Multi-cell automatic measurement method and system based on combination of YOLOv11 and EMA

Abstract

The application discloses a multi-cell automatic measurement method and system based on YOLOv11 combined with EMA, and the method comprises the following steps: obtaining a microscope cell image to be processed, and performing image data preprocessing to obtain a preprocessed microscope cell image; based on YOLOv11, an EMA attention unit is introduced to construct a YOLOv11-EMA instance cell segmentation model, and the preprocessed microscope cell image is subjected to instance segmentation to obtain a cell initial segmentation mask; the cell initial segmentation mask is subjected to false positive filtering and morphological refinement processing in sequence to obtain a cell refined mask; and the cell refined mask is subjected to adhesion cell separation and multi-parameter morphological measurement in sequence to obtain a cell morphological parameter measurement result. The application can improve the segmentation accuracy of cells by enhancing the multi-scale cell feature extraction capability, and realize the measurement accuracy of cell morphology. The application can be widely applied to the technical field of computer vision as a multi-cell automatic measurement method and system based on YOLOv11 combined with EMA.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a multi-cell automated measurement method and system based on the combination of YOLOv11 and EMA. Background Technology

[0002] Microscopic cell image observation and analysis is a core component of biology laboratory teaching in junior and senior high schools, and a crucial pathway for students to develop an understanding of cell structure and master microscopic observation techniques. However, the traditional methods of manual visual counting and measurement commonly used in experimental teaching have significant limitations: counting and statistical analysis are labor-intensive, individual sample analysis is time-consuming, and measurement results are highly subjective and easily influenced by differences in students' operational experience, making it difficult to meet the teaching needs of large sample sizes and rapid feedback in classroom teaching. Furthermore, the limited imaging conditions and significant image noise interference of conventional optical microscopes in secondary school laboratories further increase the difficulty of accurate analysis.

[0003] Early cell segmentation methods were primarily based on traditional image processing techniques, including thresholding, edge detection, region growing, and watershed algorithms. While computationally simple, these methods are computationally inefficient, require high image quality, lack robustness, struggle with complex cell morphologies and overlapping structures, and necessitate extensive manual parameter adjustments. In recent years, deep learning-based cell segmentation methods have made significant progress. Common network architectures include U-Net and its variants, Mask R-CNN, and the YOLO series. However, these technologies suffer from the following limitations: 1) Insufficient multi-scale feature extraction; cell size varies greatly, and existing networks struggle to capture cell features at different scales simultaneously.

[0004] 2) The contradiction between segmentation accuracy and computational efficiency: high-precision models usually have high computational complexity and are difficult to meet real-time requirements.

[0005] 3) Poor robustness to noise; inherent noise and artifacts in microscope images severely affect segmentation results.

[0006] 4) Adhesive cells are difficult to handle, and there is a lack of effective automated methods for separating adhesive cells.

[0007] 5) Lack of end-to-end solutions: Existing systems often only focus on the segmentation process and fail to provide a complete process from identification to measurement. Summary of the Invention

[0008] To address the aforementioned technical problems, the present invention aims to provide an automated multi-cell measurement method and system based on the combination of YOLOv11 and EMA, which can improve cell segmentation accuracy by enhancing multi-scale cell feature extraction capabilities and achieving precise cell morphology measurement.

[0009] The first technical solution adopted in this invention is: a multi-cell automated measurement method based on the combination of YOLOv11 and EMA, comprising the following steps: Acquire the microscopic cell image to be processed and perform image data preprocessing to obtain the preprocessed microscopic cell image; Based on YOLOv11, an EMA attention unit is introduced to construct a YOLOv11-EMA instance cell segmentation model, and instance segmentation is performed on the preprocessed microscope cell image to obtain the initial cell segmentation mask. The initial cell segmentation mask was sequentially subjected to false positive filtering and morphological refinement to obtain the refined cell mask. The cell refinement mask was used to separate adherent cells and perform multi-parameter morphological measurements to obtain the cell morphology parameter measurement results.

[0010] Furthermore, the step of acquiring the microscope cell image to be processed and performing image data preprocessing to obtain the preprocessed microscope cell image specifically includes: Acquire the microscope cell image to be processed and perform conversion processing to obtain a grayscale image of the microscope cell; Morphological black-hat transformation was performed on the grayscale images of microscope cells using elliptical structuring elements to extract dark spot regions from the grayscale images of microscope cells. Thresholding is performed on the dark spot regions in the grayscale image of cells under a microscope to generate a dust mask; The dust mask is dilated, and the dust area is filled using an image inpainting algorithm to obtain the restored microscope cell image. Based on the bilateral filtering method, the repaired microscope cell image is subjected to edge-preserving denoising to obtain the preprocessed microscope cell image.

[0011] Furthermore, the YOLOv11-EMA instance cell segmentation model specifically includes a channel compression and feature segmentation module, a cascaded EMA feature enhancement module, and a feature aggregation and output recovery module, wherein: The cascaded EMA feature enhancement module is an EMA attention unit connected to the end of the c2f of the YOLOv11 model. The EMA attention unit includes a dual-branch parallel feature extraction module, a cross-dimensional interaction and adaptive weight generation module, and a weighted feature fusion and output projection module. The feature aggregation and output recovery module consists of a neck network and a detection head network.

[0012] Furthermore, the step of performing instance segmentation on the preprocessed microscope cell image to obtain the initial cell segmentation mask specifically includes: The preprocessed microscope cell images were input into the YOLOv11-EMA instance cell segmentation model; The channel compression and feature segmentation module based on the YOLOv11-EMA instance cell segmentation model is used to segment the preprocessed microscope cell image to obtain the segmented microscope cell image. The cascaded EMA feature enhancement module based on the YOLOv11-EMA instance cell segmentation model performs feature enhancement processing on the segmented microscope cell images to obtain enhanced microscope cell feature images. Based on the feature aggregation and output recovery module of the YOLOv11-EMA instance cell segmentation model, feature stitching is performed on the enhanced microscope cell feature image to obtain the initial cell segmentation mask.

[0013] Furthermore, the cascaded EMA feature enhancement module based on the YOLOv11-EMA instance cell segmentation model performs feature enhancement processing on the segmented microscope cell image to obtain an enhanced microscope cell feature image. This step specifically includes: The segmented microscope cell images are input into the cascaded EMA feature enhancement module based on the YOLOv11-EMA instance cell segmentation model; The dual-branch parallel feature extraction module based on the cascaded EMA feature enhancement module performs global semantic feature extraction and local structural feature extraction on the segmented microscope cell images to obtain global feature images and local structural images of microscope cells. The cross-dimensional interaction and adaptive weight generation module based on the cascaded EMA feature enhancement module performs global average pooling on the global feature image and the local structure image of the microscope cell to generate global descriptors. The correlation matrix between the global descriptors is calculated by matrix operation and cross-dimensional interaction is performed to obtain the weight values ​​of the global feature image and the local structure image of the microscope cell. The weighted feature fusion and output projection module based on the cascaded EMA feature enhancement module linearly combines the global feature image and the local structure image of the microscope cell based on the weight values ​​of the global feature image and the local structure image of the microscope cell to obtain the enhanced microscope cell feature image.

[0014] Furthermore, the step of sequentially performing false positive filtering and morphological refinement on the initial cell segmentation mask to obtain the refined cell mask specifically includes: The initial cell segmentation mask is processed by external contour lookup, and the contour with the largest area is selected as the target contour. Obtain the area, perimeter, convex hull area, and minimum bounding rectangle of the target contour, and calculate the circularity, solidity, and aspect ratio. Based on the preset area range, preset minimum circularity, preset solidity, and preset aspect ratio range, the area, circularity, solidity, and aspect ratio of the initial cell segmentation mask are filtered to obtain the refined cell mask.

[0015] Furthermore, the step of sequentially separating adherent cells and performing multi-parameter morphological measurements on the cell refinement mask to obtain cell morphological parameter measurement results specifically includes: Perform a distance transformation on the cell refinement mask and calculate the Euclidean distance from each foreground pixel to the nearest background pixel; Select the maximum value of the Euclidean distance and perform thresholding processing according to a preset ratio to determine the foreground seed point region; Perform an iterative dilation operation on the refined cell mask to determine the background region; The unknown region is determined by subtracting the background region from the foreground seed point region. Perform connected component analysis on the foreground seed point region, add initial label values ​​to the foreground seed point region, and then increment each initial label value by one. The foreground seed point region is labeled with the initial label value incremented by one, the background region is labeled with 1, and the unknown region is labeled with 0, thus constructing a label graph. The cell refinement mask is converted into a three-channel image, and the three-channel image and the label map are segmented by the watershed algorithm to obtain the segmented label map; Extracting the marker values ​​greater than in the segmented marker map All regions are used as individual cell masks, and masks with areas smaller than the preset minimum area are filtered out to output the final individual cell mask. Multi-parameter morphological measurements were performed on the final single-cell mask to obtain the cell morphology parameter measurement results.

[0016] Furthermore, the cell morphology parameter measurement results include calculating the area, perimeter, equivalent diameter, minimum circumscribed rectangle, minimum circumscribed circle, aspect ratio, roundness, solidity, eccentricity, and centroid position of each cell.

[0017] The second technical solution adopted in this invention is: a multi-cell automated measurement system based on the combination of YOLOv11 and EMA, comprising: The first module is used to acquire the microscope cell image to be processed and perform image data preprocessing to obtain the preprocessed microscope cell image. The second module is used to construct a YOLOv11-EMA instance cell segmentation model based on YOLOv11 by introducing EMA attention units, and to perform instance segmentation on the preprocessed microscope cell images to obtain the initial cell segmentation mask. The third module is used to sequentially perform false positive filtering and morphological refinement on the initial cell segmentation mask to obtain the refined cell mask. The fourth module is used to sequentially separate adherent cells and perform multi-parameter morphological measurements on the cell refinement mask to obtain the cell morphology parameter measurement results.

[0018] The beneficial effects of the method and system of this invention are as follows: This invention acquires microscopic cell images to be processed and performs image data preprocessing to obtain preprocessed microscopic cell images, effectively solving the noise interference such as dust and dark spots in microscopic images, and maintaining the integrity of cell boundaries while removing noise; further, based on YOLOv11, an EMA attention unit is introduced to construct a YOLOv11-EMA instance cell segmentation model, and instance segmentation is performed on the preprocessed microscopic cell images to obtain an initial cell segmentation mask. By introducing the efficient multi-scale attention mechanism of EMA, the model can better capture the features of cells of different sizes; then, false positive filtering and morphological refinement processing are performed sequentially on the initial cell segmentation mask to obtain a refined cell mask; finally, adhering cell separation and multi-parameter morphological measurement are performed sequentially on the refined cell mask to obtain cell morphological parameter measurement results. The adhesion separation strategy based on distance transformation and watershed algorithm can effectively identify and separate tightly adhered cells, avoiding errors in cell counting and morphological analysis, and improving cell segmentation accuracy by enhancing multi-scale cell feature extraction capabilities, thereby achieving accurate cell morphology measurement. Attached Figure Description

[0019] Figure 1 This is a flowchart of the steps of the multi-cell automated measurement method based on the combination of YOLOv11 and EMA of the present invention; Figure 2 This is a block diagram of the multi-cell automated measurement system based on the combination of YOLOv11 and EMA of the present invention; Figure 3 This is a schematic diagram of the automated multi-cell measurement process provided in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the EMA attention unit provided in a specific embodiment of the present invention; Figure 5 This is a schematic diagram of the YOLOv11-EMA instance cell segmentation model structure provided in a specific embodiment of the present invention; Figure 6 This is a schematic diagram of the image preprocessing process provided in a specific embodiment of the present invention; Figure 7 This is a schematic diagram of the process for separating adhesion cells provided in a specific embodiment of the present invention; Figure 8 This is a schematic diagram of multi-parameter morphological measurement provided in a specific embodiment of the present invention; Figure 9 This is a schematic diagram showing the comparison of experimental results provided in a specific embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0021] Reference Figure 1 This invention provides an automated multi-cell measurement method based on the combination of YOLOv11 and EMA, which includes the following steps: S100: Acquire the microscope cell image to be processed and perform image data preprocessing to obtain the preprocessed microscope cell image; First, it should be noted that the input image is subjected to morphological black hat transformation to detect dark spot regions, a dust mask is generated by setting a threshold, and an image inpainting algorithm is used to fill the dust regions; contrast enhancement includes: applying CLAHE adaptive histogram equalization to the grayscale image, and selecting the appropriate edge enhancement strategy according to the cell type.

[0022] Specifically, the process involves acquiring and converting a microscope cell image to obtain a grayscale image of the microscope cells; performing a morphological black hat transformation on the grayscale image of the microscope cells using an elliptical structuring element to extract dark spot regions from the grayscale image; thresholding the dark spot regions in the grayscale image of the microscope cells to generate a dust mask; dilating the dust mask and filling the dust region using an image inpainting algorithm to obtain a repaired microscope cell image; and then performing edge-preserving denoising on the repaired microscope cell image using a bilateral filtering method to obtain a preprocessed microscope cell image.

[0023] In some specific embodiments, the color input image is converted into a grayscale image; a morphological black hat transform is performed on the grayscale image using an elliptical structuring element to extract dark spot regions in the image; the black hat transform result is thresholded to generate a dust mask; the dust mask is dilated to ensure complete coverage of the dust area; the Telea image inpainting algorithm or fast traversal method is used to fill the dust area using pixel information around the mask; bilateral filtering is applied to the inpainted image to preserve edge-preserving noise and retain cell edge information.

[0024] Among them, contrast enhancement employs different edge enhancement strategies based on different cell types, specifically including: 1) For onion epidermal cells, a cell wall enhancement strategy is employed: the Sobel operator is used to compute the image's... direction and Directional gradient, calculate the gradient magnitude and normalize it to Range, compare edge information with the original image Weight fusion; 2) For blood cells, a circular edge enhancement strategy is adopted: Gaussian blur with different parameters is used for difference operations (DoG), the difference results are contrast-enhanced, and the enhanced results are compared with the original image. Weight fusion; 3) For common cell types, a common edge enhancement strategy is adopted: edge detection is performed using the Laplacian operator, the edge results are slightly enhanced and fused with the original image.

[0025] S200: Based on YOLOv11, an EMA attention unit is introduced to construct a YOLOv11-EMA instance cell segmentation model, and instance segmentation is performed on the preprocessed microscope cell image to obtain the initial cell segmentation mask. First, it should be noted that the construction process of the efficient multi-scale attention (EMA) module includes: grouping and reshaping the input feature map according to the grouping factor; processing through a two-branch structure, where the first branch adopts a coordinate attention structure, performing global average pooling in the height and width directions on the grouped features respectively, segmenting and recombining them after 1 × 1 convolution, calculating coordinate attention and applying group normalization; the second branch uses 3 × 3 convolution for feature extraction; performing cross-dimensional interaction on the outputs of the two branches, obtaining channel descriptors through global average pooling, calculating the cross-branch interaction weight matrix, multiplying the attention weights with the original grouped features; restoring the shape of the original feature map, and outputting the attention-enhanced feature map.

[0026] Furthermore, the YOLOv11-EMA instance cell segmentation model specifically includes a channel compression and feature segmentation module, a cascaded EMA feature enhancement module, and a feature aggregation and output recovery module. The cascaded EMA feature enhancement module connects EMA attention units to the end of the c2f layer of the YOLOv11 model, that is, embeds EMA modules in multiple scale layers (P4 / 16, P5 / 32) of the YOLOv11 backbone network to enhance multi-scale feature expression capabilities. The EMA attention unit includes a dual-branch parallel feature extraction module, a cross-dimensional interaction and adaptive weight generation module, and a weighted feature fusion and output projection module. The feature aggregation and output recovery module consists of a neck network and a detection head network.

[0027] In this embodiment, the preprocessed microscope cell image is input into the YOLOv11-EMA instance cell segmentation model; the channel compression and feature segmentation module based on the YOLOv11-EMA instance cell segmentation model performs segmentation processing on the preprocessed microscope cell image to obtain the segmented microscope cell image; the cascaded EMA feature enhancement module based on the YOLOv11-EMA instance cell segmentation model performs feature enhancement processing on the segmented microscope cell image to obtain the enhanced microscope cell feature image; the feature aggregation and output recovery module based on the YOLOv11-EMA instance cell segmentation model performs feature stitching on the enhanced microscope cell feature image to obtain the initial cell segmentation mask.

[0028] More specifically, the segmented microscope cell images are input into a cascaded EMA feature enhancement module based on the YOLOv11-EMA instance cell segmentation model. A dual-branch parallel feature extraction module based on the cascaded EMA feature enhancement module performs global semantic feature extraction and local structural feature extraction on the segmented microscope cell images, obtaining a global feature image and a local structural image of the microscope cells. A cross-dimensional interaction and adaptive weight generation module based on the cascaded EMA feature enhancement module performs global average pooling on both the global feature image and the local structural image of the microscope cells, generating global descriptors. The correlation matrix between the global descriptors is calculated through matrix operations, and cross-dimensional interaction is performed to obtain the weight values ​​of the global feature image and the local structural image of the microscope cells. A weighted feature fusion and output projection module based on the cascaded EMA feature enhancement module performs a linear combination of the global feature image and the local structural image of the microscope cells based on their weight values, obtaining the enhanced microscope cell feature image.

[0029] It should also be noted that the EMA attention unit is used to process the input feature map. By grouping factor The system performs grouping and reshaping, employs a dual-branch structure for cross-dimensional interactive computation, and outputs attention-enhanced feature maps. The dual-branch structure includes a coordinate attention branch and... Convolutional branch; the formula for calculating the cross-dimensional interaction weights is: in and These are the normalized weight matrices for the two branches, respectively. and These are the feature outputs of the two branches, respectively; The output of the EMA attention unit is: The YOLOv11 segmentation unit is used to integrate the EMA attention unit to perform instance segmentation on the preprocessed image and output the initial segmentation mask for cells.

[0030] S300. The initial cell segmentation mask is sequentially subjected to false positive filtering and morphological refinement to obtain the refined cell mask. First, it should be noted that the morphological features of each mask are calculated, including area, roundness, solidity, and aspect ratio; based on the set cell type parameter threshold, noise detection results that do not conform to the cell features are filtered out; the morphological refinement includes: performing opening and closing operations on the mask to remove small holes and burrs, and using an edge smoothing algorithm to optimize the mask boundary.

[0031] Specifically, the initial cell segmentation mask undergoes external contour lookup processing, and the contour with the largest area is selected as the target contour. The area, perimeter, convex hull area, and minimum bounding rectangle of the target contour are obtained, and the circularity, solidity, and aspect ratio are calculated. Based on the preset area range, preset minimum circularity value, preset solidity, and preset aspect ratio range, the area, circularity, solidity, and aspect ratio of the initial cell segmentation mask are filtered to obtain the refined cell mask.

[0032] In some specific embodiments, including: 1) For each mask, find the outer contour and take the contour with the largest area as the target contour; 2) Calculate the area of ​​the target contour. ,perimeter convex hull area and the minimum bounding rectangle; 3) Calculate roundness Solidity and aspect ratio : in and These are the width and height of the minimum bounding rectangle, respectively; 4) Determine the effectiveness of the mask based on cell type: area within Within the range, roundness Solidity Aspect Ratio ; More specifically, adaptive threshold filtering is applied based on cell type: for blood cells, roundness ≥ 0.6 and solidity ≥ 0.8; for onion epidermal cells, roundness ≥ 0.1 and solidity ≥ 0.6; and for general cells, roundness ≥ 0.2 and solidity ≥ 0.5.

[0033] 5) Filter out all masks that do not meet the conditions and retain the valid cell masks.

[0034] S400: The cell refinement mask is used to sequentially separate adherent cells and perform multi-parameter morphological measurements to obtain the cell morphology parameter measurement results.

[0035] First, it needs to be explained that the process involves: performing a distance transformation on the mask to determine the foreground seed point by setting a threshold ratio; performing a dilation operation on the mask to determine the background region; calculating the difference between the background and foreground for unknown regions; labeling the foreground seed point as a connected component, setting the background label to 1 and the unknown region label to 0; applying the watershed algorithm for segmentation to extract the cell mask for each segmented region; then calculating the area, perimeter, equivalent diameter, minimum bounding rectangle, minimum bounding circle, aspect ratio, circularity, solidity, eccentricity, and centroid position of each cell; converting pixel units to physical units according to preset pixel size parameters; and generating a structured data report containing all measurement parameters.

[0036] Specifically, a distance transformation is performed on the refined cell mask to calculate the Euclidean distance from each foreground pixel to the nearest background pixel; the maximum value of the Euclidean distance is selected and thresholded according to a preset ratio to determine the foreground seed point region; an iterative dilation operation is performed on the refined cell mask to determine the background region; the unknown region is determined by subtracting the background region from the foreground seed point region; connected component analysis is performed on the foreground seed point region, an initial label value is added to the foreground seed point region, and then each initial label value is incremented by one, with the foreground seed point region labeled as the initial label value incremented by one, the background region labeled as 1, and the unknown region labeled as 0, thus constructing a label map; the refined cell mask is converted into a three-channel image, and the three-channel image and the label map are segmented using a watershed algorithm to obtain the segmented label map; the label values ​​greater than 1 in the segmented label map are extracted. All regions are used as individual cell masks, and masks with areas smaller than a preset minimum area are filtered out to output the final individual cell mask; multi-parameter morphological measurements are performed on the final individual cell mask to obtain the cell morphology parameter measurement results.

[0037] In some specific embodiments, including: 1) Input mask Perform distance transformation Calculate the Euclidean distance from each foreground pixel to the nearest background pixel; 2) Take the maximum value of the distance transformation. According to the set ratio Thresholding is performed to determine the foreground seed point region: 3) Perform an expansion operation on the original mask, iterating... Next, determine the background area. ; 4) Calculate the unknown region as the background region minus the foreground seed point region: ; 5) Perform connected component analysis on the foreground seed points, assign a unique label value to each connected region, and obtain the label map. ; 6) Add the marker value Make the background marked as Set the unknown area marker to : 7) Convert the original mask into a three-channel image. The watershed algorithm is applied for segmentation: ; 8) Extracting marker values ​​greater than All regions are used as a single cell mask, with the filtered area being less than a threshold. The mask.

[0038] Furthermore, it should be noted that multi-parameter morphological measurements include: calculating the area, perimeter, equivalent diameter, minimum circumscribed rectangle, minimum circumscribed circle, aspect ratio, roundness, solidity, eccentricity, and centroid position of each cell.

[0039] In some specific embodiments, the specific components include: For area: calculated using the contour area function, in units of pixel squared, based on pixel size. ( m / pixel) converted to : For the perimeter: it is calculated using the contour arc length function, in pixels, and then converted according to the pixel size. m: For the equivalent diameter: the formula for calculating the diameter of a circle with equal area is: For the minimum bounding rectangle: calculate the center using the minimum bounding rectangle function. ,width ,high and rotation angle The aspect ratio is: For the smallest circumcircle: the center is obtained by calculating using the smallest circumcircle function. and radius .

[0040] For roundness: the calculation formula is: The range of values ​​is , It represents a perfect circle.

[0041] For solidity: the calculation formula is: in This represents the area of ​​the convex hull.

[0042] For eccentricity: calculated through ellipse fitting, let the major axis be... The minor axis is ,but: For the centroid: calculated using profile moments, let... It is the zeroth moment. and Let be a first-order moment, then: The scale calibration method includes: obtaining the physical length of the scale. ( m) and the corresponding number of pixels Calculate pixel size: Convert all measurement parameters from pixel units to physical units. m).

[0043] Finally, as Figure 3 As shown in the accompanying drawings, the embodiments of the present invention will be described and explained in conjunction with the following: First, such as Figure 6 As shown, the first step is data acquisition and preprocessing. The first step is image acquisition, which involves acquiring cell images taken under a microscope. Supported image formats include JPG, PNG, BMP, and TIFF. The images should meet the following basic requirements: 1) Resolution: It is recommended to have a resolution of at least 512×512 pixels.

[0044] 2) Bit depth: 8-bit grayscale or 24-bit RGB color image.

[0045] 3) Magnification: Record the magnification of the microscope objective lens for subsequent size calibration.

[0046] Secondly, noise removal is performed. For dust and dark spot noise in microscope images, this invention first converts the image to grayscale, uses morphological black hat transformation combined with elliptical structure kernel to locate dark features, and generates an optimized noise mask through thresholding and dilation. Then, the Telea fast traversal method is used to fill and repair the noise area based on the mask. Finally, bilateral filtering is applied to remove residual noise while preserving edge details, achieving high-quality noise reduction of the image.

[0047] Further contrast enhancement is performed (customized according to cell type). For the general cell enhancement strategy, the CLAHE algorithm is first applied to the input image for local histogram equalization to improve contrast and suppress noise amplification. Then, the second derivative of Laplacian is calculated using a 3×3 convolution kernel and the absolute value is taken to extract high-frequency edge details. Finally, a weighted fusion strategy is adopted to synthesize the CLAHE-enhanced image and the edge detail map according to preset weights, which significantly improves the edge sharpness of the image while maintaining the overall tone.

[0048] For the blood cell enhancement strategy, Gaussian blur with different standard deviations is applied to the input image, and a Gaussian difference map is constructed by calculating the difference to extract edge features matching the blood cell scale and suppress background noise. Then, the map is normalized. Finally, a weighted fusion strategy is adopted to superimpose the edge feature map onto the original image with preset weights, which enhances the edge contour and texture details of blood cells while preserving the grayscale characteristics of the original image.

[0049] For the onion epidermal cell enhancement strategy, the Euclidean norm of the image gradient is calculated using the 3×3 Sobel operator to extract the edge information of linear structures such as cell walls; the gradient magnitude map is normalized to unify the data range; then a weighted fusion strategy is adopted to superimpose the normalized edge feature map onto the original image according to the preset weight, thereby enhancing the contour contrast of the cell wall while maintaining the texture of the cell interior.

[0050] Further design of the EMA attention module, such as Figure 4 As shown, the EMA (Efficient Multi-Scale Attention) module is the core innovation of this invention, and its structure is as follows: 1) Dual-branch parallel feature extraction: The module divides the input feature map into two parallel branches for processing. Branch 1 (Global Semantic Branch): After dimensionality reduction using 1×1 convolution, a coordinate attention module is connected. This branch focuses on capturing long-range dependencies and global location information in the feature maps, enhancing the ability to locate key regions.

[0051] Branch 2 (Local Structure Branch): Local feature extraction is performed using 3×3 convolutions (with padding to maintain size invariance). This branch focuses on preserving the local texture details and spatial structure information of the image.

[0052] 2) Cross-dimensional interaction and adaptive weight generation: First, global average pooling is applied to the outputs of both branches to compress the features into global descriptors. Then, matrix operations (transpose and add) are used to calculate the correlation matrix between these two descriptors. The sigmoid activation function is then used to map the correlation matrix into dynamic attention weights. This process enables deep interaction between features of different dimensions (global and local), adaptively adjusting the contribution of the two branches based on the input content.

[0053] 3) Weighted Feature Fusion and Output Projection: Based on the generated attention weights, an adaptive weighted fusion strategy is used to linearly combine the outputs of the two branches. The attention weights determine the balance between preserving global semantics (branch 1) and preserving local structure (branch 2). Finally, a 1×1 convolutional layer is used to project and integrate the fused features, outputting the final enhanced feature map.

[0054] Further integration of the EMA module with YOLOv11 includes: 1) Channel compression and feature segmentation: First, the number of channels in the input feature map is compressed to a specified width (c2 * e) using a 1×1 convolutional layer (cv1), and the output is divided into two parts along the channel dimension. This design draws inspiration from ELAN (Efficient Layer Aggregation Network), which enriches the network's feature learning flow by segmenting gradient paths.

[0055] 2) Cascaded EMA Feature Enhancement: The module contains a sequence list of n EMAModules. During processing, the input features are processed sequentially through a subset of the segmented features as initial input, and then processed through these n EMA modules. At each EMA module, the features undergo multi-scale attention enhancement based on cross-dimensional interaction, thereby gradually extracting more abstract and discriminative semantic information.

[0056] 3) Feature aggregation and output restoration: All intermediate processed features (including the two parts of the initial segmentation and the n parts of features output by the EMA module) are concatenated along the channel dimension. This dense connection method enables the reuse and fusion of features at different depths. Finally, another 1×1 convolutional layer (cv2) is used to perform channel blending and dimension adjustment on the concatenated features, outputting the final target feature map, maintaining the same number of channels as the input (c2) to adapt to the subsequent network structure.

[0057] like Figure 5 The diagram shows the YOLOv11-EMA network structure. The complete YOLOv11-EMA segmentation network structure is shown in Table 1 below.

[0058] Table 1. Structure of YOLOv11-EMA Segmentation Network The process then proceeds to the instance segmentation inference flow, which is used to perform object detection and instance segmentation inference tasks based on the YOLOv11-EMA model. Specifically, this includes: 1) Model inference and parameter configuration: The trained YOLOv11-EMA model is used to predict the preprocessed image. Key optimization parameters are set during the inference process: retina_masks=True: Enables a high-resolution mask generation mechanism, ensuring that the output instance segmentation mask preserves the fine edge details of the original image.

[0059] conf=0.25: Sets the confidence threshold to 0.25, filtering out target boxes with low confidence and reducing false detections.

[0060] iou=0.7: Sets the cross-union ratio threshold of non-maximum suppression (NMS) to 0.7, which is used to eliminate redundant detections with excessive overlap.

[0061] imgsz=640: Specifies the input image size for inference as 640 pixels to ensure a balance between computational efficiency and detection accuracy.

[0062] 2) Detection result parsing and traversal: Traverse the result set output by the model and parse each detected target. This process simultaneously extracts bounding box information (boxes) and mask information (masks), achieving precise alignment between the detection boxes and the segmentation masks.

[0063] 3) Structured data extraction and encapsulation: The prediction information for each target is encapsulated into a structured dictionary format. The key information extracted includes: class_id: The index of the target category.

[0064] confidence: prediction confidence score.

[0065] bbox: The coordinates of the target bounding box (in the format [x1, y1, x2, y2]).

[0066] mask: Binary mask data for instance segmentation (needs to be transferred to the CPU and converted to a NumPy array).

[0067] area: The pixel area covered by the mask, used for subsequent target scale analysis or filtering.

[0068] 4) Returns the final detection result. The function returns a dictionary list containing all valid detection information, providing a standardized data interface for downstream tasks such as target counting, area measurement, or visualization.

[0069] To further filter false positives, morphological features are first calculated. For each detected cell region, the following morphological features are calculated: Contour extraction and main target localization first utilize a contour detection algorithm to extract the contours of all connected components from a binary mask. By comparing contour areas, the contour with the largest area is selected and locked as the main target, thereby eliminating the influence of background noise or minor interference and ensuring the stability of subsequent calculations.

[0070] Then, geometric primitives and convex hull calculations are performed: Basic geometric quantities: The area and perimeter of the calculated profile serve as the basis for calculating derived features.

[0071] Convex hull analysis: Constructs the convex hull of the contour and calculates the convex hull area. The convex hull reflects the external bounding shape of the target and is used to measure the degree of concavity and convexity of the shape.

[0072] Boundary rectangle: Obtain the coordinates, width, and height of the bounding rectangle of the outline, used to describe the basic spatial occupancy of the target.

[0073] Ellipse fitting and axial parameters are used to fit the maximum contour with a minimum bounding ellipse (if the contour has fewer than 5 points, it degenerates to using the width and height of the boundary rectangle), extracting the lengths of the major and minor axes of the ellipse. These two parameters are used to describe the extension direction and flatness of the target.

[0074] In performing morphological feature quantification: Based on the above geometric primitives, the following key morphological characteristic indicators are calculated: Circularity: Calculated using the formula Calculation. The closer the value is to 1, the closer the target shape is to a perfect circle; the smaller the value, the more irregular the shape or the rougher the edges.

[0075] Solidity: This is the ratio of the target area to its convex hull area. This indicator measures the degree of indentation of the target. A low solidity indicates that there are significant indentations or burrs on the target's edges.

[0076] Aspect ratio: The ratio of the width to the height of the bounding rectangle. Used to describe the overall shape tendency of the target, such as slender or approximately square.

[0077] Eccentricity: The eccentricity calculated based on the major and minor axes. It describes the flatness of an ellipse, with 0 representing a perfect circle and close to 1 representing an extremely flat, linear shape.

[0078] Finally, the feature vector is output. The function encapsulates the calculated geometric parameters into a dictionary format and returns it, forming a complete multidimensional description of the target's shape. This facilitates downstream algorithms to perform accurate screening and judgment based on these features.

[0079] Adaptive threshold filtering. False positive filtering is performed based on morphological features to improve detection accuracy. The algorithm configures differentiated morphological threshold parameters according to different application scenarios such as blood cells, onion cells, and general types; it iterates through the detection results and extracts feature indicators such as the roundness, solidity, and pixel area of ​​the target; through multi-condition joint judgment, only valid targets whose indicators all meet the preset range are retained, and the filtered detection results are finally output.

[0080] Furthermore, such as Figure 7 As shown, adhesion cell separation is performed. First, adhesion is determined. Based on a two-dimensional mask of the target area, the pixel area and the minimum circumscribed convex hull area are calculated. The ratio of the two, i.e., the solidity, is used as the core indicator to quantify the geometric morphological features. By comparing the solidity with a preset threshold, edge depressions are identified according to the degree of deviation between the target shape and the ideal convex body, thereby determining whether cells are adhered.

[0081] Then, a watershed separation algorithm is performed. Based on the watershed algorithm, a distance transformation is applied to the binary mask to quantify the geometric features of the foreground. The foreground is determined by adaptive threshold extraction and a marker map is constructed by combining morphological operations to divide the unknown region. The marker map is used to guide watershed segmentation to generate topological geodesics to separate the sticky targets. Finally, noise is filtered by minimum area constraint to output an independent cell segmentation mask.

[0082] Separation parameters are adaptively configured based on cell type, and watershed segmentation is performed on the adhesion target based on the adhesion determination result. Successfully separated sub-regions are reconstructed into independent objects through logical verification, and the processed results are integrated with the unprocessed objects to output a unified cell detection list.

[0083] Furthermore, such as Figure 8 As shown, automatic measurement of cell morphology parameters was performed: 1) Pixel size calibration: Construct a pixel size calibration model and realize static mapping based on hardware parameters through a preset objective magnification lookup table; For images containing a ruler, use edge detection and contour parsing to extract the ruler pixel width, and calculate the mapping factor based on known physical length to realize dynamic calibration based on image features.

[0084] 2) Morphological parameter measurement: The outer contour of the binary mask is extracted, and the pixel-level area and perimeter are calculated. A calibration factor is introduced to convert the measurements into actual physical quantities. A minimum circumscribed convex hull and a rotated rectangle are constructed through geometric fitting to obtain spatial morphological features such as equivalent diameter and aspect ratio. Based on this, dimensionless shape factors such as roundness, solidity, eccentricity, and compactness are derived to comprehensively quantify the geometric complexity and topological properties of the target. Finally, multidimensional parameters and centroid coordinates are integrated to generate a standardized structured dictionary, providing accurate data support for cell morphology research.

[0085] Finally, the results are output and visualized: 1) Export results: Traverse the list of detection results, extract and flatten the recombinant cell IDs, metadata and morphological indicators; perform differential encoding according to the specified format, convert the data into a DataFrame table or JSON string, achieve persistent file storage and return a structured dataset.

[0086] 2) Visualization generation: Create image copies and use color coding to distinguish targets. Use alpha blending to achieve semi-transparent mask overlay. Draw bounding boxes and high-contrast outlines, and label the targets with index numbers and area information. Finally, generate a visualization image and perform persistent storage.

[0087] Finally, the implementation effect is verified, such as... Figure 9 As shown, to verify the effectiveness of the present invention, tests were conducted on various microscopic cell image datasets, and the test data are shown in Tables 2 and 3.

[0088] Table 2 Test Dataset Data Table Table 3 Performance Comparison Data Table In summary, compared with the original YOLOv11, the (YOLOv11-EMA) provided by this embodiment of the invention improves mAP@0.5 by about 1.8%. Compared with other attention mechanisms (such as CBAM), EMA achieves a better balance between accuracy and efficiency, with a computational cost increase of less than 1% (GFLOPs increase from 286.8 to 289.2), and significantly improves segmentation performance on small cells and cells with blurred boundaries.

[0089] In summary, the embodiments of the present invention are applicable to images of plant and animal cells acquired under conventional optical microscopes (such as 4×-100× magnification). Typical applications include commonly used biological samples in secondary school experiments such as epidermal cells of onion scale leaves, oral epithelial cells, and human blood cells. The methodology includes the following steps: First, image preprocessing is performed, using morphological top-hat transformation to remove dust particles and background noise introduced during optical imaging, and combining this with the CLAHE (Contrast-Limited Adaptive Histogram Equalization) algorithm to enhance image contrast and improve target-background separability. Second, the model building stage is entered, integrating an efficient multi-scale attention (EMA) module into the YOLOv11 backbone network. This module captures cross-dimensional pixel-level feature dependencies through a grouped channel strategy and a dual-branch parallel structure, enhancing the model's ability to express features of multi-scale cellular targets. Third, the post-processing stage is further entered, where false positive suppression and morphological refinement operations are performed on the segmentation mask, and the automatic separation of adherent cells is achieved based on an improved watershed algorithm. Finally, the morphological parameter quantification stage is performed, calculating more than 10 morphological feature parameters, including area, perimeter, equivalent diameter, circularity, aspect ratio, and concavity-convexity ratio, and generating a structured measurement report and visualization analysis results. This technology solves the problems of high noise interference in microscope images, difficulty in segmenting adherent cells, and low segmentation accuracy due to insufficient extraction of multi-scale cell features in existing technologies. It improves mAP by about 1.8%, noise filtering rate >95%, and adhesion separation success rate >90%, significantly improving the automation level and measurement accuracy of cell morphology analysis for basic education applications.

[0090] Therefore, this invention proposes a microscopic multi-cell segmentation and automatic measurement method that combines the YOLOv11 target detection and segmentation model with the EMA (Efficient Multi-Scale Attention) algorithm. This method is applicable to the intelligent identification, precise segmentation, and automatic measurement of morphological parameters of commonly used biological samples in middle school experiments, such as onion epidermal cells and human oral epithelial cells, under conventional optical microscopes. It can be widely applied in middle school cell observation experiments, exploratory experimental data analysis, and student lab report generation, effectively improving experimental teaching efficiency and analytical accuracy.

[0091] The embodiments of the present invention have the following advantages compared with the prior art: 1) Significantly improved segmentation accuracy: By introducing the efficient multi-scale attention mechanism EMA, the model can better capture the features of cells of different sizes. Under the same training conditions, compared with the original YOLOv11 model, the mAP (mean accuracy) of this invention is improved by about 1.8%, especially in the segmentation of small cells and cells with blurred boundaries.

[0092] 2) Computational efficiency remains high. The EMA module adopts a grouping strategy and a dual-branch parallel structure, which improves performance while keeping the increase in computational complexity manageable. The increase in GFLOPs (billion floating-point operations) is less than 1%, ensuring the model's real-time processing capability, with a single image processing time of less than 1 second (GPU environment).

[0093] 3) Strong noise resistance: A dedicated noise removal preprocessing module effectively solves noise interference such as dust and dark spots in microscope images. Morphological black hat transformation combined with image inpainting algorithms can maintain the integrity of cell boundaries while removing noise.

[0094] 4) Cell type adaptation: The system provides optimized configurations for different cell types (blood cells, onion epidermal cells, general cells), including differentiated preprocessing strategies, postprocessing thresholds, and separation parameters, which improves the universality and accuracy of the method.

[0095] 5) Automatic separation of adherent cells: The adhesion separation strategy based on distance transformation and watershed algorithm can effectively identify and separate tightly adherent cells, avoiding errors in cell counting and morphological analysis.

[0096] 6) End-to-end automation: This invention provides a complete automated process from image input to measurement parameter output. Users can obtain accurate cell identification, segmentation and measurement results without complicated parameter adjustments, which greatly improves work efficiency.

[0097] 7) The system offers a wide range of measurement parameters, supporting the automatic measurement of more than 10 morphological parameters, covering multiple dimensions such as area, perimeter, shape, and roundness, providing comprehensive data support for cell morphology analysis.

[0098] 8) Easy to integrate and expand: The modular design makes the system easy to integrate into existing microscope image analysis platforms, while supporting custom cell type configurations, which facilitates expansion into new application areas.

[0099] Reference Figure 2 A multi-cell automated measurement system based on YOLOv11 and EMA, including: The first module 201 is used to acquire the microscope cell image to be processed and perform image data preprocessing to obtain the preprocessed microscope cell image. The second module 202 is used to construct a YOLOv11-EMA instance cell segmentation model based on YOLOv11 by introducing EMA attention units, and to perform instance segmentation on the preprocessed microscope cell image to obtain the initial cell segmentation mask. The third module 203 is used to sequentially perform false positive filtering and morphological refinement on the initial cell segmentation mask to obtain the refined cell mask. The fourth module 204 is used to sequentially separate adherent cells and perform multi-parameter morphological measurements on the cell refinement mask to obtain the cell morphological parameter measurement results.

[0100] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0101] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A multi-cell automated measurement method based on the combination of YOLOv11 and EMA, characterized in that, Includes the following steps: Acquire the microscopic cell image to be processed and perform image data preprocessing to obtain the preprocessed microscopic cell image; Based on YOLOv11, an EMA attention unit is introduced to construct a YOLOv11-EMA instance cell segmentation model, and instance segmentation is performed on the preprocessed microscope cell image to obtain the initial cell segmentation mask. The initial cell segmentation mask was sequentially subjected to false positive filtering and morphological refinement to obtain the refined cell mask. The cell refinement mask was used to separate adherent cells and perform multi-parameter morphological measurements to obtain the cell morphology parameter measurement results.

2. The automated multi-cell measurement method based on the combination of YOLOv11 and EMA according to claim 1, characterized in that, The step of acquiring the microscope cell image to be processed and performing image data preprocessing to obtain the preprocessed microscope cell image specifically includes: Acquire the microscope cell image to be processed and perform conversion processing to obtain a grayscale image of the microscope cell; Morphological black-hat transformation of grayscale images of microscopic cells is performed using elliptical structuring elements to extract dark spot regions in the grayscale images of microscopic cells. Thresholding is performed on the dark spot regions in the grayscale image of cells under a microscope to generate a dust mask; The dust mask is dilated, and the dust area is filled using an image inpainting algorithm to obtain the restored microscope cell image. Based on the bilateral filtering method, the repaired microscope cell image is subjected to edge-preserving denoising to obtain the preprocessed microscope cell image.

3. The automated multi-cell measurement method based on the combination of YOLOv11 and EMA according to claim 2, characterized in that, The YOLOv11-EMA instance cell segmentation model specifically includes a channel compression and feature segmentation module, a cascaded EMA feature enhancement module, and a feature aggregation and output recovery module, wherein: The cascaded EMA feature enhancement module is an EMA attention unit connected to the end of the c2f of the YOLOv11 model. The EMA attention unit includes a dual-branch parallel feature extraction module, a cross-dimensional interaction and adaptive weight generation module, and a weighted feature fusion and output projection module. The feature aggregation and output recovery module consists of a neck network and a detection head network.

4. The automated multi-cell measurement method based on the combination of YOLOv11 and EMA according to claim 3, characterized in that, The step of performing instance segmentation on the preprocessed microscope cell image to obtain the initial cell segmentation mask specifically includes: The preprocessed microscope cell images were input into the YOLOv11-EMA instance cell segmentation model; The channel compression and feature segmentation module based on the YOLOv11-EMA instance cell segmentation model is used to segment the preprocessed microscope cell image to obtain the segmented microscope cell image. The cascaded EMA feature enhancement module based on the YOLOv11-EMA instance cell segmentation model performs feature enhancement processing on the segmented microscope cell images to obtain enhanced microscope cell feature images. Based on the feature aggregation and output recovery module of the YOLOv11-EMA instance cell segmentation model, feature stitching is performed on the enhanced microscope cell feature image to obtain the initial cell segmentation mask.

5. The automated multi-cell measurement method based on the combination of YOLOv11 and EMA according to claim 4, characterized in that, The cascaded EMA feature enhancement module based on the YOLOv11-EMA instance cell segmentation model performs feature enhancement processing on the segmented microscope cell images to obtain enhanced microscope cell feature images. This step specifically includes: The segmented microscope cell images are input into the cascaded EMA feature enhancement module based on the YOLOv11-EMA instance cell segmentation model; The dual-branch parallel feature extraction module based on the cascaded EMA feature enhancement module performs global semantic feature extraction and local structural feature extraction on the segmented microscope cell images to obtain global feature images and local structural images of microscope cells. The cross-dimensional interaction and adaptive weight generation module based on the cascaded EMA feature enhancement module performs global average pooling on the global feature image and the local structure image of the microscope cell to generate global descriptors. The correlation matrix between the global descriptors is calculated by matrix operation and cross-dimensional interaction is performed to obtain the weight values ​​of the global feature image and the local structure image of the microscope cell. The weighted feature fusion and output projection module based on the cascaded EMA feature enhancement module linearly combines the global feature image and the local structure image of the microscope cell based on the weight values ​​of the global feature image and the local structure image of the microscope cell to obtain the enhanced microscope cell feature image.

6. The automated multi-cell measurement method based on the combination of YOLOv11 and EMA according to claim 5, characterized in that, The step of sequentially performing false positive filtering and morphological refinement on the initial cell segmentation mask to obtain the refined cell mask specifically includes: The initial cell segmentation mask is processed by external contour lookup, and the contour with the largest area is selected as the target contour. Obtain the area, perimeter, convex hull area, and minimum bounding rectangle of the target contour, and calculate the circularity, solidity, and aspect ratio. Based on the preset area range, preset minimum circularity, preset solidity, and preset aspect ratio range, the area, circularity, solidity, and aspect ratio of the initial cell segmentation mask are filtered to obtain the refined cell mask.

7. The automated multi-cell measurement method based on the combination of YOLOv11 and EMA according to claim 6, characterized in that, The step of sequentially separating adherent cells and performing multi-parameter morphological measurements on the cell refinement mask to obtain cell morphological parameter measurement results specifically includes: Perform a distance transformation on the cell refinement mask and calculate the Euclidean distance from each foreground pixel to the nearest background pixel; Select the maximum value of the Euclidean distance and perform thresholding processing according to a preset ratio to determine the foreground seed point region; Perform an iterative dilation operation on the refined cell mask to determine the background region; The unknown region is determined by subtracting the background region from the foreground seed point region. Perform connected component analysis on the foreground seed point region, add initial label values ​​to the foreground seed point region, and then increment each initial label value by one. The foreground seed point region is labeled with the initial label value incremented by one, the background region is labeled with 1, and the unknown region is labeled with 0, thus constructing a label graph. The cell refinement mask is converted into a three-channel image, and the three-channel image and the label map are segmented by the watershed algorithm to obtain the segmented label map; Extracting the marker values ​​greater than in the segmented marker map All regions are used as individual cell masks, and masks with areas smaller than the preset minimum area are filtered out to output the final individual cell mask. Multi-parameter morphological measurements were performed on the final single-cell mask to obtain the cell morphology parameter measurement results.

8. The automated multi-cell measurement method based on the combination of YOLOv11 and EMA according to claim 7, characterized in that, The cell morphology parameter measurement results include the calculation of the area, perimeter, equivalent diameter, minimum circumscribed rectangle, minimum circumscribed circle, aspect ratio, roundness, solidity, eccentricity, and centroid position of each cell.

9. A multi-cell automated measurement system based on the combination of YOLOv11 and EMA, characterized in that, Includes the following modules: The first module is used to acquire the microscope cell image to be processed and perform image data preprocessing to obtain the preprocessed microscope cell image. The second module is used to construct a YOLOv11-EMA instance cell segmentation model based on YOLOv11 by introducing EMA attention units, and to perform instance segmentation on the preprocessed microscope cell images to obtain the initial cell segmentation mask. The third module is used to sequentially perform false positive filtering and morphological refinement on the initial cell segmentation mask to obtain the refined cell mask. The fourth module is used to sequentially separate adherent cells and perform multi-parameter morphological measurements on the cell refinement mask to obtain the cell morphology parameter measurement results.

Images