A high-throughput root in-situ imaging device and image analysis method

Abstract

This invention discloses a high-throughput in-situ root imaging device and image analysis method. The imaging device includes an imaging tower frame, which consists of an upper tower and a lower tower. An imaging unit is installed in the upper tower, and a culture medium plate positioning groove is located at the bottom of the lower tower. An adjustable-brightness ring LED lighting system is installed below the lower tower. The image analysis method includes: Step 1, culturing plant roots; Step 2, acquiring high-resolution images of plant roots; Step 3, segmenting the acquired images to obtain the root RoL region; Step 4, image recognition, identifying root branches and bifurcation points based on a deep learning algorithm; and Step 5, extracting and quantifying phenotypic traits of the plant roots. This invention enables high-throughput, standardized, and high-quality imaging of culture dishes, providing strong phenotypic data support for plant science research.

Description

Technical Field

[0001] This invention relates to the field of plant phenotypic image acquisition and processing, specifically a high-throughput root in-situ imaging device and image analysis method. Background Technology

[0002] Plant roots, as key organs for absorbing water and nutrients, anchoring the plant, and participating in various physiological and biochemical processes, have important phenotypic indicators reflecting a plant's physiological state, stress resistance, and overall health, including their morphological structure and dynamic growth characteristics (i.e., root architecture). High-throughput and precise acquisition of root phenotypic information is crucial for plant genetics and breeding, functional genomics, nutritional physiology, stress resistance research, and precision agriculture.

[0003] In recent years, the rapid advancements in plant phenomics have greatly propelled botanical research, particularly in the study of plant root morphology based on image analysis technology, which has become a powerful force in the field. Image analysis technology not only enables high-throughput data processing, ensuring the accuracy and non-destructive nature of research, but also facilitates integration and expansion with other technology platforms, playing a crucial role in the discovery of key plant phenotypic traits. In resource-constrained agricultural ecosystems, optimizing crop root structure, as a key strategy for increasing yield, is attracting researchers' attention. However, revealing the genetic basis and developmental mechanisms of root architecture faces numerous challenges, requiring a multidisciplinary and diversified research strategy that closely integrates field observations, greenhouse controlled experiments, and precise laboratory analysis. Although field studies can directly reflect the actual growth of plants in their natural environment, their inherent limitations, such as the destructive nature of root excavation, the uncontrollable and variable nature of environmental factors, and the heterogeneity of soil, severely restrict the acquisition of accurate data on the genetic characteristics and developmental features of root architecture. Petri dish culture (such as agar plate culture) has become a commonly used standardized method for plant (especially seedling) root research due to its advantages such as controllable environmental conditions (light, temperature, humidity, nutrients), relatively simple operation, ease of aseptic operation, and relatively clear visibility of roots on a two-dimensional plane. This culture method is particularly suitable for early root development studies of model plants (such as Arabidopsis thaliana and rice seedlings) or crops such as forest trees and vegetables, providing an ideal platform for high-throughput phenotypic analysis. However, existing techniques for imaging and analyzing roots in petri dishes still have significant limitations.

[0004] 1. Imaging Device Adaptability and Efficiency Issues: While conventional flatbed scanners or ordinary cameras can be used to image petri dishes, they often lack specially designed fixing devices for the size, shape (especially circular), and transparency of the dishes. Manual placement and positioning of petri dishes is inefficient and cannot meet high-throughput requirements. 2. Limited Image Quality: The contrast between roots and transparent culture media (such as agar) is usually low. Ordinary equipment is insufficient in terms of light uniformity and background suppression, easily leading to blurred root edges, loss of details (such as fine roots and root hairs), or interference from reflections from the dish lid and water vapor condensation. 3. Lack of Automation: Existing methods mostly rely on manual operation to image one or a small number of petri dishes at a time, making it difficult to achieve automated, continuous, and standardized imaging of large batches of petri dishes, becoming a bottleneck for phenomics research. 4. Insufficient Support for Transmission Imaging: Imaging using transmitted light from the bottom of the petri dish (which can significantly improve the contrast between roots and background) requires a specialized bottom light source and optical path design, which general-purpose equipment usually lacks or does not achieve good results.

[0005] 2. Image Analysis Accuracy and Automation Challenges: Background Segmentation Difficulty: Accurately segmenting complex root networks from transparent or semi-transparent culture medium backgrounds, especially when the root network is similar in color to the medium or contains water stains or impurities, presents a significant challenge. Insufficient Complex Structure Resolution: For root systems with complex topologies, such as numerous lateral roots, existing methods struggle to comprehensively and accurately quantify diverse morphological and structural parameters such as total length, projected area, average diameter, number of root tips, and angles. Low Automation Level: Semi-automated image analysis methods like IMAGEJ still require significant manual intervention in many analysis steps (e.g., manual annotation and parameter adjustment), resulting in slow analysis speed, high subjectivity, and difficulty meeting the speed and consistency requirements of high-throughput phenotypic analysis.

[0006] Therefore, there is an urgent need to develop a highly adaptable, integrated, dedicated root imaging device and a matching automated, high-precision image analysis method for in situ culture in petri dishes. Summary of the Invention

[0007] The purpose of this invention is to provide a high-throughput in-situ root imaging device and image analysis method to achieve high-throughput, standardized, and high-quality (especially high-contrast) imaging of root culture, and to overcome complex root backgrounds to achieve rapid, accurate, and comprehensive extraction of root morphology and configuration parameters, providing strong phenotypic data support for plant science research.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A high-throughput in-situ root imaging device includes a 3D-printed lightweight modular imaging tower frame, which consists of an upper tower and a lower tower. The upper and lower towers are detachably assembled together. An imaging unit is installed in the upper tower and is controlled by a Raspberry Pi controller. The bottom of the lower tower is provided with a culture medium plate positioning groove, and an adjustable-brightness ring LED lighting system is installed below the lower tower.

[0010] Furthermore, the imaging unit includes a high-sensitivity CMOS imaging sensor and a high-definition telephoto lens (5), and the imaging resolution is not less than 5μm / pixel.

[0011] Furthermore, the 3D-printed imaging tower frame is made of polylactic acid and uses an upper and lower combined structure to achieve light-shielding imaging.

[0012] A root system image analysis method based on the imaging device includes the following steps:

[0013] Step 1: Cultivate plant roots in a colloidal medium;

[0014] Step 2: Place the plant to be tested in the positioning slot of the culture medium plate, control the imaging unit to take pictures through the Raspberry Pi controller, and use the ring LED lighting system for supplemental lighting to obtain high-definition images of the plant roots;

[0015] Step 3: Segment the acquired image to obtain the root region (Rol region);

[0016] Step 4, Image recognition: Identify the branches and forks of the root system based on deep learning algorithms;

[0017] Step 5: Extract phenotypic traits from plant roots to quantify phenotypic traits.

[0018] Furthermore, in step 1, the plant root system is cultured in a colloidal medium. The specific steps are as follows: the plant seeds are soaked in a 70% ethanol solution for 1 minute, then soaked in a 15% sodium hypochlorite solution for 30 minutes, and the seeds are washed with sterile deionized water at least five times to remove the sodium hypochlorite. The plant seeds are then germinated in a vertically oriented sterile culture dish in the dark at 30°C. When the newly emerging radicle reaches a length of about 1 cm, the sterile seedlings are transplanted into a transparent growth medium in a square culture dish for cultivation. The growth medium consists of ½ Kimura B nutrient solution with a pH of 5.8 and 1% agar powder.

[0019] Furthermore, in step 2, the control program set by the Raspberry Pi controller is: image capture time interval of 3 seconds / time, and the captured images are named with timestamps and stored locally.

[0020] Furthermore, in step 3, the specific steps of image segmentation include: converting the root system image into a grayscale image, using Otsu's method to obtain a binarized image of the root system, i.e., the image background is black and the root system part is white; using digital morphology and connected component algorithms to filter out small connected components and retain the root system part.

[0021] Furthermore, in step 4, the YOLOv5 algorithm architecture is used, and a lightweight GHOST-YOLOv5 model is constructed to identify apical and bifurcation points; the Ghost Bottleneck module is used to replace the Bottlenecks module in the original YOLOv5 C3 module to achieve model lightweighting; at the same time, a multi-scale attention mechanism module is added after the three feature output layers P1, P2 and P3 of the model backbone feature extraction network. This multi-scale attention mechanism module retains the information of each channel and reduces the computational cost by reshaping some channels into batch dimensions and grouping the channel dimensions into sub-features. For any given input feature map, it is grouped into G sub-features according to the channel dimensions to learn various semantic features.

[0022] Furthermore, in step 4, the multi-scale attention mechanism module uses three parallel paths to capture the attention weights of a set of internal feature maps. Two paths are set on the 1×1 convolution branch, and the other path is located in the 3×3 convolution branch. In the 1×1 convolution branch, global average pooling is applied along the X and Y spatial directions respectively to encode the channel information. In the 3×3 convolution branch, a strategy of stacking single 3×3 convolution kernels is adopted. At the same time, the multi-scale attention mechanism module introduces two different sizes of convolution branches, 1×1 and 3×3, and uses global average pooling on the output of the 1×1 branch to encode the global spatial information, thereby realizing cross-space aggregation of information in the X and Y spatial directions.

[0023] Furthermore, in step 5, the phenotypic traits include basic root morphology and branching traits; the basic root morphology traits are calculated by measuring the overall outline or skeleton, including maximum depth, perimeter, surface area, maximum root length, length-to-width ratio, longest width, average diameter, and convex hull area; the root branching traits include the number of branches and the included angle.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] (1) Imaging efficiency is greatly improved. Through precise timing scheduling of the Raspberry Pi controller and the coordination of pulsed LED lighting, the dynamic process of the root system can be captured at the millisecond level. The efficiency is more than 10 times higher than that of traditional manual sampling (≥30 minutes / time) or low-speed imaging equipment (≥10 seconds / time), which completely solves the problem of loss of growth dynamic information caused by excessively long sampling interval.

[0026] (2) Automated analysis of multidimensional phenotypic parameters, integrating skeleton topology analysis (branch angle quantification) and contour geometry algorithm (convex hull area), automatically extracts 8 basic morphological items + 2 branch traits in two-dimensional images, breaking through the limitation of existing technologies that can only measure basic parameters such as length and area, and providing multidimensional data support for root system architecture research.

[0027] (3) Cost and universality: The modular design of the device, which uses a Raspberry Pi controller and a 3D-printed customized shell, keeps the cost of the entire system below 1 / 10 of that of traditional commercial root imaging systems. It also supports rapid adjustment of the imaging tower structure according to the size of the petri dish and is compatible with multiple crops such as Arabidopsis thaliana, rice, and corn, significantly reducing the research threshold. These advantages make the device have broad application prospects in plant science, agricultural research and other fields. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the device of the present invention;

[0029] Figure 2 This is a schematic diagram of root cultivation according to the present invention;

[0030] Figure 3 This is an image processing diagram of the present invention.

[0031] In the diagram: 1. M2 screw; 2. CMOS sensor; 3. M2 nut; 4. Upper part of the imaging tower; 5. High-resolution telephoto lens; 6. Lower part of the imaging tower; 7. Raspberry Pi controller; 8. LED ring lighting system; 9. Culture medium plate positioning slot. Detailed Implementation

[0032] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0033] Example 1

[0034] like Figure 1 A high-throughput in-situ root imaging device includes a 3D-printed lightweight modular imaging tower frame, an imaging unit, a Raspberry Pi microcontroller, and a ring LED illumination system. The imaging tower frame consists of a detachable upper imaging tower body 4 and a lower imaging tower body 6. A Raspberry Pi controller 7 serves as the core controller, integrating autofocus and timing control functions. The imaging unit consists of a high-sensitivity CMOS imaging sensor 2 and a high-definition telephoto lens 5, with an imaging resolution of no less than 5μm / pixel. The programmable ring LED illumination system 8 has an adjustable brightness range of 0-3000 lux. This device enables high-speed, non-invasive, high-definition imaging of roots on culture medium plates at 3 seconds per scan, and the imaging effect can effectively replace traditional microscopic imaging systems.

[0035] In this embodiment, the 3D printed frame is made of polylactic acid (PLA) material, and an upper and lower combined imaging tower is used to achieve light-shielding imaging; the upper tower body 4 and the lower tower body 6 of the imaging tower are connected by a bridge structure to achieve rapid opening and closing.

[0036] In this embodiment, the upper tower body integrates a CMOS sensor 2, a high-resolution telephoto lens 5, and a Raspberry controller 7. The sensor is connected to the upper tower body by M2 screws and nuts. The lower tower body is provided with a culture medium plate positioning groove 10 and an LED ring lighting system 8.

[0037] Example 2

[0038] A high-throughput in-situ root imaging device and image analysis method are described below:

[0039] S1. Plant Rooting Culture: First, the rice seeds are surface-sterilized by sequentially treating them with 70% ethanol for 1 minute, soaking them in 15% sodium hypochlorite for 15 minutes, and then rinsing them 5 times with sterile deionized water to remove residues. The sterilized seeds are then placed in a 30℃ dark environment and vertically germinated in 13cm×13cm sterile square petri dishes. When the radicle reaches 1cm, the seedlings are transplanted into petri dishes containing a specific growth medium.

[0040] Preferably, the medium consists of ½ Kimura B nutrient solution at pH 5.8 (containing 11 components including 0.091 mM KH₂PO₄ and 0.273 mM MgSO₄·7H₂O) and 1% agar solidification agent (10 g agar / 1 L nutrient solution, autoclaved). This substrate has been solidified with agar. The agar growth medium is prepared by dissolving 10 g of agar powder in 1 ½ L of Kimura B nutrient solution and autoclaving. ½ Kimura B nutrient solution contains the following components: (0.091 mM KH2PO4, 0.273 mM MgSO4·7H2O, 0.182 mM (NH4)2SO4, 0.091 mM KNO3, 0.183 mM Ca(NO3)2·4H2O, 0.003 mM M3BO3, 0.0005 mM MnCl2·4H2O, 0.001 mM (NH4)6Mo7O24·4H2O, 0.0004 mM ZnSO4·7H2O, 0.0002 mM CuSO4·5H2O, 0.02 mM Fe(III)-EDTA).

[0041] S2. Automated Image Acquisition: After 7 days of cultivation, the culture dish containing the plant is placed in the positioning slot of the imaging tower, and the root system image is automatically acquired by the camera controlled by the Raspberry Pi. The imaging device consists of four main modules: (1) Structural module: The upper and lower tower bodies of the imaging tower are made of 3D printed PLA material and are connected by a bridge to achieve rapid opening and closing; (2) Imaging module: The CMOS sensor is fixed to the upper tower body with M21 screws and connected to a high-resolution telephoto lens by thread; (3) Control module: The Raspberry Pi controller is built into the hollow shell of the upper tower body, runs a timed capture program (capture a single frame every 3 seconds) and file management (timestamp naming storage); (4) Illumination module: The lower tower body 6 integrates an LED ring lighting system 8, which is triggered by GPIO to provide pulsed synchronous lighting.

[0042] The Raspberry Pi controller's control program includes the following core functional modules: an image acquisition driver module, which directly drives the CMOS sensor and high-resolution telephoto lens combination by calling the optimized camera library picamer2; a high-speed acquisition control module, configured to automatically trigger single-frame high-definition still image capture at precise 3-second fixed time intervals, ensuring that focusing, exposure adjustment, and image sensor data reading can be completed quickly at the predetermined time point; an image storage and management module, which automatically saves each captured high-definition image frame with a unique filename with a timestamp (such as YYYYMMDD_HHMMSS.jpg) to the Raspberry Pi's local storage microSD card or a preset network storage location; and an illumination synchronization control module, which precisely controls the on and off of the LED ring lighting system (pulsed lighting) through the Raspberry Pi's GPIO pins at the moment of each image acquisition trigger.

[0043] S3. Image segmentation processing: Perform four steps on the acquired root system RGB image: (1) Extract the ROI of the root system colloidal region; (2) Apply Otsu's method to obtain a binarized image (black background / white root system); (3) Filter out noisy small connected regions through the connected component algorithm; (4) Extract the root system foreground based on morphological operations.

[0044] S4. Image recognition: Identifying root tips and branches based on deep learning algorithms.

[0045] We improved upon the YOLOv5s algorithm architecture, building a lightweight GHOST-YOLOv5 model to identify apical and bifurcation points. The YOLO series of algorithms boasts advantages such as strong real-time performance, high accuracy, and versatility, making it widely used in dense object detection. The YOLOv5s model contains multiple Conv modules and more convolutional layers in its overall network structure, meaning a larger number of parameters and computational weights. The Ghost module, however, can reduce the number of parameters and computational weights without changing the dimension and channel size of the convolutional output feature map. Therefore, we use the Ghost Bottleneck module to replace the Bottlenecks module in the original YOLOv5's C3 module to achieve model lightweighting. Furthermore, to further enhance the model's feature extraction capabilities and improve detection accuracy while reducing the number of parameters in the model, we added a multi-scale attention mechanism (EMA) module after the three feature output layers P1, P2, and P3 of the model's backbone feature extraction network. This module preserves information from each channel and reduces computational overhead by reshaping some channels into batch dimensions and grouping channel dimensions into sub-features. For any given input feature map X∈RC×H×W, X is grouped into G sub-features according to the channel dimensions to learn various semantic features. This can be represented by the equation:

[0046]

[0047]

[0048] EMA employs three parallel paths to capture attention weights from a set of intra-feature maps. Two paths are located on a 1×1 convolutional branch, and the third path is located in a 3×3 convolutional branch. In the 1×1 convolutional branch, global average pooling is applied along the X and Y spatial directions, respectively (see Equation 2), which encodes channel information. On the other hand, in the 3×3 convolutional branch, instead of traditional GN normalization and average pooling operations, this design uses a strategy of stacking individual 3×3 convolutional kernels, significantly improving multi-scale feature extraction capabilities. Furthermore, the EMA module achieves cross-spatial aggregation of information in the X and Y spatial directions by introducing two different sized convolutional branches, 1×1 and 3×3, and encoding global spatial information using global average pooling on the output of the 1×1 branch. This design enables EMA to effectively integrate richer and more diverse features, thereby improving model performance and generalization ability.

[0049] S4. Phenotypic trait quantification: Two types of traits are extracted using the OpenCV-Python algorithm library: basic morphological features, including maximum depth (difference between the Y coordinate of the lowest point of the contour and the root neck reference line), perimeter (cumulative sum of contour chain codes), surface area (total number of pixels in the contour), convex hull area (solved by Graham scan algorithm), maximum width (maximum span of non-zero pixels in the X direction), and average diameter (weighted mean based on skeletonized Euclidean distance transformation).

[0050] Branch topology features: count the number of effective branches (based on deep learning recognition), and calculate the branch angle (the system mean is obtained by least squares fitting of the branch point direction vector and inverse cosine).

[0051] Preferably, the basic morphological characteristics of the root system are obtained by acquiring the binarized root system and then calculating the geometric features of the contour. The maximum depth is the difference between the Y coordinate of the lowest point of the contour and the root neck reference line; the perimeter is the sum of the contour chain codes; the surface area is the total number of pixels within the contour; the convex hull area is solved using the Graham scan algorithm; the maximum width is the maximum span of non-zero pixels in the X direction; the average diameter is based on the skeletonization result, calculated point-by-point along the skeleton using the normal distance (Euclidean distance transformation) and then weighted by the skeleton length. The root branching characteristics described in S5 are as follows: the number of branches is calculated using a deep learning recognition algorithm; the branch angle is calculated by extracting the direction vectors of the 10px segment at the far end of the parent root and child root at the branch point (least square fitting), calculating the inverse cosine value, and then obtaining the system mean.

[0052] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the scope of protection of the present invention in any way, and all technical solutions obtained by equivalent substitution or other means fall within the scope of protection of the present invention. Parts not covered in this invention are the same as or can be implemented using existing technology.

Claims

1. A high-throughput in-situ root imaging device, characterized in that, The invention includes a 3D-printed lightweight modular imaging tower frame, which consists of an upper tower body and a lower tower body that can be detachably assembled together. An imaging unit is installed in the upper tower body and is controlled by a Raspberry Pi controller. The bottom of the lower tower body is provided with a culture medium plate positioning groove, and an adjustable brightness ring LED lighting system is installed below the lower tower body.

2. The high-throughput in-situ root imaging device according to claim 1, characterized in that, The imaging unit includes a high-sensitivity CMOS imaging sensor and a high-definition telephoto lens (5), and the imaging resolution is not less than 5μm / pixel.

3. A high-throughput in-situ root imaging device according to claim 1, characterized in that, The 3D-printed imaging tower frame is made of polylactic acid and uses an upper and lower combined structure to achieve light-shielding imaging.

4. A root system image analysis method based on the imaging device of claim 1, characterized in that, Includes the following steps: Step 1: Cultivate plant roots in a colloidal medium; Step 2: Place the plant to be tested in the positioning slot of the culture medium plate, control the imaging unit to take pictures through the Raspberry Pi controller, and use the ring LED lighting system for supplemental lighting to obtain high-definition images of the plant roots; Step 3: Segment the acquired image to obtain the root region (Rol region); Step 4, Image recognition: Identify the branches and forks of the root system based on deep learning algorithms; Step 5: Extract phenotypic traits from plant roots to quantify phenotypic traits.

5. The root system image analysis method according to claim 4, characterized in that, In step 1, the plant roots are cultured in a colloidal medium. The specific steps are as follows: the plant seeds are soaked in a 70% ethanol solution for 1 minute, then soaked in a 15% sodium hypochlorite solution for 30 minutes, and the seeds are washed with sterile deionized water at least five times to remove the sodium hypochlorite. The plant seeds are then germinated in a vertically oriented sterile culture dish in the dark at 30°C. When the newly emerging radicle reaches a length of about 1 cm, the sterile seedlings are transplanted into a transparent growth medium in a square culture dish for cultivation. The growth medium consists of ½ Kimura B nutrient solution with a pH of 5.8 and 1.5% agar powder.

6. The root system image analysis method according to claim 4, characterized in that, In step 2, the control program set by the Raspberry Pi controller is: the image capture time interval is 3 seconds / time, and the captured images are named with timestamps and stored locally.

7. The root system image analysis method according to claim 4, characterized in that, In step 3, the specific steps of image segmentation include: converting the root system image into a grayscale image, using Otsu's method to obtain a binarized image of the root system, i.e., the image background is black and the root system part is white; using digital morphology and connected component algorithms to filter out small connected components and retain the root system part.

8. The root system image analysis method according to claim 4, characterized in that, In step 4, the YOLOv5 algorithm architecture is used, and a lightweight GHOST-YOLOv5 model is built to identify apex and branch points. Improvements are made to the original algorithm by replacing the Bottlenecks module in the original YOLOv5 C3 module with the Ghost Bottleneck module to achieve model lightweighting. Simultaneously, a multi-scale attention mechanism module is added after the three feature output layers P1, P2, and P3 of the model's backbone feature extraction network. This multi-scale attention mechanism module preserves information from each channel and reduces computational overhead by reshaping some channels into batch dimensions and grouping channel dimensions into sub-features. For any given input feature map, it is grouped into G sub-features according to channel dimensions to learn various semantic features.

9. The root system image analysis method according to claim 8, characterized in that, In step 4, the multi-scale attention mechanism module uses three parallel paths to capture the attention weights of a set of internal feature maps. Two paths are set on the 1×1 convolution branch, and the other path is located in the 3×3 convolution branch. In the 1×1 convolution branch, global average pooling is applied along the X and Y spatial directions respectively to encode the channel information. In the 3×3 convolution branch, a strategy of stacking single 3×3 convolution kernels is adopted. At the same time, the multi-scale attention mechanism module introduces two different sizes of convolution branches, 1×1 and 3×3, and uses global average pooling on the output of the 1×1 branch to encode the global spatial information, thereby realizing cross-space aggregation of information in the X and Y spatial directions.

10. The root system image analysis method according to claim 4, characterized in that, In step 5, phenotypic traits include basic root morphology and branching traits; basic root morphology traits are calculated by measuring the overall outline or skeleton, including maximum depth, perimeter, surface area, maximum root length, length-to-width ratio, longest width, average diameter, and convex hull area; root branching traits include the number of branches and the included angle.

Images