Cerebrospinal fluid cell intelligent analysis and diagnosis system and method based on deep learning

Abstract

The invention belongs to the technical field of medical image analysis, and particularly relates to a cerebrospinal fluid cell intelligent analysis and diagnosis system and method based on deep learning. The method comprises the following steps: cutting an original image through a multi-scale segmentation model, and recording and storing position information of each cell image; segmenting all single cell and cell cluster images by a multi-scale segmentation model; training a cell classification model for endowing image category attributes to the single cells and the cell clusters; counting classification results output by the cell classification model to generate quantitative indexes; and generating a category visualization image and a diagnosis report through the positioning information of the single cell and cell cluster images. According to the method, a digital image processing technology and a deep learning algorithm are utilized, a traditional method is combined with an intelligent algorithm, effective image information is fully utilized, a detection result is visualized, checking by a doctor is facilitated, and finally the functions of automatic positioning, accurate classification and quantitative index of single cells and cell clusters in a cerebrospinal fluid sample are achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image analysis, and in particular relates to a cerebrospinal fluid cell intelligent analysis and diagnosis system and method based on deep learning. Background Art

[0002] Cell morphology examination plays a key role in the diagnosis of cell system diseases. In clinical laboratories, doctors usually examine cerebrospinal fluid (CSF) cells under a microscope to diagnose related cerebrospinal fluid system diseases by identifying and classifying normal or abnormal cells at various developmental stages and counting them.

[0003] Cerebrospinal fluid analysis is crucial for the diagnosis of a variety of neurological diseases, but traditional microscopic analysis methods often rely on manual observation and counting. This process is not only time-consuming and labor-intensive, but the test results are largely limited by the operator's experience and subjective judgment, and are prone to errors. With the advancement of artificial intelligence technology, automated and precise image analysis technology has become a reality. In 2023, Tianjin Medical University Cancer Hospital applied for an invention patent named "An Artificial Intelligence Automatic Analysis Method for Cerebrospinal Fluid Cells", which achieved automatic classification of cerebrospinal fluid cells. However, traditional morphological analysis was used to achieve cell extraction, and the original image was filtered and filled multiple times. The processing process was relatively cumbersome. In addition, a single density function was used as the classification feature when implementing cell classification, and the image information utilization rate was low, which may lead to low accuracy and robustness of the model.

[0004] Therefore, there is an urgent need to develop an efficient, accurate and universal intelligent cerebrospinal fluid image analysis system, which will help reduce the workload of medical staff and accelerate the diagnosis and treatment of cerebrospinal fluid system diseases. Summary of the invention

[0005] The present invention proposes a cerebrospinal fluid cell intelligent analysis and diagnosis system and method based on deep learning. The diagnostic system utilizes digital image processing technology and deep learning algorithms, combines traditional methods with intelligent algorithms, makes full use of effective image information, and visualizes the test results to facilitate doctor verification. It ultimately realizes the automatic positioning, accurate classification and quantitative indicator functions of single cells and cell clusters in cerebrospinal fluid samples, and solves the problems of low information utilization, low model accuracy and robustness that exist only in analysis based on traditional morphological analysis algorithms.

[0006] The technical solution of the present invention is achieved in this way: A cerebrospinal fluid cell intelligent analysis and diagnosis system based on deep learning, including the following parts: Digital image module, used to obtain high-resolution cerebrospinal fluid scan images; The enhanced image module segments the acquired full-view scanning image, and performs data denoising, distortion correction, deduplication, image screening and image enhancement processing on the segmented image block data; The cell labeling module locates and labels the cells to be detected in the image and obtains the labeling results; Segmentation model module for detecting single cells and cell clusters; The classification model module is used to extract single-cell images, form a single-cell database, and classify segmented cells.

[0007] A cerebrospinal fluid cell intelligent analysis and diagnosis method based on deep learning includes the following steps: Step S1: training a multi-scale segmentation model, where the trained multi-scale segmentation model is used to locate and segment single cells and cell cluster images; Step S2: Load the original image and the mask image generated by the multi-scale segmentation model, find the outline of each cell, use the bounding box extraction method to obtain the smallest rectangular area for the found cell outline, crop it on the original image, save it as a separate image file, and at the same time, record and save the position information of each cell image; Step S3: using the multi-scale segmentation model trained and optimized in step S1, inputting the cerebrospinal fluid image to be segmented, and segmenting all single cell and cell cluster images through step S2; Step S4: In the classification model module, a cell classification model is trained, and the trained cell classification model is used to assign category attributes to single cell and cell cluster images; Step S5: counting the classification results output by the cell classification model to generate quantitative indicators; Step S6: Read the location information of the saved single cell and cell cluster images, return the classification labels output by the cell classification model to the corresponding cell positions in the original image, generate category visualization images, select cells and cell cluster images that have a greater impact on the final diagnosis results to generate a diagnosis report.

[0008] Through the above technical solution, the above steps S1-S6 can realize the positioning and identification of single cells and cell clusters in cerebrospinal fluid, provide diagnostic information, and adopt multi-feature fusion and serial two-stage analysis to make full use of image information. At the same time, the separate segmentation and classification models avoid information mixing during optimization.

[0009] Optionally, in step S1, the method for training the multi-scale segmentation model includes the following steps: Step S1.1: In the digital image module, the cerebrospinal fluid smear is scanned based on a high-resolution optical microscope system to obtain a large number of full-slide scan image samples, including tumor cases and normal cases; Step S1.2: In the enhanced image module, the full-slice scanned image samples in step S1.1 are segmented according to a preset fixed size, and the segmented images are divided into two types: tumor patients and non-tumor patients, and the segmented images are subjected to denoising and enhancement preprocessing; Step S1.3: In the cell labeling module, select 50-100 representative cerebrospinal fluid image block samples in step S1.2, use labeling software to label the selected cerebrospinal fluid image block samples, label the positions of single cells and cell clusters in each image, and obtain the labels required for target detection; Step S1.4: Select the U-Net3+ model architecture as the basic model for multi-scale segmentation, and use dilated convolution instead of the original standard convolution to increase the receptive field without reducing the image resolution and capture a wider range of contextual information; Step S1.5: In the segmentation model module, the cerebrospinal fluid sample image is used as the input of the multi-scale segmentation model, and the drawn mask image is used as the output of the multi-scale segmentation model to train and optimize the multi-scale segmentation model.

[0010] Through the above technical solution, the above steps S1.1-S1.5 can realize the extraction of single cells and cell clusters of interest, retaining only the boundaries and internal information of cells and clusters, and completely eliminating the interference of background information, which can make it easier for the second-stage classification model to focus on the internal details of the cells and achieve accurate classification.

[0011] Optionally, in step S4, the method for training the cell classification model specifically includes the following steps: Step S4.1: The input image is processed through step S2 to obtain single cell and cell cluster segmentation images, from which 50% of the images are randomly selected. The selected images should cover various types of cells and cell clusters and are handed over to professional doctors for accurate classification; Step S4.2: Select an existing classification model as the cerebrospinal fluid cell classification network model architecture, and obtain a high-accuracy recognition model that distinguishes different categories of cerebrospinal fluid single cells and cell cluster images by designing a training method for the neural network model.

[0012] Through the above technical solution, the above steps S4.1-S4.2 can achieve accurate classification of the extracted cells and clusters. Based on the existing classification model architecture, the feasibility of feature extraction can be guaranteed. The training method can be optimized according to the classification object and classification results to achieve higher accuracy and robustness.

[0013] Optionally, in step S4.2, the classification network model architecture used is a ResNeXt50 network, and z-score normalization is used in data preprocessing to adjust the scale of the feature, and each pixel value is normalized using the following formula: ; in, is the original pixel value, is the mean of all input image data, is the variance of all input image data.

[0014] Through the above technical solution, the above formula can realize the standardization of segmented data and make the data at the same order of magnitude. Because when training data with different orders of magnitude, attributes with large orders of magnitude dominate, causing the algorithm to be very sensitive to the order of magnitude instead of focusing on various morphological information inside the image. In order to ensure the reliability of the results, the data needs to be normalized. Compared with other normalization methods, z-score normalization is less affected by outliers and is suitable for situations where the maximum and minimum values ​​are unknown.

[0015] Optionally, in step S4.2, the L2 norm of the weight is added to the loss function of the model training, and the calculation formula is as follows: ; in, is the original loss term, is the L2 regularization term, is the true label, is the classification model prediction label, is the weight decay coefficient, which controls the strength of regularization. Represents the weights in the model.

[0016] Through the above technical solution, the above formula will weight The L2 norm is added to the loss function, that is, the sum of the minimization of the prediction loss and the penalty term is taken as the training goal to achieve weight decay and prevent the model from overfitting. Another common regularization is the L1 norm. Compared with the L2 norm, the L1 norm causes the model to concentrate the weights on a small number of features and clear other weights to zero, while the L2 norm imposes a huge penalty on large components of the weights, making the learning algorithm biased towards a model that evenly distributes weights on a large number of features, making the model more stable to observation errors in a single variable.

[0017] Optionally, in step S4.2, after the loss function is determined, the back propagation algorithm is used to calculate the gradient of the loss function relative to each weight of the model to update the model parameters. Here, the stochastic gradient descent algorithm is used, and the calculation formula is as follows: ; in, represents the model parameters, represents the learning rate, represents the loss function, is a randomly selected sample.

[0018] Through the above technical solution, the above formula continuously updates the parameters of the classification model to minimize the value of the loss function and improve the classification accuracy. Since only one sample is used to calculate the gradient each time, compared with the batch gradient descent algorithm, the stochastic gradient descent algorithm has a smaller amount of calculation, a faster convergence speed, and can jump out of the local optimal solution, making it more likely to find the global optimal solution.

[0019] Optionally, in step S4.2, when training the classification model, cosine annealing learning rate scheduling is used to adjust the learning rate according to a periodically changing cosine function. The expression at the tth round is usually defined as: ; in, is the minimum value of the learning rate, is the maximum value of the learning rate, t is the current round number, is the length of the learning rate cycle.

[0020] Through the above technical solution, the above formula can realize dynamic learning rate scheduling, adjusting the learning rate according to the cosine curve to fine-tune the learning rate for better performance and convergence. Compared with other learning rate scheduling techniques such as step decay or exponential decay, cosine annealing has the characteristics of smooth transition and adaptability, which can ensure the smooth transition of learning rate, which is crucial to maintain the stability of the training process.

[0021] Optionally, in step S5, the specific steps include the following steps: Step S5.1: Count single cells and cell clusters separately, calculate the average area of ​​all single cell images, and then estimate the total number of cells in the cell cluster by measuring the total area of ​​the entire cell cluster; Step S5.2: Take the cell cluster category as the category of each cell therein, count all categories of single cells and cell clusters, analyze the concentration of cells of each category, and generate digital indicators.

[0022] Through the above technical solution, the above steps S5.1-S5.2 can realize the digital analysis of a glass slide scan image, and finally generate a diagnosis report based on the analysis results. After communicating with experienced clinicians, the feasibility of the estimation method is determined. The use of the estimation method to calculate the total number of cells has almost no effect on the final indicator, while greatly improving the calculation efficiency and the speed of generating the diagnosis report.

[0023] After adopting the above technical solution, the beneficial effects of the present invention are: The present invention realizes precise positioning, accurate segmentation and classification of cerebrospinal fluid cells through multi-feature fusion and a serial two-stage model. The flexible design of the classification model training strategy can be realized through step S4. The data passes through step S1 and step S4 in sequence. The segmentation and classification models are independent of each other, avoiding information mixing, and have high accuracy and robustness. Step S5 visualizes the detection results to facilitate doctors' verification and calibration.

[0024] In the present invention, multi-feature fusion means that each decoder layer of U-Net3+ fuses the smaller and same-scale feature maps from the encoder and the larger-scale feature maps from the decoder, combining the low-level details from the feature maps of different scales with the high-level semantics to achieve a full-scale aggregated feature map. The serial two-stage model means that the segmentation model and the classification model are connected in series, the image capture information is input sequentially, and the feature extraction parameters between the models are independent of each other, which can avoid information mixing. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0026] Figure 1 This is the architecture diagram of the cerebrospinal fluid cell intelligent analysis and diagnosis system based on deep learning; Figure 2 It is a flowchart of the classification model training strategy. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The embodiments of the present application disclose a system and method for intelligent analysis and diagnosis of cerebrospinal fluid cells based on deep learning.

[0029] Examples, such as Figure 1-Figure 2 As shown in the figure, a cerebrospinal fluid cell intelligent analysis and diagnosis system based on deep learning includes the following parts: Digital image module: collect visible elements in cerebrospinal fluid through cell smear centrifuge, make smears through Wright or Wright-Gieve staining, and automatically scan and collect images with an optical microscope. The cell morphology, cell membrane, and internal details of the cell nucleus in the collected images are relatively intact, and high-resolution cerebrospinal fluid scanning images are obtained.

[0030] Enhanced image module: Segment the collected full-field scanning image, and perform data denoising, distortion correction, deduplication, image screening, image enhancement and other processing methods on the segmented image block data to enhance the image quality.

[0031] Single-cell labeling module: Use professional labeling tools to locate and label the cells to be detected in the image, and double labeling is performed by physicians with more than 10 years of experience in cerebrospinal fluid cytology testing to obtain high-quality labeling results.

[0032] Segmentation model module: The U-Net3+ model is selected as the detection model for cerebrospinal fluid cells. In order to better realize the detection of multi-scale cells, the original standard convolution is replaced by dilated convolution to capture a wider range of contextual information. The training data set is input to update the model parameters. After multiple training and optimization, the model achieves high-accuracy detection of single cells and cell clusters.

[0033] Classification model module: extract single-cell images of multiple patients to form a single-cell database, and professional pathologists assign each cell category label. All data are divided, and the training set data is used to train the classification model. The classification model is based on the ResNeXt50 network architecture, optimizes the training strategy, introduces the L2 norm of the weight to prevent the model from overfitting, and uses cosine annealing learning rate scheduling during the gradient descent process to control the speed of weight adjustment and avoid falling into the local optimum. The model ultimately achieves accurate classification of segmented cells.

[0034] A method for intelligent analysis and diagnosis of cerebrospinal fluid cells based on deep learning, comprising the following steps: Step S1: training a multi-scale segmentation model. The trained multi-scale segmentation model is used to locate and segment single cell and cell cluster images. The method for training a multi-scale segmentation model includes the following steps: Step S1.1: In the digital image module, the cerebrospinal fluid smear is scanned based on a high-resolution optical microscope system to obtain a large number of full-slice scanned image samples, including tumor cases and normal cases; for example, the cerebrospinal fluid smear is placed in a designated area of ​​the microscope stage, the camera is started, the microscope automatically scans the cerebrospinal fluid smear in its entirety, and all images of the cerebrospinal fluid smear collected by the microscope are saved for later use, and the above steps are repeated to achieve the acquisition of a large number of cerebrospinal fluid smear image samples.

[0035] Step S1.2: In the enhanced image module, the full-slice scanned image samples in step S1.1 are segmented according to a preset fixed size (800 pixels × 800 pixels) to obtain cerebrospinal fluid image block samples of a suitable uniform size, and the segmented image block samples are divided into two types: tumor patients and non-tumor patients. The segmented image block samples are subjected to denoising and enhancement preprocessing to improve image quality; Step S1.3: In the cell labeling module, select about 50-100 representative cerebrospinal fluid image block samples in step S1.2, use labeling software to label the selected cerebrospinal fluid image block samples, label the positions of single cells and cell clusters in each image, and obtain the labels required for target detection; Step S1.4: Select the U-Net3+ model architecture as the basic model for multi-scale segmentation, and use dilated convolution instead of the original standard convolution to increase the receptive field without reducing the image resolution and capture a wider range of contextual information; Step S1.5: In the segmentation model module, the cerebrospinal fluid image block sample is used as the input of the multi-scale segmentation model, and the mask image drawn by the annotation software is used as the output of the multi-scale segmentation model to train and optimize the multi-scale segmentation model. The drawn mask image refers to the binary image of the cell boundary and background generated after annotation.

[0036] Step S2: Load the cerebrospinal fluid image block in step 1.2 and the mask image generated by the multi-scale segmentation model, find the outline of each cell, use the bounding box extraction method to obtain the smallest rectangular area for the found cell outline, crop it on the original image, save it as a separate image file, and at the same time, record and save the position information of each cell image; Step S3: Use the multi-scale segmentation model trained and optimized in step S1, input the cerebrospinal fluid image to be segmented, and segment all single cells and cell cluster images through step S2; step S2 crops and saves the detected cell images, thereby segmenting all single cells and cell cluster images.

[0037] Step S4: In the classification model module, a cell classification model is trained. The trained cell classification model is used to assign category attributes to single cell and cell cluster images to assist doctors in clinical diagnosis. The method for training a cell classification model specifically includes the following steps: Step S4.1: The input image is processed through step S2 to obtain single cell and cell cluster segmentation images, from which 50% of the images are randomly selected. The selected images should cover various types of cells and cell clusters and are handed over to professional doctors for accurate classification; Step S4.2: Select an existing classification model as the cerebrospinal fluid cell classification network model architecture, and obtain a high-accuracy recognition model that distinguishes different categories of cerebrospinal fluid single cells and cell cluster images by designing a training method for the neural network model.

[0038] In step S4.2, the classification network model architecture used is the ResNeXt50 network. Different from the original ResNeXt50 training, z-score normalization is used in data preprocessing, the scale of features is adjusted, and the L2 norm of weights is added when calculating the loss function to prevent model overfitting. After determining the loss function, the model parameters are optimized based on the stochastic gradient descent (SGD) algorithm. In the process of gradient descent, cosine annealing learning rate scheduling is used to control the speed of weight adjustment and avoid falling into local optimality. During the training process, z-score normalization is used to calculate the global mean and standard deviation of the pixel values ​​of the image, convert the data into values ​​within a certain range, eliminate the dimensional differences between different features, make different features comparable, and thus improve the performance of the deep learning algorithm. The following formula is used to normalize each pixel value: ; in, is the original pixel value, is the mean of all input image data, is the variance of all input image data.

[0039] In step S4.2, the L2 norm of the weight is added to the loss function of the model training. A higher weight decay is used in the early stage of training to prevent overfitting and help the model stabilize. As the training progresses, the weight decay coefficient is gradually reduced to encourage the model to learn smaller or simpler weights, allowing the model to adjust the weights more flexibly to optimize performance. The calculation formula is as follows: ; in, is the original loss term, is the L2 regularization term, is the true label, is the classification model prediction label, is the weight decay coefficient, which controls the strength of regularization. Represents the weights in the model.

[0040] In step S4.2, after the loss function is determined, the back propagation algorithm is used to calculate the gradient of the loss function relative to each weight of the model to update the model parameters. Here, the stochastic gradient descent (SGD) algorithm is used. Compared with the gradient descent algorithm, it uses only one sample (or a small batch of samples) in the data set in each iteration to update the model parameters according to the gradient and learning rate, making the training process more efficient. The calculation formula is as follows: ; in, represents the model parameters, represents the learning rate, represents the loss function, is a randomly selected sample.

[0041] In step S4.2, unlike the original network, the learning rate is no longer fixed when training the classification model. Instead, the cosine learning rate decay is used to adjust the learning rate according to a periodically changing cosine function. At the beginning of training, the learning rate starts from a higher initial value and gradually decreases to a very small value as the training progress increases. The cooling process in the simulated annealing process is used to optimize the training process and help the model converge better. The learning rate The expression at the tth round (epoch) is usually defined as: ; in, is the minimum value of the learning rate, is the maximum value of the learning rate, t is the current round number, is the length of the learning rate cycle.

[0042] Step S5: Count the nine classification results (granulocytes, lymphocytes, plasma cells, mononuclear tissue macrophages, proleukocytes, tumor cells, cerebrospinal fluid exfoliated cells, fungi and unclassified cells) output by the cell classification model to generate quantitative indicators of the concentration of each type of cell; the specific steps include the following steps: Step S5.1: Count single cells and cell clusters separately, calculate the average area of ​​all single cell images, and then estimate the total number of cells in the cell cluster by measuring the total area of ​​the entire cell cluster; Step S5.2: Taking the cell cluster category as the category of each cell therein, counting all categories of single cells and cell clusters, analyzing the concentration of cells of each category, and generating digital indicators, which specifically refer to the concentration data of each type of cell.

[0043] Step S6: Read the location information of the saved single-cell and cell cluster images, return the classification label (classification label refers to the cell category name) output by the cell classification model to the corresponding cell position in the original image, generate a category visualization image, select cells that have a greater impact on the final diagnosis result and cell cluster images to generate a diagnosis report so that doctors can check and verify to improve the accuracy of the diagnosis.

[0044] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. 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 intelligent analysis and diagnosis of cerebrospinal fluid cells based on deep learning, characterized in that: The following steps are involved: Step S1: training a multi-scale segmentation model. The trained multi-scale segmentation model is used to locate and segment single cell and cell cluster images. The details are as follows: Step S1.1: In the digital image module, the cerebrospinal fluid smear is scanned based on a high-resolution optical microscope system to obtain a large number of full-slide scan image samples, including tumor cases and normal cases; Step S1.2: In the enhanced image module, the full-slice scanned image samples in step S1.1 are segmented according to a preset fixed size, and the segmented images are divided into two types: tumor patients and non-tumor patients, and the segmented images are subjected to denoising and enhancement preprocessing; Step S1.3: In the cell labeling module, select 50-100 representative cerebrospinal fluid image block samples in step S1.2, use labeling software to label the selected cerebrospinal fluid image block samples, label the positions of single cells and cell clusters in each image, and obtain the labels required for target detection; Step S1.4: Select the U-Net3+ model architecture as the basic model for multi-scale segmentation, and use dilated convolution instead of the original standard convolution to increase the receptive field without reducing the image resolution and capture a wider range of contextual information; Step S1.5: In the segmentation model module, the cerebrospinal fluid sample image is used as the input of the multi-scale segmentation model, and the drawn mask image is used as the output of the multi-scale segmentation model to train and optimize the multi-scale segmentation model.

2. According to claim 1, a method for intelligent analysis and diagnosis of cerebrospinal fluid cells based on deep learning is characterized in that: The following steps are also included: Step S2: Load the original image and the mask image generated by the multi-scale segmentation model, find the outline of each cell, use the bounding box extraction method to obtain the smallest rectangular area for the found cell outline, crop it on the original image, save it as a separate image file, and at the same time, record and save the position information of each cell image; Step S3: using the multi-scale segmentation model trained and optimized in step S1, inputting the cerebrospinal fluid image to be segmented, and segmenting all single cell and cell cluster images through step S2; Step S4: In the classification model module, a cell classification model is trained, and the trained cell classification model is used to assign category attributes to single cell and cell cluster images; Step S5: counting the classification results output by the cell classification model to generate quantitative indicators; Step S6: Read the location information of the saved single cell and cell cluster images, return the classification labels output by the cell classification model to the corresponding cell positions in the original image, generate category visualization images, select cells and cell cluster images that have a greater impact on the final diagnosis results to generate a diagnosis report.

3. According to claim 2, a method for intelligent analysis and diagnosis of cerebrospinal fluid cells based on deep learning is characterized in that: In step S4, the method for training the cell classification model specifically includes the following steps: Step S4.1: The input image is processed through step S2 to obtain single cell and cell cluster segmentation images, from which 50% of the images are randomly selected. The selected images should cover various types of cells and cell clusters and are handed over to professional doctors for accurate classification; Step S4.2: Select an existing classification model as the cerebrospinal fluid cell classification network model architecture, and obtain a high-accuracy recognition model that distinguishes different categories of cerebrospinal fluid single cells and cell cluster images by designing a training method for the neural network model.

4. The method for intelligent analysis and diagnosis of cerebrospinal fluid cells based on deep learning according to claim 3 is characterized in that: In step S4.2, the classification network model architecture used is the ResNeXt50 network. During data preprocessing, z-score normalization is used to adjust the scale of the feature. The following formula is used to normalize each pixel value: ； in, is the original pixel value, is the mean of all input image data, is the variance of all input image data.

5. The method for intelligent analysis and diagnosis of cerebrospinal fluid cells based on deep learning according to claim 3, characterized in that: In step S4.2, the L2 norm of the weight is added to the loss function of the model training, and the calculation formula is as follows: ； in, is the original loss term, is the L2 regularization term, is the true label, is the classification model prediction label, is the weight decay coefficient, which controls the strength of regularization. Represents the weights in the model.

6. The method for intelligent analysis and diagnosis of cerebrospinal fluid cells based on deep learning according to claim 5, characterized in that: In step S4.2, after the loss function is determined, the back propagation algorithm is used to calculate the gradient of the loss function relative to each weight of the model to update the model parameters. Here, the stochastic gradient descent algorithm is used, and the calculation formula is as follows: ; in, represents the model parameters, represents the learning rate, represents the loss function, is a randomly selected sample.

7. The method for intelligent analysis and diagnosis of cerebrospinal fluid cells based on deep learning according to claim 3, characterized in that: In step S4.2, when training the classification model, the cosine annealing learning rate scheduling is used to adjust the learning rate according to a periodically changing cosine function. The expression at the tth round is defined as: ； in, is the minimum value of the learning rate, is the maximum value of the learning rate, t is the current round number, is the length of the learning rate cycle.

8. The method for intelligent analysis and diagnosis of cerebrospinal fluid cells based on deep learning according to claim 2, characterized in that: Step S5 specifically includes the following steps: Step S5.1: Count single cells and cell clusters separately, calculate the average area of ​​all single cell images, and then estimate the total number of cells in the cell cluster by measuring the total area of ​​the entire cell cluster; Step S5.2: Take the cell cluster category as the category of each cell therein, count all categories of single cells and cell clusters, analyze the concentration of cells of each category, and generate digital indicators.

9. A cerebrospinal fluid cell intelligent analysis and diagnosis system based on deep learning, including the following parts: Digital imaging module, used to obtain high-resolution cerebrospinal fluid scan images; The enhanced image module segments the acquired full-view scanning image, and performs data denoising, distortion correction, deduplication, image screening and image enhancement processing on the segmented image block data; The cell labeling module locates and labels the cells to be detected in the image and obtains the labeling results; Segmentation model module for detecting single cells and cell clusters; The classification model module is used to extract single-cell images, form a single-cell database, and classify segmented cells; The intelligent analysis and diagnosis system operates according to the deep learning-based cerebrospinal fluid cell intelligent analysis and diagnosis method described in claims 1-7.

Images