Adjustable annular diaphragm cell imaging system and dye-free deep learning identification method

Abstract

The present application relates to a tunable annular diaphragm cell imaging system and a dye-free deep learning identification method, relating to the technical field of optical imaging and image processing, comprising an annular diaphragm illumination module; the present application establishes a quantitative mapping relationship between the outer diameter of the annular diaphragm parameter and the image quality index, and adaptively finds the best bright field and dark field illumination conditions for different cell types, densities and overall profiles and internal structures, so as to obtain an original image with sharp edges and distinct internal structures without using any dye, the improved deep learning segmentation model has a two-way attention gate module, which can capture spatial details and texture semantic information important for cell recognition in parallel and differentially, realize high-accuracy dye-free cell dead or alive automatic identification, make the segmentation result more consistent with the biophysical characteristics of cells, improve the segmentation accuracy and recognition accuracy, and automatically distinguish the state of cells with complex background without fluorescent labeling.

Description

Technical Field

[0001] This invention relates to the field of imaging and image processing technology, and in particular to an adjustable annular aperture cell imaging system and a deep learning image recognition method. Background Technology

[0002] In biomedical research, drug screening, and clinical testing, it is crucial to observe colorless and transparent living cells without damage or dyes and with high contrast. Such cells, such as lymphocytes and stem cells, have small differences in refractive index with their surrounding culture medium. Under traditional bright-field microscopes, their edges are blurred and their internal structures are difficult to distinguish. This makes it impossible to clearly observe cell morphology, internal structure, and dynamics without the use of exogenous labels such as fluorescent dyes, and it is also difficult to reliably and efficiently distinguish between cell live and dead states.

[0003] Ring illumination, as an important optical contrast enhancement technology, forms a ring beam through an adjustable aperture, which can change the incident angle and spatial distribution of the illumination light, thereby improving the edge contrast and overall contrast of an image to a certain extent. However, existing ring illumination systems have significant limitations in practical applications: Adjusting illumination parameters, including the inner and outer diameters of the annular aperture and the position of the light source, heavily relies on the operator's experience, and the manual trial process is tedious and time-consuming. Secondly, due to the diversity of cell samples and the specificity of the observation target, the combination of parameters constitutes a complex multidimensional space, and it is difficult to find the optimal imaging conditions for the current sample and observation task by relying solely on manual methods. The performance of deep learning models largely depends on the quality of the input image. If the original image has uneven lighting, low contrast, or noise interference, it will directly affect the accuracy of the model segmentation and the reliability of the classification, and may even lead to the model overfitting to image artifacts under specific lighting conditions. To address the aforementioned technical deficiencies, a solution is proposed. Summary of the Invention

[0004] The purpose of this invention is to construct a collaborative framework from physical imaging optimization to intelligent information extraction. Through the automatic optimization of the illumination system, it can provide high-quality input with clearer features and less noise for deep learning models, achieving alignment between hardware adjustment and software analysis, thereby comprehensively improving the automation level and overall performance of dye-free cell detection.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an adjustable annular aperture cell imaging system, including an annular aperture illumination module, an illumination analysis module, an illumination control module, and a dye-free imaging evaluation module; The annular aperture illumination module provides adjustable annular illumination for dye-free cell samples placed on the stage, and includes LED beads, an adjustable annular aperture, and a condenser lens. The imaging acquisition module is used to image dye-free cell samples. By setting the difference between the inner and outer diameters of the adjustable annular aperture through the objective lens and image sensor, bright field and dark field images are obtained. The cell image processing module sets initial control parameters for bright field and dark field images, performs preprocessing with Gaussian filtering, obtains preliminary candidate cell region masks through an adaptive threshold algorithm, performs morphological opening and closing operations on the candidate cell region masks to remove noise and fill holes, and obtains refined cell region masks. The dye-free imaging evaluation module calculates key features of the image based on a refined mask. It uses the internal structure dispersion coefficient, which reflects the clarity of the internal structure of the cell, as the first indicator, and the boundary contrast, which reflects the cell edge contour, as the second indicator to construct a comprehensive evaluation index. The aperture parameter control module updates the parameters of the adjustable annular aperture based on an iterative optimization algorithm to optimize the comprehensive evaluation index and outputs the optimal imaging parameters of the adjustable annular aperture. The optimized optimal annular aperture parameters are then applied to the imaging system.

[0006] Furthermore, the annular aperture illumination module is equipped with an objective lens, the specific configuration of which is as follows: A 10x objective lens is used in the optical path, with a numerical aperture (NA) of 0.25. According to the formula: It is the numerical aperture of the condenser lens. It is the aperture of the condenser lens. This refers to the focal length of the condenser lens, from which one selects a suitable focal length and numerical aperture range for the condenser lens. (Condenser lens focal length) ,diameter , set here .

[0007] Furthermore, by adjusting the difference between the inner and outer diameters of the adjustable annular aperture, bright-field and dark-field images are obtained. The specific setting process is as follows: Based on the effective numerical aperture of the annular aperture, the parameter ranges for the bright field and dark field are designed separately. The specific design parameter formulas are as follows: It is the effective numerical aperture of the annular aperture. It is the numerical aperture corresponding to the inner diameter of the annular aperture. It is the numerical aperture corresponding to the outer diameter of the annular aperture; The inner diameter of the annular aperture is fixed as The outer diameter adjustment range is set to Within this range, the convergence angle of the illumination light allows direct light to enter the objective lens, forming a bright-field image with a bright background and high contrast in cell imaging; The inner diameter of the annular aperture is fixed as The outer diameter adjustment range is set to Within this range, the ring illumination light shines on the sample at a high angle, allowing only the light scattered by the cells to enter the objective lens, while direct light is excluded, thus forming a dark-field image with a dark background and bright cell edges.

[0008] Furthermore, an adaptive thresholding algorithm is used to initially obtain candidate cell region masks. Morphological opening and closing operations are then performed on these masks to remove noise and fill holes, resulting in a refined cell region mask. The specific process is as follows: Set the initial value of the outer diameter of the annular aperture for both bright-field and dark-field modes according to the imaging target; Initial value of outer diameter in bright field mode Set as parameter range The lower limit and median; Initial value of outer diameter in dark field mode Set the parameter range to The lower limit and median; Two-dimensional Gaussian filters are used to convolve the acquired original bright field image and dark field image to suppress high-frequency noise and obtain a smooth image. The smooth image is then binarized using a local adaptive thresholding algorithm. A threshold based on the local neighborhood gray-level characteristics is calculated for each pixel in the image. The pixel gray level is compared with the threshold to obtain a preliminary candidate cell region mask. Morphological closing and opening operations are sequentially performed on the preliminary candidate cell region mask to optimize the region shape and remove noise. A small circular structuring element is used to expand the mask, filling the tiny holes inside the cell region and connecting adjacent tiny fracture regions. Subsequently, the same structuring element is used to erode the region to restore its approximate original boundary, resulting in an intermediate mask. The intermediate mask is eroded using a structuring element to eliminate isolated noise points and small particle artifacts. Then, it is expanded to repair the cell region boundaries that have shrunk due to erosion, resulting in a refined cell region mask.

[0009] Furthermore, the internal structure dispersion coefficient, reflecting the clarity of the cell's internal structure, is used as the first indicator, and the boundary contrast, reflecting the cell's edge contour, is used as the second indicator to construct a comprehensive evaluation index. The specific construction and calculation process is as follows: within the refined cell region mask M, a smooth image is extracted. Calculate the standard deviation of the gray values ​​in that area. Compared with the average The internal structure discrete coefficients are obtained from the following formula: Using the Canny operator from Edge detection in the middle Perform a logical AND operation with the boundary region of M to obtain the cell outline pixel set; For each contour pixel, in the original image The average gray level of each cell's neighborhood is calculated. and the average gray level of the external background ; The boundary contrast is obtained by averaging the local contrast of all contour pixels. : ; Preset weights based on imaging focus and By weighting and summing the first and second indicators, a single indicator for comprehensively evaluating image quality is obtained:

[0010] In the above formula, N is the total number of cell outline pixels, and i is the index of the outline pixels, from 1 to N. Let be the average gray value of all pixels located inside the cell in the neighborhood of the i-th contour pixel. For the i-th contour pixel, the average gray value of all background pixels located outside the cell in the neighborhood.

[0011] An adjustable annular aperture cell imaging dye-free deep learning recognition method, the method includes the following steps: Step 1: Based on the optimal imaging parameters of the adjustable annular aperture, acquire bright-field and dark-field images of dye-free cell samples respectively, and call the improved deep learning segmentation model to process the bright-field and dark-field images to obtain accurate cell region masks. Step 2: Based on a precise cell region mask, extract the morphological features, texture features, and optical features of each cell, and then stitch and fuse the cell features from the bright field image and the dark field image to form a cell feature vector; Step 3: Input the cell feature vector into the pre-trained deep learning segmentation model, output the probability that each cell belongs to a live cell or a dead cell, and determine the classification result of the live / dead state of each cell by comparing the probability with the preset decision threshold. Step 4: Based on the classification results of live and dead cells, count the number of live cells and dead cells respectively, and generate a count report containing the distribution of cell live and dead status and the total number of cells.

[0012] Furthermore, the improved deep learning segmentation model is described in the following steps: A dual-path attention gating module is embedded in the skip connection between the encoder and decoder. The dual-path attention gating module includes a spatial detail path, a texture semantic path, an attention generation unit, and a feature weighting unit. The spatial detail path generates a spatial detail weight map that highlights cell boundary information by performing directional gradient enhancement on the input encoder feature map. ; The texture semantic pathway generates a texture semantic weight map that represents information about the internal structure of cells by performing local texture encoding on the input encoder feature map. ; Attention generation unit fusion spatial detail weight map Texture semantic weight graph In addition to upsampled features from the decoder, a spatial-channel joint attention weight map is generated. ; Feature weighting units are based on the spatial-channel joint attention weight map. The encoder features are selectively enhanced and suppressed, and the processed features are fused with the decoder features.

[0013] Furthermore, the improved deep learning segmentation model is trained in the following manner, the specific process of which is as follows: S1: Prepare the training dataset: The training dataset includes dye-free bright-field microscopic images and dark-field microscopic images of cells acquired under various annular aperture parameters, as well as the corresponding cell region masks. S2: Embed the dual-path attention gating module into the connection layer of the deep learning segmentation model; S3: During training, an illumination consistency loss term is introduced, which is calculated by computing the attention weight map generated by the dual-path attention gating module. The difference between the heatmap and the image quality assessment heatmap generated during adaptive lighting optimization is used to construct a model that constrains the areas of interest of the model and optimizes the lighting parameters.

[0014] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This adjustable annular aperture cell imaging system and deep learning image recognition method establish a quantitative mapping relationship between the outer diameter of the annular aperture parameter and image quality indicators, and design comprehensive evaluation indicators to guide the search, achieving fully automatic and precise optimization of illumination parameters. It adaptively finds the optimal bright and dark field illumination conditions for different cell types, densities, and observations of overall contours and internal structures, thus obtaining original images with sharp edges and distinct internal structures without the use of any dyes. The improved deep learning segmentation model's dual-path attention gating module can capture spatial details and textural semantic information crucial for cell identification in parallel and differentiated ways. By fusing bright and dark field multimodal image features, it enhances the visual ability of key cell biological features, achieving high-accuracy automatic identification of cell life and death without dyes. This makes the segmentation results more consistent with the biophysical characteristics of cells, improving segmentation accuracy and recognition accuracy. It can automatically determine the state of cells with complex backgrounds without fluorescent labeling. Attached Figure Description

[0015] Figure 1A schematic diagram of the system structure of the present invention is shown; Figure 2 A schematic diagram of the overall optical path structure of the system of the present invention is shown; Figure 3 A schematic diagram of the system light source structure of the present invention is shown; Figure 4 A schematic diagram of the adjustable annular aperture structure of the present invention is shown; Figure 5 A schematic diagram of the parameter optimization process structure of the present invention is shown; Figure 6 A schematic diagram of the cell image segmentation and recognition process of the present invention is shown.

[0016] Illustrations: 1. LED light bead; 2. Adjustable annular aperture; 3. Condenser lens; 4. Sample; 5. Counting plate. Detailed Implementation

[0017] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0018] 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.

[0019] Example 1: like Figure 1-5 As shown, the adjustable annular aperture cell imaging system includes an annular aperture illumination module, an illumination analysis module, an illumination control module, and a dye-free imaging evaluation module. The annular aperture illumination module provides adjustable annular illumination for dye-free cell samples placed on the stage, and includes LED beads, an adjustable annular aperture, and a condenser lens. The imaging acquisition module is used to image dye-free cell samples. By setting the difference between the inner and outer diameters of the adjustable annular aperture through the objective lens and image sensor, bright field and dark field images are obtained. The cell image processing module sets initial control parameters for bright field and dark field images, performs preprocessing with Gaussian filtering, obtains preliminary candidate cell region masks through an adaptive threshold algorithm, performs morphological opening and closing operations on the candidate cell region masks to remove noise and fill holes, and obtains refined cell region masks. The dye-free imaging evaluation module calculates key features of the image based on a refined mask. It uses the internal structure dispersion coefficient, which reflects the clarity of the cell's internal structure, as the first indicator, and the boundary contrast, which reflects the cell's edge contour, as the second indicator to construct a comprehensive evaluation index.

[0020] The annular aperture illumination module has an internal objective lens configuration, specifically as follows: a 10x objective lens is used in the optical path, with a numerical aperture (NA) of 0.25, according to the formula: It is the numerical aperture of the condenser lens. It is the aperture of the condenser lens. This refers to the focal length of the condenser lens, from which one selects a suitable focal length and numerical aperture range for the condenser lens. (Condenser lens focal length) ,diameter , set here .

[0021] By setting the difference between the inner and outer diameters of the adjustable annular aperture, bright-field and dark-field images can be obtained. The specific setting process is as follows: Based on the effective numerical aperture of the annular aperture, the parameter ranges for the bright field and dark field are designed separately. The specific design parameter formulas are as follows: It is the effective numerical aperture of the annular aperture. It is the numerical aperture corresponding to the inner diameter of the annular aperture. It is the numerical aperture corresponding to the outer diameter of the annular aperture; The inner diameter of the annular aperture is fixed as The outer diameter adjustment range is set to Within this range, the convergence angle of the illumination light allows direct light to enter the objective lens, forming a bright-field image with a bright background and high contrast in cell imaging; The inner diameter of the annular aperture is fixed as The outer diameter adjustment range is set to Within this range, the ring illumination light shines on the sample at a high angle, allowing only the light scattered by the cells to enter the objective lens, while direct light is excluded, thus forming a dark-field image with a dark background and bright cell edges.

[0022] Candidate cell region masks are initially obtained through an adaptive thresholding algorithm. Morphological opening and closing operations are then performed on these masks to remove noise and fill holes, resulting in a refined cell region mask. The specific process is as follows: Set the initial value of the outer diameter of the annular aperture for both bright-field and dark-field modes according to the imaging target; Initial value of outer diameter in bright field mode Set as parameter range The lower limit and median; Initial value of outer diameter in dark field mode Set the parameter range to The lower limit and median; Two-dimensional Gaussian filters are used to convolve the acquired original bright field image and dark field image to suppress high-frequency noise and obtain a smooth image. The smooth image is then binarized using a local adaptive thresholding algorithm. A threshold based on the local neighborhood gray-level characteristics is calculated for each pixel in the image. The pixel gray level is compared with the threshold to obtain a preliminary candidate cell region mask. Morphological closing and opening operations are sequentially performed on the preliminary candidate cell region mask to optimize the region shape and remove noise. A small circular structuring element is used to expand the mask, filling the tiny holes inside the cell region and connecting adjacent tiny fracture regions. Then, the same structuring element is used to erode the mask to restore the approximate original boundary of the region, thus obtaining the intermediate mask. The intermediate mask is eroded using a structuring element to eliminate isolated noise points and small particle artifacts. Then, it is expanded to repair the cell region boundaries that have shrunk due to erosion, resulting in a refined cell region mask.

[0023] The internal structure dispersion coefficient, reflecting the clarity of the cell's internal structure, is used as the first indicator, and the boundary contrast, reflecting the cell's edge contour, is used as the second indicator to construct a comprehensive evaluation index. The specific construction and calculation process is as follows: Extract a smoothed image within a refined cellular region mask M. Calculate the standard deviation of the gray values ​​in that area. Compared with the average The internal structure discrete coefficients are obtained from the following formula:

[0024] Using the Canny operator from Edge detection in the middle Perform a logical AND operation with the boundary region of M to obtain the cell outline pixel set; For each contour pixel, in the original image The average gray level of each cell's neighborhood is calculated. and the average gray level of the external background ; The boundary contrast is obtained by averaging the local contrast of all contour pixels. : ; Preset weights based on imaging focus and By weighting and summing the first and second indicators, we obtain a single indicator, F, for comprehensively evaluating image quality.

[0025] In the above formula, N is the total number of cell outline pixels, and i is the index of the outline pixels, from 1 to N. Let be the average gray value of all pixels located inside the cell in the neighborhood of the i-th contour pixel. For the i-th contour pixel, the average gray value of all background pixels located outside the cell in the neighborhood.

[0026] The aperture parameter control module updates the parameters of the adjustable annular aperture based on an iterative optimization algorithm to optimize the comprehensive evaluation index and outputs the optimal imaging parameters of the adjustable annular aperture. The optimized optimal annular aperture parameters are then applied to the imaging system.

[0027] In bright-field mode, the annular aperture parameters are initialized to... ; After 10 to 15 iterations, the algorithm obtains the optimal parameters. At this time, the corresponding , In weight of , In this case, F=0.24; In dark mode, the parameters are initialized to After 10 to 15 iterations, the algorithm obtains the optimal parameters as follows: At this time, the corresponding , In weight of , In this case, F=0.73.

[0028] In this scheme, through the digital precision control of the annular aperture illumination module, the system can dynamically adapt to different cell types and observation needs. In bright field mode, by optimizing the outer diameter of the annular aperture, the phase difference contrast between the overall cell outline and the background is significantly improved. In dark field mode, by precisely controlling the illumination angle, the scattered light signals of the fine internal structure and edges of the cell are maximized. Based on the adaptive adjustment of physical optics, colorless and transparent living cells can present high signal-to-noise ratio and high-definition image details without any external staining labels. The dye-free imaging evaluation module transforms subjective perceptions of image sharpness into objective, calculable mathematical indicators by calculating the internal structure dispersion coefficient CV and boundary contrast BC. By constructing a comprehensive evaluation index F and optimizing it, the system can automatically and accurately pinpoint the optimal combination of illumination parameters for the current sample.

[0029] Example 2: like Figure 6As shown, an adjustable annular aperture cell imaging dye-free deep learning recognition method includes the following steps: Step 1: Based on the optimal imaging parameters of the adjustable annular aperture, acquire bright-field and dark-field images of dye-free cell samples respectively, and call the improved deep learning segmentation model to process the bright-field and dark-field images to obtain accurate cell region masks. Step 2: Based on a precise cell region mask, extract the morphological features, texture features, and optical features of each cell, and then stitch and fuse the cell features from the bright field image and the dark field image to form a cell feature vector; Step 3: Input the cell feature vector into the pre-trained deep learning segmentation model, output the probability that each cell belongs to a live cell or a dead cell, and determine the classification result of the live / dead state of each cell by comparing the probability with the preset decision threshold. Step 4: Based on the classification results of live and dead cells, count the number of live cells and dead cells respectively, and generate a count report containing the distribution of cell live and dead status and the total number of cells.

[0030] The improved deep learning segmentation model is described in the following steps: A dual-path attention gating module is embedded in the skip connection between the encoder and decoder. The dual-path attention gating module includes a spatial detail path, a texture semantic path, an attention generation unit, and a feature weighting unit. The spatial detail pathway generates a spatial detail weight map that highlights cell boundary information by performing directional gradient enhancement on the input encoder feature map. ; The texture semantic pathway generates a texture semantic weight map that represents information about the internal structure of cells by performing local texture encoding on the input encoder feature map. ; Attention generation unit fusion spatial detail weight map Texture semantic weight graph In addition, upsampled features from the decoder are used to generate a joint spatial and channel attention weight map. ; Feature-weighted units are based on the joint attention weight map of spatial and channel dimensions. The encoder features are selectively enhanced and suppressed, and the processed features are fused with the decoder features.

[0031] The spatial detail pathway simulates the visual system's keen perception of edges, specifically enhancing gradient changes related to cell membranes and cell walls in images, enabling the network to have extremely high resolution of weak boundaries. The texture semantic pathway simulates the ability to analyze the texture of internal structures. Through gray-level co-occurrence matrix features, the network can effectively distinguish between regions inside cells filled with organelles and homogeneous background or necrotic regions. The synergistic work of these two pathways allows the improved deep learning segmentation model to focus on both the morphological outline and internal texture of cells, just like an experienced microscopist. The improved deep learning segmentation model is trained in the following manner, the specific process of which is as follows: S1: Prepare the training dataset: The training dataset includes dye-free bright-field microscopic images and dark-field microscopic images of cells acquired under various annular aperture parameters, as well as the corresponding cell region masks. S2: Embed the dual-path attention gating module into the connection layer of the deep learning segmentation model; S3: During training, an illumination consistency loss term is introduced, which is calculated by computing the attention weight map generated by the dual-path attention gating module. The difference between the heatmap and the image quality assessment heatmap generated during adaptive lighting optimization is used to construct a model that constrains the areas of interest of the model and optimizes the lighting parameters.

[0032] In this scheme, the specific framework and connection relationships of the improved deep learning segmentation model are as follows: Basic framework: The encoder consists of four layers. Each layer contains two consecutive 3x3 convolutional layers. Each convolution is followed by a batch normalization layer and an activation function, as well as a 2x2 max pooling layer for downsampling. The number of channels increases progressively, such as 64, 128, 256, and 512. Decoder: Also consists of 4 layers. Each layer first performs 2x2 transposed convolution upsampling on the input feature map, and then concatenates it with the corresponding layer encoder features after optimization. The concatenated features then go through two 3x3 convolutional layers, including batch normalization and activation functions, for fusion and refinement. The number of channels decreases step by step. Output layer: At the end of the decoder, a 1x1 convolutional layer is used to map the number of channels to the desired number of categories, which is 2 in this example, including background and cells, and outputs a probability map of each pixel belonging to the cell region through an activation function; Detailed structure of the dual-channel attention gating module: It includes spatial detail pathways, texture semantic pathways, attention generation units, and feature weighting units to process encoder features. upsampled features from the decoder For example, the specific working process is as follows: Spatial detail pathways from Extracting and enhancing cell boundary contour information, for Each channel uses a horizontal direction operator. and vertical direction operator Perform convolution to obtain the gradient map. and ; Calculate the edge intensity of each pixel. ; The calculated edge intensity map E is input into a light quantum network, typically a 1x1 convolutional layer followed by an activation function, to generate a spatial detail weight map. The value is in the interval [0, 1], and the high score area corresponds to a strong edge; Texture semantic pathway Within a local neighborhood, such as a 3x3 or 5x5 window, calculate the gray-level co-occurrence matrix and extract features that reflect texture roughness and uniformity, such as contrast, from the gray-level co-occurrence matrix. The contrast value and homogeneity value calculated for each local window are concatenated and mapped to a scalar texture saliency score through a fully connected network. The fraction is copied spatially to generate a result similar to... Texture semantic weight graphs with the same spatial dimensions ; The attention generation unit fuses boundary and texture information, and combines it with the context provided by the decoder to generate the final attention guidance. , and The feature maps are concatenated along the channel dimension and fed into a lightweight network consisting of 1x1 convolutions, batch normalization, activation functions, and a final 1x1 convolution followed by the activation function. The output is a joint attention weight map of the spatial and channel dimensions. Each element value is between [0, 1], indicating the importance of the corresponding position and channel features to the final cell segmentation task.

[0033] The feature weighting unit adaptively filters encoder features. Specifically, the process involves: adjusting the attention weight map... With encoder features Perform element-wise multiplication to obtain The operation enhances key features related to cell boundaries and internal structures while suppressing irrelevant background or noise features, resulting in weighted features. Passed to the decoder, and The parts are then assembled and proceed to subsequent processing. Hierarchical differentiation design: In shallow skip connections, such as layers 1 and 2, spatial detail paths use smaller-scale gradient operators, such as optimized 3x3 operators, and improve... The initial weight bias in the final attention is used to focus on capturing the delicate cell membrane edges; In deep skip connections, such as layers 3 and 4, the texture semantic pathway uses a larger texture analysis window, such as 5x5, and improves... The weight bias is used to better capture the semantic information of the overall cell morphology and internal regions; The training steps, parameters, and loss function for the improved deep learning segmentation model are as follows: Training data preparation: Dataset: A large number of dye-free bright-field images of cells were collected under different annular aperture parameters, covering the effective range of bright and dark fields. Each image was labeled with the corresponding binarized cell region mask by experts. The dataset was expanded using methods such as random rotation, flipping, and brightness / contrast fine-tuning to improve the model's generalization ability. Loss function design: The model's total loss function It consists of three parts: Segmentation loss: Dice loss or cross-entropy loss is used to measure the overall similarity between the predicted mask and the real mask;

[0034] in To predict probabilities, For real labels, For smoothing terms, and These are the weighting coefficients for edge loss and illumination loss, used to balance the contribution of each part of the loss to the total loss; Edge enhancement loss: Encourages the model to make clearer predictions at cell boundaries. It can calculate the loss between the predicted mask edge and the real mask edge, such as binary cross-entropy. Loss of lighting consistency : To enhance the attention of deep learning segmentation models Align with physical imaging optimization results; Input source: Image quality assessment heatmaps calculated using the optimal imaging parameters for the corresponding training images obtained during the adaptive illumination parameter optimization process. It is generated by weighted fusion of the boundary contrast BC of the local region and the discrete coefficient CV of the internal structure. The bright area represents the cell structure highlighted by high-quality lighting. During training forward propagation, the attention weight map generated by the dual-path attention gating module in the deepest layer is recorded. Calculation method: Minimization The difference between the aggregated A and the aggregated A, such as using KL divergence: Hyperparameters The strength of this loss is set to 0.1 to 0.5 to control its intensity. Training steps and parameters: using the Adam optimizer, with an initial learning rate set to... ; Training process: Initialize model weights. For each training batch, input the image into the network and perform forward propagation to obtain the prediction mask and attention maps for each layer. Calculate the total loss Perform backpropagation, compute gradients, and use the Adam optimizer to update all weight parameters of the improved deep learning segmentation model; The segmentation performance metrics are monitored on the validation set. When the performance no longer improves, the learning rate is reduced to obtain the final improved deep learning segmentation model.

[0035] The illumination consistency loss term utilizes this heatmap when training the improved deep learning segmentation model; Input alignment: Images of the same cell sample acquired under optimal lighting parameters are input into the deep learning segmentation model; Attention comparison: Extract the attention weight map generated by the dual-path attention gating module; Loss calculation: through calculation and The differences between them are used to construct a loss term, which forces the attention mechanism of the deep learning segmentation model to learn to focus on the cellular structure regions that are best revealed by high-quality physical lighting, thus achieving a deep fusion and alignment of computational vision attention and physical optics optimization results.

[0036] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value.

[0037] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation. In the two embodiments provided in this application, it should be understood that the disclosed apparatus and system can be implemented in other ways; for example, the apparatus embodiments described above are merely illustrative, for example, the division of modules is merely a logical functional division, and there may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed; another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the apparatus or module can be electrical, mechanical or other forms. The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An adjustable annular aperture cell imaging system, characterized in that, It includes a ring aperture illumination module, an imaging acquisition module, a cell image processing module, a dye-free imaging evaluation module, and an aperture parameter control module; The annular aperture illumination module provides adjustable annular illumination for dye-free cell samples placed on the stage, and includes LED beads, an adjustable annular aperture, and a condenser lens. The imaging acquisition module is used to image dye-free cell samples. By setting the difference between the inner and outer diameters of the adjustable annular aperture through the objective lens and image sensor, bright field and dark field images are obtained. The cell image processing module sets initial control parameters for bright field and dark field images, performs preprocessing with Gaussian filtering, obtains preliminary candidate cell region masks through an adaptive threshold algorithm, performs morphological opening and closing operations on the candidate cell region masks to remove noise and fill holes, and obtains refined cell region masks. The dye-free imaging evaluation module calculates key features of the image based on a refined mask, uses the internal structure dispersion coefficient, which reflects the clarity of the internal structure of the cell, as the first indicator, and the boundary contrast, which reflects the cell edge contour, as the second indicator, to construct a comprehensive evaluation index. The aperture parameter control module updates the parameters of the adjustable annular aperture based on an iterative optimization algorithm to optimize the comprehensive evaluation index and outputs the optimal imaging parameters of the adjustable annular aperture. The optimized optimal annular aperture parameters are then applied to the imaging system.

2. The adjustable annular aperture cell imaging system according to claim 1, characterized in that, The annular aperture illumination module is equipped with objective lens a, with the following specific configuration: A 10x objective lens is used in the optical path, with a numerical aperture (NA) of 0.

25. According to the formula: It is the numerical aperture of the condenser lens. It is the aperture of the condenser lens. This refers to the focal length of the condenser lens, from which one selects a suitable focal length and numerical aperture range for the condenser lens. (Condenser lens focal length) ,diameter , set here .

3. The adjustable annular aperture cell imaging system according to claim 1, characterized in that, By setting the difference between the inner and outer diameters of the adjustable annular aperture, bright-field and dark-field images can be obtained. The specific setting process is as follows: Based on the effective numerical aperture of the annular aperture, the parameter ranges for the bright field and dark field are designed separately. The specific design parameter formulas are as follows: It is the effective numerical aperture of the annular aperture. It is the numerical aperture corresponding to the inner diameter of the annular aperture. It is the numerical aperture corresponding to the outer diameter of the annular aperture; The inner diameter of the annular aperture is fixed as The outer diameter adjustment range is set to Within this range, the convergence angle of the illumination light allows direct light to enter the objective lens, forming a bright-field image with a bright background and high contrast in cell imaging; The inner diameter of the annular aperture is fixed as The outer diameter adjustment range is set to Within this range, the ring illumination light shines on the sample at a high angle, allowing only the light scattered by the cells to enter the objective lens, while direct light is excluded, thus forming a dark-field image with a dark background and bright cell edges.

4. The adjustable annular aperture cell imaging system according to claim 1, characterized in that, A preliminary candidate cell region mask is obtained through an adaptive thresholding algorithm. Morphological opening and closing operations are then performed on the candidate cell region mask to remove noise and fill holes, resulting in a refined cell region mask. The specific process is as follows: Set the initial value of the outer diameter of the annular aperture for bright field and dark field modes according to the imaging target; Initial value of outer diameter in bright field mode Set as parameter range The lower limit and median; Initial value of outer diameter in dark field mode Set the parameter range to The lower limit and median; Two-dimensional Gaussian filters are used to convolve the acquired original bright field image and dark field image to suppress high-frequency noise and obtain a smooth image. The smooth image is then binarized using a local adaptive thresholding algorithm. A threshold based on the local neighborhood gray-level characteristics is calculated for each pixel in the image. The pixel gray level is compared with the threshold to obtain a preliminary candidate cell region mask. Morphological closing and opening operations are sequentially performed on the preliminary candidate cell region mask to optimize the region shape and remove noise. A small circular structuring element is used to expand the mask, filling the tiny holes inside the cell region and connecting adjacent tiny fracture regions. Then, the same structuring element is used to erode the mask to restore the approximate original boundary of the region, thus obtaining the intermediate mask. The intermediate mask is eroded using a structuring element to eliminate isolated noise points and small particle artifacts. Then, it is expanded to repair the cell region boundaries that have shrunk due to erosion, resulting in a refined cell region mask.

5. The adjustable annular aperture cell imaging system according to claim 1, characterized in that, The internal structural dispersion coefficient, reflecting the clarity of the cell's internal structure, is used as the first indicator, and the boundary contrast, reflecting the cell's edge contour, is used as the second indicator to construct a comprehensive evaluation index. The specific construction and calculation process is as follows: Within the refined cell region mask M, a smooth image is extracted. Calculate the standard deviation of the gray values ​​in that area. Compared with the average The internal structure discrete coefficients are obtained from the following formula: Using the Canny operator from Edge detection in the middle Perform a logical AND operation with the boundary region of M to obtain the cell outline pixel set; For each contour pixel, in the original image The average gray level of each cell's neighborhood is calculated. and the average gray level of the external background ; The boundary contrast is obtained by averaging the local contrast of all contour pixels. : ; Preset weights based on imaging focus and By weighting and summing the first and second indicators, a single indicator for comprehensively evaluating image quality is obtained: In the above formula, N is the total number of cell outline pixels, and i is the index of the outline pixels, from 1 to N. Let be the average gray value of all pixels located inside the cell in the neighborhood of the i-th contour pixel. For the i-th contour pixel, the average gray value of all background pixels located outside the cell in the neighborhood.

6. A dye-free deep learning recognition method for adjustable annular aperture cell imaging, characterized in that the method is applied to an adjustable annular aperture cell imaging system as described in any one of claims 1 to 5, and the method includes the following steps: Step 1: Based on the optimal imaging parameters of the adjustable annular aperture, acquire bright-field and dark-field images of dye-free cell samples respectively, and call the improved deep learning segmentation model to process the bright-field and dark-field images to obtain accurate cell region masks. Step 2: Based on a precise cell region mask, extract the morphological features, texture features, and optical features of each cell, and then stitch and fuse the cell features from the bright field image and the dark field image to form a cell feature vector; Step 3: Input the cell feature vector into the pre-trained deep learning segmentation model, output the probability that each cell belongs to a live cell or a dead cell, and determine the classification result of the live / dead state of each cell by comparing the probability with the preset decision threshold. Step 4: Based on the classification results of live and dead cells, count the number of live cells and dead cells respectively, and generate a count report containing the distribution of cell live and dead status and the total number of cells.

7. The adjustable annular aperture cell imaging dye-free deep learning recognition method according to claim 6, characterized in that, The improved deep learning segmentation model is described in the following steps: A dual-path attention gating module is embedded in the skip connection between the encoder and decoder. The dual-path attention gating module includes a spatial detail path, a texture semantic path, an attention generation unit, and a feature weighting unit. The spatial detail pathway generates a spatial detail weight map that highlights cell boundary information by performing directional gradient enhancement on the input encoder feature map. ; That The texture semantic pathway generates a texture semantic weight map that represents information about the internal structure of cells by performing local texture encoding on the input encoder feature map. ; Attention generation unit fusion spatial detail weight map Texture semantic weight graph In addition to upsampled features from the decoder, a spatial-channel joint attention weight map is generated. ; Feature weighting units are based on the spatial-channel joint attention weight map. The encoder features are selectively enhanced and suppressed, and the processed features are fused with the decoder features.

8. The adjustable annular aperture cell imaging dye-free deep learning recognition method according to claim 7, characterized in that, The improved deep learning segmentation model is trained in the following way, and the specific process is as follows: S1: Prepare the training dataset: The training dataset includes dye-free bright-field microscopic images and dark-field microscopic images of cells acquired under various annular aperture parameters, as well as the corresponding cell region masks. S2: Embed the dual-path attention gating module into the connection layer of the deep learning segmentation model; S3: During training, an illumination consistency loss term is introduced, which is calculated by computing the attention weight map generated by the dual-path attention gating module. The difference between the heatmap and the image quality assessment heatmap generated during adaptive lighting optimization is used to construct a model that constrains the areas of interest of the model and optimizes the lighting parameters.

Images