An optical visual layering system and method for blood separation

Abstract

This invention relates to the field of blood separation technology, and more particularly to an optically visualized layering system and method for blood separation. This method aims to solve the problems of low separation efficiency, insufficient visualization capabilities, and poor layering accuracy in existing centrifugation techniques. The system includes an optical microscopy imaging module, a preprocessing module, a layering recognition module, and a 3D visualization module. A three-stage training strategy (pre-training, fine-tuning, and dynamic enhancement) improves the robustness of the layering recognition module model, achieving a layering error rate of ≤3% and a total process time of <5 minutes. This significantly improves the accuracy, efficiency, and visualization effects of blood separation, making it suitable for clinical diagnosis and biomedical research.

Description

An optically visualized layering system and method for blood separation Technical Field

[0001] This invention relates to the field of blood separation technology, and more particularly to an optically visualized layering system and method for blood separation. Background Technology

[0002] The information disclosed in the background section of this invention is intended only to enhance the understanding of the overall background of the invention and is not necessarily to be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

[0003] Blood separation is a crucial step in clinical diagnostics and biomedical research, aiming to efficiently and accurately separate whole blood into different components such as plasma, red blood cells, white blood cells, and platelets to support applications such as disease screening, transfusion therapy, and drug development. Currently, centrifugation remains the mainstream separation method, which achieves physical stratification through high-speed rotation based on the density differences of blood components. Although centrifugation technology is mature and widely used, it still has several limitations in practical operation: First, the entire process is usually time-consuming (generally exceeding 30 minutes), making it difficult to meet the rapid response requirements of point-of-care diagnosis or emergency medical scenarios; second, the separation effect is easily affected by centrifuge parameter settings, operator proficiency, and sample characteristics, easily leading to problems such as unclear stratification interfaces and cross-contamination of components, resulting in low recovery rates or insufficient purity of target components; in addition, traditional centrifugation processes lack real-time monitoring and visualization methods, making it impossible for operators to intuitively assess the separation status or adjust conditions in a timely manner, thus introducing a high risk of human error.

[0004] The rapid development of optical imaging technology, especially the advancements in microscopy and spectroscopy, has provided new technical pathways for blood separation. Optical methods, relying on high-resolution image sensors and fluorescent labeling, can achieve label-free or labeled real-time imaging of blood samples. They can also be combined with image processing algorithms to identify and locate specific cell types (such as leukocytes), to some extent compensating for the shortcomings of centrifugation in terms of visibility and real-time performance. However, existing optical separation strategies still face challenges: First, spatial resolution is limited, especially in imaging small components such as platelets, where occlusion or overlap can lead to identification difficulties. Second, most systems are still in the "observation rather than manipulation" stage, lacking an automatic separation mechanism synchronized with imaging, making it difficult to directly output sortable components or quantitative analysis results. Third, the image processing algorithms used lack robustness and are easily affected by factors such as sample impurities, illumination fluctuations, focal length drift, or cell movement, significantly impacting the accuracy and consistency of separation interpretation, and have not yet reached the high precision and reliability standards required for clinical applications. Summary of the Invention

[0005] In view of this, the present invention provides an optically visualized layering system and method for blood separation. The method provided by the present invention addresses the shortcomings of existing technologies, such as low separation efficiency, insufficient visualization, and poor accuracy. It also integrates high-resolution optical imaging and machine learning algorithms, achieving automated layering and three-dimensional visualization through real-time acquisition of blood images and feature extraction, thereby improving separation accuracy and reducing operation time, and has significant clinical application value.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0007] In a first aspect, the present invention provides an optically visualized layering system for blood separation, comprising the following modules:

[0008] The optical microscopy imaging module uses a high-resolution microscope camera and a reflective illumination system to acquire multifocal dynamic sequence images of blood samples;

[0009] The preprocessing module performs noise reduction, motion correction, and illumination equalization on the original image to eliminate sample impurities and interference.

[0010] The hierarchical recognition module uses a dual-path convolutional neural network (CNN) model for automatic recognition and segmentation of blood components;

[0011] The 3D visualization module maps the 2D semantic mask to a 3D coordinate system based on the segmentation results, generates a 3D pseudo-color layered map through a deep fusion algorithm, and outputs a quantitative report on the volume ratio, layer thickness, and distribution uniformity index of each component.

[0012] The innovation of this invention lies in "optical visual layering (layer recognition and quantization)," which belongs to the interpretation stage of the separation process rather than physical separation. Its low time consumption stems from the high efficiency of image acquisition and algorithm recognition. The dual-path CNN model has the advantages of local receptive field, parameter sharing, and multi-scale feature extraction, and can simultaneously capture cell morphology, texture differences, and boundary changes. It has higher recognition accuracy for subtle features of plasma, red blood cells, white blood cells, and platelets in microscopic images, especially under low contrast, noise interference, and cell overlap conditions, showing significant robustness and generalization compared to traditional threshold-based or edge detection-based algorithms. The dual-path structure adopted in this invention includes a semantic segmentation feature extraction path and a boundary optimization path: the feature extraction path generates multi-scale semantic segmentation heatmaps based on the ResNet-34 backbone network, which can accurately extract the global morphological features of various cells; the boundary optimization path introduces a spatial attention mechanism, which significantly improves the segmentation ability of weak boundary targets such as platelets and small white blood cells by enhancing the feature response of cell boundary regions. Meanwhile, the segmentation loss function uses weighted cross-entropy, enabling the model to assign higher weights to difficult-to-segment components (such as platelets), thus improving overall recognition accuracy. The combined effect of this module reduces the blood component stratification error to less than 3% (this stratification error is calculated based on a manually labeled validation set, using the Dice coefficient and layer thickness difference), achieving high-precision automatic stratification of whole blood samples.

[0013] The 3D visualization module utilizes spatial location information from multifocal imaging sequences to reconstruct the layer thickness, distribution gradient, and volume distribution of each blood component. It then uses different pseudo-color channels to express the spatial concentration and hierarchical relationships of plasma, erythrocytes, leukocytes, and platelets. 3D mapping not only provides a three-dimensional tomographic structure that traditional planar images cannot display, but also automatically calculates the volume percentage, layer thickness, and uniformity index (UI) of each component, outputting a standardized quantitative report. Compared to existing optical technologies that only provide two-dimensional observation results, this module can accurately present the true stacking and interface structure of blood components in three-dimensional space, supporting rapid interpretation and clinical quantitative analysis, thus improving the interpretability and medical application value of the separation results.

[0014] Preferably, multifocal dynamic sequence images of blood samples are acquired at a rate of 8 to 12 frames per second.

[0015] Preferably, the resolution of the high-resolution microscope camera is ≥ 0.5 μm.

[0016] Preferably, the denoising method of the preprocessing module is adaptive median filtering; the motion correction method is optical flow motion artifact correction.

[0017] Preferably, the algorithm for illumination equalization in the preprocessing module is as follows:

[0018] ;

[0019] in, and The mean and standard deviation of the local window. To prevent division by zero errors by using extremely small constants.

[0020] Preferably, the dual-path CNN model of the hierarchical recognition module consists of a feature extraction path and a boundary optimization path.

[0021] Preferably, the feature extraction path is to use a ResNet-34 backbone network to output a multi-scale semantic segmentation heatmap.

[0022] Preferably, the boundary optimization path involves introducing a spatial attention module to enhance the feature response of the component boundary region, and the segmentation loss function L... seg The weighted cross-entropy is calculated using the following formula:

[0023] ;

[0024] in, The weighting coefficients are assigned according to the importance of the components (platelets have the highest weight). For real labels, To predict probabilities.

[0025] Preferably, the deep fusion algorithm is based on the depth-of-focus variation and segmentation mask of multi-focal images, and uses voxel fusion to perform 3D reconstruction.

[0026] Preferably, the mapping rule for the pseudo-color layered image of the 3D visualization module is as follows:

[0027] Plasma appears as a blue channel, and its transparency decreases as concentration increases.

[0028] Red blood cells are represented by red channels, and saturation indicates cell density.

[0029] White blood cells are represented by purple channels, and the width of the boundary line reflects the degree of aggregation.

[0030] Platelets are represented by yellow channels, and their brightness is positively correlated with their distribution density.

[0031] Preferably, the formula for calculating the uniformity index (UI) is:

[0032] ;

[0033] in, The standard deviation of component concentration, The UI value represents the average concentration; the closer the UI value is to 1, the more uniform the distribution.

[0034] Preferably, the system further includes a training module, which trains a dual-path CNN model and uses the trained dual-path CNN model to identify and segment components such as plasma, red blood cells, white blood cells, and platelets through semantic segmentation and boundary optimization algorithms.

[0035] Preferably, the training module employs a phased training strategy during its training process, specifically as follows:

[0036] (1) Pre-training phase: Initialize model parameters on a public blood dataset;

[0037] (2) Fine-tuning stage: Optimize platelet and small white blood cell recognition capabilities using expert-annotated sample data;

[0038] (3) Dynamic enhancement stage: Synthesize interference samples containing bubbles, impurities or changes in illumination by SimGAN or other generative adversarial networks to expand the diversity of training data and enable the model to have better adaptability when facing different imaging conditions.

[0039] Preferably, the publicly available blood dataset includes BBBC041.

[0040] Secondly, the present invention provides an optically visualized layering method for blood separation, comprising the following steps:

[0041] Step S1: Acquire dynamic sequence images of blood samples using optical microscopy, with each sample covering 8-12 focal planes;

[0042] Step S2: Perform denoising, motion correction, and illumination equalization on the original image;

[0043] Step S3: Input the processed image from step S2 into the dual-path CNN model, and output the semantic segmentation mask of plasma, red blood cells, white blood cells, and platelets to perform automatic identification and segmentation of each component.

[0044] Step S4: Generate a three-dimensional pseudo-color layer map based on the segmentation results, calculate the volume ratio, layer thickness and distribution uniformity index of each component; output a quantitative report on the volume ratio, layer thickness and distribution uniformity of each component.

[0045] This invention provides an optically visualized layering method for blood separation. This method integrates high-resolution optical imaging with deep learning algorithms, achieving precise separation and three-dimensional visualization of whole blood components (plasma, red blood cells, white blood cells, and platelets) through real-time image acquisition, dynamic feature extraction, and intelligent layering, significantly improving the efficiency and reliability of clinical blood analysis.

[0046] Preferably, in step S1, multifocal dynamic sequence images of blood samples are acquired at a rate of 8 to 12 frames per second.

[0047] Preferably, the stepping accuracy of the focal plane in step S1 is 1~3 μm.

[0048] Preferably, in step S2, the denoising method is adaptive median filtering. Compared with traditional median filtering, adaptive median filtering can avoid incomplete denoising caused by an excessively small window size, and can also avoid image blurring caused by an excessively large window size. It is more suitable for preserving details of tiny structures such as platelets.

[0049] Preferably, in step S2, the motion correction method is optical flow motion artifact correction.

[0050] Preferably, in step S2, the processing formula for illumination equalization is as follows:

[0051] ;

[0052] in, and The mean and standard deviation of the local window. To prevent division by zero errors by using extremely small constants.

[0053] Preferably, the dual-path CNN model of the hierarchical recognition module in step S3 consists of a feature extraction path and a boundary optimization path. The feature extraction path uses a ResNet-34 backbone network to output a multi-scale semantic segmentation heatmap; the boundary optimization path introduces a spatial attention module to enhance the feature response of the component boundary region, and the segmentation loss function L... seg The weighted cross-entropy is calculated using the following formula:

[0054] ;

[0055] in, The weighting coefficients are assigned according to the importance of the components (platelets have the highest weight). For real labels, To predict probabilities.

[0056] Preferably, the formula for calculating the distribution uniformity index in step S4 is:

[0057] ;

[0058] in, The standard deviation of component concentration, The UI value represents the average concentration; the closer the UI value is to 1, the more uniform the distribution.

[0059] Preferably, the mapping rule for the pseudo-color layered image is as follows:

[0060] Plasma appears as a blue channel, and its transparency decreases as concentration increases.

[0061] Red blood cells are represented by red channels, and saturation indicates cell density.

[0062] White blood cells are represented by purple channels, and the width of the boundary line reflects the degree of aggregation.

[0063] Platelets are represented by yellow channels, and their brightness is positively correlated with their distribution density.

[0064] Preferably, the method further includes training the dual-path CNN model in step S3, wherein the training adopts a staged training strategy, and the specific steps are as follows:

[0065] Pre-training phase: Training on publicly available blood or cell imaging datasets to learn general texture and morphological features;

[0066] Fine-tuning phase: Optimize boundary recognition capabilities using sample data collected by this system, with a focus on enhancing the detection sensitivity of platelets and small-sized white blood cells;

[0067] Dynamic enhancement stage: SimGAN generative adversarial network is used to synthesize interference samples containing bubbles, impurities or changes in lighting to improve the robustness of the model.

[0068] Preferably, the publicly available blood dataset includes BBBC041.

[0069] Compared with the prior art, the present invention has achieved the following beneficial effects:

[0070] The optically visualized layering system and method for blood separation provided by this invention have the following advantages:

[0071] The optical visual stratification system and method for blood separation provided by this invention have high accuracy, with a stratification error rate of ≤ 3% (compared to 15% for traditional centrifugation), and a platelet identification accuracy of up to 98%. It is also highly efficient, with the entire process taking less than 5 minutes (compared to over 30 minutes for centrifugation), and supports real-time analysis. Enhanced visualization, the three-dimensional pseudo-color stratification map intuitively displays the spatial distribution of components, assisting doctors in rapid diagnosis. Furthermore, it has strong anti-interference capabilities; the dynamic preprocessing algorithm effectively overcomes interference from sample flow and impurities. Compared to baseline models without dynamic enhancement strategies, this method shows approximately 40% improved robustness in identifying samples containing impurities. Attached Figure Description

[0072] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0073] Figure 1 is a system flowchart of the present invention;

[0074] Figure 2 is a diagram of the hierarchical recognition convolutional neural network structure of the present invention;

[0075] Figure 3 is a flowchart of the process for generating a three-dimensional pseudo-color layered map according to the present invention;

[0076] Figure 4 is a schematic diagram of the visualization effect of the present invention. Detailed Implementation

[0077] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0078] The technical solution of the present invention will be further described below with reference to specific embodiments.

[0079] Example 1

[0080] An optically visualized layering system for blood separation, the system flowchart of which is shown in Figure 1, includes the following modules:

[0081] Optical microscopy imaging module:

[0082] Optical microscopy imaging module: Employing a high-resolution microscope camera (spatial resolution ≥ 0.5 μm) and a reflective illumination system, this module captures multi-focal dynamic sequence images of blood samples in real time. The image acquisition rate is set to 10 frames / second in this embodiment to balance temporal resolution and exposure noise. The multi-focal scanning is configured to cover 10 focal planes, with a focal plane stepping accuracy of 1 μm to accommodate small components (such as platelets) and the mechanical feasibility of the equipment. This configuration ensures the detail of platelet and small leukocyte outlines while avoiding reduced signal-to-noise ratio per frame due to excessively high frame rates or loss of layer information due to excessive stepping.

[0083] Preprocessing module: Performs adaptive median filtering for noise reduction, optical flow motion artifact correction, and illumination equalization on the original image to eliminate sample impurities or bubble interference.

[0084] An adaptive median filter is applied to the original image for denoising. This method addresses common impulse noise and highlight artifacts caused by tiny particles and optical sensor noise in microscopic images. The filter window size is adaptively adjusted based on the relationship between the minimum, maximum, and median values ​​of pixels within a local window. When a fixed window fails to effectively suppress outliers, the window size is automatically expanded, thus effectively removing isolated noise points while preserving cell edges and texture. Compared to traditional median filtering, adaptive median filtering avoids incomplete denoising due to an excessively small window size and image blurring due to an excessively large window size, making it more suitable for preserving details in minute structures such as platelets.

[0085] Motion artifact correction using optical flow methods addresses the issue of inter-frame displacement and blurring caused by slow flow of blood samples or slight translation introduced by microscope focal length adjustments during imaging. This system employs dense optical flow methods, such as multi-scale optical flow estimation based on the Farnebäck algorithm, to estimate the displacement field between adjacent frames pixel-by-pixel. Based on the obtained optical flow field, sub-pixel alignment is performed on subsequent frames to correct motion artifacts and improve the temporal consistency and spatial stability of dynamic sequences. This approach effectively reduces component misalignment caused by local motion during 3D reconstruction, ensuring a consistent spatial reference for subsequent dual-path CNN model inputs.

[0086] Illumination equalization is performed to eliminate inconsistent brightness distribution caused by uneven light sources, excessive brightness at the center of the field of view, or background shadows during microscopic imaging. It employs a local statistical normalization method for illumination equalization. Specifically, the image is divided into local windows, the mean μ and standard deviation σ within each window are calculated, and a normalization formula is used:

[0087] ;

[0088] Where ε is a small constant to prevent division by zero. This processing can make the brightness distribution of the entire image more uniform, especially reducing the shadow effect in areas with dense red blood cells, while improving the contrast of smaller targets such as platelets and white blood cells, making subsequent semantic segmentation and boundary recognition more stable and reliable, and eliminating sample impurities or bubble interference.

[0089] The algorithm for illumination equalization is as follows:

[0090] ;

[0091] in, and The mean and standard deviation of the local window. To prevent division by zero errors by using extremely small constants.

[0092] Layer recognition module:

[0093] A hierarchical model was constructed based on a convolutional neural network. The plasma, red blood cells, white blood cells and platelet components were identified and segmented through feature extraction and boundary optimization algorithms. The structure diagram of the hierarchical recognition convolutional neural network is shown in Figure 2.

[0094] The dual-path CNN model of the hierarchical recognition module consists of a feature extraction path and a boundary optimization path.

[0095] The feature extraction path uses a ResNet-34 backbone network to output a multi-scale semantic segmentation heatmap.

[0096] The boundary optimization path introduces a spatial attention module to enhance the feature response of the component boundary region, and the segmentation loss function L... seg The weighted cross-entropy is calculated using the following formula:

[0097] ;

[0098] in, The weighting coefficients are assigned according to the importance of the components (platelets have the highest weight). For real labels, To predict probabilities.

[0099] 3D visualization module: Maps the layering results to a 3D coordinate system, generates a 3D pseudo-color layered map, uses color depth to indicate component concentration and spatial distribution, outputs a quantitative report on the volume percentage, layer thickness and distribution uniformity of each component, and generates a 3D pseudo-color layered map flowchart as shown in Figure 3, and a visualization effect diagram as shown in Figure 4.

[0100] The mapping rules for 3D pseudo-color layered images are as follows:

[0101] Plasma appears as a blue channel, and its transparency decreases as concentration increases.

[0102] Red blood cells are represented by red channels, and saturation indicates cell density.

[0103] White blood cells are represented by purple channels, and the width of the boundary line reflects the degree of aggregation.

[0104] Platelets are represented by yellow channels, and their brightness is positively correlated with their distribution density.

[0105] The system also includes a training module, which trains a dual-path CNN model. The trained dual-path CNN model is used to identify and segment plasma, red blood cells, white blood cells and platelets through semantic segmentation and boundary optimization algorithms.

[0106] The training module employs a phased training strategy, specifically as follows:

[0107] (1) Pre-training phase: Initialize model parameters on a public blood dataset (such as BBBC041);

[0108] (2) Fine-tuning stage: Optimize platelet and small white blood cell recognition capabilities using expert-annotated sample data;

[0109] (3) Dynamic augmentation stage: SimGAN or other generative adversarial networks are used to synthesize perturbation samples containing bubbles, impurities or different lighting conditions to improve the robustness of the model; at the same time, data augmentation strategies such as random cropping, brightness / contrast perturbation, and rotation are used during training to further improve generalization. During training, Dice loss and weighted cross-entropy can be combined to balance class imbalance and boundary consistency.

[0110] Example 2

[0111] An optically visualized layering method for blood separation includes the following steps:

[0112] S1. Dynamic Image Acquisition: In this embodiment, dynamic sequence images of blood samples are acquired at a rate of 10 frames per second using optical microscopy. Each sample is set to cover 10 focal planes (stepping accuracy is set to 1 μm in this embodiment) to ensure complete capture of different component depths of field while taking into account the imaging signal-to-noise ratio.

[0113] S2. Preprocessing: Adaptive median filtering for noise reduction, optical flow-based motion correction, and illumination equalization are performed on the original image to eliminate errors caused by motion artifacts, bubbles, and uneven illumination. Illumination equalization is performed by normalizing based on local statistics (local mean / standard deviation) (using a small constant to avoid division by zero). The illumination equalization formula is as follows:

[0114] ;

[0115] in, and The mean and standard deviation of the local window. To prevent division by zero errors by using extremely small constants.

[0116] S3, Layered Recognition: Input the processed image from step S2 into the trained dual-path CNN model (as described above, ResNet-34 feature extraction + spatial attention boundary optimization path), and output semantic segmentation masks for plasma, red blood cells, white blood cells, and platelets; to reduce the impact of class imbalance, the segmentation adopts a combination of weighted cross-entropy and boundary-aware loss training.

[0117] The model employs a dual-path design:

[0118] Path 1 (Feature Extraction): Use the ResNet-34 backbone network to extract multi-scale features and output semantic segmentation heatmaps for plasma, red blood cells, white blood cells, and platelets.

[0119] Path Two (Boundary Optimization): Introducing a Spatial Attention Module to enhance the feature responses of component boundary regions and reduce overlap misclassification. Segmentation loss function L... seg The weighted cross-entropy is calculated using the following formula:

[0120] ;

[0121] in, The weights are categorical (platelets have the highest weight because they are easily ignored). For real labels, To predict probabilities.

[0122] S4. 3D Reconstruction and Quantization: Generate a 3D pseudo-color layered map based on the segmentation mask and depth of focus information, calculate and output quantitative reports such as the volume percentage, layer thickness distribution, and distribution uniformity index of each component; at the same time, generate an interactive 3D view for the operator to view and export.

[0123] This method achieves real-time stratification, precise segmentation, and three-dimensional visualization of whole blood components through the above steps. It is suitable for real-time clinical analysis and visualization assessment before automated laboratory sorting. The stratification error rate is ≤ 3% (compared to 15% for traditional centrifugation), and the accuracy rate for platelet identification reaches 98%. It is highly efficient, with the entire process taking less than 5 minutes (compared to more than 30 minutes for centrifugation), and supports real-time analysis.

[0124] Pseudo-color mapping: Assign color levels based on component concentration (plasma is blue, red blood cells are red, white blood cells are purple, and platelets are yellow), and use the alpha channel transparency to identify spatial layering relationships.

[0125] The formula for calculating the distribution uniformity index in step S4 is as follows:

[0126] ;

[0127] in, The standard deviation of component concentration, The UI value represents the average concentration; the closer the UI value is to 1, the more uniform the distribution.

[0128] The method further includes training the dual-path CNN model in step S3, inputting the processed image from step S2 into the trained dual-path CNN model, and outputting semantic segmentation masks for plasma, red blood cells, white blood cells, and platelets. The training employs a phased training strategy, with the following specific steps:

[0129] Pre-training phase: Train the basic segmentation network on a public blood dataset (such as BBBC041) to learn general features;

[0130] Fine-tuning phase: Optimize boundary recognition capabilities using sample data collected by this system, with a focus on enhancing the detection sensitivity of platelets and small-sized white blood cells;

[0131] Dynamic Augmentation Phase: SimGAN is used to generate samples containing impurities, bubbles, and illumination variations to simulate complex imaging scenarios and increase the diversity of training data, thereby enabling the model to adapt to various conditions. During training, conventional data augmentation methods such as random cropping, rotation, and brightness perturbation are combined to improve the model's generalization ability to different samples.

[0132] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An optically visualized layering system for blood separation, characterized in that, It includes the following modules: an optical microscopy imaging module, which uses a high-resolution microscope camera and a reflective illumination system to acquire multi-focal dynamic sequence images of blood samples; and a preprocessing module, which performs noise reduction, motion correction, and illumination equalization on the raw images to eliminate sample impurities. The hierarchical recognition module uses a dual-path CNN model for automatic identification and segmentation of blood components; the dual-path CNN model of the hierarchical recognition module consists of a feature extraction path and a boundary optimization path. The feature extraction path employs a ResNet-34 backbone network to output multi-scale semantic segmentation heatmaps; the boundary optimization path introduces a spatial attention module to enhance the feature response of component boundary regions, and the segmentation loss function L... seg The weighted cross-entropy is calculated using the following formula: ;in, The weighting coefficients are assigned according to the importance of the components, with platelets having the highest weight. For real labels, To predict probabilities; the 3D visualization module maps the 2D semantic mask to the 3D coordinate system based on the segmentation results, generates a 3D pseudo-color layered map through a deep fusion algorithm, and outputs a quantitative report on the volume ratio, layer thickness, and distribution uniformity index of each component; the deep fusion algorithm is based on the depth of focus changes and segmentation mask of multi-focal images, and uses voxel fusion to perform 3D reconstruction.

2. The system as described in claim 1, characterized in that, Acquire multi-focal dynamic sequence images of blood samples at a rate of 8-12 frames per second; or, use a high-resolution microscope camera with a resolution ≥ 0.5 μm; or, use adaptive median filtering for denoising in the preprocessing module; use optical flow motion artifact correction for motion correction; the algorithm for illumination equalization in the preprocessing module is: ;in, and The mean and standard deviation of the local window. To prevent division by zero errors by using extremely small constants.

3. The system as described in claim 1, characterized in that, The formula for calculating the evenness index UI is: ;in, The standard deviation of component concentration, The UI value represents the average concentration; the closer the UI value is to 1, the more uniform the distribution.

4. The system as described in claim 1, characterized in that, The mapping rules for the pseudo-color layered image of the 3D visualization module are as follows: plasma is represented by the blue channel, and its transparency decreases as the concentration increases; red blood cells are represented by the red channel, and saturation indicates cell density; white blood cells are represented by the purple channel, and the width of the boundary line reflects the degree of aggregation; platelets are represented by the yellow channel, and their brightness is positively correlated with their distribution density.

5. The system as described in claim 1, characterized in that, The system also includes a training module, which trains a dual-path CNN model. The trained dual-path CNN model is used to identify and segment plasma, red blood cells, white blood cells and platelets through semantic segmentation and boundary optimization algorithms.

6. The system as described in claim 5, characterized in that, The training process of the training module adopts a phased training strategy, specifically: (1) Pre-training stage: Initialize model parameters on a public blood dataset; (2) Fine-tuning stage: Optimize the recognition ability of platelets and small white blood cells using expert-annotated sample data; (3) Dynamic enhancement stage: Synthesize interference samples containing bubbles, impurities or changes in illumination through SimGAN or other generative adversarial networks to expand the diversity of training data and enable the model to have better adaptability when facing different imaging conditions.

7. An optically visualized layering method for blood separation, characterized in that, Includes the following steps: Step S1: Acquire dynamic sequence images of blood samples using optical microscopy, with each sample covering 8-12 focal planes; Step S2: Perform denoising, motion correction, and illumination equalization on the original images; Step S3: Input the processed images from Step S2 into a dual-path CNN model, which outputs semantic segmentation masks for plasma, red blood cells, white blood cells, and platelets, enabling automatic identification and segmentation of each component; The dual-path CNN model in the hierarchical recognition module of Step S3 consists of a feature extraction path and a boundary optimization path; the feature extraction path uses a ResNet-34 backbone network to output multi-scale semantic segmentation heatmaps; the boundary optimization path introduces a spatial attention module to enhance the feature response of the component boundary region, with a segmentation loss function L... seg The weighted cross-entropy is calculated using the following formula: ;in, These are the weighting coefficients assigned according to the importance of the components. For real labels, To predict the probability; Step S4: Generate a three-dimensional pseudo-color layer map based on the segmentation results, calculate the volume ratio, layer thickness and distribution uniformity index of each component; output a quantitative report of the volume ratio, layer thickness and distribution uniformity of each component.

8. The method as described in claim 7, characterized in that, In step S1, multi-focal dynamic sequence images of blood samples are acquired at a rate of 8-12 frames per second; or, in step S1, the focal plane stepping accuracy is 1-3 μm; or, in step S2, the denoising method is adaptive median filtering; or, in step S2, the motion correction method is optical flow motion artifact correction; or, in step S2, the illumination equalization processing formula is as follows: ;in, and The mean and standard deviation of the local window. To prevent division by zero errors by using extremely small constants.

9. The method as described in claim 7, characterized in that, The method also includes training the dual-path CNN model in step S3. The training adopts a phased training strategy, and the specific steps are as follows: Pre-training phase: train the basic segmentation network on a public blood dataset to learn general features; Fine-tuning stage: Optimize boundary recognition capabilities using sample data collected by this system, with a focus on enhancing the detection sensitivity of platelets and small-sized white blood cells; Dynamic enhancement stage: Improve model robustness by synthesizing interfering samples containing bubbles, impurities, or changes in lighting through SimGAN generative adversarial network.

10. The method as described in claim 7, characterized in that, The formula for calculating the distribution uniformity index UI in step S4 is as follows: ;in, The standard deviation of component concentration, For average concentration, the closer the UI value is to 1, the more uniform the distribution; or, the mapping rules for the pseudo-color stratified image are: plasma is the blue channel, and the transparency decreases as the concentration increases; red blood cells are the red channel, and the saturation indicates the cell density; white blood cells are the purple channel, and the width of the boundary line reflects the degree of aggregation; platelets are the yellow channel, and the brightness is positively correlated with the distribution density.

Images