Deep learning-based woody plant vascular tissue cell attribute statistical method

Abstract

The invention discloses a woody plant vascular tissue cell attribute statistical method based on deep learning, and the method comprises the following steps: obtaining a plant material, and obtaining a plant vascular tissue image through processing; performing format conversion and labeling on the plant vascular tissue image to obtain a labeling file; after format conversion is conducted on the annotation file again, the annotation file is input into YOLOv8 models of different sizes to be trained; and respectively slicing different types of cells in different slicing modes, photographing by adopting different photographing multiples, inputting photographed pictures into the optimized YOLOv8 models with different sizes for detection and analysis, obtaining the optimal YOLOv8 model under different slicing modes and different photographing multiples, and carrying out statistics on plant vascular tissue cell attributes. The data set manufactured by the method can be used for model training, the YOLOv8 model obtained by training can identify different types of xylem cells in image data, and accurate statistics of cell attributes is facilitated.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning and woody plant cell property statistics, and in particular relates to a method for woody plant vascular tissue cell property statistics based on deep learning. Background Art

[0002] Microscopic image analysis of woody plant vascular tissues, such as vessels, rays, and fibers, is of great significance in plant physiology and wood science. Traditional analysis methods rely on manual counting and measurement under a microscope, which is inefficient and susceptible to human error. This is especially true when processing large-scale images, where it is difficult to balance accuracy and speed. While existing partially automated image analysis tools can replace manual operations, they lack accuracy when used with different cell types, often leading to false or missed detections.

[0003] With the development of deep learning technology, the YOLO (You Only Look Once) series of models has achieved remarkable performance in object detection, enabling rapid and accurate identification of objects in images. However, there is a lack of excellent slice images and annotated datasets for woody plant vascular tissue. Currently, there is no deep learning annotated dataset for woody plant vascular tissue cells.

[0004] Therefore, it is of great practical significance to annotate a dataset for woody plant vascular tissue and develop an automated detection method based on YOLO to meet the needs of large-scale cell analysis in the fields of biology and forestry. Summary of the Invention

[0005] To solve the above technical problems, the present invention proposes a statistical method for the properties of woody plant vascular tissue cells based on deep learning, provides a production process based on woody plant vascular tissue datasets, and trains the best detection model for different types of woody plant cells and under different magnifications based on the YOLOv8 deep learning model. It can quickly and accurately quantify the properties of different types of xylem cells in image data, helping researchers to conveniently, quickly and accurately count the various properties of different types of xylem cells in woody plants, so as to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above objectives, the present invention provides a method for statistically analyzing the properties of vascular tissue cells in woody plants based on deep learning, comprising the following steps:

[0007] Obtain plant materials and process them to obtain images of plant vascular tissues;

[0008] performing format conversion and annotation on the plant vascular tissue image to obtain an annotation file;

[0009] After the format of the annotation file is converted again, it is input into the YOLOv8 model of different sizes for training to obtain optimized YOLOv8 models of different sizes;

[0010] Different cell types were sliced ​​in different ways and photographed at different magnifications. The photographs were then fed into optimized YOLOv8 models of different sizes for detection and analysis, yielding the optimal YOLOv8 models for different slicing methods and magnifications.

[0011] The properties of plant vascular tissue cells are statistically analyzed based on the optimal YOLOv8 model under different slicing methods and different shooting magnifications.

[0012] Optionally, the process of obtaining plant material and processing it to obtain an image of plant vascular tissue includes:

[0013] Plant materials were sliced ​​using a fully automatic vibrating microtome. The intact tissue was placed on a glass slide and stained with toluidine blue to obtain chromogenic cell sections. The chromogenic cell sections at different angles and magnifications were photographed, and the obtained slice images were named and stored in ascending order of the photographing sequence as plant vascular tissue images.

[0014] Optionally, the process of converting the format of the plant vascular tissue image includes:

[0015] The plant vascular tissue images are saved in tiff format. The plant vascular tissue images are classified into different folders based on different shooting targets to obtain several target cell detection folders; the target cell detection folders are format converted using python code to obtain target cell detection folders in jpg format.

[0016] Optionally, the process of labeling the target cell detection folder in jpg format includes:

[0017] Based on the Labelme annotation tool and the Segment-Anything model, the lumens of different types of cells in the target cell detection folder are circled, and the cell types and distinction codes are marked; among them, cells with unclear outlines are manually annotated by creating polygons.

[0018] Optionally, after the format of the annotation file is converted again, before inputting it into a YOLOv8 model of different sizes for training, the following steps are included:

[0019] A preset random seed is used to traverse the annotation file, the frequency of occurrence of each label is counted and sorted, and a configuration file in YOLO format is created based on the sorting result, and the image paths and labels of the training set and test set are defined; the annotation file is shuffled according to the preset random seed, and divided into training set and test set according to a preset ratio, and each annotation file in the training set and test set is converted into YOLO format and saved in the corresponding configuration file based on the defined image path and label.

[0020] Optionally, the training set and the test set are input into YOLOv8 models of different sizes for training, and the process of obtaining optimized YOLOv8 models of different sizes includes:

[0021] Set the initial parameters of the YOLOv8 model network model of different sizes;

[0022] Image enhancement is performed on the annotated images and annotated files in the training set and test set. The enhanced training set and test set are input into YOLOv8 models of different sizes for training until the loss function converges, obtaining optimized YOLOv8 models of different sizes.

[0023] Optionally, the initial parameters include at least image size, batch size, initial learning rate, intersection-over-union ratio, confidence threshold, number of categories, category name, and YOLO format dataset file paths for different types of cells.

[0024] Optionally, the process of statistically analyzing plant vascular tissue cell attributes based on the optimal YOLOv8 model includes:

[0025] Based on different slicing methods and the optimal YOLOv8 model under different shooting magnifications, different types of slice images are predicted to obtain prediction results;

[0026] Sort and number the prediction boxes of the prediction results or mark them with asterisks;

[0027] Based on the properties of the predicted box and the detection object, the cell properties of each predicted box are obtained and counted; wherein the properties include cell diameter, area, length and width, etc.

[0028] Compared with the prior art, the present invention has the following advantages and technical effects:

[0029] Compared with the existing technology, the plant xylem cell attribute statistical method proposed in the present invention can extract image slices of plant xylem cells and convert them into a data set readable by the YOLOv8 network model through data conversion and annotation. This data set can be directly used to train the plant xylem cell detection model.

[0030] The present invention uses the optimal YOLOv8 model trained in different slicing methods and at different shooting magnifications to quickly and accurately identify different types of xylem cells in image data, helping researchers to conveniently, quickly, and accurately count the length, width, area, and other attributes of cells.

[0031] The method of the present invention does not require significant adjustments for different magnifications or slice angles, and can consistently and reliably provide accurate cell detection results. Its wide applicability will bring great convenience and value to researchers and professionals, both in scientific research and practical applications.

[0032] The present invention uses the YOLOv8 network model to detect plant cells. The model has excellent performance, fast detection speed, is easy to deploy quickly, and can effectively detect small and dense target cells. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0034] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0035] Figure 2 These are slices at different angles and magnifications of the embodiment of the present invention;

[0036] Figure 3 Schematic diagrams of cell labeling in an embodiment of the present invention, wherein (a) is a labeled diagram of ductal cells in a cross section, (b) is a labeled diagram of fiber cells in a cross section, (c) is a labeled diagram of ray cells in a cross section, (d) is a labeled diagram of ray cell types in a chord section, and (e) is a labeled diagram of ray cell numbers in a chord section;

[0037] Figure 4 Schematic diagram of chord ray cell number prediction according to an embodiment of the present invention. (a) is a predicted mapping diagram of the number of chord ray cells, chord ray cell numbers, and the number of chord ray columns. (b) is a mapping diagram of the number of different types of chord ray cells and the predicted labels of different types of ray cells.

[0038] Figure 5 Schematic diagram of cross-section cell prediction according to an embodiment of the present invention, (a) is the original cross-section slice image, (b) is the predicted image of the cross-section ductal cells, (c) is the predicted image of the cross-section fiber cells, and (d) is the predicted image of the cross-section ray cells. DETAILED DESCRIPTION

[0039] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0040] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0041] Example 1

[0042] like Figure 1 As shown, this embodiment takes Poplar 717 as an example to provide a method for statistically analyzing the properties of vascular tissue cells in woody plants based on deep learning, including the following steps:

[0043] S1. Using a fully automatic vibrating microtome, slice, stain, and photograph the plant material to obtain vascular tissue cross-section images. Name each cross-section image in ascending order according to the sequence in which the images were taken, and save them as plant vascular tissue images in tiff format.

[0044] Furthermore, the steps for obtaining the 717 poplar vascular tissue slices in this embodiment are:

[0045] First, fresh stem sections of Populus 717 were cut into approximately 0.5 cm high tissue blocks using a sharp blade and fixed to the microtome tray with glue. After setting the slicing program, the sections were sliced ​​using a fully automatic vibrating microtome. When relatively intact tissue was cut, soft tweezers were used to collect it onto a clean slide that had been pre-added with a drop of clean water to obtain the slide containing the sections.

[0046] Then, the water on the slide was aspirated and 0.02% toluidine blue stain was added with a rubber-tipped dropper for 23 seconds. Then, the slide was rinsed with distilled water, leaving only a small amount of distilled water on the slide. Finally, a coverslip was placed and excess water was removed along the slide with absorbent paper to obtain a fully stained cell section.

[0047] Then, use an upright fluorescence microscope and cellSens Standard software to take photos of the slices at different angles and magnifications, name each slice in ascending order according to the sequence of photos, and save it as a tiff format image, such as Figure 2 shown.

[0048] In this embodiment, the naming format for each slice image is "image" + photo sort number + ".tiff", such as "image0001.tiff", "image0002.tiff", and so on. The first slice image "image0001.tiff" will be used as a representative to illustrate the changes in image naming after each processing step.

[0049] The microtome used was a LEIC AVT1200S fully automatic oscillating microtome, a slicing tool designed to meet the high-quality sectioning requirements of neurophysiology, neuropathology, experimental pathology, botany (roots and plants), and industry (foam plastics). The fluorescence microscope used was an Olympus DP80 fluorescence microscope, which is widely used in various fields to capture high-quality images of tiny samples such as cells. In fluorescence microscopy, samples are treated with fluorophores, causing them to fluoresce when excited by a light source, generating high-resolution fluorescence images.

[0050] This example uses an automated vibratome and standardized staining and photography processes to produce a large, high-quality woody plant vascular tissue dataset in less time and cost.

[0051] S2. Convert the tiff format image to a jpg format image. Specifically:

[0052] Plant vascular tissue images saved in tiff format are classified into different folders according to different shooting targets to obtain high-quality and clear target cell detection folders; Python code is used for format conversion to batch convert the images in each target cell detection folder from tiff format to jpg format pre-processed images.

[0053] As a specific example, the slice image "image 0001.tiff" is format converted, and the tiff format images are uniformly converted into jpg format, so that the original image is reduced from 35.8MB to about 1MB, thereby reducing the computing resources required for model training and inference, speeding up the model training process, and at the same time, the image is removed from the Chinese name to obtain "0001.jpg".

[0054] S3. Perform AI+ manual annotation on the JPG format image to obtain the annotated image and JSON annotation file;

[0055] The manual annotation uses the Labelme annotation tool and imports the Segment-Anything model as an AI aid. When annotating, the lumen of different types of xylem cells are circled instead of the cell wall, and the cell type and distinction code are marked. When annotating the image dataset in jpg format, cells with unclear outlines are manually annotated by manually creating polygons. Finally, the corresponding JSON file is obtained, and cells of the same type are grouped into the same cell folder.

[0056] Furthermore, the steps for performing Ai+ manual labeling on “0001.jpg” are as follows:

[0057] First, download the Labelme image annotation software from the Labelme official website. The Labelme image annotation software is version 5.3.1.

[0058] Then download the six official .onnx format files of Segment-Anything in labelme-5.3.1\labelme\ai\__init__.py in the official code of Labelme-5.3.1, create a new folder: model_file under labelme-5.3.1\labelme, put the above six files in it, and modify the six official .onnx format file links of Segment-Anything in labelme-5.3.1\labelme\ai\__init__.py to the path of the above model_file file, so as to import the Segment-Anything model for auxiliary labeling;

[0059] Finally, import the "0001.jpg" image into the Labelme software, perform Ai-assisted annotation on cells with obvious outlines, and manually create polygon annotations for cells with unclear outlines. After annotating the cells corresponding to the entire image, the corresponding "0001.json" file is obtained, as shown in the following example: Figure 3 shown.

[0060] S4. Convert the JSON annotation files into YOLO format annotation txt files in batches, and randomly divide the annotated images and YOLO format annotation txt files into training sets and test sets according to the set ratio;

[0061] Furthermore, the steps for converting batches of JSON files into YOLO format and dividing them into training sets and test sets in proportion are as follows:

[0062] The program then sets a random seed to ensure consistent random results across runs. It then traverses the same cell folder, finding all JSON files, counting the number of occurrences of each label, such as the object category, and sorting the labels by frequency. Based on the generated label list (sorted by label frequency), the program creates the yolo.yaml configuration file required for the YOLO format, defining the image paths and labels for the training and validation sets. The program then shuffles the paths of all JSON files according to the random seed and divides the dataset into training and validation sets in an 8:2 ratio. Each JSON file in the training and validation sets is processed separately, converting the annotation information into YOLO format (normalized coordinates and label indexes). The corresponding image files are also found and copied to the YOLO output directory, placing them in different folders depending on the training or validation set. The converted YOLO label files are saved in the labels directory, and the image files are saved in the images directory. The training and validation sets are stored in the train and val folders, respectively, to obtain YOLO-format dataset files for different cell types.

[0063] As a specific example, first repeat the operation in S3 to obtain a large number of "XXXX.jpg" images and their corresponding "XXXX.json" file folders. In this example, each folder is named in the format of "cross-section / chord-section" + "cell type" +, such as "cross-section ray cells", "chord-section ray cells", and so on. The first slice folder "chord-section ray cells" will be used as a representative to illustrate the changes in the folders after each step of processing.

[0064] Then, using Python code, read the Labelme format annotation file and the corresponding image file in the "Chord-tangent ray cell" and set the random seed to ensure that the random results generated in each run are consistent, that is, repeatable. Then, traverse the "Chord-tangent ray cell" and find all "XXXX.json" files. Count the number of occurrences of each label, such as the cell category, and sort the labels by frequency. Based on the generated label list (sorted by label frequency), the program creates the yolo.yaml configuration file required by YOLO format and defines the image paths and labels of the training set and validation set.

[0065] Finally, the paths of all "XXXX.json" files are shuffled according to a random seed, and the dataset is divided into training and validation sets in an 8:2 ratio. Each "XXXX.json" file in the training and validation sets is processed separately, and the annotation information is converted to YOLO format. The normalized coordinates and label index are named "XXXX.txt". The corresponding image files are also found and copied to the YOLO output directory, that is, placed in different folders according to the training set or validation set. The converted YOLO label files are saved in the "labels" directory, and the image files are saved in the "images" directory. The training set and validation set are stored in the "train" and "val" folders, respectively, to obtain YOLO format dataset files for different cell types.

[0066] In this embodiment, 700 717 poplar xylem transverse slice images and 400 717 poplar xylem tangential slice images are selected as sample images. The sample images are further enlarged several times through random transformations such as random scaling, cropping, flipping, color jittering, Mosaic and MixUp operations.

[0067] S5. Input the training set and the test set into YOLOv8 models of different sizes for training. For each cell and each size of YOLOv8 model, when the accuracy is the highest and not overfitting, obtain the final detection model weight corresponding to the cell corresponding to the YOLOv8 model of different sizes;

[0068] Download five different sizes of YOLOv8 model weights and their source code from the YOLOv8 official website on GitHub, including YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x.

[0069] Set the initial parameters of different YOLOv8 network models; the initial parameters include image size, batch size, initial learning rate, intersection-over-union ratio, confidence threshold, number of classes, class name, and YOLO format dataset file paths for different types of cells;

[0070] Based on the official code, we modified different model paths and epochs, and performed image enhancement on the labeled images and annotation files for each type of cell. We then fed the enhanced training and test sets into YOLOv8 models of different sizes for training until the model triggered early stopping and the loss function converged or reached the preset accuracy. This yielded the best.pt file for each YOLOv8 model size.

[0071] As a specific embodiment, the training set and the test set are input into the YOLOv8 model of different sizes for training. For each cell, 3000 epoch training is set under the YOLOv8 model of each size (YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, YOLOv8x) until the model is early stopped. Finally, the final detection model weight corresponding to the YOLOv8 model of different sizes of the cell is obtained. In this embodiment, the naming format of each weight file image is "cell type English name-" + "cross / chord" + "-model abbreviation (n, s, m, l, x)" + ".pt", such as "ray-cross-n.pt", "ray-chord-n.pt", etc., and so on. The cross-sectional weight of the ray will be represented by "ray-cross-(n\s\m\l\x).pt" to characterize the changes in weight screening after each step of processing.

[0072] S6. For the actual application of each final detection model weight, the accuracy of each cell type, different magnifications, and different slice angles is screened out for each cell type under different magnifications and different slice angles for subsequent predictions. Figure 5 As shown;

[0073] Furthermore, in this embodiment, the steps for screening the most suitable ray models for different cell types are:

[0074] Using ImageJ software, we manually counted 10 slices of tangential ray cells (ray) using the tangential cutting method and different shooting magnifications as the validation set: we used the five different size model weight files of this type of cell trained to make predictions and obtain the predicted values; finally, we compared the error between the true value on the validation set and the model predicted value to screen out the optimal model weight for this type of cell under the same slicing method and different shooting magnifications, and used this weight for detection to obtain the results shown below. Figure 4 shown.

[0075] As an implementable method, first, we manually counted 10 images of ray cells in the tangential section at different magnifications using ImageJ software, and used this as a validation set to obtain the true number of ray cells in the tangential section. The software used was ImageJ version 2.9.0.

[0076] Then, the five weight files of different sizes of this type of cells are trained, ray-cross-(n\s\m\l\x).pt, to test the validation set and obtain the predicted value, such as Figure 4 As shown;

[0077] Finally, by comparing the true values ​​on the validation set and the errors between the predicted values ​​of different models, the optimal model weights on ray cells under different magnifications were screened out.

[0078] S7. Utilize the optimal model weights at different magnifications and slice angles, combined with Python code, to solve the detection results of different aspects such as the count and cell properties of different types of cells at different magnifications and slice angles;

[0079] Furthermore, in this example, the steps for obtaining the detection results of different aspects such as the count and cell properties of different types of cells at different slice angles are as follows:

[0080] For ray cells with ordered chord sections, the specific operations to obtain the number of cells, the number of ray columns, the number of ray cells in each column, the number of cells of each type, the length of each ray cell, the width of each ray cell, the area of ​​each ray cell, and the overall area of ​​ray cells are as follows:

[0081] a). For slice images with different magnifications (10X, 20X, 40X), different conversion units are used to convert pixels to real physical units;

[0082] b) To solve the problems of counting and sorting tangent ray cells and obtaining various cell properties, two empty DataFrames are created to save the detection results. One is the overall information table of cells in a single image, as shown in Table 1, and the other is the detailed information table of each cell in a single image, as shown in Table 2.

[0083] c) Use the most suitable YOLOv8 model for the selected magnification to perform detection. The data is stored as a Numpy array, including the coordinates (x1, y1, x2, y2), confidence level, and category ID of each detection box;

[0084] d) Sort the detection boxes by y1 coordinate and classify them by row. Introduce the rows variable to store the detection boxes for each row.

[0085] e) Determine whether the detection frames belong to the same row by comparing their y1 coordinate differences. Sort the detection frames in each row from left to right by x1, and display the number (row number - frame number) and the type of ray cell (class 1, class 2, class 3, class 4) in the center of the frame. Calculate the total number of all detection frames and plot this statistical information on the image. Also calculate the width and height (in micrometers) and area (width * height) of each detection frame and store this information in a DataFrame.

[0086] f). After processing each image, save the processed image to the output folder and save all detection data (DataFrame) to an Excel file;

[0087] Table 1

[0088]

[0089] Table 2

[0090] Sample_name Cell number Cell length (μm) Cell width (μm) <![CDATA[Cell area (μm 2 )]]> image0001.tif 1-1 11.09 9.89 109.78 image0001.tif 1-2 17.91 10.40 186.51 image0001.tif 1-3 11.26 8.70 98.02 image0001.tif 1-4 11.26 9.55 107.63 image0001.tif 1-5 16.04 12.28 197.09

[0091] For disordered ductal cells, fiber cells, and ray cells in cross-section, the specific operations to obtain their properties such as cell number, cell diameter, cell area, and cell ratio are as follows:

[0092] a). For slice images with different magnifications, different conversion units are used to convert pixels to real physical units;

[0093] b) To solve the problems of cell counting, sorting, and obtaining various cell attributes of cross-section cells, two empty DataFrames were created to save the test results, as shown in Tables 3 and 4;

[0094] c) For different cell types and different magnifications, the most appropriate YOLOv8 model weights are used. The data is stored as a Numpy array, including the coordinates (x1, y1, x2, y2) of each detection box, the confidence score, and the category ID.

[0095] d) Sort the detection boxes from the top left to the bottom right according to the coordinate difference of (x1, y1), and display the number in the center of the box (1, 2, etc. in the order of sorting). If the cells are too dense, they can be directly marked with *. Finally, the statistical information is plotted on the image, as shown in the following example. Figure 5 As shown, the width and height (in micrometers) and area (width * height) of each detection box are stored in a DataFrame;

[0096] e). After processing each image, save the processed image to the output folder and save all detection data (DataFrame) to an Excel file.

[0097] Table 3

[0098]

[0099] Table 4

[0100]

[0101] Through testing, using the traditional method for manual counting, the average time for counting only the number of transverse ductal cells, the number of fiber cells, and the number of ray cells under a 20x microscope is 5-6 minutes, 60-90 minutes, and 20-30 minutes, respectively. The average time for manually counting the number of tangential ray cells and the types of tangential ray cells under a 20x microscope is 5-10 minutes and 10-15 minutes, respectively. Using the method of this embodiment, the time for counting the number of transverse ductal cells, the number of fiber cells, the number of ray cells, the number of tangential ray cells, and the types of tangential cells under a 20x microscope is within 0.1-0.3 seconds. At the same time, various required attributes such as cell length, width, cell area, etc. can also be counted, which greatly improves the statistical speed and the detection accuracy reaches more than 98%.

[0102] This example uses the YOLOv8 network model to detect plant cells. The model boasts excellent performance, fast detection speed, and ease of rapid deployment, effectively detecting small and densely packed target cells. The test results can be used to assist in analyzing the diverse structures of plant xylem cells, further tracking and exploring the growth and development of cells. This provides a deeper understanding of plant growth mechanisms, xylem development, and stress resistance, aiding in the improvement of forest species, enhancing wood quality and yield, and increasing the economic value of forestry products.

[0103] It should be noted that, for the sake of convenience, the aforementioned method embodiments are all expressed as a series of action combinations, but those skilled in the art should know that the present invention is not limited to the described order of actions, because according to the present invention, certain steps can be performed in other orders or simultaneously.

[0104] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for statistically analyzing the properties of vascular tissue cells in woody plants based on deep learning, characterized in that: The following steps are involved: Obtain plant materials and process them to obtain images of plant vascular tissues; performing format conversion and annotation on the plant vascular tissue image to obtain an annotation file; After the format of the annotation file is converted again, it is input into the YOLOv8 model of different sizes for training to obtain optimized YOLOv8 models of different sizes; Different cell types were sliced ​​in different ways and photographed at different magnifications. The photographs were then fed into optimized YOLOv8 models of different sizes for detection and analysis, yielding the optimal YOLOv8 models for different slicing methods and magnifications. The properties of plant vascular tissue cells are statistically analyzed based on the optimal YOLOv8 model under different slicing methods and different shooting magnifications.

2. The method for statistically analyzing the properties of vascular tissue cells of woody plants based on deep learning according to claim 1, wherein: The process of obtaining plant material and processing it to obtain images of plant vascular tissue includes: Plant materials were sliced ​​using a fully automatic vibrating microtome. The intact tissue was placed on a glass slide and stained with toluidine blue to obtain chromogenic cell sections. The chromogenic cell sections at different angles and magnifications were photographed, and the obtained slice images were named and stored in ascending order of the photographing sequence as plant vascular tissue images.

3. The method for statistically analyzing the properties of vascular tissue cells of woody plants based on deep learning according to claim 1, wherein: The process of converting the plant vascular tissue image into a format includes: The plant vascular tissue images are saved in tiff format. The plant vascular tissue images are classified into different folders based on different shooting targets to obtain several target cell detection folders; the target cell detection folders are format converted using python code to obtain target cell detection folders in jpg format.

4. The method for statistically analyzing the properties of vascular tissue cells of woody plants based on deep learning according to claim 3, wherein: The process of labeling the target cell detection folder in jpg format includes: Based on the Labelme annotation tool and the Segment-Anything model, the lumens of different types of cells in the target cell detection folder are circled, and the cell types and distinction codes are marked; among them, cells with unclear outlines are manually annotated by creating polygons.

5. The method for statistically analyzing the properties of vascular tissue cells of woody plants based on deep learning according to claim 1, wherein: After the annotation file is formatted again, it is input into the YOLOv8 model of different sizes for training, including: A preset random seed is used to traverse the annotation file, the frequency of occurrence of each label is counted and sorted, and a configuration file in YOLO format is created based on the sorting result, and the image paths and labels of the training set and test set are defined; the annotation file is shuffled according to the preset random seed, and divided into training set and test set according to a preset ratio, and each annotation file in the training set and test set is converted into YOLO format and saved in the corresponding configuration file based on the defined image path and label.

6. The method for statistically analyzing the properties of vascular tissue cells of woody plants based on deep learning according to claim 5, characterized in that: Inputting the training set and the test set into YOLOv8 models of different sizes for training, and obtaining optimized YOLOv8 models of different sizes includes: Set the initial parameters of the YOLOv8 model network model of different sizes; Image enhancement is performed on the annotated images and annotated files in the training set and test set. The enhanced training set and test set are input into YOLOv8 models of different sizes for training until the loss function converges, and optimized YOLOv8 models of different sizes are obtained.

7. The method for statistically analyzing the properties of vascular tissue cells of woody plants based on deep learning according to claim 6, characterized in that: The initial parameters include at least image size, batch size, initial learning rate, intersection-over-union ratio, confidence threshold, number of categories, category name, and YOLO format dataset file paths for different types of cells.

8. The method for statistically analyzing the properties of vascular tissue cells of woody plants based on deep learning according to claim 1, wherein: The process of statistically analyzing plant vascular tissue cell properties based on the optimal YOLOv8 model includes: Based on different slicing methods and the optimal YOLOv8 model under different shooting magnifications, different types of slice images are predicted to obtain prediction results; Sort and number the prediction boxes of the prediction results or mark them with asterisks; Based on the properties of the predicted box and the detected object, the cell properties of each predicted box are obtained and statistically analyzed.

Images