Urine cytology image processing method and system

Abstract

The invention belongs to the technical field of medical image processing, and relates to a urine cytology image processing method and system, and the method comprises the steps: obtaining a data set of a urine cytology image, carrying out the preprocessing of the data set, and dividing the data set into a training data set and a test data set according to a proportion; a training task is established, a Yov5-OG model is established in the training task, the Yov5-OG model is used for training, and a trained weight file is stored for a test task; establishing a test task, utilizing the Yov5-OG model trained by the training task, and adopting the test data set to evaluate the performance of the Yov5-OG model; establishing a detection task, and detecting the urine cytology image after image cutting by using the trained Yolov5-OG model to obtain a plurality of images after slicing; and splicing the image with the size before the back-cut image. The problem that a diagnosis system is complex due to the fact that convolution operation and a sub-attention mechanism are too complex and training time is long in a current mainstream method is solved, and the method has the capability of improving the accuracy rate of urine cytology image recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a urine cytology image processing method and system. Background Art

[0002] Bladder cancer is a common cancer and the eighth leading cause of cancer-related death. Urine cytology is a simple, inexpensive, and effective method for bladder cancer screening. Due to its non-invasive nature, it is widely used to detect and follow up patients with high-grade urinary tract malignancies. However, there are several major challenges with urine cytology, including relatively low sensitivity compared to cystoscopy, a high rate of uncertain results, and reliance on manual observation for test results.

[0003] With the rapid development of artificial intelligence, deep learning has been widely used in different fields. Practice has shown that the high efficiency and strong expressive power of deep learning enable it to be widely used in various fields, especially in the field of computer vision, which enables it to perform well in improving the performance of image recognition systems. Through the study of deep learning technology and its application in the examination of urine components, it is expected to assist doctors in analyzing urine test results, improve recognition accuracy, and solve the problem of heavy workload and missed diagnosis and misdiagnosis of doctors in urine cytology testing.

[0004] However, previous deep learning studies have mostly used convolutional neural networks (CNNs) to process urine cytology images. CNNs use fixed convolution kernels to extract features. Since fixed convolution kernels can only extract a specific type of feature, complex image structures and diverse features cannot be fully expressed. CNNs require a large number of parameters to represent, and the convolution operation is relatively complex, which leads to high model complexity. For small target detection, traditional convolution is not flexible enough, and the training and inference costs are too high, which is not conducive to deployment in auxiliary diagnosis systems. Summary of the invention

[0005] The first aspect of the present invention provides a urine cytology image processing method to solve the problems of high model complexity, overly complex convolution operations, and long training time in the current mainstream algorithm CNN.

[0006] Another aspect of the present invention provides a urine cytology image processing system.

[0007] The present invention is achieved in that: In one aspect, the present invention provides a method for processing urine cytology images, comprising: Acquire urine cytology images to be detected and data sets of urine cytology images for training and testing, and divide the data sets into training data sets and test data sets according to proportion after preprocessing; Establishing a training task, establishing a Yolov5-OG model in the training task, cutting each image in the training data set proportionally, marking abnormal cells, cropping the marked cells to obtain a new sample image, and training the new sample image using the Yolov5-OG model, and saving the trained weight file for the test task; Establish a test task, use the Yolov5-OG model trained by the training task, and use the test data set to evaluate the performance of the Yolov5-OG model; Establish a detection task, slice the urine cytology image to be processed, and use the trained Yolov5-OG model to detect the sliced ​​urine cytology image to obtain multiple sliced ​​images, which contain three kinds of marking information: whether the cells are darkly stained, whether the nuclear membrane is regular, and the nuclear-cytoplasmic ratio; Reassemble the sliced ​​images back to the original size; The Yolov5-OG model includes a backbone network, a neck network and a head network, wherein: Backbone network: responsible for extracting features from the image, extracting edge, texture and shape features of the image layer by layer, and encoding these features into feature maps of different scales; Neck network: The neck network is located between the backbone network and the head network, and fuses the feature maps of different scales extracted by the backbone network; Head network: predicts the location and category of the target based on the fused feature map, and finally outputs the category, confidence level, and bounding box coordinates of the target.

[0008] Further, the backbone network includes in sequence: a Focus module, a first Ghost module, a first C3Ghost module, a second Ghost module, a second C3Ghost module, a third Ghost module, a third C3Ghost module, an SPP module and a fourth C3Ghost module; The neck network includes: The Concat module is connected to the second C3Ghost module of the backbone network; The fourth ODConv convolution module is connected to the fourth C3Ghost module, and the fourth ODConv convolution module is connected to the Concat module in sequence through the first upsampling module, the first concatenation module, the sixth C3Ghost module, the second ODConv convolution module and the second upsampling module; The second ODConv convolution module is sequentially connected to the second concatenation module, the first ODConv convolution module and the fifth C3Ghost module. After the Concat module is connected to the fifth C3Ghost module, the fifth C3Ghost module forms a first output to the head network. The fourth ODConv convolution module is also connected to a third splicing module, and the third splicing module is connected to the eighth C3Ghost module to form a second output to the head network; The third splicing module is also connected to the third ODConv convolution module and the seventh C3Ghost module to form a third output to the head network, and the seventh C3Ghost module is also connected to the output of the second splicing module; The C3Ghost modules in the backbone network and the neck network include: a fourth Ghost module, a fifth Ghost module, a sixth Ghost module, an addition module, a fourth splicing module and a seventh Ghost module connected in sequence, the eighth Ghost module connects the fourth Ghost module and the fourth splicing module, and the fourth Ghost module also outputs to the addition module.

[0009] Furthermore, the SPP module includes a first convolution block, a maximum pooling layer, a fifth splicing module, and a second convolution block connected in sequence, wherein the maximum pooling layer includes a first maximum pooling layer, a second maximum pooling layer, and a third maximum pooling layer in parallel.

[0010] Furthermore, the Focus module includes four parallel slicing modules and a fifth splicing module connected to the output of the slicing module, and the output of the fifth splicing module is connected to the third convolution block.

[0011] Furthermore, the abnormal cell marking includes four classifications based on whether the cell is darkly stained or lightly stained and whether the nuclear membrane is regular, under the premise that the ratio of the area of ​​the cell nucleus to the cytoplasm is greater than 0.5; the four classifications include: dark staining and light staining of the cell, and whether the nuclear membrane of the cell nucleus is regular, and four categories are obtained by combination and arrangement: dark staining - regular nuclear membrane of the cell nucleus, light staining - regular nuclear membrane of the cell nucleus, dark staining - irregular nuclear membrane of the cell nucleus, and light staining - irregular nuclear membrane of the cell nucleus.

[0012] Another aspect of the present invention provides a urine cytology image processing system, the system comprising: A data acquisition module, which acquires the urine cytology images to be detected and the data sets of urine cytology images for training and testing, and divides the data sets into training data sets and test data sets according to proportion after preprocessing the data sets; The image preprocessing module proportionally slices each image in the training data set, marks abnormal cells, and crops the marked cells to obtain new sample images; it proportionally slices the images in the test data set and the urine cytology images that need to be tested; Task establishment module, to establish training tasks, test tasks and detection tasks; Yolov5-OG model, perform training tasks, use Yolov5-OG model to train new sample images obtained in the training data set, and obtain a trained weight file for the test task; perform test tasks, use the test data set to evaluate the performance of the Yolov5-OG model; perform detection tasks, use the trained Yolov5-OG model to process the urine cytology images to be detected, and obtain multiple sliced ​​images, wherein the sliced ​​images contain three kinds of labeling information, namely, whether the cells are darkly stained, whether the nuclear membrane is regular, and the nuclear-cytoplasmic ratio; Image post-processing module, which stitches multiple sliced ​​images back to the image size before slicing; The Yolov5-OG model includes a backbone network, a neck network and a head network, wherein: Backbone network: responsible for extracting features from the image, extracting edge, texture and shape features of the image layer by layer, and encoding these features into feature maps of different scales; Neck network: The neck network is located between the backbone network and the head network, and fuses the feature maps of different scales extracted by the backbone network; Head network: predicts the location and category of the target based on the fused feature map, and finally outputs the category, confidence level, and bounding box coordinates of the target.

[0013] Furthermore, the abnormal cell marking includes four classifications based on whether the cell is darkly stained or lightly stained and whether the nuclear membrane is regular, under the premise that the ratio of the area of ​​the cell nucleus to the cytoplasm is greater than 0.5; the four classifications include: dark staining and light staining of the cell, and whether the nuclear membrane of the cell nucleus is regular, and four categories are obtained by combination and arrangement: dark staining - regular nuclear membrane of the cell nucleus, light staining - regular nuclear membrane of the cell nucleus, dark staining - irregular nuclear membrane of the cell nucleus, and light staining - irregular nuclear membrane of the cell nucleus.

[0014] Further, the backbone network includes in sequence: a Focus module, a first Ghost module, a first C3Ghost module, a second Ghost module, a second C3Ghost module, a third Ghost module, a third C3Ghost module, an SPP module and a fourth C3Ghost module; The neck network includes: The Concat module is connected to the second C3Ghost module of the backbone network; The fourth ODConv convolution module is connected to the fourth C3Ghost module, and the fourth ODConv convolution module is connected to the Concat module in sequence through the first upsampling module, the first concatenation module, the sixth C3Ghost module, the second ODConv convolution module and the second upsampling module; The second ODConv convolution module is sequentially connected to the second concatenation module, the first ODConv convolution module and the fifth C3Ghost module. After the Concat module is connected to the fifth C3Ghost module, the fifth C3Ghost module forms a first output to the head network. The fourth ODConv convolution module is also connected to a third splicing module, and the third splicing module is connected to the eighth C3Ghost module to form a second output to the head network; The third splicing module is also connected to the third ODConv convolution module and the seventh C3Ghost module to form a third output to the head network, and the seventh C3Ghost module is also connected to the output of the second splicing module; The C3Ghost modules in the backbone network and the neck network include: a fourth Ghost module, a fifth Ghost module, a sixth Ghost module, an addition module, a fourth splicing module and a seventh Ghost module connected in sequence, the eighth Ghost module connects the fourth Ghost module and the fourth splicing module, and the fourth Ghost module also outputs to the addition module.

[0015] Furthermore, the SPP module includes: a first convolution block, a maximum pooling layer, a fifth splicing module, and a second convolution block connected in sequence, wherein the maximum pooling layer includes a parallel first maximum pooling layer, a second maximum pooling layer, and a third maximum pooling layer.

[0016] Furthermore, the Focus module includes four parallel slicing modules and a fifth splicing module connected to the output of the slicing module, and the output of the fifth splicing module is connected to the third convolution block.

[0017] The present invention has the following beneficial effects: The present invention is applicable to urine cytology image detection tasks. The Yolov5-OG model used combines the ODConv convolution module and the Ghost module in the case of a large amount of labeled data to perform small target detection on urine cytology images. The Yolov5-OG model used reduces the complexity of the model used, reduces the computational cost of training and reasoning, and improves the accuracy of small target detection. The method of the present invention is superior in target detection speed. In addition, experiments have shown that the use of the Yolov5-OG model of the present invention has the ability to improve the accuracy of identifying urine cytology images.

[0018] The ODConv convolution module in the Yolov5-OG model proposes a novel multi-dimensional attention mechanism that can learn four types of attention in the four dimensions of the convolution kernel space. These four types of attention complement each other and are gradually applied to the corresponding convolution kernels, which can significantly enhance the feature extraction ability of the convolution operation. Even with only one convolution kernel, it can compete or surpass the existing dynamic convolution methods with multiple convolution kernels, thereby greatly reducing additional parameters, which enables deep learning models to achieve a better balance between accuracy and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic diagram of the structure of a Yolov5-OG model in a urine cytology image processing method provided in an embodiment of the present invention; Figure 2 A structural diagram of the ODConv convolution module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0021] See also Figure 1 As shown, a urine cytology image processing system comprises: The data acquisition module acquires the urine cytology images to be detected and the data sets of urine cytology images for training and testing. The image preprocessing module cuts each image in the training data set proportionally, marks abnormal cells, and cuts the marked cells to obtain a new sample image; cuts the images in the test data set and the urine cytology images to be tested proportionally; and divides the data set into a training data set and a test data set proportionally after preprocessing the data set; The data set here is: the image data collected by urine cytology. The preprocessing here refers to data preprocessing of the data set and the urine cytology image to be detected, and cutting the image size into 1280×1280 size.

[0022] The task establishment module establishes training tasks, test tasks and detection tasks; the training tasks and test tasks are used to obtain the trained Yolov5-OG model, and the detection task is used to use the trained Yolov5-OG model to detect a urine cytology image that needs to be detected.

[0023] Yolov5-OG model, executes training task, cuts each image in the training data set of the image preprocessing module in proportion, marks abnormal cells, cuts the marked cells to obtain new sample images, and then trains the new sample images using Yolov5-OG model to obtain trained weight files for testing tasks; executes testing tasks, uses the testing data set to evaluate the performance of the Yolov5-OG model; executes detection tasks, uses the trained Yolov5-OG model to obtain multiple sliced ​​images, the sliced ​​images contain three kinds of marking information, namely, whether the cells are darkly stained, whether the nuclear membrane is regular, and the nuclear-cytoplasmic ratio; the detection task uses the urine cytology images to be detected acquired by the data acquisition module; Image post-processing module, which stitches multiple sliced ​​images back to the image size before slicing; Among them, see Figure 1 As shown, the Yolov5-OG model includes a backbone network, a neck network and a head network, wherein: Backbone network: responsible for extracting features from the image, extracting edge, texture and shape features of the image layer by layer, and encoding these features into feature maps of different scales; Neck network: The neck network is located between the backbone network and the head network. It is used to further process the feature maps to adapt to different detection tasks and fuse the feature maps of different scales extracted by the backbone network. Head network: predicts the location and category of the target based on the fused feature map, and finally outputs the target category, confidence level, and bounding box coordinates to complete the target detection task.

[0024] Among them, each marked cell after cutting is cropped out to obtain a new sample image, including: each marked cell is cropped into an image of 128×128 size, which is convenient for training in the Yolov5-OG model, and the result is restored to the original image size with the mark through puzzle.

[0025] The labeled abnormal cells include four categories based on whether the cell is darkly stained or lightly stained and whether the nuclear membrane is regular, under the premise that the area ratio of the cell nucleus to the cytoplasm is greater than 0.5; the four categories include: dark staining and light staining of the cell, and whether the nuclear membrane of the cell nucleus is regular, and four categories are obtained by combination and arrangement: dark staining - regular nuclear membrane of the cell nucleus, light staining - regular nuclear membrane of the cell nucleus, dark staining - irregular nuclear membrane of the cell nucleus, and light staining - irregular nuclear membrane of the cell nucleus.

[0026] In the present invention, the backbone network includes in sequence: a Focus module, a first Ghost module, a first C3Ghost module, a second Ghost module, a second C3Ghost module, a third Ghost module, a third C3Ghost module, an SPP module and a fourth C3Ghost module; The neck network includes: The Concat module is connected to the second C3Ghost module of the backbone network; The fourth ODConv convolution module is connected to the fourth C3Ghost module, and the fourth ODConv convolution module is connected to the Concat module in sequence through the first upsampling module, the first concatenation module, the sixth C3Ghost module, the second ODConv convolution module and the second upsampling module; The second ODConv convolution module is sequentially connected to the second concatenation module, the first ODConv convolution module and the fifth C3Ghost module. After the Concat module is connected to the fifth C3Ghost module, the fifth C3Ghost module forms the fourth convolution block of the first output to the head network; The fourth ODConv convolution module is also connected to a third splicing module, and the third splicing module is connected to the eighth C3Ghost module to form a second output to the fifth convolution block of the head network; The third splicing module further connects the third ODConv convolution module and the seventh C3Ghost module to form a sixth convolution block of the third output to the head network, and the seventh C3Ghost module is further connected to the output of the second splicing module; The C3Ghost modules in the backbone network and the neck network include: a fourth Ghost module, a fifth Ghost module, a sixth Ghost module, an addition module, a fourth splicing module and a seventh Ghost module connected in sequence, the eighth Ghost module connects the fourth Ghost module and the fourth splicing module, and the fourth Ghost module also outputs to the addition module.

[0027] The SPP module includes: a first convolution block, a maximum pooling layer, a fifth splicing module, and a second convolution block connected in sequence, wherein the maximum pooling layer includes a first maximum pooling layer, a second maximum pooling layer, and a third maximum pooling layer in parallel. Pooling kernels of different sizes are used to perform pooling operations on feature maps, and multi-scale feature information is integrated to improve the robustness of target detection for targets of different sizes.

[0028] The Focus module includes four parallel slicing modules and a fifth splicing module connected to the output of the slicing module. The output of the fifth splicing module is connected to the third convolution block. The four parallel slicing modules include: the first slicing module, the second slicing module, the third slicing module and the fourth slicing module. The main function is to perform slicing operations and rearrange channels to increase the number of channels and resolution of feature maps, promote information interaction between different channels, and thus enhance the model's ability to detect small targets.

[0029] All convolution blocks in the present invention include: 2D convolution module + BN layer + SiLU activation function, which is a common preprocessing path for extracting low-level features of images.

[0030] All Ghost modules in the present invention include 2D convolution module + deep convolution layer + BN layer + SiLU activation function, which are used to remove redundant features and reduce the amount of calculation. Ghost modules can use custom convolution kernel sizes, not limited to 1x1 convolution kernels. It uses traditional convolution method to generate some intrinsic feature maps, and then uses low-cost and diversified linear operations to enhance features and expand the number of channels. While retaining the intrinsic feature map, it is also synchronized with linear transformation to enhance the expression effect of the model.

[0031] The function of all upsampling modules in the present invention is to amplify the resolution of the feature map so as to merge it with the feature map of the backbone path.

[0032] All ODConv convolution modules in the present invention: contain a multi-dimensional attention mechanism to enhance feature extraction capabilities, and can learn four types of attention in the four dimensions of the convolution kernel space. These four types of attention complement each other and are gradually applied to the corresponding convolution kernels, which can significantly enhance the feature extraction capabilities of the convolution operation. Even with only one convolution kernel, it can compete or surpass existing dynamic convolution methods with multiple convolution kernels, thereby greatly reducing additional parameters, which allows the model to achieve a better balance between accuracy and efficiency.

[0033] For details, see Figure 2 As shown in the figure, the ODConv convolution module first compresses the input feature X into a feature vector of a certain length through a global average pooling layer. The compressed feature vector is then processed by a fully connected layer and a rectified linear unit. The fully connected layer maps the feature vector to a low-dimensional space at a compression rate of 1 / 16 to simplify the complexity of dynamic convolution. The rectified linear unit sets the negative values ​​in the feature vector to zero. Finally, four head branches are created, each of which includes a fully connected layer and a Softmax activation function or a Sigmoid activation function. Then, four types of attention of the ODConv convolution module are generated, and the final output data is represented as Y. Among them, the data of the head branch are represented as follows: ; ; ; and . Indicates the number of data.

[0034] All the splicing modules in the present invention splice the feature maps of the trunk path and the branch path together to fuse feature information of different scales.

[0035] The maximum pooling layer in the present invention downsamples the feature map to reduce the amount of calculation and increase the receptive field.

[0036] In the present invention, two complementary methods are used to calculate the interval of the nuclear-cytoplasmic ratio: pixel counting method and ellipse fitting method. The pixel counting method identifies the cell nucleus, cytoplasm and background through K-Means clustering and calculates the ratio of the number of pixels; while the ellipse fitting method fits an ellipse to the cell contour and calculates the area ratio. By averaging the results of the two methods, the defects of each method can be compensated to obtain a more comprehensive and accurate nuclear-cytoplasmic ratio calculation result.

[0037] The depth of nuclear staining is an important indicator for judging the state of the cell. Darkly stained nuclei usually indicate that the DNA content in the nucleus is high, and the cell may be in an active proliferation state; while lightly stained nuclei may mean that the cell is in a dormant or apoptotic state. Whether the nuclear membrane is regular or not is the second indicator. Regular nuclear membranes usually indicate that the cell is in a healthy state and has normal metabolic activity. Irregular nuclear membranes may indicate that the cell is abnormal. The area ratio of the nucleus to the cytoplasm is another important parameter that reflects the state of the cell. A higher nuclear-cytoplasm ratio usually indicates that the nuclear volume is enlarged, and the cell may be in an abnormal proliferation state; while a lower nuclear-cytoplasm ratio may mean that the cell is in a normal state or a degenerate state.

[0038] In the embodiment of the present invention, the Yolov5-OG model first performs a cut operation on the image data in the training data set, and then uses dark staining, light staining and whether the nuclear membrane is regular to mark the cells with possible lesions, that is, the cells with a nuclear-cytoplasmic ratio greater than 0.5, to perform four-category marking, and then cuts out each marked cell to obtain a new sample image. After training, the result is restored to the original image size with the mark by jigsaw puzzle; Specifically, image slicing is to generate new sample images by slicing each image in the training dataset proportionally; Marking is to classify possible abnormal cells into four categories using dark staining, light staining and calculating the nuclear-cytoplasmic ratio; Cropping is to crop each marked cell after cutting the image to obtain a new sample image.

[0039] In the embodiment of the present invention, a training task is established to crop each marked cell after image slicing to obtain a new sample image, including: cropping each marked cell into an image of 128×128 size to facilitate training in the Yolov5-OG model.

[0040] In the embodiment of the present invention, a test task is established, the Yolov5-OG model trained by the training task is used, and the test data set is used to evaluate the performance of the Yolov5-OG model.

[0041] The image post-processing module reassembles the multiple sliced ​​images output by the Yolov5-OG model back to the image size before slicing.

[0042] In one embodiment, there is a large amount of local information in the data. In order to enable the Yolov5-OG model to fully capture the local information, the input image is first subjected to an image cutting operation, thereby improving the learning efficiency of the model through limited data. In this embodiment of the present invention, the training process may include: Step S1: Cut the images of the training data set into several images of 1280×1280 size; Step S2: For abnormal cell images, i.e. cells with a nuclear-cytoplasmic ratio greater than 0.5, four-category marking is performed using dark staining, light staining, and whether the nuclear membrane is regular; Step S3: Cut each labeled cell image into 128×128 size images as the input of Yolov5-OG model; Step S4: Put the image obtained in step S3 into the Yolov5-OG model for training, and the Yolov5-OG model outputs a weight file; Step S5: By continuously repeating steps S3 and S4, the difference between the output of the Yolov5-OG model and the given labeled image is reduced, so as to achieve the effect of Yolov5-OG model training.

[0043] Step S6: Apply the weight file of the trained Yolov5-OG model to the test task and restore the result to the original image size with the markup through jigsaw puzzle.

[0044] In this embodiment, the performance of the Yolov5-OG model is evaluated by testing tasks: the performance of the Yolov5-OG model is evaluated using the training loss index. The performance of the Yolov5-OG model is evaluated using three evaluation indicators: accuracy, sensitivity, and specificity.

[0045] Another embodiment of the present invention provides a urine cytology image processing method, which uses the above-mentioned urine cytology image processing system and includes: Acquire urine cytology images to be detected and data sets of urine cytology images for training and testing, and divide the data sets into training data sets and test data sets according to proportion after preprocessing; Establishing a training task, establishing a Yolov5-OG model in the training task, cutting each image in the training data set proportionally, marking abnormal cells, cropping the marked cells to obtain a new sample image, and training the new sample image using the Yolov5-OG model, and saving the trained weight file for the test task; Establish a test task, use the Yolov5-OG model trained by the training task, and use the test data set to evaluate the performance of the Yolov5-OG model; Establish a detection task, slice the urine cytology images that need to be processed, and use the trained Yolov5-OG model to detect the sliced ​​urine cytology images to obtain multiple sliced ​​images. The sliced ​​images contain three types of labeling information: whether the cells are darkly stained, whether the nuclear membrane is regular, and the nuclear-cytoplasmic ratio. Reassemble multiple sliced ​​images back to the original size.

[0046] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for processing urine cytology images, characterized in that: include: Acquire a urine cytology image to be detected and a data set of urine cytology images for training and testing, and divide the data set and the urine cytology image to be detected into a training data set and a test data set according to a ratio after preprocessing; Establishing a training task, establishing a Yolov5-OG model in the training task, cutting each image in the training data set proportionally, marking abnormal cells, cropping the marked cells to obtain a new sample image, and training the new sample image using the Yolov5-OG model, and saving the trained weight file for the test task; Establish a test task, use the Yolov5-OG model trained by the training task, and use the test data set to evaluate the performance of the Yolov5-OG model; Establish a detection task, slice the urine cytology image to be processed, and use the trained Yolov5-OG model to detect the sliced ​​urine cytology image to obtain multiple sliced ​​images, which contain three kinds of marking information: whether the cells are darkly stained, whether the nuclear membrane is regular, and the nuclear-cytoplasmic ratio; Reassemble the sliced ​​images back to the original size; The Yolov5-OG model includes a backbone network, a neck network and a head network, wherein: Backbone network: responsible for extracting features from the image, extracting edge, texture and shape features of the image layer by layer, and encoding these features into feature maps of different scales; Neck network: The neck network is located between the backbone network and the head network, and fuses the feature maps of different scales extracted by the backbone network; Head network: predicts the location and category of the target based on the fused feature map, and finally outputs the target category, confidence level, and bounding box coordinates.

2. The urine cytology image processing method according to claim 1, characterized in that: The backbone network includes, in sequence: a Focus module, a first Ghost module, a first C3Ghost module, a second Ghost module, a second C3Ghost module, a third Ghost module, a third C3Ghost module, an SPP module and a fourth C3Ghost module; The neck network includes: The Concat module is connected to the second C3Ghost module of the backbone network; The fourth ODConv convolution module is connected to the fourth C3Ghost module, and the fourth ODConv convolution module is connected to the Concat module in sequence through the first upsampling module, the first concatenation module, the sixth C3Ghost module, the second ODConv convolution module and the second upsampling module; The second ODConv convolution module is sequentially connected to the second concatenation module, the first ODConv convolution module and the fifth C3Ghost module. After the Concat module is connected to the fifth C3Ghost module, the fifth C3Ghost module forms a first output to the head network. The fourth ODConv convolution module is also connected to a third splicing module, and the third splicing module is connected to the eighth C3Ghost module to form a second output to the head network; The third splicing module is also connected to the third ODConv convolution module and the seventh C3Ghost module to form a third output to the head network, and the seventh C3Ghost module is also connected to the output of the second splicing module; The C3Ghost modules in the backbone network and the neck network include: a fourth Ghost module, a fifth Ghost module, a sixth Ghost module, an addition module, a fourth splicing module and a seventh Ghost module connected in sequence, the eighth Ghost module connects the fourth Ghost module and the fourth splicing module, and the fourth Ghost module also outputs to the addition module.

3. The urine cytology image processing method according to claim 2, characterized in that: The SPP module includes a first convolution block, a maximum pooling layer, a fifth splicing module, and a second convolution block connected in sequence, wherein the maximum pooling layer includes a first maximum pooling layer, a second maximum pooling layer, and a third maximum pooling layer in parallel.

4. The urine cytology image processing method according to claim 2, characterized in that: The Focus module includes four parallel slicing modules and a fifth splicing module connected to the output of the slicing module, and the output of the fifth splicing module is connected to the third convolution block.

5. The urine cytology image processing method according to claim 2, characterized in that: The labeled abnormal cells include four categories based on whether the cell is darkly stained or lightly stained and whether the nuclear membrane is regular, under the premise that the area ratio of the cell nucleus to the cytoplasm is greater than 0.5; the four categories include: dark staining and light staining of the cell, and whether the nuclear membrane of the cell nucleus is regular, and four categories are obtained by combination and arrangement: dark staining - regular nuclear membrane of the cell nucleus, light staining - regular nuclear membrane of the cell nucleus, dark staining - irregular nuclear membrane of the cell nucleus, and light staining - irregular nuclear membrane of the cell nucleus.

6. A urine cytology image processing system, characterized in that: The system comprises: A data acquisition module, which acquires urine cytology images to be detected and a data set of urine cytology images for training and testing; The image preprocessing module cuts each image in the training data set proportionally, marks abnormal cells, and cuts the marked cells to obtain a new sample image; cuts the images in the test data set and the urine cytology images to be tested proportionally; and divides the data set into a training data set and a test data set proportionally after preprocessing the data set; Task establishment module, to establish training tasks, test tasks and detection tasks; Yolov5-OG model, perform training tasks, use Yolov5-OG model to train new sample images obtained in the training data set, and obtain a trained weight file for the test task; perform test tasks, use the test data set to evaluate the performance of the Yolov5-OG model; perform detection tasks, use the trained Yolov5-OG model to process the urine cytology images to be detected, and obtain multiple sliced ​​images, wherein the sliced ​​images contain three kinds of labeling information, namely, whether the cells are darkly stained, whether the nuclear membrane is regular, and the nuclear-cytoplasmic ratio; Image post-processing module, which stitches multiple sliced ​​images back to the image size before slicing; The Yolov5-OG model includes a backbone network, a neck network and a head network, wherein: Backbone network: responsible for extracting features from the image, extracting edge, texture and shape features of the image layer by layer, and encoding these features into feature maps of different scales; Neck network: The neck network is located between the backbone network and the head network, and fuses the feature maps of different scales extracted by the backbone network; Head network: predicts the location and category of the target based on the fused feature map, and finally outputs the target category, confidence level, and bounding box coordinates.

7. The urine cytology image processing system according to claim 6, characterized in that: The labeled abnormal cells include four categories based on whether the cell is darkly stained or lightly stained and whether the nuclear membrane is regular, under the premise that the area ratio of the cell nucleus to the cytoplasm is greater than 0.5; the four categories include: dark staining and light staining of the cell, and whether the nuclear membrane of the cell nucleus is regular, and four categories are obtained by combination and arrangement: dark staining - regular nuclear membrane of the cell nucleus, light staining - regular nuclear membrane of the cell nucleus, dark staining - irregular nuclear membrane of the cell nucleus, and light staining - irregular nuclear membrane of the cell nucleus.

8. The urine cytology image processing system according to claim 6, characterized in that: The backbone network includes, in sequence: a Focus module, a first Ghost module, a first C3Ghost module, a second Ghost module, a second C3Ghost module, a third Ghost module, a third C3Ghost module, an SPP module and a fourth C3Ghost module; The neck network includes: The Concat module is connected to the second C3Ghost module of the backbone network; The fourth ODConv convolution module is connected to the fourth C3Ghost module, and the fourth ODConv convolution module is connected to the Concat module in sequence through the first upsampling module, the first concatenation module, the sixth C3Ghost module, the second ODConv convolution module and the second upsampling module; The second ODConv convolution module is sequentially connected to the second concatenation module, the first ODConv convolution module and the fifth C3Ghost module. After the Concat module is connected to the fifth C3Ghost module, the fifth C3Ghost module forms a first output to the head network. The fourth ODConv convolution module is also connected to a third splicing module, and the third splicing module is connected to the eighth C3Ghost module to form a second output to the head network; The third splicing module is also connected to the third ODConv convolution module and the seventh C3Ghost module to form a third output to the head network, and the seventh C3Ghost module is also connected to the output of the second splicing module; The C3Ghost modules in the backbone network and the neck network include: a fourth Ghost module, a fifth Ghost module, a sixth Ghost module, an addition module, a fourth splicing module and a seventh Ghost module connected in sequence, the eighth Ghost module connects the fourth Ghost module and the fourth splicing module, and the fourth Ghost module also outputs to the addition module.

9. The method for constructing a urine cytology image processing system according to claim 6, characterized in that: The SPP module includes: a first convolution block, a maximum pooling layer, a fifth splicing module, and a second convolution block connected in sequence, wherein the maximum pooling layer includes a first maximum pooling layer, a second maximum pooling layer, and a third maximum pooling layer in parallel.

10. The method for constructing a urine cytology image processing system according to claim 6, characterized in that: The Focus module includes four parallel slicing modules and a fifth splicing module connected to the output of the slicing module, and the output of the fifth splicing module is connected to the third convolution block.

Images