Single-cell printing device control method and system based on convolutional neural network optimization

Abstract

The application discloses a single-cell printing equipment control method and system based on convolutional neural network optimization, and relates to the technical field of intelligent control. The control method comprises the following steps: before performing a jet printing operation on a cell suspension, continuously collecting cell suspension images containing target single cells to determine cell main area images, and after inputting a shallow convolutional network, respectively constructing a pixel gray parameter set, a pixel response parameter set and a pixel mutation parameter set; analyzing the morphological fluctuation degree of each cell main area image to screen out a standard morphological feature image; analyzing the structural integrity degree of the target single cells contained in the standard morphological feature image to determine a cell integrity index, so as to judge whether the structural integrity degree of the target single cells in the current standard morphological feature image meets the printing requirements, and to issue corresponding level printing control instructions and perform the printing control instructions. The application realizes accurate identification of complete cells in the cell suspension, and improves the yield rate and stability of printing.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, specifically to a control method and system for a single-cell printing device based on convolutional neural network optimization. Background Technology

[0002] With the rapid development of cutting-edge research in life sciences and cell engineering, single-cell printing technology, as an important supporting means for microscale biomanufacturing, has been widely used in cell array construction, tissue structure generation, and personalized drug screening. In particular, with the assistance of microfluidic platforms, the accurate identification and printing of target single cells through image-driven control has become a key link in achieving high-throughput and high-precision biomanufacturing. Therefore, how to accurately identify target single cells and assess their structural integrity before printing has become one of the core issues in improving printing quality and stability.

[0003] For the aforementioned single-cell printing task, most existing single-cell recognition and printing control methods rely on traditional image processing techniques for target extraction and structure judgment, making it difficult to accurately analyze the minute deformations and complex structures of damaged edges of cells in the image. Moreover, existing printing control usually makes printing decisions based on a single image frame, ignoring the dynamic fluctuations of cell morphology over time, which can easily lead to judgment errors. At the same time, the integrity of cell structure lacks multi-dimensional and quantitative indicators before actual printing, making it difficult to achieve adaptive control decisions based on multi-parameter fusion. These shortcomings directly affect the recognition accuracy and consistency of the printed product during the printing process, limiting the widespread deployment of single-cell printing in high-precision application scenarios.

[0004] The root cause of the aforementioned situation and shortcomings lies in the fact that existing single-cell recognition and printing control methods lack deep structural perception capabilities in image parameter modeling, making it difficult to construct a model mechanism that reflects the synergistic relationship between pixel structure response, edge mutations, and grayscale characteristics. Especially in single-cell suspensions, the differences in transparency between cells are small, the boundaries are blurred, and there is interference from cell fragments. Conventional image discrimination models struggle to distinguish the subtle differences between intact single cells and damaged structures, making it highly prone to misjudging the target structure before the printing command is issued. The consequences include: damaged cells being printed incorrectly during the spraying process, leading to tissue construction failure; or intact target cells being misjudged as damaged, requiring frequent changes to the cell suspension, resulting in decreased operational efficiency, longer printing cycles, and system-level failures such as nozzle blockage and microvalve misopening due to delayed control. Therefore, constructing a printing control method that can achieve quantitative evaluation of cell integrity based on a convolutional neural network-optimized intelligent recognition mechanism has become an urgent need to improve the stability and accuracy of single-cell printing. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a control method and system for single-cell printing equipment based on convolutional neural network optimization, which solves the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a single-cell printing device control method based on convolutional neural network optimization, comprising the following steps:

[0007] S1. Before performing jet printing on the cell suspension, continuously acquire images of the cell suspension containing the target single cell to determine the main cell region image of each frame. After inputting into a shallow convolutional network, construct pixel grayscale parameter set, pixel response parameter set and pixel mutation parameter set respectively.

[0008] S2. Based on the pixel grayscale parameter set, pixel response parameter set and pixel mutation parameter set, analyze the degree of morphological fluctuation of the main cell region image in each frame in order to screen out standard morphological feature images.

[0009] S3. Based on the selected standard morphological feature images, analyze the integrity of the target single-cell structure contained in the standard morphological feature images and determine the cell integrity index.

[0010] S4. Compare and analyze the cell integrity index to determine whether the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, and issue and execute the corresponding level of printing control command.

[0011] Preferably, step S1 specifically includes:

[0012] S11. Before performing jet printing on the cell suspension, an industrial camera set in the center of the field of view of the microfluidic channel continuously acquires images of the cell suspension containing the target single cell at a preset sampling frequency. After grayscale processing, a cell suspension image frame sequence is constructed, and the main cell region image of each frame is determined by the projection image range of the target single cell in each frame.

[0013] S12. Based on the pixel resolution of the image, divide the main cell region image of each frame into pixels and record the gray value of each pixel in each frame image to construct a pixel gray value parameter set, wherein the pixel gray value dataset includes the gray value of each pixel in each frame of the main cell region image.

[0014] S13. Input the cell main region images of each frame into a pre-trained shallow convolutional network, and perform structural scanning on the cell main region images through multiple convolutional kernels respectively. Extract the maximum value of the structural response output of all convolutional kernels at the corresponding pixel positions as the structural response value of the corresponding pixel, and construct a pixel response parameter set. The pixel response dataset includes the structural response value of each pixel in each cell main region image.

[0015] Preferably, in step S14, based on the constructed pixel grayscale dataset, a pixel abrupt change parameter set is constructed by performing local gradient analysis on the grayscale distribution of pixels surrounding each pixel, specifically including:

[0016] Centered on the current pixel, a 3×3 pixel matrix surrounding the current pixel is selected as the local region input. Convolution operations are performed with Sobel operators in two standard directions. Specifically, the gray values ​​of each pixel in the pixel matrix are multiplied by their corresponding Sobel weights and then summed to obtain the horizontal gray value change Gx and the vertical gray value change Gy of the current pixel. The standard directions include the horizontal and vertical directions.

[0017] The horizontal grayscale change value Gx and the vertical grayscale change value Gy of the current pixel are correlated. After calculating the two-dimensional gradient magnitude, the local edge abrupt change coefficient of the current pixel is determined, specifically as follows: In the formula, This represents the local abrupt change coefficient of the current pixel's edge.

[0018] Based on the determined edge local mutation coefficient of the current pixel, the edge local mutation coefficient of each pixel in the main cell region image of each frame is calculated sequentially, and a pixel mutation parameter set is constructed.

[0019] Preferably, step S2 specifically includes:

[0020] S21. Perform feature recognition on the constructed pixel grayscale parameter set, extract the grayscale value of each pixel in the main cell region image of each frame, and combine the statistical mean value algorithm to obtain the grayscale mean value of the main cell region image of each frame.

[0021] S22. Based on the constructed pixel grayscale parameter set, pixel response parameter set, and pixel mutation parameter set, correlation is performed, and after dimensionless processing, the morphological fluctuation degree of the main cell region image in each frame is analyzed to determine the morphological tension value of the main cell region image in each frame, specifically: In the formula, This represents the morphological tension value of the main cell region in the i-th frame image. This represents the grayscale value of the j-th pixel in the i-th frame of the main cell region image. This represents the average grayscale value of the main cell region image in the i-th frame. This represents the structural response value of the j-th pixel in the i-th frame of the main cell region image. This represents the local abrupt change coefficient of the edge at the j-th pixel in the i-th frame of the main cell region image. This indicates the number of pixels.

[0022] Preferably, in step S23, the standard morphological feature image is determined by comparing and analyzing the morphological tension values ​​of the acquired cell main region images, specifically including:

[0023] Based on the morphological tension values ​​of each frame of the main cell region image, and combined with a statistical averaging algorithm, the mean morphological tension value is obtained. The difference between the morphological tension value of each frame of the main cell region image and the mean morphological tension value is calculated to obtain the morphological tension difference value of each frame of the main cell region image. After feature extraction, the main cell region image corresponding to the smallest morphological tension difference value is selected as the standard morphological feature image.

[0024] Preferably, step S3 specifically includes:

[0025] S31. Based on the determined standard morphological feature image, combined with the pixel response parameter set, extract the structural response value of each pixel in the standard morphological feature image, and after performing the extreme value search operation, calculate the difference between the maximum and minimum structural response values ​​in the obtained standard morphological feature image to determine the structural response amplitude of the standard morphological feature image.

[0026] S32. Based on the standard morphological feature image and combined with the pixel mutation parameter set, extract the edge local mutation coefficient of each pixel in the standard morphological feature image, compare the edge local mutation coefficient of each pixel with the preset mutation threshold, filter out the corresponding pixel that the edge local mutation coefficient exceeds the mutation threshold, mark it as edge pixel, and record the gray value of each edge pixel.

[0027] S33. After performing a second-order Laplacian transform on the gray values ​​of each edge pixel, obtain the second-order transform value of each edge pixel, and combine it with a statistical mean-calculation algorithm to determine the boundary sharpening index of the standard morphological feature image.

[0028] Preferably, in step S34, the morphological tension value, structural response amplitude, and boundary sharpening index of the standard morphological feature image are correlated. After dimensionless processing, the integrity of the target single-cell structure contained in the standard morphological feature image is analyzed to determine the cell integrity index, specifically: In the formula, Indicates the cell integrity index. Indicates the morphological tension value. Indicates the amplitude of the structural response. This indicates the boundary sharpening index.

[0029] Preferably, step S4 specifically includes:

[0030] S41. Compare and analyze the cell integrity index with the preset integrity threshold to determine whether the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, and issue the corresponding level of printing control command, specifically including:

[0031] If the cell integrity index exceeds the integrity threshold, it means that the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, and at this time, a first-level printing control command is issued.

[0032] If the cell integrity index does not exceed the integrity threshold, it means that the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, and at this time, a secondary printing control command is issued.

[0033] Preferably, S42, the corresponding printing control operation is executed according to the received primary and secondary printing control instructions, specifically including:

[0034] When a first-level printing control command is received, the following actions are executed: This indicates that the target single-cell structure in the cell suspension has not been damaged, and the nozzle structure is adjusted by the piezoelectric microvalve to perform a jet printing operation on the cell suspension containing the complete target single cell;

[0035] When a secondary printing control command is received, the following actions are taken: if the target single-cell structure in the cell suspension is damaged, the cell suspension is replaced, and the S1-S4 process is repeated until a primary printing control command is issued. The nozzle structure is then adjusted via a piezoelectric microvalve to perform a jet printing operation on the cell suspension containing the complete target single cell.

[0036] The control system for a single-cell printing device optimized based on a convolutional neural network includes a pixel parameter acquisition module, a feature selection module, a structure analysis module, and a printing control module.

[0037] The pixel parameter acquisition module is used to continuously acquire images of cell suspension containing target single cells before the cell suspension is sprayed and printed, in order to determine the main cell region image of each frame. After being input into a shallow convolutional network, the pixel grayscale parameter set, pixel response parameter set and pixel mutation parameter set are constructed respectively.

[0038] The feature filtering module is used to analyze the degree of morphological fluctuation in the main cell region images of each frame based on the pixel grayscale parameter set, pixel response parameter set, and pixel mutation parameter set, so as to filter out standard morphological feature images.

[0039] The structural analysis module is used to analyze the integrity of the target single-cell structure contained in the selected standard morphological feature images and determine the cell integrity index.

[0040] The printing control module is used to compare and analyze the cell integrity index to determine whether the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, so as to issue and execute the corresponding level of printing control command.

[0041] This invention provides a control method and system for a single-cell printing device based on convolutional neural network optimization, which has the following beneficial effects:

[0042] (1) By constructing a control method and system for single-cell printing equipment based on convolutional neural network optimization, the accuracy of target single-cell structure recognition and the adaptive capability of printing control are significantly improved. This method integrates the multi-kernel structure perception mechanism of shallow convolutional network. Before printing, based on the continuous image frame sequence, it constructs three types of pixel parameter sets: grayscale characteristics, structural response and edge mutation. It also realizes the dynamic quantitative expression of cell morphological tension through dimensionless processing. On this basis, it can not only accurately screen standard morphological feature images, but also construct a cell integrity index that reflects the integrity of single-cell structure, and issue hierarchical printing control commands accordingly. Overall, it effectively breaks through the limitations of traditional methods that rely on single-frame images and are difficult to evaluate structural integrity. It realizes accurate identification and intelligent decision control of intact cells in single-cell suspensions with strong structural complexity, blurred boundaries and a lot of interference information, thereby improving the accuracy, stability and operating efficiency of printing, and providing key technical support for single-cell printing control and its high-precision application.

[0043] (2) The proposed pixel parameter acquisition and convolution feature extraction mechanism forms three nonlinear structure perception channels—grayscale, response, and mutation—during the image acquisition stage, significantly enhancing the ability to distinguish microstructures in cell images, especially forming a clear discrimination interface between the target single cell and its surrounding background, damaged fragments, or blurred edge areas. Traditional image processing methods generally rely on low-dimensional features such as grayscale enhancement and edge sharpening. When faced with image information with high cell contour transparency and strong morphological fluctuations, edge overgeneralization or structural mis-extraction may occur, leading to incorrect identification of target units. However, this invention introduces a shallow convolutional network structure and combines it with local gradient mutation calculation, which can extract the mutation coefficient of cell edges at the pixel level, thereby constructing a more differentiated set of pixel mutation parameters. This process significantly improves the system's ability to model the small deformations and damaged edge details of intact cells, avoiding the misidentification problem caused by unclear boundary judgment in traditional methods, and achieving highly reliable and accurate locking of target single cells, laying a stable foundation for subsequent integrity judgment and printing control.

[0044] (3) By constructing a cell integrity index, the morphological tension value, structural response amplitude and boundary sharpening index are normalized and fused to form a quantifiable and comparable printing decision-making basis parameter, which effectively makes up for the defects of single control strategy and insufficient feedback response in the existing technology. In traditional methods, printing execution mostly relies on single frame image judgment, which ignores the dynamic fluctuation of cell morphology in the time sequence and does not evaluate the cell structure state, resulting in printing failure, nozzle mis-spraying and the need to frequently replace cell suspension. In contrast, this invention, while constructing cell integrity index parameters, also sets integrity threshold, divides the first and second level of printing control, and automatically executes printing or replacement operations based on control commands to realize the dynamic closed loop of the printing control link. This mechanism can not only continuously evaluate the target cell state under high frequency sampling, but also make intelligent adjustments according to the current recognition quality, ensuring that the execution of printing commands has higher stability, robustness and real-time response capability, and meets the control accuracy requirements in precision manufacturing scenarios. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the control method for a single-cell printing device based on convolutional neural network optimization according to the present invention;

[0046] Figure 2 This is a block diagram of the control system for the single-cell printing device optimized based on convolutional neural networks according to the present invention;

[0047] Figure 3 This is a logic diagram of the single-cell printing device control method based on convolutional neural network optimization according to the present invention.

[0048] Figure 4 This is a simplified flowchart of the control process of the single-cell printing device optimized based on convolutional neural networks according to the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Example 1

[0051] Please see Figure 1 , Figure 3 and Figure 4 This invention provides a control method for a single-cell printing device based on convolutional neural network optimization, comprising the following steps:

[0052] S1. Before performing jet printing on the cell suspension, continuously acquire images of the cell suspension containing the target single cell to determine the main cell region image of each frame. After inputting into a shallow convolutional network, construct pixel grayscale parameter set, pixel response parameter set and pixel mutation parameter set respectively.

[0053] S2. Based on the pixel grayscale parameter set, pixel response parameter set and pixel mutation parameter set, analyze the degree of morphological fluctuation of the main cell region image in each frame in order to screen out standard morphological feature images.

[0054] S3. Based on the selected standard morphological feature images, analyze the integrity of the target single-cell structure contained in the standard morphological feature images and determine the cell integrity index.

[0055] S4. Compare and analyze the cell integrity index to determine whether the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, and issue and execute the corresponding level of printing control command.

[0056] In this embodiment, the single-cell printing device control method based on convolutional neural network optimization effectively solves the key technical bottlenecks of existing technologies in single-cell recognition and printing control, such as insufficient recognition accuracy, one-sided structural judgment, and lack of multi-parameter fusion decision-making. First, by introducing continuous frame image acquisition and shallow convolutional network analysis before printing through S1, the method no longer relies on single-frame image judgment, effectively improving the comprehensive perception capability of cell morphological temporal fluctuations. On this basis, the pixel grayscale parameter set, pixel response parameter set, and pixel mutation parameter set are constructed to form a three-dimensional feature expression channel for cell structure, significantly enhancing the ability to identify and distinguish blurred cell boundary regions, minor deformations, and damaged structures. Through S2, the above three types of parameter sets are fused and analyzed, and the image can be quantitatively extracted. The method first extracts the morphological tension value and selects the most representative standard morphological feature images, avoiding judgment bias caused by local abnormal frames. Then, S3 further introduces the construction mechanism of cell integrity index, realizing the quantitative evaluation of structural integrity before printing for the first time, providing a stable and reliable judgment basis for subsequent printing operations. Finally, in S4, the integrity index is compared and analyzed with a preset threshold, and hierarchical printing control commands are issued, enabling closed-loop regulation capability based on adaptive decision-making according to cell state. Overall, this method significantly improves the accuracy and stability of pre-printing recognition, makes up for the shortcomings of traditional methods in terms of insufficient image modeling depth, lagging judgment of structural damage, and single control decision, and provides a systematic and intelligent printing control path for high-precision single-cell printing.

[0057] Example 2

[0058] Please refer to Figure 1 and Figure 3 Specifically, the steps in S1 include:

[0059] S11. Before performing jet printing on the cell suspension, an industrial camera set in the center of the field of view of the microfluidic channel continuously acquires images of the cell suspension containing the target single cell at a preset sampling frequency. After grayscale processing, a cell suspension image frame sequence is constructed, and the main cell region image of each frame is determined by the projection image range of the target single cell in each frame.

[0060] The main cell region image in each frame refers to the local image region containing the main structure of the target single cell extracted from each frame of a continuously acquired sequence of cell suspension images. It represents the spatial region in the image most closely related to the actual existence and structural features of the target single cell. Its function is to remove irrelevant background or other interfering cell components from the image, so that subsequent image parameter extraction and structural analysis are more focused on the target unit, thereby improving recognition accuracy and processing efficiency. The method of obtaining this region is as follows: First, an industrial camera set in the center of the microfluidic channel is used to continuously capture the suspension at high frequency to obtain a series of grayscale image frames. Then, in each frame, the projection range of the target single cell is identified and located. The entire frame image is cropped according to the projection range to form the main cell region image. This processing not only ensures that the image parameter construction and convolutional network analysis focus on the actual effective cell structure, but also provides a unified spatial reference for the subsequent construction of grayscale, structural response, and mutation parameter sets.

[0061] S12. Based on the pixel resolution of the image, divide the main cell region image of each frame into pixels and record the gray value of each pixel in each frame image to construct a pixel gray value parameter set, wherein the pixel gray value dataset includes the gray value of each pixel in each frame of the main cell region image.

[0062] The grayscale value of each pixel in the main cell region image of each frame refers to the brightness intensity value corresponding to each pixel in each frame of the main cell region image. It is usually represented by an integer between 0 and 255 and is used to reflect the brightness of the pixel in the image. The role of these grayscale values ​​is to accurately characterize the internal density, contour boundary and morphological changes of the target single cell structure. It is the basic data source for subsequent image structure recognition, response analysis and mutation detection. The acquisition process is as follows: First, the main region image of each frame is divided into grids according to the pixel resolution of the image, and it is decomposed into a matrix composed of several pixels. Then, the original color image is converted into a single-channel grayscale image through grayscale processing algorithm. The grayscale value of each pixel is read point by point and recorded in sequence. Finally, a pixel grayscale parameter set containing the grayscale information of all pixels in all frames of images is constructed. This parameter set not only reflects the intensity distribution characteristics of the cell image in space, but also provides a stable numerical basis for subsequent morphological fluctuation analysis and standard image selection.

[0063] S13. Input the cell main region images of each frame into a pre-trained shallow convolutional network, and perform structural scanning on the cell main region images through multiple convolutional kernels respectively. Extract the maximum value of the structural response output of all convolutional kernels at the corresponding pixel position as the structural response value of the corresponding pixel, and construct a pixel response parameter set. The pixel response dataset includes the structural response value of each pixel in each cell main region image.

[0064] A pre-trained shallow convolutional network is used to extract multi-scale structural features from the main cell region images of each frame. Its core function is to perceive and quantify the local structural patterns, edge continuity, and texture features inside the cell by sliding the convolutional kernel across the image, thus providing a highly sensitive basis for the subsequent generation of structural response values. Compared with traditional image processing methods, shallow convolutional networks have the advantages of being lightweight, low-latency, and suitable for real-time processing. Moreover, its network depth is controlled within a limited range, which is more conducive to preserving subtle structural changes in cell images rather than generalization learning. The training process of this convolutional network involves collecting a large number of single-cell image samples with labeled information. Under the supervised learning framework, the structural recognition accuracy or feature response consistency is used as the optimization objective. Combining cross-entropy or mean squared error, the convolutional kernel weights are iteratively updated through the gradient backpropagation algorithm. Finally, a stable set of network parameters that can effectively capture the typical structural response features of single-cell images is obtained, thus demonstrating strong recognition adaptability and robustness for unknown images in practical applications.

[0065] The structural response value of each pixel in the main cell region image of each frame is formed by inputting the main cell region image into a pre-trained shallow convolutional neural network. For each pixel location, multiple convolutional kernels perceive and extract features from its surrounding structure. The maximum value is taken from the output of a set of structural feature responses, reflecting the response intensity of the pixel to the structural pattern in the local region. Its function is to capture the deep features such as potential structural boundaries, morphological contours, and organelle distribution in cell images, effectively overcoming the limitations of traditional grayscale values ​​that can only reflect brightness and lack structural perception. It provides more discriminative feature support for recognizing cell integrity, boundary continuity, and degree of deformation. The process of obtaining the structural response value is as follows: First, the main region image of each frame is input into the pre-trained shallow convolutional network model; then, the image is slid-scanned by multiple convolutional kernels of different sizes or different orientation perception capabilities. Each convolutional kernel calculates a response value for each pixel and its neighborhood; finally, at each pixel location, the maximum value is selected from the output responses of all convolutional kernels as the structural response value of that point, and a pixel response parameter set is constructed accordingly, thereby realizing the structural perception modeling of all pixels in the cell region.

[0066] S14. Based on the constructed pixel grayscale dataset, a pixel abrupt change parameter set is constructed by performing local gradient analysis on the grayscale distribution of pixels surrounding each pixel, specifically including:

[0067] Centered on the current pixel, a 3×3 pixel matrix surrounding the current pixel is selected as the local region input. Convolution operations are performed with Sobel operators in two standard directions. Specifically, the gray values ​​of each pixel in the pixel matrix are multiplied by their corresponding Sobel weights and then summed to obtain the horizontal gray value change Gx and the vertical gray value change Gy of the current pixel. The standard directions include the horizontal and vertical directions.

[0068] The horizontal grayscale change value Gx and the vertical grayscale change value Gy of the current pixel are calculated based on the spatial distribution of grayscale values ​​within its 3×3 neighborhood, centered on that pixel. These values ​​represent the intensity of brightness changes in the horizontal and vertical directions, respectively, and are key indicators for measuring the magnitude of the grayscale gradient of the pixel in the two main directions. They reflect the degree of edge changes and directional characteristics of local regions in an image, particularly useful in cell images for identifying cell boundaries, damaged edges, or mutation points, effectively assisting in determining the integrity and edge continuity of target structures. The method for obtaining these values ​​is as follows: First, extract the grayscale value matrix of the surrounding 3×3 pixels centered on the current pixel. Then, convolve this matrix with the Sobel operators (classic convolution kernels used for edge detection) in two standard directions, multiplying the grayscale value of each pixel in the matrix by the corresponding value in the Sobel weight matrix and summing the results to obtain the grayscale change values ​​of the current pixel in the horizontal direction Gx and the vertical direction Gy. Combining these two values ​​can be further used to calculate the degree of local mutation, providing basic data for the subsequent construction of mutation parameter sets.

[0069] The horizontal grayscale change value Gx and the vertical grayscale change value Gy of the current pixel are correlated. After calculating the two-dimensional gradient magnitude, the local edge abrupt change coefficient of the current pixel is determined, specifically as follows: In the formula, This represents the local abrupt change coefficient of the current pixel's edge.

[0070] It should be noted that the above formula is an important process for calculating the edge local mutation coefficient of the current pixel in the image. Here, Gx and Gy represent the gray-level change values ​​of the pixel in the horizontal and vertical directions, respectively. This formula comprehensively evaluates the intensity of pixel gray-level change through two-dimensional gradient modulus, thereby accurately characterizing its local edge mutation characteristics. This calculation plays a key role in this invention: by accurately extracting the gray-level mutation features of the boundary region in the image, it significantly improves the ability to identify edge structures in single-cell images. Especially in the context of blurred cell suspension boundaries and complex interference from damage, as mentioned in the background technology, it can effectively solve the problem that traditional methods have difficulty distinguishing between intact cells and broken structures. As a parameter characterizing the sensitivity of pixel boundaries in cell images, the edge local mutation coefficient can be used to screen and locate structural edge pixels, and can also provide a quantitative basis for the subsequent construction of boundary sharpening index and cell integrity evaluation. Thus, it plays a fundamental supporting role in achieving high-precision structure recognition, controlling jetting decisions, and improving printing consistency. It is a key parameter that runs through the core image analysis process of this printing control method.

[0071] Based on the determined edge local mutation coefficient of the current pixel, the edge local mutation coefficient of each pixel in the main cell region image of each frame is calculated sequentially, and a pixel mutation parameter set is constructed.

[0072] In this embodiment, through the multi-step design of S1, a high-dimensional pixel parameter construction mechanism integrating continuous image acquisition, deep structure recognition, and edge mutation analysis was established before single-cell printing. This significantly improved the image expression quality and microstructure recognition capability of the target cells, overcoming the problem that existing technologies rely solely on single-frame images and lack deep structural perception. First, S11 uses an industrial camera to perform continuous high-frequency sampling in the core region of the microfluidic channel, effectively capturing the overall picture and dynamic features of the cells during their movement in the suspension, providing a temporally stable image foundation for subsequent structural analysis. Next, S12 constructs a pixel grayscale parameter set to fully record the grayscale distribution details within the main cell region, laying a statistical foundation for subsequent morphological change trend analysis. S13 introduces a pre-trained shallow convolutional network, using a multi-convolutional kernel structure scanning mechanism to improve... By capturing the response intensity of each pixel at the structural level, the ability to identify local structural features, contour clarity, and morphological differences of cells is significantly enhanced. In S14, the Sobel operator is combined to perform bidirectional gradient convolution analysis on a local 3×3 pixel region, accurately capturing the mutation behavior of cell edges and ultimately forming a set of pixel mutation parameters, which makes up for the deficiency of traditional grayscale analysis in quantifying local boundary changes. Overall, this group of steps not only breaks through the technical limitations of extracting cell information from a single image dimension or static frame, but also establishes an image parameter system that integrates grayscale characteristics, structural response, and edge mutation. It can still achieve high-precision modeling of target single-cell structures under complex background interference and high-transparency cell boundary blurring conditions, providing a reliable, accurate, and multi-source feature basis for subsequent integrity judgment and printing control.

[0073] Example 3

[0074] Please refer to Figure 1 and Figure 3 Specifically, the steps in S2 include:

[0075] S21. Perform feature recognition on the constructed pixel grayscale parameter set, extract the grayscale value of each pixel in the main cell region image of each frame, and combine the statistical mean value algorithm to obtain the grayscale mean value of the main cell region image of each frame.

[0076] S22. Based on the constructed pixel grayscale parameter set, pixel response parameter set, and pixel mutation parameter set, correlation is performed, and after dimensionless processing, the morphological fluctuation degree of the main cell region image in each frame is analyzed to determine the morphological tension value of the main cell region image in each frame, specifically: In the formula, This represents the morphological tension value of the main cell region in the i-th frame image. This represents the grayscale value of the j-th pixel in the i-th frame of the main cell region image. This represents the average grayscale value of the main cell region image in the i-th frame. This represents the structural response value of the j-th pixel in the i-th frame of the main cell region image. This represents the local abrupt change coefficient of the edge at the j-th pixel in the i-th frame of the main cell region image. This indicates the number of pixels.

[0077] It should be noted that the formula in S22 is the core expression used to calculate the morphological tension value of the main cell region image in each frame. It forms a quantitative measure of the overall morphological stability and variation amplitude of the image by weighting and coupling pixel grayscale deviation, structural response value, and edge local abrupt change coefficient. In this formula, This represents the average grayscale value of the main cell region image in the i-th frame, reflecting the degree of grayscale shift of that pixel. This represents the edge local mutation coefficient of the j-th pixel in the i-th frame of the cell's main region image. It reflects the image's edge sensitivity and structural mutation status. Through the composite calculation of these three types of parameters, the resulting morphological tension value accurately characterizes the stability and dynamic fluctuation of the microstructure in each frame of the image. This calculation logic is closely related to the problem pointed out in the background technology that traditional methods are difficult to capture the dynamic fluctuation of cell morphology within a time series. This formula can identify small but crucial deformation trends in multi-frame sequence images, effectively extract morphological fluctuation feature images, and provide a precise basis for subsequent selection of the most stable and representative standard morphological images. At the same time, the morphological tension value, as a measure of the degree of structural change in a single frame image, can not only help identify abnormal image frames, but also serve as a core intermediate quantity for multi-parameter fusion, playing a key supporting role in improving the overall recognition accuracy and the intelligence of control decisions. It is an important bridge connecting dynamic structure analysis and judgment execution.

[0078] In image processing and cell morphology recognition research, traditional methods rely solely on grayscale values ​​or edge detection metrics to measure structural changes. However, in complex cell images, single grayscale statistics or boundary sharpening often fail to simultaneously reflect the overall structural balance and the intensity of local mutations. Therefore, a new metric is needed that can comprehensively reflect local pixel deviation, structural response, and boundary perturbation. The morphological tension value was proposed in this context. The theoretical basis of this formula comes from the idea of ​​coupling tensor fields with local image responses. Pixel grayscale deviation is used as tensor strain, convolution response results are mapped as structural forces, and boundary perturbation coefficients are used as local constraints. The product term formed by the coupling of these three factors is averaged after overall pixel normalization to obtain the morphological tension value, which is used to describe the morphological stability of the main cell region image in each frame. Therefore, by analogy with the idea of ​​mechanical tension, the calculation method of the morphological tension value is derived from three dimensions: image grayscale dispersion (deviation), structural contribution (response), and boundary perturbation (constraint).

[0079] The specific application of the above-mentioned formula for calculating morphological tension values ​​in this invention is as follows:

[0080] Image acquisition stage output: After each image frame is acquired, grayscale extraction and structural analysis are performed on the main region, and morphological tension values ​​are constructed;

[0081] Tension value calculation and screening: morphological tension values ​​are calculated for each frame of the main cell region image, which are then used for screening standard morphological feature images;

[0082] The standard morphological feature image is selected based on the following criteria: the morphological tension value of each frame's main cell region image is calculated by comparing it with the mean morphological tension value, and the cell region image corresponding to the smallest morphological tension difference is selected as the standard morphological feature image.

[0083] Specifically, S23, based on the comparative analysis of the morphological tension values ​​of the acquired cell main region images in each frame, standard morphological feature images are determined, including:

[0084] Based on the morphological tension values ​​of each frame of the main cell region image, and combined with a statistical averaging algorithm, the mean morphological tension value is obtained. The difference between the morphological tension value of each frame of the main cell region image and the mean morphological tension value is calculated to obtain the morphological tension difference value of each frame of the main cell region image. After feature extraction, the main cell region image corresponding to the smallest morphological tension difference value is selected as the standard morphological feature image.

[0085] The standard morphological feature image is selected from multiple consecutively acquired images of the main cell region. After analyzing and comparing morphological tension values, the image that best represents the overall stable morphological trend and is most representative of the structure is chosen. This image reflects the most complete and morphologically balanced state of the target single cell in the current time series. Its purpose is to provide a unified and reliable image basis for subsequent structural integrity assessment, avoiding judgment bias caused by sudden blurring, edge interference, or local anomalies in a single frame image. The specific selection process is as follows: first, the morphological tension value of all frames is calculated and the mean is obtained. Then, the tension value of each frame image is subtracted from the mean. The smaller the difference, the closer the morphology of the image is to the overall trend and the less fluctuation there is. Finally, the image with the smallest morphological tension difference is selected as the standard morphological feature image. Through this selection mechanism, the best analysis object is automatically determined from the time series images, effectively improving the stability of structural analysis and the accuracy of judgment.

[0086] In this embodiment, the grayscale, response, and mutation fusion analysis and morphological tension evaluation mechanism set in S2 effectively overcomes the key technical shortcomings of traditional single-cell printing control, which relies solely on a single image frame for judgment and ignores temporal morphological fluctuations. This achieves quantitative evaluation of the dynamic stability of cell structures and automatic selection of the optimal image. Specifically, S21 establishes an overall brightness baseline for each frame image through full-frame extraction and statistical mean analysis of the pixel grayscale parameter set, serving as the initial parameter for measuring changes in cell morphological distribution. In S22, for the first time, grayscale values, structural response values, and edge mutation coefficients are normalized under the same index system, and a morphological tension value is comprehensively constructed, achieving a unified measure of the intensity of image morphological fluctuations. This value can sensitively reflect the presence of cells in consecutive frame images. This module identifies anomalous behaviors such as mutations, blurring, damage, or shifts, overcoming the limitations of traditional methods that rely on local edge features and are susceptible to misjudgment due to background interference. The subsequent S23 uses the statistical mean as a reference benchmark to compare the morphological tension differences between each frame, automatically selecting the frame with the most stable morphology and structure that best meets the desired standard as the standard morphological feature image. This step chain not only enables automatic identification of the highest-quality frame within an image group but also effectively avoids the risk of erroneous decisions caused by single-frame anomalies. It exhibits stronger robustness and accuracy in practical applications involving complex cell morphology and significant structural fluctuations. Therefore, this module demonstrates unique technical advantages in addressing the challenges of identifying structural damage and lacking quantitative evidence for dynamic image changes, significantly improving the decision-making basis for subsequent structural judgment and printing control.

[0087] Example 4

[0088] Please refer to Figure 1 and Figure 3 Specifically, the S3 steps include:

[0089] S31. Based on the determined standard morphological feature image, combined with the pixel response parameter set, extract the structural response value of each pixel in the standard morphological feature image, and after performing the extreme value search operation, calculate the difference between the maximum and minimum structural response values ​​in the obtained standard morphological feature image to determine the structural response amplitude of the standard morphological feature image.

[0090] The structural response amplitude is the difference between the maximum and minimum structural response values ​​of all pixels in a standard morphological feature image. This value reflects the intensity fluctuation range of structural features in the spatial distribution of the image and is an important indicator for measuring the overall structural clarity and configurational integrity of the image. Its function is to characterize whether the target single cell has clear and continuous structural boundaries and significant internal configurational differences in the current image, thereby helping to determine whether the cell has the structural integrity suitable for printing. The acquisition method is as follows: First, in the standard morphological feature image, combined with the pixel response parameter set, extract the structural response values ​​of all pixels; then, through extreme value search, find the maximum and minimum response values, and calculate the difference between the two, which is the structural response amplitude. The larger the value, the richer the structural details and the clearer the boundaries in the image, which helps to accurately distinguish the target cell from the surrounding background or interfering structures. It is one of the key components of the cell integrity index.

[0091] S32. Based on the standard morphological feature image and combined with the pixel mutation parameter set, extract the edge local mutation coefficient of each pixel in the standard morphological feature image, compare the edge local mutation coefficient of each pixel with the preset mutation threshold, filter out the corresponding pixel that the edge local mutation coefficient exceeds the mutation threshold, mark it as edge pixel, and record the gray value of each edge pixel.

[0092] The mutation threshold is set adaptively by statistical analysis of the gray-level gradient features of edge and non-edge regions in a large number of cell images, combined with the principle of optimal edge recognition accuracy, and using empirical judgment or optimization algorithms.

[0093] Edge pixels refer to pixels in a standard morphological feature image that exhibit drastic local grayscale changes and significant edge features. Specifically, they are identified by pixels whose local mutation coefficient exceeds a preset mutation threshold, typically located at the boundary contour of the target single-cell structure. Their function is to accurately locate the outer boundary region of the cell, facilitating the assessment of cell structure integrity and continuity, and providing crucial information for subsequent boundary sharpening analysis and integrity calculation. The acquisition process involves: extracting the local mutation coefficients of all pixels in the standard image from the pixel mutation parameter set, i.e., the local change intensity index calculated through grayscale gradient analysis; then comparing the mutation coefficient of each pixel with the system-set mutation threshold, identifying pixels with local mutation coefficients exceeding the threshold as having significant edge mutations and thus being determined as edge pixels; finally, recording the grayscale values ​​of these pixels for subsequent boundary sharpening index calculation. This step effectively eliminates interference from homogeneous regions within the image, focusing on extracting contour region change features, and is a crucial foundation for identifying cell structure damage, boundary discontinuities, or abnormal deformations.

[0094] S33. After performing a second-order Laplacian transform on the gray values ​​of each edge pixel, obtain the second-order transform value of each edge pixel, and combine it with a statistical mean-calculation algorithm to determine the boundary sharpening index of the standard morphological feature image.

[0095] The boundary sharpening index is an index obtained by averaging the second-order Laplacian transform values ​​of edge pixels in a standard morphological feature image. It quantifies the clarity and sharpness of cell edge contours in an image. Its core function is to assess the integrity and continuity of the target single-cell structure boundary, determining whether it has a clear and undamaged geometric boundary shape. It is a key factor in determining whether a cell structure is suitable for printing. The index is obtained as follows: First, the edge pixels selected in the previous step are extracted, and their corresponding gray values ​​are obtained. Then, a second-order Laplacian transform is performed on these gray values, utilizing its sensitivity to the second derivative of gray-level changes to enhance the response of the contour region in the image, thereby highlighting the structural boundary. Finally, the average value of the Laplacian transform results of all edge pixels is calculated, and the resulting value is the boundary sharpening index. The larger the value, the more drastic the gray-level changes in the edge region and the clearer and more complete the boundary; conversely, a smaller value indicates blurred cell edges and potential damage, helping the system automatically identify printing risks and make accurate control decisions.

[0096] Specifically, S34, the morphological tension value, structural response amplitude, and boundary sharpening index of the standard morphological feature image are correlated. After dimensionless processing, the integrity of the target single-cell structure contained in the standard morphological feature image is analyzed to determine the cell integrity index, specifically: In the formula, Indicates the cell integrity index. Indicates the morphological tension value. Indicates the amplitude of the structural response. This represents the boundary sharpening index, where, This is expressed as a logarithmic function reflecting cell morphological tension, used to quantify the degree of structural fluctuation. The exponential decay factor represents the amplitude of the structural response. A larger structural response amplitude indicates more drastic differences in internal structure and lower cell integrity. The inverse measure of the boundary sharpness index is the larger the boundary sharpness index, the clearer the boundary and the higher the cell integrity.

[0097] It should be noted that the formula in S34 is a key quantitative indicator for calculating cell structural integrity, namely the cell integrity index. This index assesses whether the target single-cell structure in a standard morphological feature image is printable. The formula comprehensively considers three dimensions: the degree of image morphological change, structural response intensity, and edge sharpening quality, forming the basis for intelligent structural recognition. As mentioned in the background technology, traditional recognition methods struggle to quantitatively assess the integrity of single-cell structures, often leading to misjudgments due to a lack of multi-dimensional data fusion. This formula couples and calculates multiple key visual parameters, effectively solving the problems of difficulty in determining damaged cell structures and misjudging complete structures. By comprehensively measuring the morphological tension, structural response, and edge sharpness of cells in the image, this index accurately reflects the cell structural integrity, thus serving as the core quantitative basis for printing decisions. The cell integrity index is a core indicator for comprehensively evaluating whether the identified target single cell in a standard morphological feature image possesses structural integrity, edge sharpness, and texture stability. This indicator provides a quantitative standard before determining whether to proceed with printing. This mechanism significantly improves recognition accuracy and printing control intelligence, and is a fundamental evaluation mechanism supporting high-throughput, high-reliability single-cell printing processes.

[0098] The formula for the cell integrity index is based on the technical requirement of parallel evaluation of three key factors in the cell structure integrity analysis process: first, the overall gray-level tension of the image, i.e., whether the pixel gray-level fluctuation amplitude within the cell region reflects continuous and regular changes; second, the perception of deep image structure, i.e., whether the activation response of the convolutional neural network to the input structure in the middle and later layers is stable and balanced; and third, the sharpness level of the boundary, i.e., whether the cell edge has a continuous and clear trend of pixel gradient change. Among these three, gray-level tension is a feature indicator at the static image level, the structural response amplitude reflects the characteristics of the deep feature extraction module, and the edge sharpness indicator is a feature indicator at the image level. Like the analytical performance of microscopic local features; in traditional image evaluation, these three are often judged separately, while the fusion algorithm of cell integrity index is used to respond to the above three core characteristics simultaneously, and complete the index construction without relying on weights or introducing normalization coefficients; the formula introduces the natural logarithm function, exponential function and fractional expression, so that the effect of each parameter on the final result can maintain the nonlinear response characteristics, and will not dominate the output in absolute value; such function structure has balance and decay characteristics, which is the theoretical mathematical basis for realizing reasonable evaluation of the response of complex image structures, and has the composite processing capability of fusing neural network feature layer output and low-order image features;

[0099] The three parts constitute a stepwise suppression type product function structure, with all terms being positive and exhibiting a nonlinear response relationship; its derivation path is as follows: By performing compression mapping on the morphological tension value, its influence on the output is kept at the boundary convergence when the tension is high; the second term By using an exponential decay mechanism to constrain the integrity index through internal response mutations, the misclassification of images from structurally abrupt regions in the activation layer as complete images is intuitively suppressed; the third term The inverse expression of the edge response term is constructed using the boundary convergence ratio method, and the value of this term decreases as the edge sharpening is enhanced. By adding a constant term of 2 to the denominator, the problem of near zero caused by the Zrh minimum is avoided, ensuring that the blurred edge image is not misjudged as complete. The entire formula avoids the use of fixed weighting coefficients. All factors are independent of each other but have a combined response, which has mathematical differentiability and continuity, and can be directly used for subsequent threshold judgment and strategy selection.

[0100] In this embodiment, through deep structural analysis and edge detail enhancement processing of standard morphological feature images, a multi-dimensional quantitative expression of single-cell structural integrity is realized for the first time, providing a core evaluation index with discriminative and adaptive capabilities for high-precision printing control. In S31, extreme value search is performed on the standard image based on the pixel response parameter set to extract and calculate the difference between the maximum and minimum structural response values, i.e., the structural response amplitude. This parameter can reflect the overall structural fluctuations and morphological stability of the cell in the main region image, and is particularly suitable for identifying local ruptures, collapses, or aberrations. Subsequently, S32 and S33 further introduce the pixel mutation parameter set to extract grayscale and perform second-order Laplacian transform on edge pixels in the image that exceed the mutation threshold, ultimately generating boundary sharpening. The index, which can keenly capture the clarity and integrity of cell edges, effectively distinguishes between intact structures and damaged fragments or interference artifacts, is used in conjunction with the S34 index. This index, in a dimensionless manner, integrates morphological tension values, structural response amplitudes, and boundary sharpening indices to form a unified integrity index system, enabling numerical characterization of the target single-cell structural state. This mechanism effectively solves the problems of traditional methods, such as difficulty in quantifying structural integrity, fuzzy judgment criteria, and difficulty in integrating multiple parameters for joint decision-making. It not only makes accurate judgments about intact cells but also adapts to complex morphological changes through continuous indices, significantly improving the reliable discrimination capability and system automation level before printing. This meets the needs of high-end single-cell printing scenarios that prioritize both structural recognition accuracy and control response sensitivity.

[0101] Example 5

[0102] Please refer to Figure 1 and Figure 3 Specifically, the S4 steps include:

[0103] S41. Compare and analyze the cell integrity index with the preset integrity threshold to determine whether the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, and issue the corresponding level of printing control command, specifically including:

[0104] If the cell integrity index exceeds the integrity threshold, it means that the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, and at this time, a first-level printing control command is issued.

[0105] If the cell integrity index does not exceed the integrity threshold, it means that the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, and at this time, a secondary printing control command is issued.

[0106] The integrity threshold is a benchmark value used to determine whether the target single-cell structure meets the requirements for jet printing. Its function is to provide a quantitative decision standard to clearly classify the cell integrity index, thereby distinguishing between structurally intact and potentially damaged cell states, and guiding the issuance of printing control commands. The threshold is set based on a large amount of experimental data statistics and expert experience. The specific methods include: during the sample training phase, calculating the integrity index of multiple known structurally intact and damaged cell images, combining image quality grading and printing success rate evaluation, and using statistical analysis methods (ROC curve analysis or minimum error rate method) to select the critical value that best distinguishes the two structural states as the integrity threshold.

[0107] Specifically, S42 executes the corresponding print control operations based on the received primary and secondary print control commands, including:

[0108] When a first-level printing control command is received, the following actions are executed: This indicates that the target single-cell structure in the cell suspension has not been damaged, and the nozzle structure is adjusted by the piezoelectric microvalve to perform a jet printing operation on the cell suspension containing the complete target single cell;

[0109] When a secondary printing control command is received, the following actions are taken: if the target single-cell structure in the cell suspension is damaged, the cell suspension is replaced, and the S1-S4 process is repeated until a primary printing control command is issued. The nozzle structure is then adjusted via a piezoelectric microvalve to perform a jet printing operation on the cell suspension containing the complete target single cell.

[0110] In this embodiment, by introducing a hierarchical printing control mechanism with cell integrity index as the core criterion, an intelligent printing decision-making system with self-judgment, self-adjustment, and self-closed-loop execution capabilities is constructed. This significantly solves a series of systemic defects in the prior art caused by the lack of quantitative basis for structural state judgment, the single printing control decision, and the lag in operation after misjudgment. Specifically, S41 uses a preset integrity threshold as a benchmark to compare and analyze the cell integrity index, clearly distinguish whether the printing conditions are met, and issues a first-level or second-level printing control command accordingly. It has dynamic adjustment capabilities when facing target cells with different quality states, breaking through the technical bottleneck of traditional static judgment and difficulty in adapting to the complexity of actual suspensions. S42 further forms a closed-loop linkage between control commands and execution actions, wherein the first-level printing control command directly drives the piezoelectric microvalve injection operation. This ensures that high-quality cells are printed smoothly under structural integrity. When a secondary printing control command is issued, it automatically triggers suspension replacement and full-process re-inspection until the printing conditions are met before proceeding. This reduces the probability of damaged cells being accidentally sprayed or the cycle of printing failures being redundant. More importantly, this mechanism establishes a closed-loop path of "structural quality identification - control judgment - execution feedback - structural re-judgment," effectively avoiding system failure risks such as tissue construction failure, decreased printing efficiency, nozzle clogging, and microvalve mis-opening caused by misjudgment. It achieves a triple improvement in printing process stability, control accuracy, and response efficiency. Therefore, S4 not only completes the closed-loop "identification - judgment - control" in terms of technical path, but also endows the system with high adaptability and fault tolerance for single-cell printing tasks in practical applications. It is an important guarantee mechanism for promoting precision biomanufacturing towards high reliability and high consistency.

[0111] Example 6

[0112] Please refer to Figure 1 and Figure 2 Specifically: a single-cell printing equipment control system based on convolutional neural network optimization, including a pixel parameter acquisition module, a feature selection module, a structure analysis module, and a printing control module;

[0113] The pixel parameter acquisition module is used to continuously acquire images of cell suspension containing target single cells before the cell suspension is sprayed and printed, in order to determine the main cell region image of each frame. After being input into a shallow convolutional network, the pixel grayscale parameter set, pixel response parameter set and pixel mutation parameter set are constructed respectively.

[0114] The feature filtering module is used to analyze the degree of morphological fluctuation in the main cell region images of each frame based on the pixel grayscale parameter set, pixel response parameter set, and pixel mutation parameter set, so as to filter out standard morphological feature images.

[0115] The structural analysis module is used to analyze the integrity of the target single-cell structure contained in the selected standard morphological feature images and determine the cell integrity index.

[0116] The printing control module is used to compare and analyze the cell integrity index to determine whether the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, so as to issue and execute the corresponding level of printing control command.

[0117] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A control method for a single-cell printing device based on convolutional neural network optimization, characterized in that: Includes the following steps: S1. Before performing jet printing on the cell suspension, continuously acquire images of the cell suspension containing the target single cell to determine the main cell region image of each frame. After inputting into a shallow convolutional network, construct pixel grayscale parameter set, pixel response parameter set and pixel mutation parameter set respectively. S2. Based on the pixel grayscale parameter set, pixel response parameter set and pixel mutation parameter set, analyze the degree of morphological fluctuation of the main cell region image in each frame in order to screen out standard morphological feature images. S3. Based on the selected standard morphological feature images, analyze the integrity of the target single-cell structure contained in the standard morphological feature images and determine the cell integrity index; the specific steps of S3 include: S31. Based on the determined standard morphological feature image, combined with the pixel response parameter set, extract the structural response value of each pixel in the standard morphological feature image, and after performing the extreme value search operation, calculate the difference between the maximum and minimum structural response values ​​in the obtained standard morphological feature image to determine the structural response amplitude of the standard morphological feature image. S32. Based on the standard morphological feature image and combined with the pixel mutation parameter set, extract the edge local mutation coefficient of each pixel in the standard morphological feature image, compare the edge local mutation coefficient of each pixel with the preset mutation threshold, filter out the corresponding pixel that the edge local mutation coefficient exceeds the mutation threshold, mark it as edge pixel, and record the gray value of each edge pixel. S33. After performing a second-order Laplacian transform on the gray values ​​of each edge pixel, obtain the second-order transform value of each edge pixel, and combine it with the statistical mean calculation algorithm to determine the boundary sharpening index of the standard morphological feature image. S34. Correlate the morphological tension value, structural response amplitude, and boundary sharpening index of the standard morphological feature image. After dimensionless processing, analyze the integrity of the target single-cell structure contained in the standard morphological feature image to determine the cell integrity index, specifically: In the formula, Indicates the cell integrity index. Indicates the morphological tension value. Indicates the amplitude of the structural response. Indicates the boundary sharpness index; S4. Compare and analyze the cell integrity index to determine whether the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, and issue and execute the corresponding level of printing control command.

2. The control method for a single-cell printing device based on convolutional neural network optimization according to claim 1, characterized in that: The specific steps in S1 include: S11. Before performing jet printing on the cell suspension, an industrial camera set in the center of the field of view of the microfluidic channel continuously acquires images of the cell suspension containing the target single cell at a preset sampling frequency. After grayscale processing, a cell suspension image frame sequence is constructed, and the main cell region image of each frame is determined by the projection image range of the target single cell in each frame. S12. Based on the pixel resolution of the image, divide the main cell region image of each frame into pixels and record the gray value of each pixel in each frame image to construct a pixel gray value parameter set, wherein the pixel gray value dataset includes the gray value of each pixel in each frame of the main cell region image. S13. Input the cell main region images of each frame into a pre-trained shallow convolutional network, and perform structural scanning on the cell main region images through multiple convolutional kernels respectively. Extract the maximum value of the structural response output of all convolutional kernels at the corresponding pixel positions as the structural response value of the corresponding pixel, and construct a pixel response parameter set. The pixel response dataset includes the structural response value of each pixel in each cell main region image.

3. The control method for a single-cell printing device based on convolutional neural network optimization according to claim 2, characterized in that: S14. Based on the constructed pixel grayscale dataset, a pixel abrupt change parameter set is constructed by performing local gradient analysis on the grayscale distribution of pixels surrounding each pixel, specifically including: Centered on the current pixel, a 3×3 pixel matrix surrounding the current pixel is selected as the local region input. Convolution operations are performed with Sobel operators in two standard directions. Specifically, the gray values ​​of each pixel in the pixel matrix are multiplied by their corresponding Sobel weights and then summed to obtain the horizontal gray value change Gx and the vertical gray value change Gy of the current pixel. The standard directions include the horizontal and vertical directions. The horizontal grayscale change value Gx and the vertical grayscale change value Gy of the current pixel are correlated. After calculating the two-dimensional gradient magnitude, the local edge abrupt change coefficient of the current pixel is determined, specifically as follows: In the formula, This represents the local abrupt change coefficient of the current pixel's edge. Based on the determined edge local mutation coefficient of the current pixel, the edge local mutation coefficient of each pixel in the main cell region image of each frame is calculated sequentially, and a pixel mutation parameter set is constructed.

4. The control method for a single-cell printing device based on convolutional neural network optimization according to claim 3, characterized in that: The specific steps in S2 include: S21. Perform feature recognition on the constructed pixel grayscale parameter set, extract the grayscale value of each pixel in the main cell region image of each frame, and combine the statistical mean value algorithm to obtain the grayscale mean value of the main cell region image of each frame. S22. Based on the constructed pixel grayscale parameter set, pixel response parameter set, and pixel mutation parameter set, correlation is performed, and after dimensionless processing, the morphological fluctuation degree of the main cell region image in each frame is analyzed to determine the morphological tension value of the main cell region image in each frame, specifically: In the formula, This represents the morphological tension value of the main cell region in the i-th frame image. This represents the grayscale value of the j-th pixel in the i-th frame of the main cell region image. This represents the average grayscale value of the main cell region image in the i-th frame. This represents the structural response value of the j-th pixel in the i-th frame of the main cell region image. This represents the local abrupt change coefficient of the edge at the j-th pixel in the i-th frame of the main cell region image. This indicates the number of pixels.

5. The control method for a single-cell printing device based on convolutional neural network optimization according to claim 4, characterized in that: S23. Based on the comparative analysis of the morphological tension values ​​of the acquired cell main region images, standard morphological feature images are determined, specifically including: Based on the morphological tension values ​​of each frame of the main cell region image, and combined with a statistical averaging algorithm, the mean morphological tension value is obtained. The difference between the morphological tension value of each frame of the main cell region image and the mean morphological tension value is calculated to obtain the morphological tension difference value of each frame of the main cell region image. After feature extraction, the main cell region image corresponding to the smallest morphological tension difference value is selected as the standard morphological feature image.

6. The control method for a single-cell printing device based on convolutional neural network optimization according to claim 5, characterized in that: The specific steps in S4 include: S41. Compare and analyze the cell integrity index with the preset integrity threshold to determine whether the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, and issue the corresponding level of printing control command, specifically including: If the cell integrity index exceeds the integrity threshold, it means that the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, and at this time, a first-level printing control command is issued. If the cell integrity index does not exceed the integrity threshold, it means that the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, and at this time, a secondary printing control command is issued.

7. The control method for a single-cell printing device based on convolutional neural network optimization according to claim 6, characterized in that: S42. Execute the corresponding printing control operations according to the received primary and secondary printing control commands, specifically including: When a first-level printing control command is received, the following actions are executed: This indicates that the target single-cell structure in the cell suspension has not been damaged, and the nozzle structure is adjusted by the piezoelectric microvalve to perform a jet printing operation on the cell suspension containing the complete target single cell; When a secondary printing control command is received, the following actions are taken: if the target single-cell structure in the cell suspension is damaged, the cell suspension is replaced, and the S1-S4 process is repeated until a primary printing control command is issued. The nozzle structure is then adjusted via a piezoelectric microvalve to perform a jet printing operation on the cell suspension containing the complete target single cell.

8. A single-cell printing device control system based on convolutional neural network optimization, used to implement the single-cell printing device control method based on convolutional neural network optimization as described in any one of claims 1 to 7, characterized in that: It includes a pixel parameter acquisition module, a feature selection module, a structure analysis module, and a print control module; The pixel parameter acquisition module is used to continuously acquire images of cell suspension containing target single cells before the cell suspension is sprayed and printed, in order to determine the main cell region image of each frame. After being input into a shallow convolutional network, the pixel grayscale parameter set, pixel response parameter set and pixel mutation parameter set are constructed respectively. The feature filtering module is used to analyze the degree of morphological fluctuation in the main cell region images of each frame based on the pixel grayscale parameter set, pixel response parameter set, and pixel mutation parameter set, so as to filter out standard morphological feature images. The structural analysis module is used to analyze the integrity of the target single-cell structure contained in the selected standard morphological feature images and determine the cell integrity index. The printing control module is used to compare and analyze the cell integrity index to determine whether the integrity of the target single cell structure in the current standard morphological feature image meets the printing requirements, so as to issue and execute the corresponding level of printing control command.

Images