Cell slice image recognition method based on lightweight LiteNet model

Abstract

The invention overcomes the defects in the prior art, and provides a brand-new lightweight deep learning model for solving the problems that the current deep learning model is large in parameter quantity, long in training time, not suitable for rapid identification under the condition of resource limitation and the like, and the model is named as LiteNet. The model can be applied to light-weight slice image recognition tasks of various types of cells. According to experimental analysis, compared with an existing model, various evaluation indexes of the model in a cell slice image recognition task have the advantages that the parameter number is reduced by about 93.37%, the performance index is improved by about 20.35%, and the training time is reduced by 631 s compared with ResUNet with high performance indexes. In conclusion, the LiteNet model provided by the invention is novel and reasonable in design, accurate and efficient, has the advantages of high recognition performance, light model weight, short training time and the like, and has excellent performance in a cell slice image recognition task under the condition of insufficient hardware computing power.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and medical image recognition, and in particular to a cell slice image recognition method based on a lightweight LiteNet model. Background Art

[0002] Cell slice image recognition is a key technology in biomedical image processing, which aims to accurately separate and identify individual cells from complex slice images. In cell biology research, cell slice image recognition helps to analyze the morphology, structure and function of cells, and facilitates the study of cell growth, differentiation, migration and other processes; in the field of neuroscience, the identification of neuronal cells can help researchers understand the development of the nervous system and disease mechanisms; in pathological diagnosis, cell slice image recognition can be used to quantitatively analyze tumor cells in pathological slices, assisting doctors in determining the type, grade and prognosis of tumors.

[0003] Earlier cell slice image recognition mainly relied on manual segmentation of individual cells. Professionals observed the cell image through a microscope and used a drawing pen to manually outline the cell boundaries on the image. This method is extremely time-consuming and laborious, and the accuracy and repeatability of the recognition results are greatly affected by the experience and subjective factors of the experimenter. With the development of computer technology, some semi-automatic and automated cell slice image recognition methods have gradually emerged. For example, the cell slice image is preprocessed by digital image processing methods to enhance the contrast between the cell and the background, and then the boundary recognition and annotation are performed manually on the preprocessed image; there is also a threshold-based method, which selects a suitable threshold to divide the pixels in the image into cells and non-cells according to the grayscale difference between the cell and the background, but the effect is not good for cells with uneven grayscale or images with complex backgrounds; and the edge detection-based method determines the cell boundary by detecting the grayscale mutation of pixels in the cell image, but when the cell boundary is blurred or there is noise, the accuracy of edge detection will be greatly affected and the effect is unstable.

[0004] In recent years, with the continuous advancement of computer vision and deep learning technology, cell slice image recognition has gradually developed towards fully automatic detection based on deep learning. Cell slice image recognition based on deep learning can automatically learn cell boundary features and achieve high-precision recognition of cells. However, in actual situations, due to factors such as hardware costs, there is a contradictory relationship between recognition speed and accuracy. Summary of the invention

[0005] The technical problem to be solved by the present invention is that in view of the deficiencies in the above-mentioned prior art, the present invention proposes a new lightweight deep learning model, which can be applied to lightweight recognition tasks of various types of cell slice images.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] 1. A lightweight cell slice image recognition model, wherein the model is named LiteNet, and the specific research contents are as follows:

[0008] The model proposed in this invention is a new lightweight convolutional neural network model for image recognition. The architecture of this model is based on an encoder-decoder. The encoder part is based on a lightweight design concept and is mainly composed of a feature preprocessing module and a lightweight residual connection module:

[0009] (1) The feature preprocessing module, as the cell slice image receiving module of the entire model, is responsible for the initial feature extraction and processing of the input image, and prepares for the subsequent lightweight residual module to accurately extract pixel information of more individual cell regions.

[0010] The feature preprocessing module includes two input branches:

[0011] ① The first branch uses a 5*5 convolution kernel to extract initial features of the morphology of individual cells. A relatively large convolution kernel can cover a wider range of local areas in the cell image, which helps to capture the macroscopic features of the cell morphology. For example, it can perceive the overall contour shape, size changes, and approximate texture distribution of the cell, thereby extracting initial morphological features;

[0012] ② The second branch uses a 1*1 convolution kernel to pre-learn the features of individual cell edges, and can independently learn and extract information related to the cell edge at each pixel location, such as edge sharpness, grayscale difference of edge pixels, and other features, focusing on the key part of the cell edge to obtain more detailed local feature information. The feature preprocessing module designed in this way, on the one hand, by setting up two different input branches, excavates the relevant feature information of individual cells from different angles, laying the foundation for subsequent more in-depth feature extraction; on the other hand, the two branches have their own focus and cooperate with each other to jointly complete the preprocessing of individual cell features, which is conducive to improving the richness and effectiveness of the overall feature expression, and providing high-quality feature input for subsequent high-level feature extraction modules.

[0013] (2) The lightweight residual connection module is the main component of the encoder, and adopts the design concept of lightweight and residual connection. Each lightweight residual connection module sets two feature extraction paths, and finally obtains an enhanced feature map through weighted fusion. After the above three sets of lightweight residual connection modules, the encoder part will learn a large number of morphological related features such as cell contour shape, size change, texture distribution, boundary smoothness, overall aspect ratio of the cell, grayscale change trend of the main part of the cell, etc., and then input this information into the ASPP module and connect it to the decoder through the ASPP module.

[0014] (3) The ASPP module, as the middleware connecting the encoder and the decoder, consists of multiple dilated convolutions with different sampling rates and is responsible for the extraction and fusion of multi-scale feature information. This module uses dilated convolutions with different sampling rates to perform parallel sampling of input feature maps at multiple scales, and can capture the feature information of cells of different sizes or different parts of cells in the image. In addition, the dilated convolution can increase the receptive field of the convolution kernel by inserting holes in the convolution kernel without increasing the number of parameters and the amount of calculation. The dilated convolutions with different sampling rates in the ASPP module further expand the spatial receptive field of the network. The multi-scale feature extraction and fusion capabilities of the ASPP module enable it to automatically adapt to the scale changes of different individual cells in the image and help the decoder restore the individual cell features. After the ASPP multi-scale feature extraction and fusion module, the decoder is connected.

[0015] (4) The decoder consists of a lightweight upsampling module and an ASPP fusion output module.

[0016] ① In the lightweight upsampling module, the feature maps of each layer of the encoder's peer connection part and the feature maps after upsampling of the decoder part are fused and weighted to achieve the fusion of cell morphology information at multiple scales; then, the feature map after information fusion is input into the next lightweight upsampling module, which adopts a design concept similar to the lightweight residual connection module, and uses a dual-branch, multi-scale, lightweight module design to maximize the efficiency of feature information flow and reduce parameter redundancy; finally, the cell information restoration image after multiple lightweight upsampling modules is input into the ASPP multi-scale fusion output module;

[0017] ②The ASPP fusion output module extracts multi-scale features from the global feature image and integrates the context information, which is finally output by the model to obtain a high-quality cell slice recognition image.

[0018] 2. A lightweight cell slice image recognition method, wherein the method mainly includes model training and recognition methods, and the specific steps are as follows:

[0019] S1. Cell slice image data acquisition and preprocessing: Use an electron microscope and a macro camera to shoot and record cell slices to obtain original cell slice images. Since cell slice images are not easy to obtain and the number is insufficient to support model training, it is necessary to perform data preprocessing operations on the original cell slice images to expand the data set, mainly including: unifying image size, random scaling, random flipping, random cropping and normalization. The training data obtained through these operations provides a high-quality and diverse data foundation for subsequent model training.

[0020] S2. Label individual cells: The pixels in the cell slice image are divided into individual cells and background areas. Use image labeling software to manually label individual cell boundaries to obtain training label images.

[0021] S3. Construct a lightweight cell slice image recognition model: Based on the lightweight concept and the encoder-decoder design architecture, a lightweight cell slice image recognition model is constructed. The LiteNet neural network model proposed in the present invention reduces the number of model parameters and memory usage, and improves training efficiency; on the other hand, the model has strong generalization ability and is suitable for different types of cell slice image recognition tasks.

[0022] S4, model training: Divide the data set obtained in S2 into a training set and a test set. The training set is used for model training. By calculating the error between the label value and the true value, the model weight parameters are updated using the back propagation mechanism to improve the performance of the model. The test set is used to evaluate the model performance and verify the generalization ability of the model to ensure the effectiveness and reliability of the model in practical applications.

[0023] S5. Cell slice image recognition: Load the model trained in S4, use the cell slice image as the model input, click Run, and the model outputs the slice image cell slice image recognition result.

[0024] The present invention proposes a cell slice image recognition method based on a lightweight LiteNet model. Compared with the existing models, the LiteNet model proposed in the present invention is based on the lightweight concept and the encoder-decoder design architecture, which greatly reduces the model training time and resource occupancy rate, and efficiently separates individual cells from the background. In addition, the model has strong generalization and can be used for different types of cell slice image recognition tasks. The model has a small number of parameters, low memory usage, and strong recognition performance, which significantly improves the operating efficiency of cell slice image recognition under insufficient hardware conditions, and is particularly suitable for rapid detection scenarios of individual cells.

[0025] In summary, the present invention is novel, reasonable, accurate and efficient in design. Compared with the existing models, the present model achieves a good balance between hardware consumption and recognition performance, and performs well in cell slice image recognition tasks under the condition of insufficient hardware computing power.

[0026] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 The figure is a flow chart of the method of the present invention.

[0028] Figure 2 Dice scores of the model of the present invention and the comparison model.

[0029] Figure 3 It is the IoU between the proposed model and the comparison model.

[0030] Figure 4 It is the mIoU of the model of the present invention and the comparison model.

[0031] Figure 5 It is the accuracy of the model of the present invention and the comparison model.

[0032] Figure 6 This is the recognition result of the model of the present invention and the comparison model. DETAILED DESCRIPTION

[0033] In order to make the technical solution of the present invention clearer, the technical solution of the present invention will be described in detail below in conjunction with the accompanying drawings and embodiments.

[0034] 1. Please refer to the following Figure 1 , the present invention proposes a new lightweight cell slice image recognition model. The specific research details of this model are as follows:

[0035] The model uses an encoder-decoder structure to extract high-level texture morphology and low-level detail edge features of cell slice images to achieve high-quality cell slice image recognition. The encoder part focuses on extracting cell feature information in the image and reduces computational complexity through lightweight design, thereby improving feature extraction efficiency; the decoder part is responsible for restoring the encoded feature information to a cell slice recognition image, and restoring image detail information through multi-scale fusion and lightweight upsampling technology.

[0036] First, the encoder part consists of a feature preprocessing module, a lightweight residual connection module and an ASPP module.

[0037] (1) Feature preprocessing module: It adopts a dual-branch design. The two branches extract features from the image respectively, and then perform fusion calculation to achieve the purpose of combining detail features and macro textures.

[0038] ① After the model receives the cell slice image data, the 5*5 convolutional branch extracts the image macroscopic feature information through a relatively large receptive field. For common round or elliptical cells, it can effectively capture the overall contour shape and texture information of the cell, and roughly determine the circular boundary of the cell or the long and short axis range of the ellipse, thereby determining the size change of the cell. In terms of texture distribution feature extraction, it can sense the approximate uniformity or directionality of the distribution of substances inside the cell, such as the specific texture direction of the fibrous structure in some cells, and convert these macroscopic features of cell morphology and texture into feature representations that can be processed by subsequent modules;

[0039] ②The 1*1 convolution branch is more focused on learning the edge features of individual cells. Since a convolution operation only acts on the position of a single pixel, it can analyze the pixel characteristics of the cell boundary more carefully. It can accurately detect the grayscale difference of edge pixels and determine the sharpness of the edge, that is, a clear edge will show a large grayscale change, while a blurred edge has a small grayscale difference. This precise grasp of edge details provides accurate local feature information supplement for subsequent feature extraction.

[0040] Through the fusion calculation of these two branches, the macroscopic morphological features obtained by the 5*5 convolution branch and the edge local features extracted by the 1*1 convolution branch are integrated in subsequent processing to construct a complete initial cell feature extraction module, which covers both the overall morphological texture of the cell and the cell edge details, providing a rich and hierarchical feature input basis for the subsequent lightweight residual connection module.

[0041] (2) Lightweight residual connection module: This module is based on the architecture design of lightweight dual-path feature extraction and weighted fusion enhancement.

[0042] Two feature extraction paths are set inside each lightweight residual connection module. These two paths adopt a lightweight design concept, aiming to reduce the number of parameters and calculations while efficiently extracting cell morphological features. One path sets a 1*1 depth-separable convolution layer, focusing on extracting high-frequency detail features of cells, such as microscopic structural features such as tiny protrusions or depressions on the cell surface; the other path sets two 5*5 convolution layers, focusing more on low-frequency features of cells, such as overall morphological trends and texture changes over a large range. The features extracted by these two paths are finally obtained by weighted fusion to obtain an enhanced feature map. This fusion method is not a simple addition, but different weights are assigned according to the importance of different features and their contribution to cell recognition. For example, in some cell types, if the cell edge features are more critical to recognition, then the 1*1 convolution branch that focuses on edge feature extraction will be given a higher weight during the fusion process. After three groups of such lightweight residual connection modules, various morphological features of cells are fully learned and integrated, including accurate depiction of cell contour shape, precise quantification of size, detailed analysis of texture distribution, as well as boundary smoothness, overall aspect ratio of cells, grayscale change trend of cell body and other feature information are fully extracted and passed to the ASPP module.

[0043] (3)ASPP module: a multi-scale feature fusion calculation group composed of multiple dilated convolutions with different sampling rates.

[0044] ① In cell slice images, for cells with obvious size differences, such as large mature cells and small immature cells, dilated convolutions with different sampling rates can adapt to their morphological differences. A dilated convolution with a large sampling rate can cover the entire range of larger cells and capture their global features, such as the overall morphology of large cells and the general internal structure distribution; a dilated convolution with a small sampling rate focuses on the local detail areas of smaller cells or larger cells, such as the complete outline of smaller cells and the tiny structural details inside larger cells;

[0045] ② In addition, the atrous convolution expands the receptive field of the convolution kernel by inserting holes without significantly increasing the consumption of computing resources. This enables the network to understand the relationship between cells and the surrounding background, such as the relative position of cells in tissue slices and the spatial relationship with adjacent cells. After parallel sampling of atrous convolutions with different sampling rates, multi-scale feature information is fused to enable the model to automatically adapt to the scale changes of individual cells, ensuring that both large-scale cell population characteristics and small-scale individual cell detail characteristics can be effectively captured and integrated, providing a comprehensive and multi-scale information basis for the decoder to restore individual cell characteristics.

[0046] Secondly, the decoder part consists of a lightweight upsampling module and an ASPP fusion output module.

[0047] (1) Lightweight upsampling module: Through efficient module architecture design, lightweight and efficient multi-scale information integration and restoration are achieved.

[0048] ① The feature maps of each layer of the encoder's equivalent connection part are fused and weighted with the upsampled feature maps of the decoder part. In this process, the feature maps of different levels of the encoder contain different degrees of cell information. For example, the feature maps of the shallower layers may contain the original position and local detail information of the cells, while the feature maps of the deeper layers more reflect the semantic categories and overall morphological characteristics of the cells. When the upsampled feature maps are fused with these feature maps of different levels, different weights are assigned according to the importance of different features and their contribution to cell recognition, and a dual-branch, multi-scale, lightweight design concept similar to the lightweight residual connection module is adopted. The dual branches can extract and fuse features from different scales to improve the richness of feature information;

[0049] ② The multi-scale design can adapt to the extraction requirements of cell slice feature maps at different resolutions; lightweight design ensures efficient operation under limited computing resources. Through this design, the circulation efficiency of feature information in the decoder is improved as much as possible, the loss and redundancy in the information transmission process are reduced, and the cell feature information can be gradually and accurately restored to image information.

[0050] (2) ASPP fusion output module: performs final multi-scale feature extraction and global context information integration on the cell information restoration image processed by three lightweight upsampling modules.

[0051] In the cell slice feature map, multi-scale feature extraction will further optimize the presentation of cell details, such as clearer separation of cell gaps. Global context information integration considers the relationship between cells in the entire image scene, such as the distribution density of cells in the tissue, and the information between cells and the surrounding background. Ultimately, the model outputs a high-quality cell slice recognition image, which can not only clearly display the morphology, boundaries and other details of individual cells, but also reflect the location and relationship information of cells in the background environment, providing strong support for medical applications such as cell research and disease diagnosis.

[0052] This model, which uses encoder-decoder as the basic architecture, lightweight as the design goal, and residual connection as the design concept, can effectively process cell slice images and achieve high-precision cell slice image recognition tasks, and has important application value in the field of biomedical image analysis.

[0053] 2. Please refer to the following Figure 1 The present invention provides a lightweight cell slice image recognition method, wherein the method mainly includes a model training and recognition method:

[0054] S1. Cell slice image data acquisition and preprocessing: First, use two professional devices, an electron microscope and a macro camera, to shoot and record the cell slices. The electron microscope can present the fine features of the cell microstructure, and the macro camera can capture a clear overall picture of the cell slice from a suitable distance. Through their collaborative shooting, the original cell slice image is finally obtained. However, in actual situations, there are two problems with cell slice images. On the one hand, it is difficult to obtain, and it is not as easy to collect in large quantities as some common images; on the other hand, the number of images obtained is often limited, and it is difficult to meet the sufficient data required for subsequent model training. Based on this, it is necessary to perform a series of data preprocessing operations on the original cell slice images to expand the data set. The specific preprocessing operations and their functions are as follows:

[0055] Unify image size: Adjust original cell slice images of different specifications to a unified size, which helps standardize data processing in the subsequent model training process, avoids calculation and matching problems caused by image size differences, and ensures that the model can receive input data in a consistent format;

[0056] Random scaling: Enlarge or reduce the image at a certain random ratio. This can simulate the morphological changes of cells under different observation conditions, increase the diversity of data, enable the model to learn the characteristic performance of cells at different scales, and enhance its adaptability to various actual situations;

[0057] Random flipping: randomly flipping the image horizontally or vertically. In this way, the data can be further enriched without changing the essential characteristics of the cells, allowing the model to identify individual cells from multiple angles and avoid bias in identifying cells in a specific orientation.

[0058] Random cropping: Randomly select a certain area from the original image for cropping to generate a new image sample. The cropped part still contains information about individual cells, so that the characteristics of cells in local areas can be mined. At the same time, it also greatly expands the number of samples in the data set and improves the model's ability to recognize different local characteristics of cells;

[0059] Normalization: Normalize the pixel values ​​of the image and map them to a specific value range [0, 1]. This helps stabilize the numerical calculations during model training, improves training efficiency, and enables the model to better focus on the relative feature relationships of the image rather than the absolute pixel values, thus enhancing the generalization ability of the model.

[0060] S2. Label individual cells: In cell slice images, pixel points can be used to clearly distinguish between individual cells and background areas. In order for the model to accurately identify individual cells, professional image labeling software is required to carefully label the boundaries of individual cells manually. The labelers accurately outline the contours of each individual cell based on the cell's morphology in the image, the difference from the surrounding background, and other features, and finally form a training label image. These label images will correspond to the previously preprocessed cell slice image data, and serve as an important reference for judging whether the recognition results are correct during model training.

[0061] S3, build a lightweight cell slice image recognition model: based on the idea of ​​lightweight, the cell slice image recognition model is built to solve the problems of excessive parameters, excessive memory usage and low training efficiency that may exist in traditional models. The encoder-decoder design architecture is adopted, which has unique advantages. The encoder part is responsible for feature extraction of the input cell slice image, and gradually converts its complex image information into abstract feature representation; the decoder restores the corresponding cell individual related information, such as the position and morphology of the cell, based on the features extracted by the encoder, and then realizes cell slice image recognition. The LiteNet neural network model proposed in the present invention has significant advantages: by carefully designing the network structure, adopting the appropriate convolution kernel size and number, optimizing the inter-layer connection method and other methods, the number of parameters in the model is effectively reduced, thereby reducing the memory usage of the model when it is running. This not only enables the model to run smoothly on devices with limited resources, but also speeds up the training speed of the model, improves training efficiency, and reduces the training time cost. In addition, this model fully considers the commonalities and differences of different types of cell individuals in morphology, structure, etc. in the process of architecture design and parameter adjustment, so that it can be well adapted to different types of cell slice image recognition tasks. No matter what type of cells it is, as long as it is input in the form of a cell slice image, the model has a high probability of accurately identifying the individual cells, showing good versatility and practicality.

[0062] S4, model training: Divide the data set obtained in S2 into a training set and a test set. The training set is used for model training. By calculating the error between the label value and the true value, the model weight parameters are updated using the back propagation mechanism to improve the performance of the model. The test set is used to evaluate the model performance and verify the generalization ability of the model to ensure the effectiveness and reliability of the model in practical applications.

[0063] S5, cell slice image recognition: When the model is fully trained in step S4 and reaches the expected performance requirements, the actual cell slice image recognition task can be carried out. First, load the trained model into the corresponding operating environment, then use the cell slice image to be recognized as the input data of the model, click the run button to start the program, and the model will process and analyze the input slice image based on the individual cell feature information learned in the training phase, and finally output the recognition result of the individual cell in the slice image.

[0064] Example 1

[0065] This example is a lightweight cell slice image recognition system, which includes an electron microscope and a macro camera as image acquisition devices, and a portable computer as an image annotation and deep learning model training and reasoning processing unit. First, the original image of the cell slice is obtained through the image acquisition device; then, the original image of the cell slice is transferred to the portable computer through the USB interface for image preprocessing and annotation. The specific data preprocessing operations include: unifying the image size, random scaling, random flipping, random cropping and normalization; then, the training image data is loaded into the LiteNet model for training and outputting quantifiable training results; finally, the image to be identified is input into the trained model for cell slice image recognition to obtain the recognition result. The data set in this embodiment contains a total of 670 images, including 2400 images in the training set and 280 images in the test set.

[0066] The model of the present invention is now quantitatively evaluated and compared with the existing model, and the following evaluation indicators are used: Dice score, intersection over union (IoU), mean intersection over union (mIoU), accuracy, and macro recall. Figures 2 to 6 , all evaluation indicators of the model of the present invention are higher than those of the existing models, and the model of the present invention has fewer parameters and lower video memory usage, which means that the model of the present invention has a shorter training time than the existing models. On the whole, the LiteNet model implemented by the present invention can produce accurate and stable results in the task of cell slice image recognition. The following is a comparison of the model LiteNet of the present invention with existing models such as Fully Convolutional Networks (FCN), U-shaped Network (UNet), and Residual U-Net (ResUNet).

[0067] Table 1 Model size comparison table

[0068]

[0069] Table 2 Model performance comparison

[0070]

[0071] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent structural change made to the above embodiment based on the technical essence of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A cell slice image recognition method based on a lightweight LiteNet model, characterized in that: In particular, it involves a lightweight model LiteNet for cell slice image recognition: The model proposed by the present invention is a new lightweight image recognition convolutional neural network model. The architecture of the model is: encoder-decoder. Among them, the encoder part is mainly composed of a feature preprocessing module and a lightweight residual connection module based on the lightweight design idea. As the cell slice image receiving module of the entire model, the feature preprocessing module is responsible for the initial feature extraction and processing of the input image, and is ready for the subsequent lightweight residual module to accurately extract pixel information of more individual cell areas. As the main component of the encoder, the lightweight residual connection module adopts the design idea of ​​lightweight and residual connection. Each lightweight residual connection module sets two feature extraction paths, and finally obtains an enhanced feature map through weighted fusion. After the above three groups of lightweight residual connection modules, the encoder part will learn a large number of morphological related features such as cell contour shape, size change, texture distribution, boundary smoothness, overall aspect ratio of the cell, grayscale change trend of the main part of the cell, etc., and then input this information into the ASPP module and connect it to the decoder through the ASPP module. The decoder consists of a lightweight upsampling module and an ASPP multi-scale fusion output module. In the lightweight upsampling module, the feature maps of each layer of the encoder's equivalent connection part and the feature maps after upsampling of the decoder part are fused and weighted to achieve the fusion of cell morphology information at multiple scales; then, the feature maps after information fusion are input into the next lightweight upsampling module, which adopts a design concept similar to the lightweight residual connection module, and uses a dual-branch, multi-scale, lightweight module design to maximize the efficiency of feature information flow and reduce parameter redundancy; finally, the cell information restoration image after multiple lightweight upsampling modules is input into the ASPP multi-scale fusion output module to perform multi-scale feature extraction and integration of global context information, and finally the model outputs a high-quality cell slice recognition image.

2. A lightweight model LiteNet for cell slice image recognition as claimed in claim 1, characterized in that: The encoder part: It consists of a feature preprocessing module, a lightweight residual connection module and an ASPP module. (1) Feature preprocessing module: It adopts a dual-branch design. The two branches extract features from the image respectively, and then perform fusion calculation to achieve the purpose of combining detail features and macro textures. ① After the model receives the cell slice image data, the 5*5 convolutional branch extracts the image macroscopic feature information through a relatively large receptive field. For common round or elliptical cells, it can effectively capture the overall contour shape and texture information of the cell, and roughly determine the circular boundary of the cell or the long and short axis range of the ellipse, thereby determining the size change of the cell. In terms of texture distribution feature extraction, it can sense the approximate uniformity or directionality of the distribution of substances inside the cell, such as the specific texture direction of the fibrous structure in some cells, and convert these macroscopic features of cell morphology and texture into feature representations that can be processed by subsequent modules; ②The 1*1 convolution branch is more focused on learning the edge features of individual cells. Since a convolution operation only acts on the position of a single pixel, it can analyze the pixel characteristics of the cell boundary more carefully. It can accurately detect the grayscale difference of edge pixels and determine the sharpness of the edge, that is, a clear edge will show a large grayscale change, while a blurred edge has a small grayscale difference. This precise grasp of edge details provides accurate local feature information supplement for subsequent feature extraction. Through the fusion calculation of these two branches, the macroscopic morphological features obtained by the 5*5 convolution branch and the edge local features extracted by the 1*1 convolution branch are integrated in subsequent processing to construct a complete initial cell feature extraction module, which covers both the overall morphological texture of the cell and the cell edge details, providing a rich and hierarchical feature input basis for the subsequent lightweight residual connection module. (2) Lightweight residual connection module: This module is based on the architecture design of lightweight dual-path feature extraction and weighted fusion enhancement. Two feature extraction paths are set inside each lightweight residual connection module. These two paths adopt a lightweight design concept, aiming to reduce the number of parameters and calculations while efficiently extracting cell morphological features. One path sets a 1*1 depth-separable convolution layer, focusing on extracting high-frequency detail features of cells, such as microscopic structural features such as tiny protrusions or depressions on the cell surface; the other path sets two 5*5 convolution layers, focusing more on low-frequency features of cells, such as overall morphological trends and texture changes over a large range. The features extracted by these two paths are finally obtained by weighted fusion to obtain an enhanced feature map. This fusion method is not a simple addition, but different weights are assigned according to the importance of different features and their contribution to cell recognition. For example, in some cell types, if the cell edge features are more critical to recognition, then the 1*1 convolution branch that focuses on edge feature extraction will be given a higher weight during the fusion process. After three groups of such lightweight residual connection modules, various morphological features of cells are fully learned and integrated, including accurate depiction of cell contour shape, precise quantification of size, detailed analysis of texture distribution, as well as boundary smoothness, overall aspect ratio of cells, grayscale change trend of cell body and other feature information are fully extracted and passed to the ASPP module. (3)ASPP module: a multi-scale feature fusion calculation group composed of multiple dilated convolutions with different sampling rates. ① In cell slice images, for cells with obvious size differences, such as large mature cells and small immature cells, dilated convolutions with different sampling rates can adapt to their morphological differences. Dilated convolutions with large sampling rates can cover the entire range of larger cells and capture their global features. Such as the overall morphology and internal structure distribution of large cells; the dilated convolution with a small sampling rate focuses on the local details of smaller cells or larger cells, such as the complete outline of smaller cells and the tiny structural details inside larger cells; ② In addition, the atrous convolution expands the receptive field of the convolution kernel by inserting holes without significantly increasing the consumption of computing resources. This enables the network to understand the relationship between cells and the surrounding background, such as the relative position of cells in tissue slices and the spatial relationship with adjacent cells. After parallel sampling of atrous convolutions with different sampling rates, multi-scale feature information is fused to enable the model to automatically adapt to the scale changes of individual cells, ensuring that both large-scale cell population characteristics and small-scale individual cell detail characteristics can be effectively captured and integrated, providing a comprehensive and multi-scale information basis for the decoder to restore individual cell characteristics.

3. The LiteNet lightweight model for cell slice image recognition according to claim 1, characterized in that: The decoder part: It consists of a lightweight upsampling module and an ASPP fusion output module. (1) Lightweight upsampling module: Through efficient module architecture design, lightweight and efficient multi-scale information integration and restoration are achieved. ① The feature maps of each layer of the encoder's equivalent connection part are fused and weighted with the upsampled feature maps of the decoder part. In this process, the feature maps of different levels of the encoder contain different degrees of cell information. For example, the feature maps of the shallower layers may contain the original position and local detail information of the cells, while the feature maps of the deeper layers more reflect the semantic categories and overall morphological characteristics of the cells. When the upsampled feature maps are fused with these feature maps of different levels, different weights are assigned according to the importance of different features and their contribution to cell recognition, and a dual-branch, multi-scale, lightweight design concept similar to the lightweight residual connection module is adopted. The dual branches can extract and fuse features from different scales to improve the richness of feature information; ② The multi-scale design can adapt to the extraction requirements of cell slice feature maps at different resolutions; lightweight design ensures efficient operation under limited computing resources. Through this design, the circulation efficiency of feature information in the decoder is improved as much as possible, the loss and redundancy in the information transmission process are reduced, and the cell feature information can be gradually and accurately restored to image information. (2) ASPP fusion output module: performs final multi-scale feature extraction and global context information integration on the cell information restoration image processed by three lightweight upsampling modules. In the cell slice feature map, multi-scale feature extraction will further optimize the presentation of cell details, such as clearer separation of cell gaps. Global context information integration considers the relationship between cells in the entire image scene, such as the distribution density of cells in the tissue, and the information between cells and the surrounding background. Ultimately, the model outputs a high-quality cell slice recognition image, which can not only clearly display the morphology, boundaries and other details of individual cells, but also reflect the location and relationship information of cells in the background environment, providing strong support for medical applications such as cell research and disease diagnosis.

4. A cell slice image recognition method based on a lightweight LiteNet model, characterized in that: In particular, it relates to a lightweight cell slice image recognition method, comprising the following steps: S1. Cell slice image data acquisition and preprocessing: First, the cell slices are photographed and recorded with the help of two professional devices, an electron microscope and a macro camera. The electron microscope can present the fine features of the cell microstructure, and the macro camera can capture a clear overall picture of the cell slice from a suitable distance. Through their coordinated shooting, the original cell slice image is finally obtained. However, in reality, there are two problems with cell slice images. On the one hand, it is difficult to obtain them, and they are not as easy to collect in large quantities as some common images; on the other hand, the number of images obtained is often limited, and it is difficult to meet the sufficient data required for subsequent model training. Based on this, it is necessary to perform a series of data preprocessing operations on the original cell slice images to expand the data set, including: Unify image size, randomly scale, randomly flip, randomly crop and normalize; S2. Label individual cells: In cell slice images, pixel points can be used to clearly distinguish individual cells from background areas. In order for the model to accurately identify individual cells, professional image labeling software is needed to manually label the boundaries of individual cells and serve as an important reference for judging whether the recognition results are correct during model training. S3. Construct a lightweight cell slice image recognition model: Based on the lightweight concept, a cell slice image recognition model is constructed to solve the problems that traditional models may have, such as excessive number of parameters, excessive memory usage, and low training efficiency. The LiteNet neural network model proposed in the present invention adopts the encoder-decoder design architecture, which has significant advantages: by carefully designing the network structure, adopting the appropriate size and number of convolution kernels, optimizing the inter-layer connection method, etc., the number of parameters in the model is effectively reduced, thereby reducing the memory usage of the model when it is running. This not only enables the model to run smoothly on devices with limited resources, but also speeds up the training speed of the model, improves training efficiency, and reduces training time costs. In addition, this model fully considers the commonalities and differences of different types of cell individuals in morphology, structure, etc. during the process of architecture design and parameter adjustment, so that it can adapt well to different types of cell slice image recognition tasks. No matter what type of cell is faced, as long as it is input in the form of a cell slice image, the model has a high probability of accurately identifying the individual cells therein, showing good versatility and practicality; S4, model training: Divide the data set obtained in S2 into a training set and a test set. The training set is used for model training. By calculating the error between the label value and the true value, the model weight parameters are updated using the back propagation mechanism to improve the performance of the model. The test set is used to evaluate the model performance and verify the generalization ability of the model to ensure the effectiveness and reliability of the model in practical applications. S5, cell slice image recognition: When the model is fully trained in step S4 and reaches the expected performance requirements, the actual cell slice image recognition task can be carried out. First, load the trained model into the corresponding operating environment, then use the cell slice image to be recognized as the input data of the model, click the run button to start the program, and the model will process and analyze the input slice image based on the individual cell feature information learned in the training phase, and finally output the recognition result of the individual cell in the slice image.

Images