A deep learning-integrated phenotyping pipeline for vascular bundle phenotypes and its application in evaluating sap flow in the maize stem

Jianjun Du ,Ying Zhang ,Xianju Lu,Minggang Zhang,Jinglu Wang,Shengjin Liao,Xinyu Guo ,Chunjiang Zhao

Beijing Key Lab of Digital Plant,Research Center of Information Technology,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China

Keywords:Deep learning Maize stem Phenotyping Semantic segmentation Vascular bundle

ABSTRACT Plant vascular bundles are responsible for water and material transportation,and their quantitative and functional evaluation is desirable in plant research.At the single-plant level,the number,size,and distribution of vascular bundles vary widely,posing a challenge to automatically and accurately identifying and quantifying them.In this study,a deep learning-integrated phenotyping pipeline was developed to robustly and accurately detect vascular bundles in Computed Tomography (CT) images of stem internodes.Two semantic indicators were used to evaluate and identify a suitable feature extraction network for semantic segmentation models.The epidermis thickness of maize stem was evaluated for the first time and adjacent vascular bundles were improved using an adaptive watershed-based approach.The counting accuracy(R2)of vascular bundles was 0.997 for all types of stem internodes,and the measured accuracy of size traits was over 0.98.Combining sap flow experiments,multiscale traits of vascular bundles were evaluated at the single-plant level,which provided an insight into the water use efficiency of the maize plant.

1.Introduction

The vascular system is an essential aspect of stem structure and is responsible for both the delivery of resources (water,essential mineral nutrients,sugars,and amino acids) to the various plant organs and the provision of mechanical support,accurate acquisition of vascular bundle phenotype is the premise of its function study [1].Conventionally,observation and identification of plant tissues have relied on histological analysis of sectioned specimens.Light-based microscopy,confocal laser scanning microscope,fluorescence microscopy,and scanning electron-based microscopy techniques are the primary technical means of observing the inner tissues of plant samples.However,the required sample preparation typically alters the structure of specimens by a series of physical and chemical treatments comprising sampling,fixing,sectioning,and staining.With less preprocessing,microcomputed tomography (micro-CT) has been used extensively to generate high-resolution images of internal plant tissues non-destructively and accurately characterize their structure.Plant organs imaged and visualized by X-ray microscopy include seeds[2],leaves [3,4],and stems [5].Tissue water contents of plant tissues affect X-ray imaging quality owing to their low absorption contrast;intensity changes in CT images are associated with the degree of lignification,drying,and dyeing operations.The technique[6] using osmium tetroxide,iodinated contrast agents,inorganic iodine,and phosphotungstic acid to stain soft tissues could enhance tissue contrast in micro-CT images.For vascular bundles,because of the influence of differences in water content and material composition on X-ray attenuation,the CT value range used for image reconstruction needs to be manually adjusted to ensure the visual clarity of vascular bundles,so that the intensity values of reconstructed CT images are not uniform.

For CT images of the maize stem,the manual delineation of vascular bundles is prone to errors and inter/intra-observer variation.Because conventional image segmentation based on feature engineering is object-dependent,some empirical algorithm parameters are necessary to customize the algorithm pipeline for the specific CT image.Several studies[5,7-9]have focused on this topic.However,the segmentation and identification of vascular bundles are still a bottleneck.Deep learning based on Deep Convolutional Neural Network (CNN) is a data-driven approach that can explore multiple image features,thus it is well adapted to the size and shape changes of vascular bundles.CNN also provides an end-toend mechanism for achieving segmentation of the input image in just one forward network inference.In recent years,CNN has achieved state-of-the-art performance in object detection and image segmentation tasks and is widely used for image-based phenotyping in plant phenomics [10-13].In semantic segmentation,encoder-decoder networks such as FCN [14],SegNet [15],U-NET[16],PSPNet[17],and DeepLab[18]have become the most widely used methods.For CT image analysis,U-Net is one of the most well-known fully convolutional network(FCN)structures and uses a U-shaped encoder-decoder structure to down or up sample the original image and uses skip connection to preserve the highresolution spatial information.

The architecture of vascular bundles can be quantified by several traits including number,size,morphology,and distribution.In a previous study [19],trait evaluation of vascular bundles was performed for third internodes of maize stems in 20 maize cultivars at two growth stages [9],and 480 inbred lines were also assessed.Those results indicated that the traits of vascular bundles were associated with maize cultivar and growth stages.Further,the size and shape of vascular bundles at different internodes varied greatly from the base node to the apical node.Thus,it is a challenge to adapt an automatic and robust phenotyping pipeline based on feature engineering for CT images of diverse internodes.

As an effective long-distance source-sink transport system,vascular bundles show correlations between vascular bundle phenotypes and drought stress resistance [20-22].Crops with high potential adaptability to environments have strong plasticity in the morphological structure of their stems.High plasticity enables them to increase the transport capacity of water and material by changing their structure under drought-stress conditions,thus reducing the effect on their growth and development.Of interest is the relationship between the architecture of vascular bundles at the single-plant level and traits involved in water transportation.A systematic analysis of vascular bundles’ architecture and patterns would provide a guide for evaluating stem flow in maize plants.

The objective of this study was to develop a deep learningintegrated phenotyping pipeline to automatically extract and quantify vascular bundles for various types of stem internode.Our approach was to evaluate the architectural differences of vascular bundles at the single-plant level for multiple growth environments and investigate the relationship between flow and multiscale traits.

2.Materials and methods

2.1.Plant material,growth conditions,and data acquisition

Four maize inbred lines (B73,Zheng58,Qi319,and Ji63) were planted in the field and drought shed of the Beijing Academy of Agriculture and Forestry Sciences (39°94′N,116°29′E).The field was a natural cultivation environment,and the drought shed was covered with waterproof plastic film to induce drought stress.Sap flow was monitored for selected inbred lines at the tasseling stage in the field (July 30 to August 4,2019) and drought shed(August 13-18,2019),respectively.The sensors of stem flow gauges (Dynamax Inc.,Houston,TX,USA) were installed at the third internodes of inbred lines,as shown in Fig.1A.After the sap flow experiments,these plants were destructively sampled and each stem internode from the third to the uppermost node was cut into a segment with 1 cm thickness.All segments were then soaked in FAA solution (90:5:5 v/v/v,70% ethanol:100%formaldehyde:100% acetic acid).Detailed sample preparation protocols were as described [23].

Before CT imaging,all stem samples of the whole maize plant were sorted from the third to the apical internodes,and imaged with a digital camera(Canon 5D Mark III,Canon,Tokyo Prefecture,Japan),as shown in Fig.1B.In the laboratory,stem samples were imaged using a desktop Micro-CT device (SkyScan 1172,Bruker,Nazareth,Belgium).The unified scanning parameters were set as 40 kV/250 μA,the imaging pixel sizes as 13.55 μm,and the 2K scanning mode (2000 × 2000 pixels).The slices (cross-section images) of stem samples were reconstructed as 8-bit BMP files and visualized in sequence as Fig.1C.

2.2.Annotation dataset preparation

To achieve automatic segmentation of vascular bundles in multiple types of stem samples,we constructed a dataset of vascular bundles based on 47 CT images(of 2000×2000 pixels)for training and evaluating the semantic segmentation model.CT images were captured from maize stems of diverse growth stages,varieties,and nodes.Vascular bundles in each CT image were of diverse sizes and morphologies so that these images could represent the characteristics of varying maize cultivars and stem internodes.To facilitate the annotation of vascular bundles,a simple contour-editing tool was developed to delineate the fine contours of vascular bundles for each CT image,and then generate the corresponding mask image in PNG format.The mask image and its corresponding CT image formed an image pair for retrieval and use.The size of the CT image was much larger than the input size required by the models.To prevent information loss due to the considerable shrinking of the input image,a sliding window was used to simultaneously divide the CT and mask images into a series of small images of 288 × 288 pixels (Fig.S1A).Each image pair was evaluated under supervision,and some small images that contained no valuable information about the maize stem were removed from the annotation dataset.For each vascular bundle,a closed spline curve could be initialized and dragged to fit the boundary of the bundle,thus refining the observable boundaries of the bundles by humanmachine interaction (Fig.S1B-E).The resulting 1659 image pairs were divided randomly into training (1383),validation (227),and test (276) datasets.

2.3.Semantic segmentation

Five semantic segmentation models based on vgg16 [24],resnet-50[25],inceptionv3[26],efficientNet0[27],and inceptionresnetv2 [28] were trained using the same dataset to determine the appropriate feature extraction network for semantic segmentation of vascular bundles,as shown in Fig.S1.This network would resize input images of differing sizes to a fixed size(288×288 pixels),and the loss function was a combination of Focal loss[29]and Dice loss [30].

DSC (Dice Similarity Coefficient) and IoU (Intersection-over-Union) were the most commonly used evaluation metrics for image segmentation tasks [31].There was a slight difference between IoU and DSC in measuring the similarity between the predicted image(PR)and its ground-truth image(GT).The mean similarity degrees of all evaluation images could be calculated as follows:

where n indicates the number of images.

Fig.1.Experiment design and data acquisition.(A)Sap flow measurement at the third stem internode of each single maize plant using Flow-32 equipment.(B)Stem samples of a whole maize plant were sorted from the third to the apical internodes.(C) Each internode was reconstructed as cross-sectional images by a Micro-CT device and the generated CT images were visualized in sequence.

Two semantic metrics were used to indicate the model capacity for identifying vascular bundles.By extracting all individual connected regions from the predicted image and ground-truth image,the relative differences of object area (OA),perimeter (OP),and number (ON) could be calculated.Thus,the mean absolute percentage error (MAPE) of OA,OP,and ON was used to indicate the accuracy of semantic segmentation of vascular bundles.Because each vascular bundle was represented as a connected object,all holes in the predicted image were taken as wrong regions.Thus,the mean absolute error (MAE) of the hole area (HA),perimeter(HP),and number(HN)were also computed to represent the deviation degree of false segmentation.MAPE and MAE were calculated as follows:

where n indicates the number of images,obj is a feature of segmented objects (OA,OP,or ON),and hol is a hole feature of segmented results (HA,HP,or HN).MAPEobjwas the most important semantic indicator and represented the inference accuracy of area,perimeter and number of vascular bundles.The larger the MAPEobj,the worse is the inference of objects.MAEholwas the auxiliary semantic indicator that represented the region integrity of vascular bundles.Vascular bundles in the annotated image never contained any holes,so the smaller the MAEhol,the better.

2.4.Phenotyping pipeline

The phenotyping pipeline was employed to extract rich and diverse traits of stem internodes from CT images,as shown in Fig.2.This pipeline consisted of three steps: detecting vascular bundles,identifying zones,and phenotyping.The above semantic segmentation was first used to obtain the initial candidate regions of vascular bundles.Then,each candidate region was evaluated to determine whether it was a single valid vascular bundle.In the given CT image,the maize stem could be divided into three zones:the epidermis(EZ),periphery(PZ),and inner(IZ)zones.The definition and detection approach of these zones were as described [9].In the phenotyping step,we studied traits of maize stem and its vascular bundles from multiple perspectives,including quality,quantity,size,and shape,and calculated the corresponding traits of vascular bundles and function zones.We found that the adjacency relationship of vascular bundles in the periphery zone was for describing the degree of separation between vascular bundles.Thus,two descriptive traits were defined: area ratio of individual vascular bundles (ARIVB) and separation ratio of vascular bundles(SRVB).The complete traits are listed in Table S1.In the phenotyping pipeline,estimation of epidermis thickness and identification of vascular bundles were two challenging issues that determined the computation accuracy of function zones and vascular bundles(Fig.2).These are discussed in detail in the next two sections.

Fig.2.Phenotyping pipeline of maize stem CT images.Series of CT images at various stem internodes were collected for the batch processing of this pipeline.

2.4.1.Estimating the epidermis thickness

Many vascular bundles in the periphery zone were in close contact with the epidermis region,and the epidermis was often damaged and fractured.For this reason,it was difficult to robustly estimate the thickness of the epidermis.However,by observing various slice images,the following inferences could be made:(1) because the epidermis zone,as a ring region,showed little thickness variation,we could use a uniform value to represent its thickness;(2)in the current CT imaging mode,the epidermis thickness of the maize stem was small,and must fall within a specific range(e.g.,less than 20 pixels).(3)The brightness of the epidermis zone was always higher than that of other regions,and the mean intensity of the estimated region (ER) of the epidermis gradually decreased from the epidermis to the interior.(4) Starting from the outer boundary of the epidermis,the boundary was shrunk gradually with the smallest step size of 1,and each new region(NR) could be used to calculate the mean intensity of NR.

Based on the above assumptions,we presented a thickness estimation approach for the epidermis zone,as shown in Fig.3A.Because the thickness range was assumed as [1,N] pixels (N was a user-defined parameter that must be larger than the true epidermis thickness),an erode operation could be performed N times from the slice mask (SM) image,yielding N eroded masks (EM).By combining SM and EM,N estimated regions (ER) could be obtained;by combining the adjacent EMs,we could obtain N-1 new regions (NR).We accordingly calculated the mean intensity of each ER and NR in the CT image.Because we set N as 20,these intensity values could be plotted as two intensity curves (AIEP and AINR),as shown in Fig.3B.Owing to a large area of cavities in the periphery and inner regions,the mean intensity of NR dropped sharply once NR left the epidermis zone,meaning that the two curves could intersect.For this reason,the intersection point was calculated based on linear fitting,and taken as the best estimate of epidermis thickness.

Fig.3.Estimating the epidermis thickness of the maize stem.(A) The flow chart using continuous morphological operations and intensity calculation.(B) Finding the intersection point between two intensity curves (AIEP and AINR).

2.4.2.Identifying vascular bundles

The phenotyping of vascular bundles was divided into a twostage task.In the first stage,all candidate regions of vascular bundles were extracted by the semantic segmentation model and used to identify vascular bundles.In some cases,the individual candidate region might represent a vascular bundle but might contain numerous vascular bundles.There was no doubt that the candidate region containing multiple vascular bundles contained sufficient information in size and shape to describe the aggregation of multiple vascular bundles.Each candidate region was identified mainly in the second stage according to its geometric and morphological features.An adaptive watershed-based approach was used to identify these candidate regions,as shown in Fig.4.This approach consisted of two hyperparameters: the convex area ratio (CAR)between the number of object pixels and the convex hull area,and the minimum distance (MD) between two adjacent objects.The classical watershed-based algorithm used a unique distance parameter to separate all adjacent objects in the given connected region.However,once the connected region contained more than two objects,and the size of the objects (or the distance between the objects)varied greatly,the boundaries among multiple objects could not be adequately detected.In contrast to the classical watershed-based algorithm,MD in the present approach could be automatically adjusted to adapt to the size change of objects.

Fig.4.The adaptive watershed-based approach for identifying vascular bundles from a list of candidate regions.

The semantic segmentation of vascular bundles generated a list of candidate regions,which was used to identify vascular bundles using the adaptive watershed-based approach.The process may be described as follows.(1) The maximum inscribed circle of each given connected region was first calculated and its radius was taken as the MD parameter.(2)A distance map was built by calculating the Euclidean distance from every object pixel to the nearest zero pixels.(3) All peaks in this distance map were found,and the distances between peaks must be less than the MD.(4) For each local peak,a connected component analysis using 8-connectivity and the classical watershed-based algorithm was performed to set each pixel as a unique label value,and each unique label value was retrieved as a unique object,namely a pixel set with the same label value representing an individual object.(5)A list of individual candidate regions was generated by the classical watershed-based algorithm,the largest object region was selected for the identification of vascular bundles,and other regions were set as new candidate regions to repeat the above processes.(6) A region that was assigned as a vascular bundle was output into a new list or discarded.This list included all individual vascular bundles.

2.5.Working conditions and statistical analysis

The image analysis pipeline was implemented with Python(3.6.7) in the Windows package.The computer configuration was: Intel(R) Core i7-5930 k CPU@3.50 GHz (Intel Corp.,CA,USA),128 G memory,two 8 GB NVIDIA GeForce GTX-1080 Ti graphics cards (NVIDIA Corp.,CA,USA),2 TB hard disk,and the operating system was Windows 10.U-Nets were constructed by using TensorFlow (1.12.0) and Keras (2.2.4) as the backend on Compute Unified Device Architecture (CUDA,9.0.176) and the Deep Neural Network library (cuDNN,7.3.1).Pearson’s correlation and the group violin were implemented using Seaborn libraries in Python (0.10.0).

3.Results and discussion

3.1.Training CNN models

Five feature extraction networks (backbones),i.e.,vgg16,resnet-50,inceptionv3,efficientNet0,and inceptionresnetv2,were used to train semantic segmentation models,respectively.The same hyperparameters were configured as follows: sigmoid was used as the activation function;batch size was set as 8 to maximize the use of GPU storage;epochs was 50;an adaptive moment estimate (Adam) was an optimizer;the learning rate (LR) was set to 5×10-5;data augmentation was implemented [32],thereby contributing to both the prevention of over-fitting and enhanced performance.After 50 epochs,Loss and IoU curves of training and validation are drawn (Fig.S1F).The convergence trend of these model training was consistent.However,the training time and storage size of models with different backbones differed owing to the parameters of the feature extraction network (Table S2).In the following section,we further evaluated whether the segmentation results of these models correctly revealed the semantic information of vascular bundles,using semantic metrics(formulas 3-4).

3.2.Evaluating semantic segmentation models

We used 18 original CT image pairs to evaluate the performance of five models.In the reasoning processing of models,CT images were first scaled to their nearest multiple of 32 (a CT image with 2000 × 2000 pixels was scaled into 1984 × 1984 pixels with little information loss) as the input of the current model,and the segmented images were scaled back to their original size.In five models,the mIoU and mDSC scores of resnet-50,inceptionv3,and inceptionresnetv2 were higher than that of vgg-16 and efficientNet0 (as shown in Table 1).Although the model with resnet-50 yielded the highest mIoU and mDSC scores in the original CT images,the semantic metrics of objects gave lower scores than those of inceptionresnetv2.For MAPEobj,a model with inceptionresnetv2 achieved the highest scores,and MAPEOA,MAPEOPand MAPEONof vascular bundles were respectively 5.028%,1.631%,and 1.330%.For MAEhol,the model with resnet-50 was the most effective to maintain the integrity of vascular bundles.As a comparison,a model with inceptionresnetv2 also gave good performance(MAEHA,MAEHPand MAEHNwere less than 1),indicating that there were almost no holes in the extracted vascular bundles.The semantic metrics of objects and holes largely determined the accuracy of vascular bundle semantics.Accordingly,MAPEobjand MAEholwere selected as the criteria for determining the most appropriate backbone.Among the five models,the model with inceptionresnetv2 was most suitable for the presented phenotyping pipeline of vascular bundles.The mean inference time based on inceptionresnetv2 was relatively high (1.361 s) but acceptable.

Table 1 Model evaluation using original images (2000 × 2000 pixels).

3.3.Identification of vascular bundles

The semantic segmentation model with inceptionresnetv2 was used to automatically extract vascular bundles from CT images.The generated candidate regions of semantic segmentation needed to be further identified as vascular bundles to improve the counting accuracy of vascular bundles (VB_N).The counting error of VB_N arose mainly from the adjacent vascular bundles in the periphery zone of the maize stem.In the various internodes of maize stem,vascular bundles in the periphery zone were usually more minor and in closer contact.This distribution pattern of vascular bundles was observed in CT images for multiple growth stages and multiple internodes at the same growth stage.Typically,dozens of candidate regions that contained multiple vascular bundles could be observed in the single CT image,as shown in Fig.5A.Accordingly,candidate regions were classified into two types:individual regions (Fig.5B) and multiple regions (Fig.5C).

An adaptive watershed-based approach was utilized to identify vascular bundles by dividing candidate regions.The divided regions were labeled with discriminative colors,as shown in Fig.5D.The convex area ratio(CAR)was set as 0.9 for the identification filter of vascular bundles through extensive experiments.The minimum distance(MD)values were labeled below candidate regions,which demonstrated the adaptive changes of MD.Finally,each candidate region was divided into a series of individual regions,which were further identified as vascular bundles,and the counting accuracy of vascular bundles could be greatly improved.

Fig.5.Evaluating the adaptive watershed-based approach.(A)Semantic segmentation image with candidate regions.(B)The individual regions are extracted by meeting the identification criteria(CAR).(C)Multiple regions consist of multiple objects.(D)Objects are split by adaptive minimum distance(MD).The digit in(D)indicates the adjusted MD values.

3.4.Pipeline performance

We used 131 CT images to evaluate the computation performance of the phenotyping pipeline.Five quantity-related and size-related traits were measured manually as ground-truth values(GT_) to evaluate computation accuracy.Manual detection indicators included the cross-section’s main axis and short axis length(GT_SZ_LA and GT_SZ_SA),and the number of vascular bundles in the cross-section,periphery zone,and inner zone (GT_VB_N,GT_PZ_VB_N,and GT_IZ_VB_N).The coefficient of determination(R2)was used to assess the consistency of measured and predicted values for each trait.As shown in Fig.S2A and B,R2 of SZ_LA and SZ_SA were 0.982 and 0.984,respectively,indicating that the size-related traits predicted by this pipeline were highly believable.The predicted number of vascular bundles (VB_N) achieved extremely high counting accuracy with R2of 0.997(Fig.S2C).However,the predicted number of vascular bundles in the periphery and inner zones (PZ_VB_N and IZ_VB_N) dropped slightly to R2of 0.979 and 0.959,as shown in Fig.S2D and E.The results indicated that the trait measurements estimated by our algorithm were in good agreement with manual measurements.In contrast with previous studies based on conventional segmentation methods [9],the new phenotyping technique based on a convolutional neural network could quantify the characteristics of epidermis thickness,and showed higher robustness and accuracy in the quantitativerelated traits for multiple stem internodes.

The phenotyping pipeline could be applied to CT images of multiple internodes to obtain phenotypic information,as shown in Fig.6.The identified vascular bundles of several stem internodes were used as a starting point to quantify the distribution of vascular bundles(Fig.6A).Four function zones were used to characterize the properties of the slice and vascular bundles: slice zone (SZ),epidermis zone (EZ),periphery zone (PZ),and inner zone (IZ) are shown in Fig.6B (The green curve is the boundary between PZ and IZ,and the blue curve is the boundary between EZ and PZ).Vascular bundles were plotted in pseudo colors (Fig.6C),and further classified into the corresponding zones (Fig.6D).The traits of vascular bundles could be counted respectively in the inner,periphery,and entire slice zone,and represented as IZ_VB_N,PZ_VB_N,and VB_N.The results showed that the phenotyping pipeline could robustly and accurately detect the epidermis,function zones,and vascular bundles of multiple internodes.After the traits of vascular bundles at different internodes were calculated,we could analyze the spatial distribution,effects of drought stress,and flow behavior of vascular bundles at the single-plant level.

3.5.Distribution of vascular bundles in different internodes

Four inbred lines (Zheng58,B73,Qi319,and Ji63) were used as test material.Stem samples from the third to the uppermost internode of each inbred line were collected at silking stage.After CT scanning,the deep learning-integrated phenotyping pipeline was used to quantitatively analyze maize stem and vascular bundles at the single-plant level (Fig.7).The results showed that vascular bundle density,including VB_D,PZ_VB_D,and IZ_VB_D,increased with the increase of stem nodes.In contrast,the morphological and geometric properties of stem cross section (such as SZ_P,SZ_LA,SZ_SA,CS_A,CS_CA,and CS_CCA),the number,morphological and geometric properties of vascular bundles (such as VB_N,PZ_VBN,IZ_VBN,VB_LAave,VB_SAave,VB_Aave,VB_Pave,VB_A,PZ_VB_A,and IZ_VB_A),and the morphological and geometric properties of function zones(EZ_A,PZ_A,IZ_A,PZ_A,PZ_T,IZ_T,and IZ_T) decreased gradually with the increase in stem nodes.Generally,the number of vascular bundles in the periphery zone(PZ_VB) was more extensive than that in the inner zone (IZ_VB).However,the density of vascular bundles in the periphery zone(PZ_VB_D)was much larger than that in the inner zone(IZ_VB_D).The density of vascular bundles might be a more stable quantityrelated trait,given that it reflects both the number and area of vascular bundles.The size-related traits showed similar change trends with the number of vascular bundles.The thickness of the epidermis (EZ_T) did not show regular changes in different nodes,but showed significant differences in different inbred lines.The quality traits were associated mainly to the material composition and density in the vascular bundles and function zones.The mean intensity of epidermis zone (EZ_Imean) was highest,and that of the inner zone (IZ_Imean) lowest.The intensity of vascular bundles in the periphery zone (PZ_VB_Imean) was also higher than that in the inner zone (IZ_VB_Imean),i.e.,PZ_VB_Imean >VB_Imean >IZ_V B_Imean.

Fig.6.Automated phenotyping pipeline of vascular bundles for four internodes (5th,8th,11th,and 13th stem internode) at the single-plant level.(A) Identified vascular bundles.(B) Function zones.(C) Vascular bundles with pseudo colors.(D) Vascular bundles are classified into the corresponding zones.

3.6.Effects of drought stress on maize stem and vascular bundles

Group violin charts were used to show the trait differences of stems between the field and drought-shed environments(Fig.S3).The length and width of the cross section (SZ_LA and SZ_SA),and the epidermis thickness (EZ_T) in the field environment were much higher than those in the drought shed environment,and the difference of SZ_CA was nearly double.For vascular bundles,the total number (VB_N),the average area(VB_Aave),length (VB_LAave) and width (VB_SAave) in the field were slightly higher than that in the drought shed,and showed a larger range of variation.Similar to SZ_CA,the total area of vascular bundles (VB_A) showed a nearly twofold difference between the two environments.The area ratio of individual vascular bundles(ARIVB) and the separation ratio of vascular bundles (SRVB)showed slight differences in the two environments.

The functional zones of the stem were also heavily influenced by environmental conditions.The areas of the epidermis (EZ_CA),periphery (PZ_CA) and inner zones (IZ_CA) in the field were also higher than those in the drought shed.For the periphery and inner zones,the number (PZ_VB_N and IZ_VB_N),area (PZ_VB_A and IZ_VB_A),convex area (PZ_VB_CA and IZ_VB_CA) of vascular bundles in the field were significantly higher than that in the drought shed.However,the shape-related traits showed only slight differences,as for the convex area ratio (PZ_VB_CAR and IZ_VB_CAR).The densities of vascular bundles (PZ_VB_D and IZ_VB_D) were much smaller in the field than in the drought shed.

The mean pixel intensities of the cross-section(SZ_Imean),epidermis (EZ_Imean),periphery (PZ_Imean),and inner zones (IZ_Imean) in the field were higher than in the drought shed,and showed a consistent order: EZ_Imean >PZ_Imean >SZ_Imean >I Z_Imean.The same pattern could be found for the vascular bundles in different functional zones (PZ_VB_Imean and IZ_VB_Imean).

Fig.7.Trait comparison of stem tissue and vascular bundles at the single-plant level among four inbred lines (Zheng58,B73,Qi319,and Ji63).(A) CT scanning images for several stem internodes(from the third to top node)of four inbred lines.(B)Phenotyping processing results of vascular bundles for several internodes(from the third to top node) of four inbred lines.

3.7.Sap flow analysis

Sap flow experiments in maize plants were carried out successively in the field and the drought shed environments.Sap flow rates (g h-1) were affected not only by the hydraulic structure and physiological factors (vessels,slice area,leaf transpiration,etc.) of the maize plant but by environmental or meteorological factors (radiation,air temperature,soil moisture,wind speed)[33].Although the sap flow rate showed wide variation at the hourly scale,the statistical values for consecutive and multiple days could exclude to some extent the effects of external environmental factors,and reflect the sap transport performance of the plant itself.For this reason,the maximum and summary values of sap flow per day were used to characterize the sap behavior of the maize plant,i.e.,the capacity of sap flow rate (SFR,g h-1)and capacity of sap flow amount (SFA,g).The SFR curve between 7:00 AM and 6:00 PM every day was used to calculate the two sap flow (SF) indicators,due to the measured sap flow rate of 0 at other times.

The SFA for six consecutive days in the field and drought-shed environments are plotted in Fig.S4A and B.In the same growth environment,the stem flow curves of different plants showed similar trends,indicating that the SFA change of the same plant was affected by environmental factors,while the structure and physiological characteristics determined the SFA differences of different plants.To eliminate the influence of the growth environment,we calculated the max,mean,standard deviation(std),and min values of SFA,and obtained eight new indicators of sap flow,as shown in Fig.S4C.Several of these indicators were strongly correlated.From a quantitative statistical point of view,the SFR_mean per plant was 168.908 g h-1in the field environment and 146.519 in the drought shed,and SFA_means per plant for one day were 964.977 g and 766.672 g,respectively.We chose four representative indicators:SFR_mean,SFR_min,SFA_mean,and SFA_min,to investigate the relationship between sap flow and traits of vascular bundles.

3.8.Relationship between sap flow and the multiscale traits of vascular bundles

We calculated multiscale traits of maize stems and vascular bundles separately at the cross section,internode,and singleplant level.First,traits of the cross section were derived from the CT image of each internode.These traits included 2D size-related traits (SZ_A_mean and VB_A_mean) and quantity-related traits(VB_N_mean,PZ_VB_N_mean,IZ_VB_N_mean,PZ_VB_D_mean,and IZ_VB_D_mean).Further,we introduced the length of each internode (measured manually) to calculate 3D size-related traits of each internode and whole plant.At the internode scale,the average volumes of the stem,epidermis zone,periphery zone,inner zone and vascular bundles (S_SZ_V_mean,S_EZ_V_mean,S_PZ_V_mean,S_IZ_V_mean and S_VB_V_mean) could be calculated,respectively.The volumes of stem,epidermis,periphery zone,inner zone,and vascular bundles were further calculated at the single-plant scale,including P_SZ_V,P_EZ_V,P_PZ_V,P_IZ_V,and P_VB_V.Detailed descriptions are provided in Table S3.

Fig.8.Correlation analysis between sap flow(SF)indictors and multiscale traits of maize stems.a,density-related traits group;b,minimum sap flow indicators group;c,2D size-related traits group;d,3D size-related traits group;e,quantity-related traits group;f,maximum sap flow indicators group.

Hierarchical clustering was performed to evaluate the relationship between sap flow indicators and multiscale traits,as shown in Fig.8.By measuring the distance between each pair of traits,cluster analysis grouped traits that were close together.The investigated traits were divided into six groups for examining the differences and relationship between traits and SF indicators,namely density-related traits group(a),minimum sap flow indicators group (b),2D size-related traits group (c),3D size-related traits group (d),quantity-related traits group (e),and maximum sap flow indicators group (f).The Pearson correlation coefficients(PCCs) between traits were also calculated and annotated in each square of the heat map.Some of the valuable findings are as follows:

(1) Quantity-related traits vs.SF indicators.

Based on the cluster analysis,quantity-related traits were assigned to a,e,and f groups.The densities (a,PZ_VB_D_mean and IZ_VB_D_mean)of vascular bundles in the periphery and inner zones were negatively correlated with max sap flow indicators(with max PCC -0.63),and more weakly correlated with min sap flow indicators (with max PCC only 0.36).However,the number of vascular bundles (e,VB_N_mean and PZ_VB_N_mean;f,IZ_VB_N_mean) was strongly positively correlated with sap flow indicators (with max PCC 0.75).In particular,IZ_VB_N_mean showed the highest PCC with sap flow and accordingly was clustered with the maximum sap flow group (f).These results indicated a high correlation between vascular bundles in the inner zone and stem flow.We accordingly speculated that the vascular bundle number in the inner zone is an important factor determining stem transport capacity.

(2) Stem length vs.SF indicators.

Here we identified a relationship between stem length (P_Len)and sap flow.The trait of stem length showed a fragile relationship with SFR_mean and SFA_mean indicator(PCCs were only 0.18 and 0.03,respectively),indicating that we can hardly infer the SFR and SFA of the plant from plant height.However,P_Len was significantly correlated with SFR_min and SFA_min (their PCCs reached 0.6 and 0.67,respectively).The minimum SFR per hour for the plant and the minimum SFA per day for the plant represented the sap flow of the maize when water transport is not active.The results suggested that stem length might determine the minimum sap flow requirement of the plant.

(3) 2D size-related traits vs.SF indicators.

With respect to two-dimensional traits,the mean area of stem internodes (SZ_A_mean) and mean area of the vascular bundle(VB_A_mean) were significantly correlated with SFA_mean and SFR_mean.It seemed that maize plants’ maximum sap flow could be inferred from stem and vascular bundle area traits (Fig.8c).Moreover,the correlation coefficient between SFR_mean and VB_A_mean was the highest,which means that the size of the vascular bundle is an essential factor affecting transport efficiency.Further combined with the above results,the number and area of vascular bundles of the stem,in particular the bundles in the inner zone,had important reference significance for evaluating the water transport efficiency of the plant.

(4) 3D size-related traits vs.SF indicators.

After introducing stem height,we calculated a series of threedimensional traits.These traits were highly correlated with each other (Fig.8d),and significantly correlated with size-related traits(Fig.8c).Among them,the correlation coefficients of P_IZ_V and SFA_mean were the largest (PCC=0.67).P_VB_V and SFA_mean also showed a high correlation (PCC=0.57).Compared with the 2D size-related traits,the correlation between sap flow performance and the 3D size-related traits was weaker.

The response of crops to drought stress is a comprehensive and complex process,with corresponding changes at cell,tissue,organ,individual,and population levels.The adaptive phenotypes are also reflected in the morphological structure,physiological function,metabolic process,and gene and protein expression.Phenotypes increased in dimensionality and scales will help to comprehend the physiological traits better [34].Early studies [35,36] have shown that diverse morphological structures and water transport supply strategies formed by plants to adapt to the needs of survival competition,such as changes in vascular bundle morphological characteristics.Previous observations have shown that vascular bundle area,number,etc.influenced water transportation.In this study,multiscale traits of vascular bundles were evaluated and provided a unique insight into the water transport and use efficiency of the maize plant.

4.Conclusions

We presented a deep learning-integrated phenotyping pipeline to quantify vascular bundles of maize stems at the single-plant level.The pipeline automatically extracts traits from CT images of multiple stem internodes of diverse varieties and growth environments.The counting accuracy of vascular bundles achieved R2of 0.997,and the measured accuracy of size-related traits was R2>0.98.The processing time of a single CT image was about 3 s.Five types of traits: quality-,quantity-,size-,shape-and descriptive-related traits,were evaluated at the single-plant scale and two growth environments,and multiscale traits of vascular bundles from the slice to the whole-plant level were also calculated to evaluate the relationship with sap flow indicators.The phenotyping pipeline may allow investigation the relationship between sap flow and traits of vascular bundles.

CRediT authorship contribution statement

Jianjun Du:Conceptualization,Investigation,Methodology,Software,Writing -original draft.Ying Zhang:Conceptualization,Investigation,Resources,Validation,Writing -original draft.Xianju Lu:Data curation,Resources.Minggang Zhang:Data curation,Resources.Jinglu Wang:Data curation,Resources.Shengjin Liao:Data curation,Resources.Xinyu Guo:Funding acquisition,Project administration,Writing -review &editing.Chunjiang Zhao:Funding acquisition,Project administration,Writing -review &editing.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This study was supported by the Construction of Collaborative Innovation Center of Beijing Academy of Agriculture and Forestry Science(KJCX201917),Beijing Academy of Agriculture and Forestry Sciences Grants (QNJJ202124),the National Natural Science Foundation of China (31801254 and U21A20205),and Beijing Natural Science Foundation (5202018).

Appendix A.Supplementary data

Supplementary data for this article can be found online at https://doi.org/10.1016/j.cj.2022.04.012.