A multi-source collaborative peripheral blood cell detection method and system based on cell population structure consistency modeling

Abstract

The application discloses a multi-source collaborative peripheral blood cell detection method based on cell group structure consistency modeling. The method constructs a structure-aware collaborative detection network SC-YOLO, fuses a coordinate attention mechanism, a multi-scale dynamic convolution structure and a self-supervised distillation strategy, and realizes basic detection of multi-scale cell targets in a peripheral blood smear microscopic image. On this basis, a cell group structure consistency graph CSCG is constructed, cell instances detected are taken as graph nodes, a graph structure is established by taking spatial distance, morphological similarity and category consistency as constraints, and global structure reasoning and error correction are carried out through iterative message passing. Through the two-stage collaborative mechanism of the detection network and the structure graph reasoning, the detection accuracy and robustness of dense cell regions, small target platelets and overlapping cells are effectively improved, the generalization ability of the model under different staining conditions and imaging environments is enhanced, and the model can be applied to clinical blood routine automatic analysis and intelligent microscopic auxiliary diagnosis scenes.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and artificial intelligence detection technology, specifically relating to a multi-source collaborative peripheral blood cell detection method and system based on cell population structure consistency modeling. Background Technology

[0002] Peripheral blood cell detection is a core component of clinical hematology testing, playing an irreplaceable role in the early screening, diagnosis, and efficacy evaluation of anemia, infections, leukemia, and various hematologic disorders. Traditional manual microscopic examination relies on subjective interpretation of blood smears by laboratory physicians, which is not only time-consuming and labor-intensive but also highly susceptible to individual differences in experience, making it difficult to meet the needs of large-scale clinical screening and precision medicine. With the rapid development of digital pathology and artificial intelligence technologies, automated peripheral blood cell detection methods based on deep convolutional neural networks have gradually become a research hotspot. These methods achieve automatic cell localization and classification in images through end-to-end feature learning, and are expected to significantly improve the efficiency and objectivity of routine blood cell analysis, possessing significant clinical application value and research significance.

[0003] Existing methods for peripheral blood cell detection are mainly based on general object detection frameworks, among which single-stage detection algorithms, represented by the YOLO series, are widely used due to their good real-time performance and detection accuracy. These methods typically employ a backbone network to perform multi-level downsampling of the input image to extract visual features, then fuse semantic information at different scales through a feature pyramid structure, and finally, the detection head performs cell target category prediction and bounding box regression on the feature map. In practical clinical applications, such methods have shown some feasibility on standard datasets, but their technical implementation faces the following inherent limitations: First, in peripheral blood smear microscopic images, platelets are extremely small, red blood cells are densely distributed and largely overlap, and white blood cells have varied morphologies. Existing methods use fixed receptive fields and fixed upsampling methods in their backbone networks and feature fusion modules, lacking the ability to dynamically adapt to cell targets of different scales, resulting in prominent issues of missed detection of small platelets and localization deviations of overlapping cells. Second, existing methods treat cell detection as an independent instance recognition task, relying solely on local region features to independently distinguish each cell, lacking explicit modeling of spatial distribution relationships, morphological similarity, and population structure consistency among cells. In densely overlapping scenarios, it is difficult to use the structural information of the cell community to constrain and correct the detection results as a whole, easily leading to false detections and bounding box drift. In addition, existing methods show a significant decrease in feature representation ability and generalization performance when faced with different staining conditions, imaging equipment, and differences in data distribution between hospitals, while relying on large-scale finely labeled data to alleviate this problem is extremely costly.

[0004] Therefore, how to construct a detection method capable of adaptively sensing multi-scale cellular targets, explicitly modeling intercellular structural relationships, and maintaining good generalization ability under limited annotation conditions is a technical challenge that urgently needs to be solved in this field. This invention aims to address the limitations of the existing technology by providing an improved peripheral blood cell detection technology. Summary of the Invention

[0005] The technical problem to be solved by this invention is that existing peripheral blood cell detection methods have obvious limitations in multi-scale cell target perception, structural relationship modeling in densely overlapping scenes, and cross-domain generalization ability, which leads to missed detection of small target cells, deviation in the localization of overlapping cells, and a significant decrease in detection performance under different imaging conditions.

[0006] To address the aforementioned problems, one aspect of the present invention provides a multi-source collaborative peripheral blood cell detection method based on cell population structure consistency modeling, comprising:

[0007] S1: Construct a cell image dataset containing red blood cells, white blood cells, platelets and abnormal cells based on peripheral blood smear microscopic images, and preprocess the cell images in the cell image dataset to obtain standard cell images;

[0008] S2: Construct the structure-aware collaborative detection network SC-YOLO, input standard cell images into the SC-YOLO network for forward inference, and obtain initial cell detection results;

[0009] S3: Construct a cell population structure consistency graph (CSCG) based on the initial cell detection results, and perform iterative structural reasoning on the node features in the CSCG through a graph structure message passing mechanism to obtain the cell detection results after structural correction.

[0010] S4: Perform non-maximum suppression on the structure-corrected cell detection results to obtain the final cell detection results;

[0011] The training process of the SC-YOLO network includes: using the self-supervised pre-trained model DINOv2 as the teacher network and the SC-YOLO network as the student network; transferring the deep semantic knowledge of the teacher network to the student network through feature distillation; and using a multi-task joint loss function to supervise and fine-tune the SC-YOLO network.

[0012] Another aspect of the present invention provides a multi-source collaborative peripheral blood cell detection system based on cell population structure consistency modeling, the system comprising a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the multi-source collaborative peripheral blood cell detection method based on cell population structure consistency modeling.

[0013] Another aspect of the present invention provides a computer storage medium storing a computer program, which, when executed by a processor, implements the aforementioned method for multi-source collaborative peripheral blood cell detection based on cell population structure consistency modeling.

[0014] The present invention has at least the following beneficial effects

[0015] This invention addresses the limitations of existing peripheral blood cell detection methods in multi-scale target perception, densely overlapping scene structure modeling, and cross-domain generalization by constructing a multi-source collaborative detection framework—a structure-aware collaborative detection network—based on cell population structure consistency graph inference. Specifically, this invention introduces a coordinate attention mechanism and multi-scale dynamic receptive field collaborative convolution into the SC-YOLO detection network, enhancing the model's ability to express fine-grained features of small targets (platelets) and overlapping cells. Building upon this, a cell population structure consistency graph is further constructed, and iterative message passing is used to perform global structure inference and consistency optimization on the initial detection results, achieving false positive elimination, false negative compensation, and unified category consistency, significantly improving the accuracy and stability of cell detection in densely overlapping areas. Simultaneously, this invention introduces a self-supervised distillation training strategy, effectively transferring deep semantic knowledge from the teacher network to the student network, reducing dependence on large-scale labeled data while enhancing the model's generalization ability under different staining conditions and imaging devices. This invention significantly improves the overall performance of peripheral blood cell detection and provides reliable technical support for automated clinical blood routine analysis and intelligent microscopic-assisted diagnosis. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the structural framework of the present invention;

[0017] Figure 2 This is a schematic diagram of the self-supervised knowledge distillation process of the present invention;

[0018] Figure 3 This is a schematic diagram of the SC-YOLO network structure of the present invention;

[0019] Figure 4 This is a schematic diagram of the CSCG cell population structure consistency map reasoning module of the present invention. Detailed Implementation

[0020] 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.

[0021] Please see Figures 1-4 One aspect of the present invention provides a multi-source collaborative peripheral blood cell detection method based on cell population structure consistency modeling, comprising:

[0022] S1: Construct a cell image dataset containing red blood cells, white blood cells, platelets and abnormal cells based on peripheral blood smear microscopic images, and preprocess the cell images in the cell image dataset to obtain standard cell images;

[0023] This embodiment uses clinically collected peripheral venous blood samples. After standard blood smear preparation and Wright-Gymsa staining, microscopic images are acquired using a digital pathology scanner under 100x oil immersion. The image resolution is 2048×2048 pixels, in TIFF lossless format, ensuring clear identification of morphological details such as cell nuclei, cytoplasmic granules, and cell membrane contours. The acquired raw images were independently labeled by two laboratory physicians, covering normal red blood cells, normal white blood cells (further subdivided into neutrophils, lymphocytes, monocytes, eosinophils, and basophils), platelets, and various abnormal cells (such as target red blood cells and immature white blood cells). The labeling includes cell bounding boxes and category labels. After labeling, a third attending physician reviews and corrects the results to ensure accuracy and consistency. The final cell image dataset contains a total of 12,000 microscopic images, covering blood smear samples from different staining batches, different patient sources, and different imaging conditions, to enhance data diversity. The dataset is divided into training, validation, and test sets in a ratio of 7:1.5:1.5. The distribution of samples in each class is balanced using a combination of oversampling and undersampling to avoid class imbalance from causing bias in model training.

[0024] The preprocessing stage includes two steps: image illumination normalization and geometric normalization. For illumination normalization, the original RGB images are first converted to the LAB color space, the L luminance channel is extracted, and the mean and standard deviation of the L channels for all images in the training set are calculated. Then, the L channels of each image are aligned according to their mean and variance, and the A and B channels are scaled proportionally to eliminate color and luminance differences caused by different staining batches and imaging lighting conditions, while preserving the original color proportions of the cells. For geometric normalization, each image is scaled and filled to a uniform size of 640×640 pixels using a center-fill method. The filled pixel value is taken as the mean of the pixels in the training set images, ensuring that the filled area does not introduce additional noise and does not change the original cell morphology. During training, random horizontal flipping, random rotation (±15°), random scaling (0.8–1.2 times), and mosaic data augmentation are applied online to the training set images to expand the diversity of training samples and further improve the model's generalization ability to different imaging conditions and cell morphology changes.

[0025] Please see Figure 3 S2: Construct the structure-aware collaborative detection network SC-YOLO, input standard cell images into the SC-YOLO network for forward inference, and obtain initial cell detection results;

[0026] Preferably, the SC-YOLO network is based on the YOLOv8n network structure as its backbone, including a cascaded backbone feature extraction network, a feature fusion network, and a multi-scale detection head network.

[0027] The backbone feature extraction network is obtained by embedding a coordinate attention module at the output of each residual stage of the YOLOv8n backbone network.

[0028] The feature fusion network is obtained by replacing all bilinear interpolation upsampling modules of the YOLOv8n feature fusion network with the content-aware feature reorganization upsampling module CARAFE.

[0029] The multi-scale detector head network is obtained by replacing the convolutional layers of the YOLOv8n detector head network with the LSKNet large-selectivity convolutional kernel structure.

[0030] Preferably, the specific implementation steps of the coordinate attention module include:

[0031] Input feature map of coordinate attention module Global average pooling is performed along both the horizontal and vertical directions to obtain the horizontal feature vector. and vertical eigenvectors ;in, For feature map The number of channels, and Feature maps Height and width;

[0032] horizontal feature vector and vertical eigenvectors Concatenate to obtain concatenated feature vectors ; Concatenate feature vectors Passing in sequence After processing by convolution, batch normalization, and the SiLU activation function, the weights are re-split into horizontal and vertical weight vectors. and The weight vectors in the horizontal and vertical directions after splitting and Pass through in turn Convolution and the Sigmoid function are used to obtain the horizontal weight coefficients. and vertical weighting coefficient ;

[0033] Based on the horizontal weighting coefficient and vertical weighting coefficient For the input feature map Enhancement is performed to obtain the enhanced feature map. :

[0034]

[0035] in, Represents the enhanced feature map In position The characteristic elements; , and These represent the channel index, height index, and width index, respectively. Represents the horizontal weighting coefficient In position The characteristic elements; Indicates the vertical weighting coefficient In position The characteristic elements.

[0036] Preferably, the specific implementation steps of the content-aware feature reconstruction upsampling module CARAFE include:

[0037] For the input feature map Zero-padding is applied to the edges of the input feature map of the CARAFE module. Through convolution kernel size of Content-aware convolutional layers generate recombined weights :

[0038]

[0039] Based on restructuring weight For feature maps Dynamic recombination is performed to obtain the upsampled feature map. The calculation formula is:

[0040]

[0041] in, Represents an integer greater than 1. For feature map The number of channels, and Feature maps Height and width; Representation of feature map In position eigenvalues; , and These represent the channel index, height index, and width index, respectively. and These are the offset indices along the height and width directions within the reconstructed window, respectively. Represents the reorganization weight exist The value at; Represents the input feature map exist The value at; Indicates rounding down; This indicates the upsampling factor.

[0042] Preferably, each detector head of the multi-scale detector head network replaces the convolutional layers of the YOLOv8n detector head network with an LSKNet large-selectivity convolutional kernel structure, and the specific implementation steps of each detector head include:

[0043] The feature map of the input detection head The inputs are processed by the three convolutional branches respectively, and the output features of the three convolutional branches are obtained. and :

[0044]

[0045]

[0046]

[0047] in, , and These represent the kernel size; ;

[0048] For the input feature map Perform global average pooling to obtain the global feature description vector along the channel dimension. :

[0049]

[0050] in, Represents the global feature description vector In position The characteristic elements; , and Representing feature maps respectively The channel index, height index, and width index; For feature map The number of channels, and Feature maps Height and width; Representation of feature map In position The characteristic elements;

[0051] global feature description vector The inputs are sequentially fed into a fully connected layer and a batch normalization layer for feature compression and activation. The resulting vector is then mapped to a 3D logits vector by the output fully connected layer. Finally, the fusion weights of the three convolutional branches are obtained by normalization using the Softmax function. and ;

[0052] Based on the fusion weights of the three convolutional branches and Features and Weighted fusion is performed to obtain the output features of the detection head:

[0053]

[0054] in, This indicates the output characteristics of the detection head.

[0055] This embodiment implements the construction and inference of the SC-YOLO structure-aware collaborative detection network based on the general deep learning framework PyTorch. SC-YOLO uses YOLOv8n as its backbone, with an input image size of 640×640×3. The overall network consists of a three-level cascaded structure of a backbone feature extraction network, a feature fusion network, and a multi-scale detection head network.

[0056] The backbone feature extraction network is based on the five-stage residual structure of YOLOv8n. Each residual stage consists of 2 to 4 residual blocks, and feature downsampling is achieved through convolutional layers with a stride of 2, progressively downsampling the input image to a deep feature map of scale 20×20. A coordinate attention module is embedded at the output of each residual stage (i.e., the end feature maps of stages 1 to 5) to enhance the model's ability to perceive the spatial location of cells.

[0057] Feature map output from the first residual stage For example, the specific execution process of the coordinate attention module is as follows: First, global average pooling is performed along the horizontal and vertical directions respectively, resulting in... and ; concatenate the two in the second dimension as Then, it is reduced to dimensionality by 1×1 convolution. Then, batch normalization and the SiLU activation function are applied sequentially; subsequently, the data is re-split into horizontal directions. and vertical direction The two branches are each recovered to their original state via 1×1 convolution. The number of channels is calculated, and the horizontal weight coefficients are obtained by normalizing them using the Sigmoid function. and vertical weighting coefficient Finally, and Extending in spatial dimension to the original feature map After multiplying element-wise by the same spatial dimensions, the enhanced feature map is obtained. The output feature map sizes of the five coordinate attention modules are 320×320, 160×160, 80×80, 40×40 and 20×20, respectively, with 64, 128, 256, 512 and 1024 channels, respectively.

[0058] In this embodiment, the feature fusion network receives three input features from the backbone network: an 80×80×256 feature map output from the third C2f module, a 40×40×512 feature map output from the fourth C2f module, and a 20×20×1024 feature map output from the SPPF module. These three input features correspond to small-scale, medium-scale, and large-scale branches, respectively. After processing by the feature fusion network, the three fused features are output to the detection head.

[0059] The specific connection relationships and tensor dimension changes of the feature fusion network are as follows:

[0060] Top-down path:

[0061] First, the 20×20×1024 feature map output by the SPPF module is input into the CARAFE upsampling module, with an upsampling factor of [missing value]. The feature map size is enlarged to 40×40, while the number of channels remains unchanged at 1024. The upsampling result is concatenated with the 40×40×512 feature map output by the fourth C2f module of the backbone network in the channel dimension to obtain a 40×40×1536 concatenated feature map. This feature map is then fused and compressed by the C2f module to output a 40×40×512 feature map.

[0062] Next, the 40×40×512 feature map is input again into the CARAFE upsampling module, with an upsampling factor r=2, increasing the feature map size to 80×80 while maintaining 512 channels. The upsampling result is then concatenated with the 80×80×256 feature map output by the third C2f module of the backbone network in the channel dimension to obtain an 80×80×768 concatenated feature map. This feature map is then fused and compressed by the C2f module to output an 80×80×256 feature map, which serves as the P3 scale input for the subsequent detection head.

[0063] Bottom-up path:

[0064] The 80×80×256 feature map is downsampled to 40×40 using a convolutional layer with a stride of 2, while maintaining 256 channels. The downsampled result is then concatenated with the 40×40 C2f output (40×40×512) from the top-down path along the channel dimension to obtain a 40×40×768 concatenated feature map. The C2f module then performs feature fusion and channel compression to output a 40×40×512 feature map.

[0065] The 40×40×256 feature map is downsampled again through a convolutional layer with a stride of 2, reducing its size to 20×20 while maintaining 512 channels. The downsampled result is then concatenated with the 20×20×1024 feature map output by the SPPF module along the channel dimension to obtain a 20×20×1536 concatenated feature map. The C2f module performs feature fusion and channel compression to output a 20×20×1024 feature map, which is then used as the P5 scale input for the subsequent detection head.

[0066] Finally, the 40×40 scale C2f output (40×40×512) from the top-down path is used as the P4 scale input to the detector head. Thus, the feature fusion network provides the detector head with fused features at three scales: P3 (80×80×256), P4 (40×40×512), and P5 (20×20×1024).

[0067] In the above upsampling operation, the CARAFE content-aware feature reconstruction upsampling module is used instead of the traditional bilinear interpolation upsampling. Taking a 20×20×1024 feature map output by SPPF as an example, the specific execution process of CARAFE upsampling to 40×40×1024 is as follows: First, the input feature map is processed... The edges are zero-padded, and recombined weights are generated through a 3×3 content-aware convolutional layer. Subsequently, for the output feature map In position Each pixel in the input feature map The above corresponds to the position Within a 3×3 neighborhood window centered on the target, using The nine corresponding weight values ​​are used to sum the neighboring pixels to obtain the upsampled feature values. CARAFE effectively preserves fine-grained features such as cell nucleus contours and cytoplasmic granules through content-aware weight generation and dynamic recombination, avoiding information loss caused by traditional bilinear interpolation.

[0068] The detection head network is configured on three scale branches, corresponding to fused feature maps at scales of 80×80 (small target detection, responsible for platelets and small cells), 40×40 (medium target detection, responsible for red blood cells and medium-sized cells), and 20×20 (large target detection, responsible for white blood cells and large cells). In each detection head, the original 3×3 convolutional layers of YOLOv8n are replaced with LSKNet large-selectivity convolutional kernel structures, with the kernel size set to... , , These correspond to the three convolutional branches: large kernel, medium kernel, and small kernel. Taking a 40×40 scale detector as an example, the input feature map... The input consists of three convolution branches: a 7×7 convolution output. 5×5 convolution output 3×3 convolution output At the same time, the input feature map Perform global average pooling to obtain After passing through two fully connected layers (with a hidden layer of dimension 64), the output is a 3D logits vector, which is then normalized by Softmax to obtain the fusion weights. and ;satisfy The final detection head output features are The number of channels for the three detection heads are 256, 512, and 512, respectively, corresponding to the number of channels in their respective scale-fused feature maps.

[0069] The standardized 640×640×3 cell image is input into the SC-YOLO network. After five-stage downsampling and coordinate attention enhancement in the backbone network, enhanced feature maps at five scales are obtained. The last three scales (80×80, 40×40, and 20×20) are selected and input into the feature fusion network. After CARAFE upsampling and cross-scale fusion, fused feature maps at three scales are obtained. The three detection heads output the detection results at the corresponding scales. Each detection result includes bounding box coordinates (center point x-coordinate, center point y-coordinate, width, and height), class probability vector (covering preset categories such as red blood cells, white blood cell subtypes, platelets, and abnormal cells), and detection confidence. After summing the detection results at the three scales, a low-threshold filtering with a confidence threshold of 0.01 is initially performed to filter out background noise detection boxes, obtaining the initial cell detection results, which are used for structural correction in the subsequent CSCG graph inference module.

[0070] By constructing the SC-YOLO structure-aware collaborative detection network, forward inference and initial cell target detection of peripheral blood smear microscopic images were completed. In the backbone feature extraction stage, the coordinate attention mechanism embedded in the output of each residual stage enhances the network's ability to perceive the spatial position and boundary contour of cells through global pooling in the horizontal and vertical directions and weight recalibration, effectively alleviating the localization deviation problem in densely overlapping scenes. In the feature fusion stage, the CARAFE content-aware upsampling module replaces the traditional bilinear interpolation, accurately fusing adjacent features through dynamic weight reorganization, preserving fine-grained morphological information such as cell nuclear contours and cytoplasmic granules, and providing high-quality multi-scale feature expression for distinguishing similar cell subtypes. In the multi-scale detection head stage, the LSKNet large selective convolutional kernel structure achieves adaptive perception of cell targets at different scales through dynamic weighted fusion of large, medium and small convolutional branches, taking into account the detection accuracy of small targets such as platelets and large targets such as white blood cells. Through the synergistic enhancement of the three core modules mentioned above, SC-YOLO can output initial detection results for each input image, including bounding box coordinates, class probability, and detection confidence, while maintaining lightweight inference speed. This provides a high-recall set of candidate cell instances for subsequent CSCG cell population structure consistency map inference and structure correction.

[0071] By constructing the SC-YOLO structure-aware collaborative detection network, multi-scale feature extraction and initial target detection are performed on peripheral blood smear microscopic images, providing a high-recall set of candidate cell instances for the subsequent CSCG graph inference module. Specifically, the coordinate attention mechanism enhances the network's ability to locate cells in densely overlapping regions by explicitly encoding spatial location information; the CARAFE content-aware upsampling module effectively preserves fine-grained morphological features such as cell nuclear contours and cytoplasmic granules by dynamically recombining weights and fusing adjacent features; and the LSKNet large-selectivity convolutional kernel structure achieves adaptive multi-scale detection of small targets (platelets) and large targets (leukocytes) through dynamic weighted fusion of three convolutional branches. Through the synergistic enhancement of these modules, SC-YOLO effectively improves the initial detection accuracy and recall in densely overlapping scenes while maintaining lightweight inference speed, overcoming the limitations of existing methods in multi-scale target perception and fine-grained feature expression, and laying a reliable detection foundation for subsequent structure correction.

[0072] Please see Figure 4 S3: Construct a cell population structure consistency graph (CSCG) based on the initial cell detection results, and perform iterative structural reasoning on the node features in the CSCG through a graph structure message passing mechanism to obtain the cell detection results after structural correction.

[0073] Preferably, step S3 includes:

[0074] S31: Using each detected cell in the initial cell detection results as a cell node, extract the bounding box coordinates, unified class probability vector, detection confidence, cell size, and cell center coordinates of each cell from the SC-YOLO output as the initial features of the cell node; where, the... The bounding box of a cell is represented as Then the first The width, height, and center coordinates of each cell are as follows:

[0075]

[0076]

[0077]

[0078] in, Indicates the first The coordinates of the top-left corner of the bounding box of each cell Indicates the first The coordinates of the bottom right corner of the bounding box of each cell; and Indicates the first The width and height of each cell; Indicates the first The center coordinates of each cell; Indicates the first The center of each cell; then the initial characteristics of the node. Represented as:

[0079]

[0080] in, Indicates the first Confidence level for individual cell detection; Indicates the first A uniform category probability vector for each cell;

[0081] In this embodiment, a Cell Population Structure Consistency Graph Inference Subsystem (CSCG) is implemented based on the PyTorch framework. The input is the initial cell detection result output by the SC-YOLO network, and the output is the cell detection result after structure correction.

[0082] For each cell instance detected in the input image, it is defined as a graph node. Number of categories. It covers six types of cells: red blood cells, neutrophils, lymphocytes, monocytes, platelets, and abnormal cells.

[0083] S32: Calculate the edge weights between nodes based on the center distance, size similarity, and category similarity between cells in the initial cell detection results, and construct a cell population structure consistency graph (CSCG); each node in the CSCG corresponds to a cell node, and the edges represent the structural relationships between cell nodes;

[0084] For the The cell and the first For each cell, first calculate the center-to-center distance:

[0085]

[0086] in, Describes the Euclidean norm. Represents the straight-line distance between the centers of two cells; then the first... The cell and the first The distance similarity of the individual cells is:

[0087]

[0088] in, Indicates the first The cell and the first Distance similarity between individual cells; Represents an exponential function; Indicates the distance scale parameter;

[0089] No. The cell and the first The size similarity of the individual cells is:

[0090]

[0091] in, Indicates the first The cell and the first Cell size similarity; and Indicates the first The width and height of each cell; This indicates taking the minimum value; This indicates taking the maximum value; To prevent logarithmic operations from producing extremely small constants with zero values;

[0092] No. The cell and the first The class similarity of the individual cells is:

[0093]

[0094] in, Indicates the first The cell and the first Class similarity of individual cells; This represents the value of the independent variable that makes the function reach its maximum value; Indicates the first A uniform category probability vector for each cell;

[0095] No. The cell and the first Unnormalized boundary weight of each cell for:

[0096]

[0097] in, , and Represents the weight parameters. ,and , and All are greater than or equal to 0; for Normalize:

[0098]

[0099] in, Indicates the first The cell and the first The edge weight of each cell ; Indicates the relationship between the first and second elements in the CSCG graph. A cell has a set of all adjacent nodes connected by an edge; if the... The cell and the first Euclidean distance between the centers of cells If the distance is less than a preset threshold, then the first... The cell and the first Each cell has edge connections, and vice versa. The cell and the first Each cell has no edge connections;

[0100] In this embodiment, the core of graph structure construction is defining the edge connection relationships between nodes and the edge weight calculation method. A preset distance threshold is used. Pixels (this threshold is empirically set based on the average distance between adjacent cells in a peripheral blood smear, typically 1.5 to 2.0 times the average cell diameter), if Then establish an edge connection; This is a distance scale parameter that controls the rate at which distance similarity decreases as spatial distance increases; , , ,satisfy .

[0101] S33: Iterative structural reasoning of the graph CSCG is performed using a graph structure message passing mechanism to obtain the final features of each cell node after update:

[0102]

[0103] in, Indicates the feature retention coefficient. Indicates the first The cells in the first Features at the next iteration; when hour, Indicates the first The final characteristics of each cell; Indicates the first The cells in the first Features at the next iteration; Indicates the first The cells in the first Features at the next iteration;

[0104] In this embodiment Controlling the degree to which its own characteristics are preserved, This parameter represents the number of iterations. The value of this parameter is based on experience—too few iterations (1 time) result in insufficient propagation of neighborhood information, while too many iterations (more than 5 times) lead to overly smoothed node features, and 3 iterations can achieve a balance in most scenarios.

[0105] S34: Based on the final updated features of each cell node, perform structural correction of the cell detection results. The specific correction rules and functions are as follows:

[0106] False detection cell removal: based on the final characteristics updated for each cell node. Obtain the corrected detection confidence. If node satisfy: If the node lacks support from surrounding cellular structures, it is considered a suspected false positive and is removed from the initial detection results. This represents the preset confidence threshold. This represents the preset threshold for the number of neighboring nodes;

[0107] Missed detection area compensation: determined based on connectivity or density clustering results in the CSCG graph. Each densely populated cell region; for each densely populated region , computationally dense region Mean radius of inner cells :

[0108]

[0109] in, Indicates densely packed cell regions The number of cells inside;

[0110] In this embodiment Detection boxes below this confidence level are typically background noise or low-quality predictions; Too few neighborhood nodes mean that the cell is spatially isolated, lacks population structure support, and does not conform to the physiological characteristics of dense distribution of blood cells.

[0111] S35: In densely populated cell areas Inside the convex hull or α-shape boundary, according to the step size Generate a set of candidate centroids For each candidate center point Check all of the following conditions in sequence:

[0112] Condition 1: Candidate center point It does not fall within any existing detection bounding box, meaning there are no detected cells. , making lie in The internal area;

[0113] Condition 2: Candidate center point Distance to the nearest detected cell center satisfy ;

[0114] Condition 3: Based on candidate center points Center, radius Within the circular area, the number of cells detected is no less than indivual;

[0115] When all three conditions above are met, the candidate center point is determined. As candidate centers for missed detection; for each center identified as a candidate center for missed detection, a corresponding candidate missed cell is created in the initial cell detection results. Extract its All detected cells within the radius are considered as candidate missed cells. The neighborhood of detected cell set is used to calculate candidate missed cells. The neighborhood of the detected cell set is the weighted average of the width and height of all cells. and :

[0116]

[0117]

[0118] in, Indicating candidate missed cells Cell sets have been detected in the neighborhood; Indicating candidate missed cells With the Normalized distance weights calculated based on center distance among the detected cells in each neighborhood:

[0119]

[0120] in, Indicating candidate missed cells With the The center distance between cells in each neighborhood has been detected. For distance scale parameters;

[0121] according to and Generate candidate center points for missed detections Predicted bounding box :

[0122]

[0123] in, Indicates the candidate center point of missed detection The coordinates;

[0124] The probability vector of candidate missed cells is obtained by weighted voting of the class probabilities of neighboring cells:

[0125]

[0126]

[0127] in, Indicating candidate missed cells The class probability vector; take as candidate missed detection cells Category; Indicating candidate missed cells The detection confidence level; express The Middle Confidence level for individual cell detection;

[0128] In this embodiment, step S35 is the most critical step in the CSCG subsystem. The specific implementation of determining the K densely populated cell regions based on the connectivity or density clustering results in the CSCG graph is as follows:

[0129] Method 1 for determining dense regions (based on connectivity): Edge connections in CSCG are considered as connectivity between nodes. A Union-Find or Depth-First Search (DFS) algorithm is used to divide the graph into several connected components. Each connected component represents a dense region. The size of the connected component reflects the size of the cluster; isolated clusters with fewer than 3 nodes are removed. In this embodiment, the graph contains 168 nodes and 842 edges. After partitioning, 7 connected components are obtained. The five largest components contain more than 10 nodes each, while the two smaller components contain 2 and 1 isolated nodes respectively. After removing the smaller components, the following is obtained: One effective dense area;

[0130] Method 2 for determining dense regions (based on density clustering): Using the DBSCAN algorithm to determine the coordinates of the node centers. Clustering is performed. The neighborhood radius eps is set to 1.5 times the average cell diameter, and the minimum number of samples is [not specified]. To filter out noise points, the DBSCAN clustering results show that each cluster corresponds to a densely populated region of cells. Six clusters were obtained, with five clusters having between 8 and 32 nodes, and one cluster containing only two nodes was excluded as noise. In practice, connectivity is preferred, and density clustering is used as a supplement when the graph has poor connectivity (too many isolated nodes).

[0131] In one of the dense areas For example (located in the upper left corner of the image, containing 24 cells, including 9 platelets, 13 red blood cells, and 2 white blood cells), calculate the average radius of the cells in this region; Inside the convex hull or α-shape boundary, according to the step size A grid of candidate center points is generated for each pixel. The Graham scan algorithm is used to calculate the convex hull boundary, and the Delaunay triangulation method is used for the α-shape boundary. After candidate points are generated, three conditions are checked sequentially:

[0132] Condition 1: Candidate center point It does not fall within any existing detection bounding box, meaning there are no detected cells. , making lie in The internal area;

[0133] Condition 2: Candidate center point Distance to the nearest detected cell center satisfy ;

[0134] Condition 3: Based on candidate center points Center, radius Within the circular area, the number of cells detected is no less than One; for candidate points that simultaneously meet all three conditions Create candidate missed detection cells Its neighborhood Defined as a radius of 2.5 All detected cells within the range. Calculate the normalized distance weights. , .

[0135] S36: Select candidate missed cells The category, predicted bounding box, and confidence level are added to the initial cell detection results to obtain the structure-corrected cell detection results.

[0136] In this embodiment, the confidence level Candidate missed cells greater than 0.3 The data is then added to the detection results, and non-maximum suppression (IoU threshold 0.45) is performed to remove redundant bounding boxes. The final output is the structure-corrected cell detection result.

[0137] By constructing a Cell Population Structure Consistency Map (CSCG), the local cell detection results output by SC-YOLO are placed within the global structural relationship of the cell population for reasoning and correction, effectively solving the problem of false positives and false negatives caused by relying solely on local features in densely overlapping scenarios. Specifically, based on the final features of nodes and neighborhood consistency scores, false positives caused by background impurities and edge overlaps are accurately eliminated through dual threshold judgment of confidence and the number of neighboring nodes. Dense cell regions are determined by graph connectivity component partitioning or density clustering, and candidate center points are generated under convex hull or α-shape boundary constraints. False negative locations are screened by combining distance constraints, bounding box avoidance, and neighborhood density. Finally, the bounding boxes and category labels of false negative cells are restored using the weighted average size of neighboring cells and category voting, achieving targeted compensation for the initial detection results. Through the synergistic optimization of false positive elimination and false negative compensation, the CSCG subsystem elevates the detection results from local independent discrimination to global structural consistency constraints, significantly enhancing the integrity, category consistency, and spatial rationality of detection results in dense cell clusters, providing reliable structural-level optimization guarantees for high-precision peripheral blood cell detection in clinical scenarios.

[0138] S4: Perform non-maximum suppression on the structure-corrected cell detection results to obtain the final cell detection results;

[0139] The training process of the SC-YOLO network includes: using the self-supervised pre-trained model DINOv2 as the teacher network and the SC-YOLO network as the student network; transferring the deep semantic knowledge of the teacher network to the student network through feature distillation; and using a multi-task joint loss function to supervise and fine-tune the SC-YOLO network.

[0140] In this embodiment, platelet detection is used as an example. Among the 62 platelet detection frames, the one with the highest confidence level is located in the central region of the image, with a confidence level of 0.92 and a structural consistency score of 0.72, belonging to a high consistency region. Three detection boxes with an IoU exceeding 0.50 were found, all of which were duplicate predictions of the same platelet (due to the addition of candidate boxes in dense regions), and these were suppressed. The remaining 58 detection boxes were then processed sequentially.

[0141] exist Within the dense region, there is overlap between the four platelet candidate boxes added in step S35 and the original detection boxes. After calculation, the IoU between two of the candidate boxes and the original detection boxes are 0.48 and 0.46, respectively, which are below the dynamic threshold of 0.50, and are therefore determined to be valid missed detection compensation and retained; the other two IoUs are 0.62 and 0.55, respectively, which are above the threshold, and are therefore determined to be duplicate predictions and suppressed.

[0142] Preferably, the transfer of deep semantic knowledge from the teacher network to the student network through feature distillation includes:

[0143] S41: Obtain an unlabeled standard cell image dataset For each unlabeled standard cell image Two different data augmentation operations were performed to obtain two augmented samples. and The two augmented samples are then input into the DINOv2 teacher network encoder. The patch token output from the last layer of the encoder is then subjected to global average pooling and projection head. Processing yields a normalized global semantic feature vector. and :

[0144]

[0145]

[0146] in, This refers to the DINOv2 teacher network encoder. Indicates global average pooling. The projection head of the teacher network. This represents the vector normalization operation; the teacher network is fine-tuned by comparing the learning loss function.

[0147]

[0148] in, This indicates the number of unlabeled standard cell images. Indicates the first Normalized global semantic feature vector generated from unlabeled standard cell images. Represents cosine similarity. Indicates the temperature coefficient; Represents an exponential function; and Indicates the first Normalized global semantic feature vector generated from unlabeled standard cell images;

[0149] S42: Fix the teacher network parameters and use standard cell images. Input the teacher network and the SC-YOLO student network respectively, and extract the three-scale teacher feature set generated by the feature pyramid adaptation layer of the teacher network. :

[0150]

[0151] in, , and These correspond to the teacher features output by the small-scale, medium-scale, and large-scale detection branches, respectively.

[0152] And extracting SC-YOLO student network features and fusing the network output to obtain a three-scale student feature set. :

[0153]

[0154] in, , and These correspond to the student features output by the small-scale, medium-scale, and large-scale detection branches, respectively.

[0155] S43: Process each teacher feature and each student feature separately through an adaptation layer and vectorization:

[0156]

[0157]

[0158] in, , , and These represent teacher characteristics and student characteristics respectively in the first... Channel adaptation layers at various scales Indicates vectorization operation, Indicates the distillation temperature coefficient; and Indicates the number after processing Teacher and student characteristics at various scales; Indicates the activation function;

[0159] S44: Updating the parameters of the student network using three-scale KL divergence distillation loss:

[0160]

[0161] in, Denotes KL divergence, ; This indicates the total number of unlabeled and labeled standard cell images in the current batch; Indicates the first in the current training batch A standard cell image.

[0162] In this embodiment, the teacher network uses the DINOv2-base pre-trained model, which contains 86M parameters and has been self-supervised pre-trained on the ImageNet-22K large-scale dataset, possessing strong general visual feature extraction capabilities. Because there are domain differences (color distribution, texture features, cell morphology, etc.) between general image data and clinical peripheral blood cell images, the teacher network needs to be specifically fine-tuned using unlabeled peripheral blood cell images.

[0163] Unlabeled image dataset is For each image Perform two different data augmentation operations, a and b:

[0164] Enhancement operation a (strong enhancement) includes: random cropping, random horizontal flipping, random color dithering, random Gaussian blur, and random grayscale conversion;

[0165] Data augmentation operation b (weak augmentation) includes: random pruning and random horizontal flipping;

[0166] Two enhanced views are input into the DINOv2 teacher network encoder. The DINOv2 encoder, based on the VisionTransformer architecture, divides the image into 14×14 patches, which are then encoded using 12 Transformer layers to output a sequence of patch tokens. All patch tokens output from the last layer of the encoder are then processed using global average pooling. Obtain the global feature vector, and then use the projection head. Mapping, DINOv2 default projection head It contains 3 layers of MLP; finally, it is obtained by vector normalization Norm. and ; This is a temperature coefficient used to control the sharpness of the similarity distribution; a smaller value indicates a lower degree of similarity. The value makes the model focus more on the most similar negative samples; during fine-tuning, the LARS optimizer is used, and the learning rate is... The training ran for 2020 rounds. The fine-tuned teacher network was able to more accurately extract domain-specific features such as nuclear morphology, cytoplasmic granules, and staining distribution from peripheral blood cell images, providing high-quality feature distributions for subsequent distillation.

[0167] After the teacher network is fine-tuned, its parameters are fixed, and it no longer participates in gradient updates. Labeled standard cell images are then used. Input the teacher network DINOv2 and the student network SC-YOLO respectively, and extract their respective multi-scale features.

[0168] For the teacher network, since DINOv2 uses a standard ViT structure (single-scale output), a feature pyramid adaptation layer is needed to generate three-scale features. Specifically, the patch token sequences output from layers 6, 9, and 12 of the DINOv2 encoder are reshaped into two-dimensional feature maps (sizes 14×14, 14×14, and 14×14, respectively), and then passed through three independent adaptation layers (1×1 convolution to adjust the number of channels, bilinear interpolation upsampling to 56×56, 28×28, and 14×14) to obtain... , and ;

[0169] For the student network SC-YOLO, extract the three-scale feature maps output by its feature fusion network, with corresponding sizes of 80×80×256, 40×40×512, and 20×20×1024. The three-scale student feature set is as follows:

[0170]

[0171] Because the teacher and student networks have different structures, their feature map sizes and number of channels differ. An adaptation layer is needed to align the student features to the dimensional space of the teacher features. Specifically, for each scale... Teacher characteristics and student characteristics are adapted through their respective channel layers. and

[0172] Based on teacher characteristics and For example, firstly, bilinear interpolation is used to unify teacher and student features to a spatial size of 56×56, and then... Convolution unifies the number of channels in both to 256, and then flattens the 56×56 spatial dimension feature map. (achieved), then divided by the distillation temperature coefficient The data is softened and then normalized to a probability distribution using the Softmax function. This represents the total number of labeled and unlabeled images in the current batch.

[0173] By applying non-maximum suppression to the structurally corrected cell detection results, redundant detections caused by overlapping candidate boxes in dense cell regions are effectively eliminated, resulting in a definitive cell detection output. This processing is performed independently within each cell category, iteratively retaining the highest-confidence boxes and suppressing other highly overlapping boxes after sorting by detection confidence. This maintains high recall while further improving detection accuracy, ensuring that the final output detection results possess both structural integrity and category consistency. This approach can be directly applied to practical scenarios such as automated clinical blood routine analysis and intelligent microscopic-assisted diagnosis.

[0174] Preferably, the multi-task joint loss function is:

[0175]

[0176]

[0177]

[0178]

[0179]

[0180]

[0181]

[0182]

[0183] in, This represents the joint loss function for multiple tasks. This represents the total number of labeled cell instances participating in the loss calculation in the labeled standard cell images of the current training batch; , and Indicates the weighting coefficient; Indicates the first Classification loss for each labeled cell instance; Indicates the first Boundary regression loss for each labeled cell instance; Indicates the total number of cell types. Indicates the first The labeled cell instance in the first Label values ​​on the class, The model predicts the first... The labeled cell instance belongs to the first The probability of a class To prevent logarithmic operations from producing extremely small constants with zero values; Indicates the first Predicted bounding boxes for each labeled cell instance; Indicates the first The true bounding box of a labeled cell instance; , , and They represent the first The x-coordinate of the center point, y-coordinate of the center point, width, and height of the predicted bounding box for each labeled cell instance; , , and They represent the first The x-coordinate of the center point, y-coordinate of the center point, width, and height of the true bounding box of each labeled cell instance; Indicates the first The inner auxiliary box corresponding to the predicted bounding box of each labeled cell instance. Indicates the first The inner auxiliary box corresponding to the true bounding box of each labeled cell instance; For the first Scaling factor of the internal auxiliary box of each labeled cell instance; Indicates the base scaling factor. and These represent the lower and upper limits of the scaling factor, respectively. This represents the truncation function.

[0184] In this embodiment To prevent logarithmic operations from producing zero values. Lower bound of scaling factor Upper limit of scaling factor , The dynamic value taking logic is as follows: when the predicted box area is smaller than the actual box area, It tends to take smaller values ​​and use smaller-scale auxiliary bounding boxes to calculate the loss, which helps with fast regression of high IoU samples; when the predicted box area is larger than the true box area, A preference for larger values ​​and the use of larger-scale auxiliary bounding boxes are suitable for low IoU samples. Distillation temperature coefficient. Higher temperatures make the probability distribution of the Softmax output smoother, which is beneficial for students to learn the hidden knowledge of the teacher network. , , ,in, The value is larger because the numerical range of the bounding box regression loss is usually smaller than that of the classification loss, and a higher weight is needed to ensure that the localization accuracy dominates the total loss. To control the transfer strength of distillation knowledge and prevent distillation loss from dominating the training process and weakening the role of supervised signals, the aforementioned weight coefficients are kept fixed in the early stages of training and fine-tuned during training based on the performance on the validation set.

[0185] During training, the AdamW optimizer was used, with an initial learning rate set to... The weight decay coefficient is A cosine annealing learning rate scheduling strategy was adopted. After each training round, the model performance was evaluated on the validation set, and an early stopping strategy was triggered if the average accuracy on the validation set did not improve for 10 consecutive rounds. After 100 rounds of training, the model achieved an average accuracy of 94.2% on the test set, verifying the effectiveness of the multi-task joint loss function in optimizing the parameters of the SC-YOLO network.

[0186] By constructing a multi-task joint loss function, the classification loss, bounding box regression loss, and distillation loss are organically integrated, achieving collaborative optimization training of SC-YOLO network parameters. Specifically, the Inner-IoU auxiliary bounding box loss introduces a scale-adaptive scaling factor, dynamically adjusting the size of the internal auxiliary box based on the area ratio of the predicted box to the ground truth box. This ensures that small targets (such as platelets) focus on the main body region during regression to avoid background interference, and that large targets (such as white blood cells) maintain boundary stability during regression, effectively improving the localization accuracy and convergence speed of cell targets of different sizes. Simultaneously, the classification loss ensures the network accurately distinguishes cell categories, and the distillation loss continuously transfers deep semantic knowledge from the teacher network to enhance generalization ability. All three share the network backbone parameters, forming multi-dimensional gradient supervision during backpropagation. This avoids the local optima problem that may result from optimizing a single loss, enabling the model to maintain stable detection performance even under conditions of limited labeled data and cross-domain imaging.

[0187] Another aspect of the present invention provides a multi-source collaborative peripheral blood cell detection system based on cell population structure consistency modeling, the system comprising a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the multi-source collaborative peripheral blood cell detection method based on cell population structure consistency modeling.

[0188] Another aspect of the present invention provides a computer storage medium storing a computer program, which, when executed by a processor, implements the aforementioned method for multi-source collaborative peripheral blood cell detection based on cell population structure consistency modeling.

[0189] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0190] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-source collaborative peripheral blood cell detection method based on cell population structure consistency modeling, characterized in that, include: S1: Construct a cell image dataset containing red blood cells, white blood cells, platelets and abnormal cells based on peripheral blood smear microscopic images, and preprocess the cell images in the cell image dataset to obtain standard cell images; S2: Construct the structure-aware collaborative detection network SC-YOLO, input standard cell images into the SC-YOLO network for forward inference, and obtain initial cell detection results; S3: Construct a cell population structure consistency graph (CSCG) based on the initial cell detection results, and perform iterative structural reasoning on the node features in the CSCG through a graph structure message passing mechanism to obtain the cell detection results after structural correction. S4: Perform non-maximum suppression on the structure-corrected cell detection results to obtain the final cell detection results; The training process of the SC-YOLO network includes: using the self-supervised pre-trained model DINOv2 as the teacher network and the SC-YOLO network as the student network; transferring the deep semantic knowledge of the teacher network to the student network through feature distillation; and using a multi-task joint loss function to supervise and fine-tune the SC-YOLO network.

2. The method for detecting multi-source collaborative peripheral blood cells based on cell population structure consistency modeling according to claim 1, characterized in that, The SC-YOLO network is based on the YOLOv8n network structure, including a cascaded backbone feature extraction network, a feature fusion network, and a multi-scale detection head network. The backbone feature extraction network is obtained by embedding a coordinate attention module at the output of each residual stage of the YOLOv8n backbone network. The feature fusion network is obtained by replacing all bilinear interpolation upsampling modules of the YOLOv8n feature fusion network with the content-aware feature reorganization upsampling module CARAFE. The multi-scale detector head network is obtained by replacing the convolutional layers of the YOLOv8n detector head network with the LSKNet large-selectivity convolutional kernel structure.

3. The method for detecting multi-source collaborative peripheral blood cells based on cell population structure consistency modeling according to claim 1, characterized in that, The specific implementation steps of the coordinate attention module include: Input feature map of coordinate attention module Global average pooling is performed along both the horizontal and vertical directions to obtain the horizontal feature vector. and vertical eigenvectors ;in, For feature map The number of channels, and Feature maps Height and width; horizontal feature vector and vertical eigenvectors Concatenate to obtain concatenated feature vectors ; Concatenate feature vectors Passing in sequence After processing by convolution, batch normalization, and the SiLU activation function, the weights are re-split into horizontal and vertical weight vectors. and The weight vectors in the horizontal and vertical directions after splitting and Pass through in turn Convolution and the Sigmoid function are used to obtain the horizontal weight coefficients. and vertical weighting coefficient ; Based on the horizontal weighting coefficient and vertical weighting coefficient For the input feature map Enhancement is performed to obtain the enhanced feature map. : in, Represents the enhanced feature map In position The characteristic elements; , and These represent the channel index, height index, and width index, respectively. Represents the horizontal weighting coefficient In position The characteristic elements; Indicates the vertical weighting coefficient In position The characteristic elements.

4. The method for detecting multi-source collaborative peripheral blood cells based on cell population structure consistency modeling according to claim 1, characterized in that, The specific implementation steps of the content-aware feature reconstruction upsampling module CARAFE include: For the input feature map Zero-padding is applied to the edges of the input feature map of the CARAFE module. Through convolution kernel size of Content-aware convolutional layers generate recombined weights : Based on restructuring weight For feature maps Dynamic recombination is performed to obtain the upsampled feature map. The calculation formula is: in, Represents an integer greater than 1. For feature map The number of channels, and Feature maps Height and width; Representation of feature map In position eigenvalues; , and These represent the channel index, height index, and width index, respectively. and These are the offset indices along the height and width directions within the reconstructed window, respectively. Represents the reorganization weight exist The value at; Represents the input feature map exist The value at; Indicates rounding down; This indicates the upsampling factor.

5. The method for detecting multi-source collaborative peripheral blood cells based on cell population structure consistency modeling according to claim 1, characterized in that, Each detector head of the multi-scale detector head network replaces the convolutional layers of the YOLOv8n detector head network with an LSKNet large-selectivity convolutional kernel structure. The specific implementation steps of each detector head include: The feature map of the input detection head The inputs are processed by the three convolutional branches respectively, and the output features of the three convolutional branches are obtained. and : in, , and These represent the kernel size; ; For the input feature map Perform global average pooling to obtain the global feature description vector along the channel dimension. : in, Represents the global feature description vector In position The characteristic elements; , and Representing feature maps respectively The channel index, height index, and width index; For feature map The number of channels, and Feature maps Height and width; Representation of feature map In position The characteristic elements; global feature description vector The inputs are sequentially fed into a fully connected layer and a batch normalization layer for feature compression and activation. The resulting vector is then mapped to a 3D logits vector by the output fully connected layer. Finally, the fusion weights of the three convolutional branches are obtained by normalization using the Softmax function. and ; Based on the fusion weights of the three convolutional branches and Features and Weighted fusion is performed to obtain the output features of the detection head: in, This indicates the output characteristics of the detection head.

6. The method for detecting multi-source collaborative peripheral blood cells based on cell population structure consistency modeling according to claim 1, characterized in that, The process of transferring deep semantic knowledge from the teacher network to the student network through feature distillation includes: S41: Obtain an unlabeled standard cell image dataset For each unlabeled standard cell image Two different data augmentation operations were performed to obtain two augmented samples. and The two augmented samples are then input into the DINOv2 teacher network encoder. The patch token output from the last layer of the encoder is then subjected to global average pooling and projection head. Processing yields a normalized global semantic feature vector. and : in, This refers to the DINOv2 teacher network encoder. Indicates global average pooling. The projection head of the teacher network. This represents the vector normalization operation; the teacher network is fine-tuned by comparing the learning loss function. in, This indicates the number of unlabeled standard cell images. Indicates the first Normalized global semantic feature vector generated from unlabeled standard cell images. Represents cosine similarity. Indicates the temperature coefficient; Represents an exponential function; and Indicates the first Normalized global semantic feature vector generated from unlabeled standard cell images; S42: Fix the teacher network parameters and use standard cell images. Input the teacher network and the SC-YOLO student network respectively, and extract the three-scale teacher feature set generated by the feature pyramid adaptation layer of the teacher network. : in, , and These correspond to the teacher features output by the small-scale, medium-scale, and large-scale detection branches, respectively. And extracting SC-YOLO student network features and fusing the network output to obtain a three-scale student feature set. : in, , and These correspond to the student features output by the small-scale, medium-scale, and large-scale detection branches, respectively. S43: Process each teacher feature and each student feature separately through an adaptation layer and vectorization: in, , , and These represent teacher characteristics and student characteristics respectively in the first... Channel adaptation layers at various scales Indicates vectorization operation, Indicates the distillation temperature coefficient; and Indicates the number after processing Teacher and student characteristics at various scales; Indicates the activation function; S44: Updating the parameters of the student network using three-scale KL divergence distillation loss: in, Denotes KL divergence, ; This indicates the total number of unlabeled and labeled standard cell images in the current batch; Indicates the first in the current training batch A standard cell image.

7. The method for detecting multi-source collaborative peripheral blood cells based on cell population structure consistency modeling according to claim 6, characterized in that, The multi-task joint loss function is: in, This represents the joint loss function for multiple tasks. This represents the total number of labeled cell instances participating in the loss calculation in the labeled standard cell images of the current training batch; , and Indicates the weighting coefficient; Indicates the first Classification loss for each labeled cell instance; Indicates the first Boundary regression loss for each labeled cell instance; Indicates the total number of cell types. Indicates the first The labeled cell instance in the first Label values ​​on the class, The model predicts the first... The labeled cell instance belongs to the first The probability of a class To prevent logarithmic operations from producing extremely small constants with zero values; Indicates the first Predicted bounding boxes for each labeled cell instance; Indicates the first The true bounding box of a labeled cell instance; , , and They represent the first The x-coordinate of the center point, y-coordinate of the center point, width, and height of the predicted bounding box for each labeled cell instance; , , and They represent the first The x-coordinate of the center point, y-coordinate of the center point, width, and height of the true bounding box of each labeled cell instance; Indicates the first The inner auxiliary box corresponding to the predicted bounding box of each labeled cell instance. Indicates the first The inner auxiliary box corresponding to the true bounding box of each labeled cell instance; For the first Scaling factor of the internal auxiliary box of each labeled cell instance; Indicates the base scaling factor. and These represent the lower and upper limits of the scaling factor, respectively. This represents the truncation function.

8. The method for detecting multi-source collaborative peripheral blood cells based on cell population structure consistency modeling according to claim 1, wherein step S3 includes: S31: Using each detected cell in the initial cell detection results as a cell node, extract the bounding box coordinates, unified class probability vector, detection confidence, cell size, and cell center coordinates of each cell from the SC-YOLO output as the initial features of the cell node; where, the... The bounding box of a cell is represented as Then the first The width, height, and center coordinates of each cell are as follows: in, Indicates the first The coordinates of the top-left corner of the bounding box of each cell Indicates the first The coordinates of the bottom right corner of the bounding box of each cell; and Indicates the first The width and height of each cell; Indicates the first The center coordinates of each cell; Indicates the first The center of each cell; then the initial characteristics of the node. Represented as: in, Indicates the first Confidence level for individual cell detection; Indicates the first A uniform category probability vector for each cell; S32: Calculate the edge weights between nodes based on the center distance, size similarity, and category similarity between cells in the initial cell detection results, and construct a cell population structure consistency graph (CSCG); each node in the CSCG corresponds to a cell node, and the edges represent the structural relationships between cell nodes; For the The cell and the first For each cell, first calculate the center-to-center distance: in, Describes the Euclidean norm. Represents the straight-line distance between the centers of two cells; then the first... The cell and the first The distance similarity of the individual cells is: in, Indicates the first The cell and the first Distance similarity between individual cells; Represents an exponential function; Indicates the distance scale parameter; No. The cell and the first The size similarity of the individual cells is: in, Indicates the first The cell and the first Cell size similarity; and Indicates the first The width and height of each cell; This indicates taking the minimum value; This indicates taking the maximum value; To prevent logarithmic operations from producing extremely small constants with zero values; No. The cell and the first The class similarity of the individual cells is: in, Indicates the first The cell and the first Class similarity of individual cells; This represents the value of the independent variable that makes the function reach its maximum value; Indicates the first A uniform category probability vector for each cell; No. The cell and the first Unnormalized boundary weight of each cell for: in, , and Represents the weight parameters. ,and , and All are greater than or equal to 0; for Normalize: in, Indicates the first The cell and the first The edge weight of each cell ; Indicates the relationship between the first and second elements in the CSCG graph. A cell has a set of all adjacent nodes connected by an edge; if the... The cell and the first Euclidean distance between the centers of cells If the distance is less than a preset threshold, then the first... The cell and the first Each cell has edge connections, and vice versa. The cell and the first Each cell has no edge connections; S33: Iterative structural reasoning of the graph CSCG is performed using a graph structure message passing mechanism to obtain the final features of each cell node after update: in, Indicates the feature retention coefficient. Indicates the first The cells in the first Features at the next iteration; when hour, Indicates the first The final characteristics of each cell; Indicates the first The cells in the first Features at the next iteration; Indicates the first The cells in the first Features at the next iteration; S34: Based on the final updated features of each cell node, perform structural correction of the cell detection results. The specific correction rules and functions are as follows: False detection cell removal: based on the final characteristics updated for each cell node. Obtain the corrected detection confidence. If node satisfy: If the node lacks support from surrounding cellular structures, it is considered a suspected false positive and is removed from the initial detection results. This represents the preset confidence threshold. This represents the preset threshold for the number of neighboring nodes; Missed detection area compensation: determined based on connectivity or density clustering results in the CSCG graph. Each densely populated cell region; for each densely populated region , computationally dense region Mean radius of inner cells : in, Indicates densely packed cell regions The number of cells inside; S35: In densely populated cell areas Inside the convex hull or α-shape boundary, according to the step size Generate a set of candidate centroids For each candidate center point Check all of the following conditions in sequence: Condition 1: Candidate center point It does not fall within any existing detection bounding box, meaning there are no detected cells. , making lie in The internal area; Condition 2: Candidate center point Distance to the nearest detected cell center satisfy ; Condition 3: Based on candidate center points Center, radius Within the circular area, the number of cells detected is no less than indivual; When all three conditions above are met, the candidate center point is determined. As candidate centers for missed detection; for each center identified as a candidate center for missed detection, a corresponding candidate missed cell is created in the initial cell detection results. Extract its All detected cells within the radius are considered as candidate missed cells. The neighborhood of detected cell set is used to calculate candidate missed cells. The neighborhood of the detected cell set is the weighted average of the width and height of all cells. and : in, Indicating candidate missed cells Cell sets have been detected in the neighborhood; Indicating candidate missed cells With the Normalized distance weights calculated based on center distance among the detected cells in each neighborhood: in, Indicating candidate missed cells With the The center distance between cells in each neighborhood has been detected. For distance scale parameters; according to and Generate candidate center points for missed detections Predicted bounding box : in, Indicates the candidate center point of missed detection The coordinates; The probability vector of candidate missed cells is obtained by weighted voting of the class probabilities of neighboring cells: in, Indicating candidate missed cells The class probability vector; take as candidate missed detection cells Category; Indicating candidate missed cells The detection confidence level; express The Middle Confidence level for individual cell detection; S36: Select candidate missed cells The category, predicted bounding box, and confidence level are added to the initial cell detection results to obtain the structure-corrected cell detection results.

9. A multi-source collaborative peripheral blood cell detection system based on cell population structure consistency modeling, the system comprising a memory and a processor; characterized in that, The memory is used to store the application program; the processor is used to run the application program and execute the multi-source collaborative peripheral blood cell detection method based on cell population structure consistency modeling as described in any one of claims 1 to 8.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements a multi-source collaborative peripheral blood cell detection method based on cell population structure consistency modeling as described in any one of claims 1 to 8.

Images