the 3D semantic segmentation are available at the GitHub repository (https://github.com/st707311g/SStrainer3D/, branch 1.0). The datasets used during the current study are available from the corresponding author upon reasonable request.

Abbreviations

3D:

Three-dimensional

CT:

Computed tomography

fps:

Frames per second

GPU:

Graphics processing unit

GT:

Ground truth

MSE:

Mean squared error

ROI:

Regions of interest

RSA:

Root system architecture

References

  1. Arai-Sanoh Y, Takai T, Yoshinaga S, Nakano H, Kojima M, Sakakibara H, et al. Deep rooting conferred by DEEPER ROOTING 1 enhances rice yield in paddy fields. Sci Rep. 2014;4:5563. https://doi.org/10.1038/srep05563.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  2. Uga Y, Sugimoto K, Ogawa S, Rane J, Ishitani M, Hara N, et al. Control of root system architecture by DEEPER ROOTING 1 increases rice yield under drought conditions. Nat Genet. 2013;45:1097–102. https://doi.org/10.1038/ng.2725.

    Article  CAS  PubMed  Google Scholar 

  3. van der Bom FJT, Williams A, Raymond NS, Alahmad S, Hickey LT, Singh V, et al. Root angle, phosphorus, and water: interactions and effects on durum wheat genotype performance in drought-prone environments. Plant Soil. 2023. https://doi.org/10.1007/s11104-023-05966-z.

    Article  Google Scholar 

  4. Lynch J. Root architecture and plant productivity. Plant Physiol. 1995;109:7–13. https://doi.org/10.1104/pp.109.1.7.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  5. Uga Y. Challenges to design-oriented breeding of root system architecture adapted to climate change. Breed Sci. 2021;71:3–12. https://doi.org/10.1270/jsbbs.20118.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  6. Gowariker V, Krishnamurthy VN, Gowariker S, Dhanorkar M, Paranjape K. The fertilizer encyclopedia. Hoboken: John Wiley & Sons; 2009.

    Google Scholar 

  7. Dayoub E, Lamichhane JR, Schoving C, Debaeke P, Maury P. Early-stage phenotyping of root traits provides insights into the drought tolerance level of soybean cultivars. Agronomy. 2021;11:188. https://doi.org/10.3390/agronomy11010188.

    Article  CAS  Google Scholar 

  8. Henry A, Gowda VRP, Torres RO, McNally KL, Serraj R. Variation in root system architecture and drought response in rice (Oryza sativa): phenotyping of the OryzaSNP panel in rainfed lowland fields. F Crop Res. 2011;120:205–14. https://doi.org/10.1016/j.fcr.2010.10.003.

    Article  Google Scholar 

  9. Lopes MS, Reynolds MP. Partitioning of assimilates to deeper roots is associated with cooler canopies and increased yield under drought in wheat. Funct Plant Biol. 2010;37:147–56. https://doi.org/10.1071/FP09121.

    Article  Google Scholar 

  10. Kang J, Peng Y, Xu W. Crop root responses to drought stress: molecular mechanisms, nutrient regulations, and interactions with microorganisms in the rhizosphere. Int J Mol Sci. 2022. https://doi.org/10.3390/ijms23169310.

    Article  PubMed  PubMed Central  Google Scholar 

  11. Mairhofer S, Zappala S, Tracy S, Sturrock C, Bennett MJ, Mooney SJ, et al. Recovering complete plant root system architectures from soil via X-ray μ-computed tomography. Plant Methods. 2013;9:8. https://doi.org/10.1186/1746-4811-9-8.

    Article  PubMed  PubMed Central  Google Scholar 

  12. van Dusschoten D, Metzner R, Kochs J, Postma JA, Pflugfelder D, Bühler J, et al. Quantitative 3D analysis of plant roots growing in soil using magnetic resonance imaging. Plant Physiol. 2016;170:1176–88. https://doi.org/10.1104/pp.15.01388.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  13. Ahmed S, Klassen TN, Keyes S, Daly M, Jones DL, Mavrogordato M, et al. Imaging the interaction of roots and phosphate fertiliser granules using 4D X-ray tomography. Plant Soil. 2016;401:125–34. https://doi.org/10.1007/s11104-015-2425-5.

    Article  CAS  Google Scholar 

  14. Helen HL, Gao W, van der Bom F, Han WZ, Doolette CL, Maksimenko A, et al. Use of X-ray tomography for examining root architecture in soils. Geoderma. 2022;405:115405. https://doi.org/10.1016/j.geoderma.2021.115405.

    Article  Google Scholar 

  15. Vanhees DJ, Schneider HM, Sidhu JS, Loades KW, Bengough AG, Bennett MJ, et al. Soil penetration by maize roots is negatively related to ethylene-induced thickening. Plant Cell Environ. 2022;45:789–804. https://doi.org/10.1111/pce.14175.

    Article  CAS  PubMed  Google Scholar 

  16. Pflugfelder D, Metzner R, Dusschoten D, Reichel R, Jahnke S, Koller R. Non-invasive imaging of plant roots in different soils using magnetic resonance imaging (MRI). Plant Methods. 2017;13:102. https://doi.org/10.1186/s13007-017-0252-9.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  17. Mcneill A, Kolesik P. X-ray CT investigations of intact soil cores with and without living crop roots. SuperSoil 2004 3rd Aust New Zel Soils Conf. 2004.

  18. Wieland R, Ukawa C, Joschko M, Krolczyk A, Fritsch G, Hildebrandt TB, et al. Use of deep learning for structural analysis of computer tomography images of soil samples. R Soc Open Sci. 2021;8:201275. https://doi.org/10.1098/rsos.201275.

    Article  PubMed  PubMed Central  Google Scholar 

  19. Pfeifer J, Kirchgessner N, Colombi T, Walter A. Rapid phenotyping of crop root systems in undisturbed field soils using X-ray computed tomography. Plant Methods. 2015;11:41. https://doi.org/10.1186/s13007-015-0084-4.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  20. Teramoto S, Takayasu S, Kitomi Y, Arai-Sanoh Y, Tanabata T, Uga Y. High-throughput three-dimensional visualization of root system architecture of rice using X-ray computed tomography. Plant Methods. 2020;16:66. https://doi.org/10.1186/s13007-020-00612-6.

    Article  PubMed  PubMed Central  Google Scholar 

  21. Rogers ED, Monaenkova D, Mijar M, Nori A, Goldman DI, Benfey PN. X-ray computed tomography reveals the response of root system architecture to soil texture. Plant Physiol. 2016;171:2028–40. https://doi.org/10.1104/pp.16.00397.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  22. Booth S, Kurtz B, de Heer MI, Mooney SJ, Sturrock CJ. Tracking wireworm burrowing behaviour in soil over time using 3D X-ray computed tomography. Pest Manag Sci. 2020;76:2653–62. https://doi.org/10.1002/ps.5808.

    Article  CAS  PubMed  Google Scholar 

  23. Flavel RJ, Guppy CN, Tighe M, Watt M, McNeill A, Young IM. Non-destructive quantification of cereal roots in soil using high-resolution X-ray tomography. J Exp Bot. 2012;63:2503–11. https://doi.org/10.1093/jxb/err421.

    Article  CAS  PubMed  Google Scholar 

  24. Mairhofer S, Zappala S, Tracy SR, Sturrock C, Bennett M, Mooney SJ, et al. RooTrak: automated recovery of three-dimensional plant root architecture in soil from X-ray microcomputed tomography images using visual tracking. Plant Physiol. 2012;158:561–9. https://doi.org/10.1104/pp.111.186221.

    Article  CAS  PubMed  Google Scholar 

  25. Teramoto S, Tanabata T, Uga Y. RSAtrace3D: robust vectorization software for measuring monocot root system architecture. BMC Plant Biol. 2021;21:398. https://doi.org/10.1186/s12870-021-03161-9.

    Article  PubMed  PubMed Central  Google Scholar 

  26. Gao W, Schlüter S, Blaser SRGA, Shen J, Vetterlein D. A shape-based method for automatic and rapid segmentation of roots in soil from X-ray computed tomography images: rootine. Plant Soil. 2019;441:643–55. https://doi.org/10.1007/s11104-019-04053-6.

    Article  CAS  Google Scholar 

  27. Phalempin M, Lippold E, Vetterlein D, Schlüter S. An improved method for the segmentation of roots from X-ray computed tomography 3D images: rootine vol 2. Plant Methods. 2021;17:39. https://doi.org/10.1186/s13007-021-00735-4.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  28. Teramoto S, Uga Y. Four-dimensional measurement of root system development using time-series three-dimensional volumetric data analysis by backward prediction. Plant Methods. 2022;18:133. https://doi.org/10.1186/s13007-022-00968-x.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  29. Blaser SRGA, Schlüter S, Vetterlein D. How much is too much?—influence of X-ray dose on root growth of faba bean (Vicia faba) and barley (Hordeum vulgare). PLoS ONE. 2018;13:e0193669. https://doi.org/10.1371/journal.pone.0193669.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  30. Herrero-Huerta M, Raumonen P, Gonzalez-Aguilera D. 4DRoot: root phenotyping software for temporal 3D scans by X-ray computed tomography. Front Plant Sci. 2022;13:986856. https://doi.org/10.3389/fpls.2022.986856.

    Article  PubMed  PubMed Central  Google Scholar 

  31. Guo Y, Liu Y, Georgiou T, Lew MS. A review of semantic segmentation using deep neural networks. Int J Multimed Inf Retr. 2018;7:87–93. https://doi.org/10.1007/s13735-017-0141-z.

    Article  CAS  Google Scholar 

  32. Tracy SR, Black CR, Roberts JA, McNeill A, Davidson R, Tester M, et al. Quantifying the effect of soil compaction on three varieties of wheat (Triticum aestivum L.) using X-ray micro computed tomography (CT). Plant Soil. 2012;353:195–208. https://doi.org/10.1007/s11104-011-1022-5.

    Article  CAS  Google Scholar 

  33. Smith AG, Petersen J, Selvan R, Rasmussen CR. Segmentation of roots in soil with U-Net. Plant Methods. 2020;16:13. https://doi.org/10.1186/s13007-020-0563-0.

    Article  PubMed  PubMed Central  Google Scholar 

  34. Smith AG, Han E, Petersen J, Olsen NAF, Giese C, Athmann M, et al. RootPainter: deep learning segmentation of biological images with corrective annotation. New Phytol. 2022;236:774–91. https://doi.org/10.1111/nph.18387.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  35. Gaggion N, Ariel F, Daric V, Lambert É, Legendre S, Roulé T, et al. ChronoRoot: high-throughput phenotyping by deep segmentation networks reveals novel temporal parameters of plant root system architecture. Gigascience. 2021;10:1–15. https://doi.org/10.1093/gigascience/giab052.

    Article  Google Scholar 

  36. Shen C, Liu L, Zhu L, Kang J, Wang N, Shao L. High-throughput in situ root image segmentation based on the improved DeepLabv3+ method. Front Plant Sci. 2020;11:576791. https://doi.org/10.3389/fpls.2020.576791.

    Article  PubMed  PubMed Central  Google Scholar 

  37. Li Y, Huang Y, Wang M, Zhao Y. An improved U-Net-based in situ root system phenotype segmentation method for plants. Front Plant Sci. 2023;14:1115713. https://doi.org/10.3389/fpls.2023.1115713.

    Article  PubMed  PubMed Central  Google Scholar 

  38. Teramoto S, Uga Y. A deep learning-based phenotypic analysis of rice root distribution from field images. Plant Phenom. 2020. https://doi.org/10.3413/2020/3194308.

    Article  Google Scholar 

  39. Selzner T, Horn J, Landl M, Pohlmeier A, Helmrich D, Huber K, et al. 3D U-Net segmentation improves root system reconstruction from 3D MRI Images in automated and manual virtual reality work flows. Plant Phenom. 2023;5:0076. https://doi.org/10.3413/plantphenomics.0076.

    Article  Google Scholar 

  40. Alle J, Gruber R, Wörlein N, Uhlmann N, Claußen J, Wittenberg T, et al. 3D segmentation of plant root systems using spatial pyramid pooling and locally adaptive field-of-view inference. Front Plant Sci. 2023;14:1120189. https://doi.org/10.3389/fpls.2023.1120189.

    Article  PubMed  PubMed Central  Google Scholar 

  41. Gerth S, Claußen J, Eggert A, Wörlein N, Waininger M, Wittenberg T, et al. Semiautomated 3D root segmentation and evaluation based on X-ray CT imagery. Plant Phenom. 2021;2021:8747930. https://doi.org/10.3413/2021/8747930.

    Article  CAS  Google Scholar 

  42. He K, Zhang X, Ren S, Sun J. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Trans Pattern Anal Mach Intell. 2015;37:1904–16. https://doi.org/10.1109/TPAMI.2015.2389824.

    Article  PubMed  Google Scholar 

  43. Soltaninejad M, Sturrock CJ, Griffiths M, Pridmore TP, Pound MP. Three dimensional root CT segmentation using multi-resolution encoder-decoder networks. IEEE Trans Image Process. 2020;29:6667–79. https://doi.org/10.1109/TIP.2020.2992893.

    Article  Google Scholar 

  44. Douarre C, Schielein R, Frindel C, Gerth S, Rousseau D. Transfer learning from synthetic data applied to soil–root segmentation in X-ray tomography images. J Imaging. 2018;4:65. https://doi.org/10.3390/jimaging4050065.

    Article  Google Scholar 

  45. Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. Med Image Comput Comput Interv 2015 18th Int Conf Munich, Ger Oct 5–9, 2015, Proceedings, Part III 18. 2015;234–41. https://doi.org/10.1007/978-3-319-24574-4_28

  46. Chen H, Dou Q, Yu L, Qin J, Heng P-A. VoxResNet: deep voxelwise residual networks for brain segmentation from 3D MR images. Neuroimage. 2018;170:446–55. https://doi.org/10.1016/j.neuroimage.2017.04.041.

    Article  PubMed  Google Scholar 

  47. Fey M, Lenssen JE. Fast graph representation learning with PyTorch geometric. arXiv. 2019. https://doi.org/10.48550/arXiv.1903.02428.

    Article  Google Scholar 

  48. Van Rossum G, Drake FL. Python 3 reference manual. Scotts Valley: CreateSpace; 2009.

    Google Scholar 

  49. Sudre CH, Li W, Vercauteren T, Ourselin S, Jorge Cardoso M. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. Deep Learn Med Image Anal Multimodal Learn Clin Decis Support Third Int Work DLMIA 2017, 7th Int Work ML-CDS 2017, Held Conjunction with MICCAI 2017, Québec City, QC, Canada, Sept . 2017. p. 240–8

  50. Kingma DP, Ba JL. Adam: a method for stochastic optimization. arXiv. 2014. https://doi.org/10.48550/arXiv.1412.6980.

    Article  Google Scholar 

  51. Yang Q, Yan P, Zhang Y, Yu H, Shi Y, Mou X, et al. Low-dose CT image denoising using a generative adversarial network with Wasserstein distance and perceptual loss. IEEE Trans Med Imaging. 2018;37:1348–57. https://doi.org/10.1109/TMI.2018.2827462.

    Article  PubMed  PubMed Central  Google Scholar 

  52. Skourt BA, El Hassani A, Majda A. Lung CT image segmentation using deep neural networks. Procedia Comput Sci. 2018;127:109–13. https://doi.org/10.1016/j.procs.2018.01.104.

    Article  Google Scholar 

  53. Chen H, Zhang Y, Zhang W, Liao P, Li K, Zhou J, et al. Low-dose CT via convolutional neural network. Biomed Opt Express. 2017;8:679–94. https://doi.org/10.1364/BOE.8.000679.

    Article  PubMed  PubMed Central  Google Scholar 

  54. Chen H, Zhang Y, Kalra MK, Lin F, Chen Y, Liao P, et al. Low-dose CT with a residual encoder-decoder convolutional neural network. IEEE Trans Med Imaging. 2017;36:2524–35. https://doi.org/10.1109/TMI.2017.2715284.

    Article  PubMed  PubMed Central  Google Scholar 

  55. Wolterink JM, Leiner T, Viergever MA, Išgum I. Generative adversarial networks for noise reduction in low-dose CT. IEEE Trans Med Imaging. 2017;36:2536–45. https://doi.org/10.1109/TMI.2017.2708987.

    Article  PubMed  Google Scholar 

  56. Zhao Y, Sun C, Xu X, Chen J. RIC-Net: a plant disease classification model based on the fusion of inception and residual structure and embedded attention mechanism. Comput Electron Agric. 2022;193:106644. https://doi.org/10.1016/j.compag.2021.106644.

    Article  Google Scholar 

  57. Niu X, Yan B, Tan W, Wang J. Effective image restoration for semantic segmentation. Neurocomputing. 2020;374:100–8. https://doi.org/10.1016/j.neucom.2019.09.063.

    Article  Google Scholar 

  58. Brempong EA, Kornblith S, Chen T, Parmar N, Minderer M, Norouzi M. Denoising pretraining for semantic segmentation. Proc IEEE/CVF Conf Comput Vis Pattern Recognit. 2022. p. 4174–85. https://doi.org/10.1109/CVPRW56347.2022.00462

  59. Dey S, Bhattacharya R, Schwenker F, Sarkar R. Median filter aided CNN based image denoising: an ensemble approach. Algorithms. 2021;14:109. https://doi.org/10.3390/a14040109.

    Article  Google Scholar 

  60. Ramesh S, Sasikala S, Paramanandham N. Segmentation and classification of brain tumors using modified median noise filter and deep learning approaches. Multimed Tools Appl. 2021;80:11789–813. https://doi.org/10.1007/s11042-020-10351-4.

    Article  Google Scholar 

  61. Marmanis D, Schindler K, Wegner JD, Galliani S, Datcu M, Stilla U. Classification with an edge: improving semantic image segmentation with boundary detection. ISPRS J Photogramm Remote Sens. 2018;135:158–72. https://doi.org/10.1016/j.isprsjprs.2017.11.009.

    Article  Google Scholar 

  62. Chen L-C, Barron JT, Papandreou G, Murphy K, Yuille AL. Semantic image segmentation with task-specific edge detection using cnns and a discriminatively trained domain transform. Proc IEEE Conf Comput Vis pattern Recognit. 2016;4545–54. https://doi.org/10.1109/CVPR.2016.492

  63. Wang S, Liu W, Wu J, Cao L, Meng Q, Kennedy PJ. Training deep neural networks on imbalanced data sets. Int Jt Conf Neural Netw. 2016. https://doi.org/10.1109/IJCNN.2016.7727770.

    Article  Google Scholar 

  64. Teramoto S, Uga Y. Improving the efficiency of plant root system phenotyping through digitization and automation. Breed Sci. 2022;72:48–55. https://doi.org/10.1270/jsbbs.21053.

    Article  PubMed  PubMed Central  Google Scholar 

  65. Khoroshevsky F, Zhou K, Chemweno S, Edan Y, Bar-hillel A. Automatic root length estimation from images acquired in situ without segmentation. Plant Phenom. 2023. https://doi.org/10.3413/plantphenomics.0132.

    Article  Google Scholar 

Download references

Acknowledgements

Not applicable.

Funding

This study was supported by JSPS KAKENHI (grant number: 22K14871 to ST) and JST CREST (grant number: JPMJCR17O1 to YU), Japan.

Author information

Authors and Affiliations

Authors

Contributions

ST designed the study, cultivated and scanned the plants, trained the models, analyzed the data, and wrote the manuscript. YU managed the X-ray CT system and revised the manuscript. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Shota Teramoto.

Ethics declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Additional information

Publisher's Note

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

Supplementary Information

13007_2024_1208_MOESM1_ESM.pptx

Supplementary Material 1: Fig. S1 Dividing CT volume for the training of semantic segmentation. (a) CT volume before division. (b) CT volume divided into small sub-volumes. (c) The sub-volume comprising unique and overlapped regions.

13007_2024_1208_MOESM2_ESM.pptx

Supplementary Material 2: Fig. S2 Segmentation results of CT volume used for training. Predicted segmentation of (a) S0033, (b) S0066, (c) S0150, (d) S0300, and (e) S0600.

Supplementary Material 3: Movie S1 Animation of 3D rendering of RSAvis3D-processed S0150.

Supplementary Material 4: Movie S2 Animation of 3D rendering of predicted S0150. White: predicted segments overlapping the ground truth, blue: predicted segments not overlapping the ground truth, red: ground truth segments not predicted.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Teramoto, S., Uga, Y. Convolutional neural networks combined with conventional filtering to semantically segment plant roots in rapidly scanned X-ray computed tomography volumes with high noise levels. Plant Methods 20, 73 (2024). https://doi.org/10.1186/s13007-024-01208-0

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s13007-024-01208-0

Keywords