Target-shaped red blood cell image recognition counting and proportion method

Abstract

This invention belongs to the field of medical image processing technology, specifically relating to a method for identifying, counting, and determining the proportion of target-shaped red blood cells. The method involves acquiring high-resolution digital images of bone marrow or peripheral blood smears under an oil immersion lens; inputting these high-resolution digital images into a dual-task deep learning detection model, which simultaneously performs pixel-level semantic segmentation of individual red blood cells and their intracytoplasmic target structures; based on the localization and pixel-level semantic segmentation, determining the presence and classification of each red blood cell, counting the number of target-shaped red blood cells and the total number of red blood cells, and determining the proportion of target-shaped red blood cells using a proportioning algorithm; based on the proportion of target-shaped red blood cells, fusing red blood cell parameters to generate morphological integrated indicators, and outputting a diagnostic report. This solves the problems of lacking standardized image acquisition procedures and difficulty in accurately distinguishing between target and non-target-shaped red blood cells in existing technologies.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a method for identifying, counting, and determining the proportion of target-shaped red blood cells in images. Background Technology

[0002] Target cells, as an important abnormal indicator in clinical blood morphology examination, are crucial for the accurate detection of their quantity and proportion in the diagnosis, differential diagnosis, and disease assessment of various anemia diseases such as thalassemia and iron deficiency anemia. Currently, the detection of target cells mainly relies on manual microscopic examination or early image analysis techniques. Manual microscopic examination involves laboratory personnel observing bone marrow or peripheral blood smears under an oil immersion microscope, identifying and counting target cells based on experience, thus providing a basis for clinical diagnosis. Early image analysis techniques attempt to assist in identification through simple image processing algorithms, reducing manual labor intensity and improving detection efficiency to some extent.

[0003] Existing technologies have many inherent defects, making it difficult to meet clinical needs for accuracy and standardization in detection: They lack standardized image acquisition procedures, and systematic calibration and verification of oil immersion lens focal length, illumination intensity, and smear quality are not performed. Acquired images are easily affected by impurities and imaging noise, resulting in inconsistent image quality and introducing errors into subsequent identification and analysis. They cannot simultaneously achieve precise localization of individual red blood cells and accurate analysis of the target structure within the cytoplasm. Traditional single-task models either can only locate red blood cells but struggle to accurately segment the target structure, or their segmentation accuracy is insufficient to effectively distinguish pixel differences between the target and the cytoplasm. The determination and classification of red blood cell validity lack scientific model support, easily misclassifying non-cellular impurities, broken cells, and overlapping areas as valid red blood cells, leading to inaccurate counting results. Furthermore, the classification criteria for target red blood cells are vague, making it difficult to accurately distinguish between target and non-target red blood cells. Simply counting the proportion of target red blood cells without integrating other red blood cell parameters to generate comprehensive morphological indicators results in relatively singular detection results, failing to provide a more comprehensive diagnostic reference for clinicians, and the standardization and integration of diagnostic reports are insufficient. Summary of the Invention

[0004] This application provides a method for identifying, counting, and determining the proportion of target-shaped red blood cells in images, in order to solve the problems in the prior art, such as the lack of a standardized image acquisition process and the difficulty in accurately distinguishing between target-shaped and non-target-shaped red blood cells.

[0005] The first aspect of this application provides a method for identifying, counting, and determining the proportion of target-shaped red blood cells, comprising the following steps: acquiring a high-resolution digital image of a bone marrow or peripheral blood smear under an oil immersion lens; inputting the high-resolution digital image into a dual-task deep learning detection model, and simultaneously performing pixel-level semantic segmentation of individual red blood cells and their intracytoplasmic target structures through the dual-task deep learning detection model; based on the localization and pixel-level semantic segmentation, determining the existence and classification of each individual red blood cell, counting the number of target-shaped red blood cells and the total number of red blood cells, and determining the proportion of target-shaped red blood cells through a proportion algorithm; based on the proportion of target-shaped red blood cells, fusing red blood cell parameters to generate a morphological integrated index, and outputting a diagnostic report.

[0006] Preferably, acquiring high-resolution digital images of bone marrow or peripheral blood smears under an oil immersion microscope includes: constructing a dedicated platform for oil immersion microscope image acquisition; based on the dedicated platform, performing oil immersion objective focal length calibration, illumination intensity standardization, and specimen smear quality verification to determine image acquisition adaptation parameters; and according to the image acquisition adaptation parameters, acquiring the original image using a high-magnification microscopic imaging device, and generating a high-resolution digital image that meets analytical standards after impurity interference removal and imaging noise suppression processing.

[0007] Preferably, the dual-task deep learning detection model is used to simultaneously locate individual red blood cells and perform pixel-level semantic segmentation of their intracytoplasmic target structures. This includes: constructing a dual-task deep learning detection model framework based on the fusion of object detection and semantic segmentation; based on the dual-task deep learning detection model framework, the localization branch learns the morphological contour and boundary features of individual red blood cells through feature pyramid and adaptive anchor box matching, and outputs the bounding box coordinates, confidence level, and individual identifier of red blood cells in real time; the segmentation branch focuses on the gray-scale distribution, morphological features, and pixel differences of the intracytoplasmic target structures through multi-scale feature extraction and attention mechanisms, and simultaneously performs pixel-level semantic segmentation of the target structures to generate accurate target structure segmentation masks; through joint training of the localization branch and the segmentation branch and loss function optimization, a synchronous parsing result containing red blood cell localization information and intracytoplasmic target structure segmentation information is output.

[0008] Preferably, the existence determination and classification of each individual red blood cell includes: constructing a red blood cell existence determination and classification model; based on the red blood cell existence determination and classification model, combined with the morphological contour, boundary features and pixel distribution patterns in the red blood cell location information, determining whether the region is a valid red blood cell individual, excluding suspected areas of non-cellular impurities, severe fragmentation and excessive overlap, and obtaining valid red blood cell individuals; classifying them as target red blood cells or non-target red blood cells according to the morphological characteristics, grayscale differences and spatial distribution of the target structure in the cytoplasm of the valid red blood cell individual.

[0009] Preferably, the method of counting target red blood cells and total red blood cells includes: constructing a dedicated model for counting red blood cells; based on the dedicated model for counting red blood cells, including valid red blood cell individuals after existence determination and classification, counting the total number of red blood cells that meet the criteria within the field of view, and obtaining statistical classification results; and synchronizing the number of individuals marked as target red blood cells in the statistical classification results to form a clear and corresponding statistical data on total red blood cells and target red blood cell counts.

[0010] Preferably, the formula for the dual-task deep learning detection model is: ; ; ; ; ; ; ; ; ; ; in, For the first Layer fusion features; For upsampling operators; for Deep and strong semantic features; This is an element-wise summation operation; The first output for the backbone network High-resolution features in shallow layers; For intersection, union, and comparison; For anchor frame; The true bounding box; For intersection operations; This is a union operation; For classification loss; This represents the number of positive sample anchor frames; This represents the total number of anchor frames; For indicator functions; For anchor frame serial number; The set of positive samples; For anchor frame The probability of being predicted as a red blood cell; For focusing parameters; For bounding box regression loss; To predict bounding box parameters; These are the parameters of the actual bounding box; For smoothing L1 loss function; The difference between the predicted value and the actual value; To determine the total loss for the branch; These are the classification loss weight coefficients; These are the regression loss weight coefficients; This is an attention weight map; Use the Sigmoid activation function; This is a 1×1 convolution operation; This is a global average pooling operation; This is a feature map of the target core structure after multi-scale feature extraction; This is a global max pooling operation; The feature map is after attention weighting; For branching loss; The feature map height; The width of the feature map; For pixels The probability of predicting a target structure; For pixels Authentic labeling; It is the minimum value; For the first The probability value of each pixel being predicted as a bullseye structure; The total loss due to the combined efforts of both missions; To determine the weighting coefficients for the branch loss; The weighting coefficients for the branch loss are used.

[0011] The second aspect of this application provides a target-shaped red blood cell image recognition, counting, and proportion system, comprising: an image acquisition module for acquiring high-resolution digital images of bone marrow or peripheral blood smears under an oil immersion lens; a dual-task deep learning detection module for inputting the high-resolution digital images into a dual-task deep learning detection model, and simultaneously performing pixel-level semantic segmentation of individual red blood cells and their intracytoplasmic target structures through the dual-task deep learning detection model; a counting and proportion analysis module for determining the existence and classification of each individual red blood cell based on the location and pixel-level semantic segmentation, counting the number of target-shaped red blood cells and the total number of red blood cells, and determining the proportion of target-shaped red blood cells through a proportion algorithm; and a diagnostic report generation module for generating morphological integrated indicators based on the proportion of target-shaped red blood cells, fusing red blood cell parameters, and outputting a diagnostic report.

[0012] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the program to implement the target red blood cell image recognition, counting, and proportion method as described in the above embodiments.

[0013] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the target red blood cell image recognition, counting, and proportion method as described in the above embodiments.

[0014] A fifth aspect of this application provides a computer program product, including a computer program or instructions, for implementing the target red blood cell image recognition, counting, and proportion method as described in the above embodiments.

[0015] Therefore, this application includes the following beneficial effects: By accurately analyzing high-resolution digital images of bone marrow or peripheral blood smears under oil immersion, and using a dual-task deep learning detection model to simultaneously perform individual red blood cell localization and pixel-level semantic segmentation of the target structure, it effectively overcomes the shortcomings of traditional manual microscopic counting, such as reliance on the experience of laboratory personnel, large subjective errors, poor repeatability, and low efficiency. The simultaneous localization and segmentation operations also avoid the error accumulation problem caused by step-by-step processing, significantly improving the accuracy of target red blood cell identification and the reliability of counting results. The automated classification statistics and proportion calculation process greatly shortens the testing time for a single sample, reduces labor costs, and has the ability to process clinical samples in batches, meeting the application needs of large-scale screening. For target red blood cells with atypical morphology, pixel-level semantic segmentation can accurately capture the characteristic details of the target structure in the cytoplasm, further reducing the probability of missed detection and false detection. In addition, the morphological integrated indicators and diagnostic reports generated by integrating key parameters such as mean erythrocyte volume and hemoglobin concentration can provide objective and quantitative reference for the differential diagnosis of thalassemia, iron deficiency anemia, and other related blood diseases, helping to promote efficient decision-making in diagnosis and treatment. This solves the problems in existing technologies, such as the lack of standardized image acquisition processes and the difficulty in accurately distinguishing between target and non-target red blood cells.

[0016] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0017] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for identifying, counting, and determining the proportion of target-shaped red blood cells according to an embodiment of this application; Figure 2 This is a schematic diagram of a pixel-level semantic segmentation scenario according to an embodiment of this application; Figure 3 This is a schematic diagram of a red blood cell count scenario provided according to an embodiment of this application; Figure 4This is a flowchart of a method for identifying, counting, and determining the proportion of target-shaped red blood cells according to an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a target-shaped red blood cell image recognition, counting, and proportion system according to an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0018] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0019] The following describes a method for identifying, counting, and determining the proportion of target-shaped red blood cells according to embodiments of this application, with reference to the accompanying drawings. Addressing the problem mentioned in the background art of accurately distinguishing between target and non-target-shaped red blood cells, this application provides a method for identifying, counting, and determining the proportion of target-shaped red blood cells. In this method, through precise analysis of high-resolution digital images of bone marrow or peripheral blood smears under oil immersion, a dual-task deep learning detection model is used to simultaneously perform individual red blood cell localization and pixel-level semantic segmentation of the target structure. This effectively overcomes the shortcomings of traditional manual microscopic counting, such as reliance on the experience of laboratory personnel, large subjective errors, poor repeatability, and low efficiency. The simultaneous localization and segmentation operations also avoid the error accumulation problem caused by step-by-step processing, significantly improving the accuracy of target-shaped red blood cell identification and the reliability of the counting results. The automated classification, statistics, and proportion calculation process significantly shortens the testing time for a single sample, reduces labor costs, and has the ability to process clinical samples in batches, meeting the application needs of large-scale screening. Furthermore, for atypical target-shaped red blood cells, pixel-level semantic segmentation can accurately capture the characteristic details of the target structure within the cytoplasm, further reducing the probability of missed or false detections. Furthermore, the morphological integrated indicators and diagnostic reports generated by fusing key parameters such as mean corpuscular volume (MCV) and hemoglobin concentration can provide objective and quantitative references for the differential diagnosis of related blood diseases such as thalassemia and iron deficiency anemia, thus facilitating efficient decision-making in diagnosis and treatment. This solves the problems of existing technologies, such as the lack of standardized image acquisition procedures and the difficulty in accurately distinguishing between target and non-target red blood cells.

[0020] Specifically, Figure 1 The target red blood cell image recognition, counting, and proportion method provided in the embodiments of this application.

[0021] like Figure 1 As shown, the method for identifying, counting, and determining the proportion of target-shaped red blood cells in an image includes the following steps: In step S101, a high-resolution digital image of the bone marrow or peripheral blood smear under an oil immersion lens is acquired.

[0022] It is understood that the high-resolution digital images of bone marrow or peripheral blood smears under an oil immersion lens obtained in this application are the fundamental prerequisite and quality assurance for the target red blood cell image recognition, counting, and proportion method. It not only takes advantage of the high magnification of the oil immersion lens to clearly present the microscopic morphological characteristics of red blood cells, especially the key details such as the central target structure, cytoplasmic distribution, and cell edge contours unique to target red blood cells, but also converts the visual information under the traditional microscope into calculable digital image data through high-resolution digital acquisition, eliminating the field of view limitations and subjective judgment biases of manual microscopic examination. This provides an accurate and stable input data source for the subsequent dual-task deep learning detection model for red blood cell individual localization and pixel-level semantic segmentation of the target structure. At the same time, the digital images support long-term storage, repeated retrieval, and multi-scenario reuse, ensuring the accuracy and traceability of subsequent cell classification, quantity statistics, and proportion calculation.

[0023] In this application example, obtaining high-resolution digital images of bone marrow or peripheral blood smears under an oil immersion microscope includes: constructing a dedicated platform for oil immersion microscope image acquisition; based on the dedicated platform, performing oil immersion objective focal length calibration, illumination intensity standardization, and specimen smear quality verification to determine image acquisition adaptation parameters; and according to the image acquisition adaptation parameters, acquiring the original image using a high-magnification microscopic imaging device, and generating a high-resolution digital image that meets the analysis standards after impurity interference removal and imaging noise suppression processing.

[0024] Among them, the oil immersion objective image acquisition platform is a microscope accessory platform that integrates precision mechanical positioning, illumination system and digital imaging equipment. It is specifically designed to stably and clearly acquire high-resolution microscopic images at high magnification with oil immersion objectives.

[0025] It is understood that, through its proprietary architecture design, this application enables precise calibration of the oil immersion objective focal length, standardized illumination intensity, and rigorous verification of specimen smear quality. This effectively solves problems such as blurred image details and feature distortion caused by focal length shift, uneven illumination, and inconsistent smear quality in traditional acquisition methods. Simultaneously, it accurately determines suitable image acquisition parameters, providing stable acquisition conditions for high-magnification microscopic imaging equipment. Combined with subsequent impurity removal and noise suppression processing, it ultimately generates high-resolution digital images that meet analysis standards. This lays a solid foundation for the accurate localization, semantic segmentation, and subsequent count ratio analysis and diagnostic report generation of the dual-task deep learning detection model, significantly improving the stability, repeatability, and accuracy of the entire detection process.

[0026] It should be noted that the dedicated oil immersion lens image acquisition platform is built through a combination of hardware and software. The hardware integrates a high-magnification oil immersion lens, a high-precision microscopic imaging sensor, a standardized ring illumination system, and a motorized precision stage. An image acquisition card is used to ensure the compatibility and connection of the optical and electrical paths. Specimen fixation devices and dustproof sealing components are also included to guarantee acquisition stability. The software features a dedicated control module that integrates functions such as objective lens focus adjustment, illumination intensity control, and stage movement and positioning. It also incorporates a preliminary quality detection algorithm for specimen smears and synchronously establishes parameter storage and retrieval systems for centralized management and coordinated operation of the hardware. Through multiple optical path calibrations, equipment debugging, and performance tests, the platform ensures that it meets the accuracy, stability, and standardization requirements for image acquisition under the oil immersion lens, providing fundamental support for subsequent operations such as focus calibration and illumination standardization.

[0027] The process included calibrating the oil immersion objective's focal length, standardizing the illumination intensity, and verifying the quality of specimen smears. The oil immersion objective's focal length was calibrated using a standard calibration slide mounted on a dedicated platform. The platform software controlled the motorized stage to move the calibration slide while simultaneously adjusting the oil immersion objective's focal length. Combined with real-time image sharpness data from the imaging sensor, the adjustment was repeatedly fine-tuned until the standard graduations on the calibration slide were sharp and the edges were clear. The corresponding focal length parameters were recorded and saved as baseline values. This calibration was repeated multiple times to eliminate system errors. Illumination intensity standardization used a standard grayscale card as a reference. The adjustment was performed using the platform's ring illumination system. The system gradually adjusts the light source brightness, uses software to collect pixel values ​​of grayscale cards at different brightness levels, and calibrates them to the reference brightness range that meets the requirements of image analysis. At the same time, the uniformity compensation function is enabled to ensure that there is no difference in brightness within the field of view. The quality verification of specimen smears is carried out by acquiring local images of the smears through the platform imaging equipment. The embedded algorithm detects the thickness uniformity, cell distribution density, and whether there are obvious impurities or scratches in the smears. Simultaneously, manual review is combined to confirm and reject smears with uneven thickness, severe cell stacking, or excessive contamination. The quality parameters of qualified smears are recorded to provide a basis for determining the adaptation parameters for subsequent image acquisition.

[0028] For example, a dedicated platform for oil immersion imaging was constructed, with the core configuration of an Olympus BX53 biological microscope (equipped with a 100× oil immersion objective lens, numerical aperture 1.40), a Hamamatsu ORCA-Flash 4.0LTs CMOS camera (resolution 2048×1536 pixels, pixel size 0.102μm / pixel), a customized stage autofocus module, and an LED cold light source control system. An image processing workstation (CPU: Intel Core i9-13900K, GPU: NVIDIA RTX 4090, equipped with dedicated image acquisition software Image-ProPlus 10.0) was also integrated. Based on this platform, the oil immersion objective lens focal length was first calibrated using a standard blood cell smear as the calibration sample, setting the focus range to 0-50μm with a calibration accuracy of ±0.02μm to ensure clear imaging of the red blood cell cytoplasm and target structure. Then, the illumination intensity was standardized, adjusting the LED cold light source brightness to 1200±50 lx. The color temperature was stabilized at 5500K, and a neutral density filter was used to eliminate light source inhomogeneity, ensuring that the illumination variation coefficient within the field of view was ≤3%. Simultaneously, specimen smear quality verification was performed. Bone marrow smears with a thickness of 2-3 μm and a cell distribution density of 80-120 cells / field were selected. Specimens with damaged smears, excessive staining (staining time exceeding 12 minutes), and excessive impurity attachment (more than 5 impurity particles per field of view) were excluded. The image acquisition adaptation parameters were determined to be an exposure time of 80 ms, a gain of 1.2×, and white balance parameters R=102, G=100, and B=98. Subsequently, the original images were acquired using a high-magnification microscope with the above parameters. Impurity interference with an area <5 μm² was removed using the workstation's built-in algorithm, and Gaussian filtering (3×3 kernel size) was used to suppress imaging noise, improving the image signal-to-noise ratio to ≥35 dB. Finally, a high-resolution digital image meeting the analysis standards was generated. A single image clearly shows the morphological contours of red blood cells and the fine target structures within the cytoplasm, meeting the input requirements of the subsequent dual-task deep learning detection model.

[0029] In step S102, the high-resolution digital image is input into the dual-task deep learning detection model, and the localization of individual red blood cells and the pixel-level semantic segmentation of their intracytoplasmic target structures are performed simultaneously through the dual-task deep learning detection model.

[0030] Pixel-level semantic segmentation refers to classifying each pixel in an image according to its semantic category, thereby achieving precise region division of different objects.

[0031] It is understood that the embodiments of this application can overcome the limitations of traditional contour detection, accurately distinguishing the intracytoplasmic target structure from other regions at the pixel level, clearly capturing the gray-scale distribution pattern, subtle morphological features, and pixel-level differences of the target structure, effectively avoiding feature omissions and misjudgments caused by the small size of the target structure, its gray-scale being close to that of the cytoplasm, or its blurred edges. Simultaneously, the accurate target structure segmentation mask generated by this segmentation method can be deeply linked with the red blood cell localization branch results, providing quantitative basis for subsequent red blood cell presence determination and classification, helping to accurately distinguish target red blood cells from non-target red blood cells, eliminating suspected interference areas, ensuring the accuracy of target red blood cell count statistics, and thus providing highly reliable feature data support for the proportion algorithm, morphological integration index generation, and diagnostic report output. This significantly improves the specificity and sensitivity of the entire detection process for target red blood cell identification, ensuring the clinical reference value of the final diagnostic results.

[0032] For example, such as Figure 2 As shown, based on the dual-task model segmentation branch, pixel-level semantic segmentation is performed on the previously acquired high-resolution erythrocyte images of 2048×1536 pixels with a signal-to-noise ratio ≥35dB. First, a four-layer feature pyramid is used for multi-scale feature extraction, with the number of channels gradually increased from 256 to 1024. An SE spatial attention mechanism is embedded to strengthen the feature weight of the target region, accurately capturing the gray-level difference between the target and cytoplasm (target gray level 85-110, cytoplasm gray level 130-160) and pixel distribution patterns. The image is then segmented into 512×512 pixel segments. Block-based, 128-pixel step-size sliding segmentation was used to preserve edge information. The model was jointly optimized using cross-entropy (weight 0.4) and Dice loss function (weight 0.6). After 300 rounds of training, the model converged, and the pixel labeling accuracy of the generated segmentation mask reached ±1 pixel. The recall rate of the target structure segmentation was 93.5% and the precision rate was 92.8%. It can effectively distinguish staining residues and impurities with an area <2μm². Simultaneously, pixel-level segmentation results matching the red blood cell localization bounding box were output, providing core data support for the subsequent accurate classification of target red blood cells.

[0033] In this application example, a dual-task deep learning detection model is used to simultaneously locate individual red blood cells and perform pixel-level semantic segmentation of their intracytoplasmic target structures. This includes: constructing a dual-task deep learning detection model framework based on the fusion of object detection and semantic segmentation; based on this framework, the localization branch learns the morphological contour and boundary features of individual red blood cells through feature pyramid and adaptive anchor box matching, and outputs the bounding box coordinates, confidence scores, and individual identifiers of red blood cells in real time; the segmentation branch focuses on the grayscale distribution, morphological features, and pixel differences of the intracytoplasmic target structures through multi-scale feature extraction and attention mechanisms, and simultaneously performs pixel-level semantic segmentation of the target structures to generate accurate target structure segmentation masks; through joint training and loss function optimization of the localization and segmentation branches, a synchronous parsing result containing both red blood cell localization information and intracytoplasmic target structure segmentation information is output.

[0034] Among them, the bullseye structure segmentation mask refers to a binary image segmentation result specifically used to accurately delineate and identify "bullseye"-shaped annular or concentric circle structural regions in medical images.

[0035] It is understood that the target structure segmentation mask in this application embodiment, as the core output of pixel-level semantic segmentation, can transform the target structure features into quantifiable pixel identifiers, accurately defining their range, boundaries, and distribution, and avoiding misjudgments caused by the small size of the target and its grayscale being close to that of the cytoplasm. Simultaneously, it can be linked with red blood cell localization information to provide objective evidence for the accurate classification and counting of target-shaped red blood cells. The quantifiable features it carries can also support the generation of morphological integration indicators, improving detection accuracy and the reliability of diagnostic reports.

[0036] It should be noted that a dual-task deep learning detection model framework based on the fusion of object detection and semantic segmentation is constructed. This framework enhances the representation of image features across different dimensions through multi-scale feature fusion layers, while parallel object detection and semantic segmentation branches are built to achieve collaborative processing of the two tasks. The localization branch is designed based on an FPN structure, introducing an adaptive anchor box generation mechanism that dynamically adjusts the anchor box size and proportion based on prior knowledge of red blood cell morphology, improving the accuracy of bounding box localization. The segmentation branch adopts an improved U-Net architecture, sharing low-level features with the localization branch at the encoding end, and adding an attention gating structure at the decoding end to focus on the target structure region. Skip connections compensate for the loss of detail during feature dimensionality reduction. The framework also embeds a feature interaction mechanism to achieve bidirectional transfer and complementarity of features between the two branches. A joint loss function is designed to balance the training priorities of both tasks. After model initialization, hyperparameter tuning, and pre-training optimization, a dual-task fusion framework that balances localization speed and segmentation accuracy, and adapts to the requirements of red blood cell and target structure parsing, is finally constructed.

[0037] A precise bullseye structure segmentation mask is generated through multi-scale feature processing and precise localization of the segmentation branch. This branch adopts an improved U-Net architecture, sharing underlying image features with the localization branch at the encoding end. Simultaneously, it enhances multi-scale feature extraction for bullseye structure characteristics, capturing the grayscale distribution and morphological differences of bullseyes of different sizes. At the decoding end, an attention-gated structure is used to precisely focus on the bullseye region within the cytoplasm, suppressing interference from irrelevant regions such as red blood cell edges and background impurities. Skip connections are used to fuse the detailed features extracted at the encoding end with the features at the decoding end, compensating for the loss of detail during feature dimensionality reduction. The fused features are then classified at the pixel level to determine whether each pixel belongs to the bullseye structure. After thresholding and edge correction optimization, false bullseye regions and segmentation noise are eliminated, ultimately generating a segmentation mask with clear contours, precise boundaries, and a close fit to the actual bullseye structure.

[0038] The output includes synchronized analytical results containing individual red blood cell localization information and cytoplasmic target structure segmentation information. This output relies on the collaborative linkage and training optimization of the localization and segmentation branches. The localization branch uses feature pyramids and adaptive anchor box matching to output the bounding box coordinates, confidence level, and individual identifier for each red blood cell in real time, accurately marking the individual red blood cell position and contour range. The segmentation branch simultaneously generates a target structure segmentation mask, clearly defining the pixel distribution of the target region within the cytoplasm. The training process of the two branches is collaboratively optimized based on a joint loss function, calibrating the deviation between the localization boundary and the segmentation region, eliminating information conflicts between branches, and establishing a one-to-one association between the individual red blood cell identifier and the corresponding target segmentation mask. Bounding box information, confidence level data, individual identifier, target pixel distribution, and segmentation mask information are integrated and aligned, eliminating low-confidence localization results and invalid segmentation regions. After data format standardization, the final output is a synchronized analytical result that combines accurate red blood cell localization with complete target structure segmentation, providing unified data support for subsequent red blood cell classification.

[0039] For example, the bullseye structure segmentation mask is a binary image that perfectly matches the size of the input image (2048×1536 pixels, signal-to-noise ratio ≥35dB). It is generated by iterative optimization of the segmentation branch of the dual-task model. The pixels in the bullseye region are labeled as 1, and the pixels in the non-bullseye regions (cytoplasm, background, impurities) are labeled as 0. The pixel labeling accuracy reaches 99.2%, which can accurately cover the near-circular bullseye structure with a diameter of 2-4μm, and the edge pixel deviation is controlled within ±1 pixel. Based on this mask, the pixel proportion of each red blood cell bullseye region in a single image can be quantitatively counted (the pixel proportion of the bullseye region of a normal target-shaped red blood cell is 5%-15%, and it is judged as a non-bullseye structure if it is lower than 3% or higher than 20%). At the same time, the pixel matching degree between the mask and the red blood cell positioning bounding box is checked (matching degree ≥98% is considered effective association), eliminating the situation of segmentation and positioning misalignment. Finally, relying on the mask, the false positive rate of bullseye structure recognition is ≤2.3%, and the false negative rate is ≤1.8%, providing accurate pixel-level morphological basis for subsequent red blood cell classification and target-shaped red blood cell counting.

[0040] In this application example, the formula for the dual-task deep learning detection model is: ; ; ; ; ; ; ; ; ; ; in, For the first Layer fusion features; For upsampling operators; for Deep and strong semantic features; This is an element-wise summation operation; The first output for the backbone network High-resolution features in shallow layers; For intersection, union, and comparison; For anchor frame; The true bounding box; For intersection operations; This is a union operation; For classification loss; This represents the number of positive sample anchor frames; This represents the total number of anchor frames; For indicator functions; For anchor frame serial number; The set of positive samples; For anchor frame The probability of being predicted as a red blood cell; For focusing parameters; For bounding box regression loss; To predict bounding box parameters; These are the parameters of the actual bounding box; For smoothing L1 loss function; The difference between the predicted value and the actual value; To determine the total loss for the branch; These are the classification loss weight coefficients; These are the regression loss weight coefficients; This is an attention weight map; Use the Sigmoid activation function; This is a 1×1 convolution operation; This is a global average pooling operation; This is a feature map of the target core structure after multi-scale feature extraction; This is a global max pooling operation; The feature map is after attention weighting; For branching loss; The feature map height; The width of the feature map; For pixels The probability of predicting a target structure; For pixels Authentic labeling; It is the minimum value; For the first The probability value of each pixel being predicted as a bullseye structure; The total loss due to the combined efforts of both missions; To determine the weighting coefficients for the branch loss; The weighting coefficients for the branch loss are used.

[0041] It is understood that the simultaneous execution of individual red blood cell localization and pixel-level semantic segmentation of the target structure within the cytoplasm in this embodiment avoids the problems of error accumulation and low efficiency caused by traditional step-by-step detection. Its localization branch relies on feature pyramids and adaptive anchor box matching to accurately capture the outline boundary of red blood cells, while the segmentation branch focuses on the subtle features of the target structure through multi-scale feature extraction and attention mechanisms. The joint training of the two branches and the optimization of the loss function achieve synergistic effects between tasks, which not only improves the accuracy of red blood cell localization and the precision of target structure segmentation, but also provides highly reliable basic data for subsequent red blood cell presence determination, classification and counting. This strongly supports the calculation of the proportion of target red blood cells and the generation of morphological integration indicators, ensuring the accuracy of diagnostic reports and clinical reference value.

[0042] For example, the dual-task deep learning detection model, based on a YOLOv8 and U-Net++ fusion architecture, simultaneously performs localization and segmentation tasks on high-resolution red blood cell images of 2048×1536 pixels with a signal-to-noise ratio ≥35dB. The localization branch extracts multi-scale features through a four-layer feature pyramid, with the number of channels progressively increasing from 256 to 1024. It also uses three adaptive anchor boxes (16×16, 32×32, and 64×64 pixels) to match red blood cells of different sizes. With a confidence threshold of 0.75, the model achieves a red blood cell bounding box localization accuracy of 94.1%, with no duplicate or missing individual identifiers. The segmentation branch incorporates a SE spatial attention mechanism, focusing on the grayscale difference between the bullseye and cytoplasm (bullseye 85-110, cytoplasm 130-160). Iterative segmentation is performed using 512×512 pixel blocks. Cross-entropy and Dice loss function (weight ratio 0.4:0.6) are used for joint optimization. After 300 rounds of training, the model converges, achieving a bullseye structure segmentation recall of 93.5% and a precision of 92.8%. The model achieves collaborative inference through parameter sharing between the localization and segmentation branches, with a single image processing time of ≤0.8 seconds. It simultaneously outputs red blood cell bounding box coordinates, individual identifiers, and bullseye segmentation masks, providing accurate two-dimensional data support for subsequent classification and counting.

[0043] In step S103, based on localization and pixel-level semantic segmentation, the existence of each individual red blood cell is determined and classified, the number of target red blood cells and the total number of red blood cells are counted, and the proportion of target red blood cells is determined by a proportion algorithm.

[0044] The proportion algorithm refers to a mathematical method that calculates the ratio of a part of the value to the whole value and multiplies it by 100% to obtain the proportional relationship.

[0045] The formula is: ; ; in, The percentage of target-shaped red blood cells; The number of target red blood cells determined by the dual-task deep learning detection model; The total number of red blood cells located by the dual-task deep learning detection model; For morphological integration indicators; Mean corpuscular volume (MCV); This is the normal reference value for mean corpuscular volume (MCV). Mean corpuscular hemoglobin concentration; This represents the normal reference value for mean corpuscular hemoglobin concentration (MCH). The weighting coefficient of the proportion of target-shaped erythrocytes in the morphological integrated index; The weighting coefficient of the normalized mean corpuscular volume in the morphological integrated index; The weighting coefficient of the normalized value of mean erythrocyte hemoglobin concentration in the morphological integrated index.

[0046] It is understood that the embodiments of this application, based on the precise statistical count of target red blood cells and the total number of effective red blood cells, quickly calculate the percentage of target red blood cells, transforming qualitative morphological recognition results into quantitative indicators, avoiding subjective bias and statistical errors that are prone to occur in manual counting; at the same time, the percentage data output by the algorithm can be deeply integrated with red blood cell morphology parameters, providing core quantitative basis for the generation of morphological integrated indicators, improving the objectivity and scientificity of diagnostic indicators, ultimately supporting the accurate output of clinical diagnostic reports, and strengthening the clinical reference value of test results.

[0047] For example, peripheral blood smears were taken from clinical anemia samples. Following strict clinical blood morphology testing procedures, smear preparation, methanol fixation, and Wright-Gymza complex staining were performed. High-resolution digital images of the field of view were acquired using a 100x oil immersion lens and a high-definition digital imaging system. These images were then input into a dual-task deep learning detection model. The model first used an anchor-frame matching algorithm to accurately locate individual red blood cells. Then, a semantic segmentation module performed pixel-level analysis of the cytoplasmic structure of each red blood cell, focusing on identifying the characteristic boundaries of the target core structure. Simultaneously, an artifact interference exclusion threshold was set to filter out non-specific structures such as uneven staining and cell overlap, ensuring that only morphologically intact red blood cells were classified. After existence determination and classification, the total number of red blood cells in the field of view was found to be 1800, of which 270 were target-shaped red blood cells with typical target core depression structures. Calculations showed that the proportion of target-shaped red blood cells in this sample was 15%.

[0048] In this embodiment, the existence determination and classification of each individual red blood cell includes: constructing a red blood cell existence determination and classification model; based on the red blood cell existence determination and classification model, combined with the morphological contour, boundary features and pixel distribution patterns in the red blood cell location information, determining whether the region is a valid red blood cell individual, excluding suspected regions of non-cellular impurities, severe fragmentation and excessive overlap, and obtaining valid red blood cell individuals; classifying them into target red blood cells or non-target red blood cells according to the morphological characteristics, grayscale differences and spatial distribution of the target structure in the cytoplasm of the valid red blood cell individual.

[0049] Among them, the red blood cell individual existence determination and classification model refers to a system that uses computer vision or machine learning technology to automatically identify red blood cells in microscopic images, determine their existence, and classify them based on their morphological and other characteristics.

[0050] The formula is: ; ; ; ; ; ; ; ; ; in, The score is used to determine the overall presence of red blood cells. The morphological feature weighting coefficient; These are the boundary feature weight coefficients; These are the pixel distribution feature weighting coefficients; Normalized values ​​for morphological features; These are the normalized values ​​for boundary features; The normalized value of pixel distribution features; The roundness of red blood cells; Pi is a constant. The pixel area of ​​the red blood cell region; The perimeter of the boundary pixels of the red blood cell region; For the smoothness of red blood cell boundaries; The pixel number representing the boundary of the red blood cell; This represents the total number of boundary pixels. For the first The curvature of each boundary pixel; For the first The curvature of each boundary pixel; The average curvature of the boundary pixels; The gray-level variance of the red blood cell region; The area is defined as the region of red blood cells. The grayscale value of a pixel; This represents the average gray value within the red blood cell region; The x-coordinate of the pixel within the red blood cell region; The vertical coordinate of the pixel within the red blood cell region; The threshold for determining the presence of red blood cells; The proportion of the target area to the total surface area of ​​red blood cells; The pixel area of ​​the target structure; The total pixel area of ​​effective red blood cells; This represents the absolute difference in average grayscale between the target region and the cytoplasmic region. This represents the average grayscale value of the target area. This represents the average gray value of the cytoplasmic region. For target red blood cell classification discrimination value; The target core proportion feature weighting coefficient; The grayscale difference feature weighting coefficient; The maximum grayscale value of the image; Threshold for classifying target red blood cells.

[0051] It is understood that the embodiments of this application rely on the quantitative data of morphological contours, boundary features, and target structure segmentation masks of red blood cell localization information to accurately determine the effective red blood cell individuals within the field of view, exclude non-cellular impurities, severely fragmented and excessively overlapping suspected areas, and avoid invalid areas interfering with subsequent statistics. At the same time, the model can accurately distinguish between target and non-target red blood cells based on the morphological features, grayscale differences, and spatial distribution of the target structure within the cytoplasm of effective red blood cells, providing reliable data support for the accurate statistics of the number of target red blood cells and the total number of effective red blood cells, effectively improving the accuracy of the calculation of the proportion of target red blood cells, and strengthening the clinical reference value of morphological integrated indicators and diagnostic reports.

[0052] It should be noted that, combining the morphological contour, boundary features, and pixel distribution patterns in the red blood cell localization information, the morphological contour features of red blood cells within the localization bounding box are first extracted. By analyzing the regularity, roundness, and overall shape of the contour to determine if they conform to the basic morphology of normal red blood cells, a preliminary distinction is made between intact red blood cells and suspected broken or deformed areas. Then, the boundary features are simultaneously correlated to verify the clarity, continuity, and edge pixel gradient changes of the contour boundaries. If the boundaries are blurred, broken, or have abnormal gradients, they are identified as areas that may be affected by impurities or overlap. Finally, the pixel distribution patterns are linked to statistically analyze the mean, variance, and gray-level distribution histogram of pixels within the area. The pixel differences between the normal red blood cell cytoplasm and the background and impurities are compared to verify whether the pixel distribution within the contour conforms to the homogeneity characteristics of red blood cell cytoplasm. These three factors work together to confirm that only when the morphological contour is complete, the boundary is clear and continuous, and the pixel distribution matches the characteristics of red blood cell cytoplasm, it is determined to be a valid red blood cell individual. This also lays the foundation for subsequent classification based on target structure features, enabling precise multi-dimensional feature linkage screening.

[0053] To determine whether an area contains a valid red blood cell, non-cellular impurities, severely fragmented areas, and excessively overlapping areas are excluded. First, for non-cellular impurities, the difference between the pixel grayscale distribution of the area and the red blood cell cytoplasm is compared. Combined with features such as irregular morphological contours and unclear pixel gradients at the boundaries, areas with abnormal grayscale and disordered morphology are identified and excluded as impurities. Next, severely fragmented areas are examined. Based on the characteristics of incomplete morphological contours, broken boundaries, and inability to form a complete cell morphology, accompanied by scattered pixel distribution, fragmented red blood cells with more than 1 / 3 of their contour missing and interrupted boundary continuity are removed. Finally, excessively overlapping areas are identified. By analyzing the cross-over of boundary features, abnormal stacking peaks in pixel distribution, and multiple contour boundaries interpenetrating and making it impossible to distinguish individual cells, such areas are classified as suspected overlapping areas and excluded. After screening these three types of areas one by one, only areas with complete morphological contours, clear and continuous boundaries, and pixel distributions matching the characteristics of red blood cell cytoplasm are retained. These are ultimately identified as valid red blood cell individuals, providing accurate targets for subsequent classification.

[0054] For example, a model for determining and classifying the presence of individual red blood cells was constructed. The training set included 100,000 peripheral blood red blood cell microscopic images (including 85,000 valid red blood cell samples, 15,000 samples of non-cellular impurities, and samples of severely fragmented and excessively overlapping areas). The core determination features were the roundness of the red blood cell morphology (threshold 0.75), the gradient variance of the boundary features (threshold 0.3), and the entropy of the pixel distribution (threshold 4.2). After 50 rounds of dynamic monitoring and iteration, the model was validated on an independent test set of 10,000 peripheral blood red blood cell microscopic images. The results showed that the accuracy rate of valid red blood cell determination reached 98.2%, the exclusion rate of non-cellular impurities reached 99.1%, and the misclassification rate of severely fragmented and excessively overlapping areas was controlled within 0.8%, which can achieve accurate screening of valid red blood cell individuals.

[0055] In this embodiment of the application, the method of counting the number of target red blood cells and the total number of red blood cells includes: constructing a special model for counting red blood cells; based on the special model for counting red blood cells, including valid red blood cell individuals after existence determination and classification, counting the total number of red blood cells that meet the criteria within the field of view, and obtaining the statistical classification result; and simultaneously counting the number of individuals marked as target red blood cells in the classification result, forming a clear and corresponding statistical data on the total number of red blood cells and the number of target red blood cells.

[0056] Among them, the red blood cell count model refers to an analysis module that uses an automated blood analyzer to quickly and accurately count red blood cells in a unit volume of blood, based on the principles of electrical impedance or flow cytometry.

[0057] The formula is: ; ; Among them, indicator function The rules for determining the value are as follows: ; ; ; The total number of effective red blood cells after presence determination under a single oil immersion field of view; To determine the number of all red blood cell candidates output by the localization branch of the dual-task deep learning detection model; This represents the number of invalid candidate individuals identified by the existence determination model (including three categories: non-cellular impurities, severely fragmented red blood cells, and excessively overlapping red blood cells). The number of target red blood cells under a single oil immersion field of view; For the first A function to determine the target core structure of an effective red blood cell; The individual serial number of the effective red blood cells; The percentage of target-shaped red blood cells under a single oil immersion field of view; for Average percentage of target-shaped red blood cells per field of view; For the first Number of target red blood cells per field of view; For the first The total number of effective red blood cells per field of view; This is the sequence number of the oil immersion field of view; This represents the total number of observations in the oil immersion field of view.

[0058] It is understood that the embodiments of this application can accurately include valid red blood cell individuals after existence determination and classification, filter out non-cellular impurities, broken and overlapping invalid areas, and avoid the influence of interfering data on statistical results; the model can simultaneously count the total number of valid red blood cells and the number of target red blood cells in the field of view, generate one-to-one corresponding accurate statistical data pairs, and provide an accurate and reliable calculation basis for the proportion algorithm; at the same time, its output standardized statistical results can directly empower the generation of morphological integrated indicators and the quantitative output of diagnostic reports, significantly improving the accuracy of target red blood cell proportion calculation and the stability of the entire detection process.

[0059] It should be noted that the total number of red blood cells meeting the criteria within the field of view is counted. A unique counting index is established by combining the individual identifier and bounding box coordinates from the red blood cell localization information to avoid duplicate counting or omissions caused by cell proximity or cropping at the edge of the field of view. Simultaneously, the morphological contour integrity, boundary clarity, and pixel distribution compliance of each candidate individual are re-verified to remove potentially low-confidence individuals and suspected interference areas. By scanning all imaging areas within the field of view point by point, valid red blood cells meeting all criteria are counted and accumulated one by one. Field of view parameters are recorded simultaneously to ensure statistical consistency. Finally, the total number of red blood cells meeting the criteria within the field of view is output, providing accurate baseline data for subsequent calculation of the target red blood cell percentage.

[0060] For example, such as Figure 3 As shown, in constructing a dedicated model for red blood cell count, 12,000 clinically labeled peripheral blood smear images were included in the training phase. Among them, 856,000 were labeled normal red blood cell samples and 123,000 were target red blood cell samples. After 50 rounds of iteration and optimization, the model achieved an accuracy of 98.7% in identifying individual red blood cells and a recall rate of 97.2% for classifying target red blood cells. In actual clinical testing, a statistical analysis was conducted on a peripheral blood smear field of view of a thalassemia patient. The model first determined the presence of cells in the field of view, removing interfering targets such as platelets and white blood cells. Then, it classified the individuals determined to be valid red blood cells. Finally, the total number of red blood cells meeting the standard in the field of view was 215, and 43 target red blood cells were simultaneously marked, forming a one-to-one correspondence between the total number and the target red blood cell count statistics, providing accurate support for the subsequent calculation of the target red blood cell percentage (approximately 20.0%).

[0061] In step S104, based on the proportion of target-shaped red blood cells, red blood cell parameters are fused to generate morphological integrated indicators and output a diagnostic report.

[0062] Among them, the morphological integration index is a comprehensive measurement standard used to quantify the degree of synergistic change and unity of morphological features among different structures or modules of an organism.

[0063] Understandably, this application's real-time example constructs a multi-dimensional, comprehensive red blood cell morphology assessment system by integrating target red blood cell percentage data with multiple parameters such as red blood cell morphology, size, and grayscale. This transforms isolated counting results into quantitative indicators with clinical interpretability. This indicator can more comprehensively reflect the morphological characteristics and abnormal patterns of the red blood cell population, effectively improving the accuracy of identifying morphological abnormalities related to hematological diseases. At the same time, it provides well-structured and sufficiently supported core assessment content for diagnostic reports, strengthening the correlation between test results and clinical diagnosis, and providing a more scientific and reliable reference for disease screening and condition assessment.

[0064] It should be noted that the fusion of red blood cell parameters, with the proportion of target red blood cells as the core benchmark, first incorporates key quantitative parameters such as red blood cell count, mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), and red blood cell distribution width (RDW). Simultaneously, it combines auxiliary red blood cell morphology parameters. A weighted fusion algorithm prioritizes each parameter, with the proportion of target red blood cells and RWD having the highest weights. This approach also considers the synergistic influence of MCV and morphological abnormality rate, eliminating redundant parameters. The quantitative parameters, morphological parameters, and the proportion of target red blood cells are integrated through multi-dimensional data correlation to construct a morphological integration index that combines quantitative accuracy with morphological specificity. This deep fusion of single proportion data with multi-dimensional red blood cell parameters provides comprehensive and accurate morphological evidence for diagnostic reports.

[0065] The system generates integrated morphological indicators and outputs a diagnostic report. These indicators are generated by algorithmically transforming the proportion of fused target red blood cells, quantitative red blood cell parameters, and morphological auxiliary parameters. This process condenses multidimensional data into quantifiable numerical indicators, accurately mapping the degree of target red blood cell abnormalities to overall red blood cell morphological characteristics, and correlating them with clinical diagnostic threshold calibration indicators to ensure their effectiveness. The diagnostic report output integrates these integrated indicators, target red blood cell count and proportion data, core segmentation and localization images, and red blood cell parameter details according to clinical treatment guidelines. It generates a standardized report containing data tables, morphological feature descriptions, and preliminary diagnostic references, which can be directly exported to clinical treatment systems or printed out, providing an intuitive basis for clinical decision-making.

[0066] For example, in the peripheral blood smear examination of a patient suspected of having β-thalassemia, high-resolution digital images under oil immersion were first acquired. A dual-task deep learning detection model was then used to perform pixel-level semantic segmentation of individual red blood cells and target structures within the cytoplasm. The total number of red blood cells in the field of view was found to be 2300, with 644 target cells, resulting in a target cell percentage of 28.0%. Simultaneously, core red blood cell parameters measured by a hematology analyzer were analyzed: mean corpuscular volume (MCV) was 59.8 fl (normal reference range 80-100 fl), with a deviation of -25.25%, and mean corpuscular hemoglobin was also measured. The target erythrocyte concentration (MCH) was 18.2 pg (normal reference range 27-34 pg), and its deviation was calculated to be -32.35%. Based on the above parameters, a morphological integration index (MI) was constructed, and the calculation formula was set as MI = target erythrocyte percentage + |MCV deviation| + |MCH deviation|. Substituting the data, the morphological integration index MI value of this patient was 85.6%, which far exceeded the upper limit of the MI value reference for healthy people (20%). Combined with clinical symptoms and medical history, the diagnostic report indicated that the patient's erythrocyte morphological characteristics were highly consistent with the hematological manifestations of β-thalassemia. It is recommended to further improve hemoglobin electrophoresis and gene testing to clarify the diagnosis.

[0067] The following will illustrate the method for identifying, counting, and determining the proportion of target-shaped red blood cells through a specific embodiment, such as... Figure 4 As shown, it includes: A 28-year-old male patient was admitted to the hematology department of a tertiary hospital. He complained of fatigue, dizziness, and pallor for three months, palpitations and shortness of breath after exertion, and occasional mild jaundice of the sclera. He had a family history of thalassemia. Outpatient blood routine examination showed a red blood cell count of 3.2×10¹² / L, hemoglobin of 82g / L, mean corpuscular volume (MCV) of 65fL, and mean corpuscular hemoglobin content (MCH) of 22pg, suggesting microcytic hypochromic anemia. To further clarify the type of anemia and assess the severity of the condition, it was necessary to detect the specific number and proportion of target red blood cells in bone marrow and peripheral blood samples using target red blood cell image recognition counting and proportion methods, so as to provide accurate morphological evidence for clinical diagnosis.

[0068] This test strictly followed the target-shaped red blood cell image recognition, counting, and proportion method. First, sample preparation and high-resolution digital image acquisition were performed. 2 ml of bone marrow aspiration and 1 ml of peripheral venous blood were collected from the patient. After anticoagulation, a uniformly thick bone marrow smear was prepared using the spreading method. Similarly, a peripheral blood smear was prepared after anticoagulation with EDTA-K2. After air-drying, the smears were stained using the Wright-Giemsa staining method. The staining time was strictly controlled at 15 minutes and the staining temperature at 25℃. After staining, the smears were rinsed with phosphate buffer until no floating stain remained, and then air-dried again before being placed under an optical microscope at 100× oil pressure. An oil immersion lens (numerical aperture 1.4) was used in conjunction with a high-resolution charge-coupled device (CCD) image sensor (2048×2048 pixels, pixel size 0.5μm) to construct an image acquisition system for the oil immersion field of view. During the acquisition process, the microscope focus was adjusted until the red blood cell morphology was clearly distinguishable, avoiding ghosting, blurring, and reflection interference. For each sample, 50 high-resolution digital images of non-overlapping oil immersion fields of view were randomly acquired, covering different areas of the smear to ensure the representativeness of the sample. After acquisition, the images were stored in TIFF format, and the acquisition parameters (focal length, exposure time, magnification) of each field of view were recorded to provide basic data for subsequent model processing. The stored high-resolution digital images were then input into a pre-trained dual-task deep learning detection model. This model was built by fusing the YOLOv8 object detection algorithm and the U-Net semantic segmentation algorithm, and it has the ability to simultaneously achieve individual red blood cell localization and pixel-level semantic segmentation of the target structure. During the model training process, 10,000 clinically annotated oil immersion images of bone marrow and peripheral blood smears were used as the training set, which included images of red blood cells of various morphologies such as normal red blood cells, target red blood cells, oval red blood cells, and sickle red blood cells. The annotations included red blood cell bounding boxes (localization annotations) and pixel regions of the target structure within the cytoplasm of target red blood cells (semantic segmentation annotations). After 50 rounds of iterative training using the adaptive moment estimation (Adam) optimizer, the model loss function converged to 0.02, the red blood cell localization accuracy reached 98.5%, and the target structure segmentation IoU value (Intersection over Union) reached 97.2%, which can effectively distinguish target red blood cells from other abnormal red blood cells and impurities.

[0069] After the image is input into the model, the model first processes the image, including grayscale enhancement, Gaussian denoising (standard deviation 0.8), and adaptive contrast adjustment, to eliminate background noise, uneven staining, and other interference factors in the image. Then, the target detection branch traverses the image pixels to identify and locate each individual red blood cell, generating the red blood cell bounding box coordinates. At the same time, the semantic segmentation branch performs pixel-level scanning on the cytoplasmic region of each located red blood cell to identify whether there is a target structure characteristic of target red blood cells (i.e., a circular or near-circular densely stained area in the center of the red blood cell, surrounded by a lightly stained band, and an outer ring of densely stained edges, resembling a "target" shape). The pixel region of the target structure is accurately segmented, and the range, area, and location information of the target structure are marked. The localization results of each red blood cell and the target structure segmentation results are output simultaneously to form an intermediate data file containing the red blood cell position, boundary, and target structure parameters. The entire model processing takes about 8 minutes for every 50 images, which is significantly more efficient than traditional manual detection.Based on the localization and pixel-level semantic segmentation results output by the model, the system automatically determines and classifies each individual red blood cell. The criteria are as follows: if the red blood cell boundary is intact, the morphology is regular, or although there are slight abnormalities, it can be clearly identified as a red blood cell, and no target structure is detected after segmentation, it is classified as a normal red blood cell; if the red blood cell boundary is intact, and a clear target structure is detected after segmentation (the area of ​​the target structure accounts for 10%-30% of the total red blood cell area, which meets the morphological definition of a target red blood cell), it is classified as a target red blood cell; if the red blood cell boundary is broken, the morphology is blurred and cannot be identified, or although there is an abnormal morphology, there is no target structure, it is classified as other abnormal red blood cells (not counted in the normal red blood cell and target red blood cell counts). At the same time, a manual review step is set up to manually review the images of red blood cells that the system has identified as target red blood cells and suspected target red blood cells (target structure area accounts for 8%-12%) to exclude false positive annotations (such as staining impurities, false target structures caused by red blood cell overlap). In this detection, a total of 1286 red blood cells were located in 50 fields of view of the bone marrow sample. After system evaluation and manual review, 132 abnormal red blood cells were excluded, 924 normal red blood cells and 230 target-shaped red blood cells were identified. In the peripheral blood sample, 1158 red blood cells were located in 50 fields of view. After excluding 105 abnormal red blood cells, 893 normal red blood cells and 160 target-shaped red blood cells were identified. The total number of red blood cells in the bone marrow sample was 1154 (normal + target-shaped), with 230 target-shaped red blood cells. The total number of red blood cells in the peripheral blood sample was 1053, with 160 target-shaped red blood cells. These were then calculated using a percentage algorithm. The percentage of target red blood cells was calculated using the formula: Target red blood cell percentage = (Number of target red blood cells ÷ Total number of red blood cells) × 100%. Substituting the data, the percentage of target red blood cells in the bone marrow sample was approximately 230 ÷ 1154 × 100% ≈ 20.0%, and the percentage in the peripheral blood sample was approximately 160 ÷ 1053 × 100% ≈ 15.2%. The percentage error between the two repeated tests was also calculated: 0.3% for the bone marrow sample and 0.2% for the peripheral blood sample, ensuring the accuracy and repeatability of the counting and percentage results.

[0070] Based on the above target cell percentage results, the patient's red blood cell parameters (MCV, MCH, erythrocyte distribution width (RDW), hemoglobin concentration (Hb) obtained from routine blood tests) were further integrated to generate a morphological integrated index. The integrated index was constructed using a weighted calculation method, selecting the target cell percentage (weight 0.4), MCV (weight 0.2), MCH (weight 0.2), and RDW (weight 0.2) as core parameters. The formula was: Morphological integrated index = 0.4 × target cell percentage + 0. 2 × (MCV / mean normal MCV) + 0.2 × (MCH / mean normal MCH) + 0.2 × (RDW / mean normal RDW), where the mean normal MCV is set at 82-100 fL (median 91 fL), the mean normal MCH is set at 27-34 pg (median 30.5 pg), and the mean normal RDW is set at 11.5%-14.5% (median 13.0%). Substituting patient data, the calculated bone marrow sample morphological integration index = 0.4 × 20.0% + 0.2×(65 / 91)+0.2×(22 / 30.5)+0.2×(16.8% / 13.0%)≈0.08+0.2×0.714+0.2×0.721+0.2×1.292≈0.08+0.143+0.144+0.258≈0.625; Peripheral blood sample morphological integration index = 0.4×15.2%+0.2×(65 / 91)+0.2×(22 / 30.5)+0.2×(16.8% / 13.0%)≈0. 061+0.143+0.144+0.258≈0.606. Referring to the clinical morphological integrated index threshold (normal range 0.3-0.5, exceeding 0.5 indicates moderate to severe erythrocyte morphological abnormalities, combined with an increased proportion of target red blood cells, highly suggestive of hemolytic anemia such as thalassemia), the patient's morphological integrated index in both samples exceeded 0.6, and the proportion of target red blood cells in the bone marrow sample reached 20.0%, and in the peripheral blood it reached 15.2%, both significantly higher than the normal reference value (normal target red blood cell proportion <2%).

[0071] Finally, the system automatically generates a morphological diagnostic report, which includes basic sample information (patient name, gender, age, sample type, and testing time), testing method (target red blood cell image recognition, counting, and percentage method; dual-task deep learning model model; and image acquisition parameters), counting results (total number of red blood cells in bone marrow and peripheral blood, number of target red blood cells, and percentage of target red blood cells), red blood cell parameter fusion results (morphological integrated index values ​​and interpretation), morphological diagnostic conclusion (based on family history and complete blood count results, β-thalassemia (intermediate type) is highly suspected, and further thalassemia gene testing and hemoglobin electrophoresis are recommended to clarify the subtype), and clinical recommendations (regularly review complete blood counts and red blood cell morphology, monitor the degree of anemia, avoid fatigue and infection, and provide blood transfusions when necessary). The report also includes typical target red blood cell images (labeled with the target core structure segmentation region and bounding box) and a model processing flowchart for easy viewing by clinicians. The results of this test provided precise morphological evidence for clinical diagnosis. Subsequent gene testing of the patient showed that they were heterozygous for β-thalassemia, and hemoglobin electrophoresis showed elevated HbA2, confirming a diagnosis of β-thalassemia (intermediate type). This result was consistent with the target red blood cell image recognition, counting, and morphological integrated index analysis results, validating the clinical applicability and accuracy of this method. Compared to traditional manual counting methods (manual counting of 50 oil immersion fields takes approximately 40 minutes, with a counting error of about 5%-8%, and target red blood cell identification is greatly affected by subjective factors), this method not only significantly improves detection efficiency and counting accuracy but also achieves fusion analysis of target red blood cell proportion and red blood cell parameters through morphological integrated indexes, providing more comprehensive information for clinical disease diagnosis. It can be widely applied to red blood cell morphology detection in diseases such as thalassemia, iron deficiency anemia, and hemolytic anemia, and is especially suitable for clinical scenarios requiring large-sample screening and precise morphological analysis. Furthermore, the standardized procedure of this method reduces reliance on the professional skills of testing personnel, contributing to the standardization and homogenization of red blood cell morphology detection.

[0072] In subsequent clinical applications, this method can be further improved by continuously expanding the training dataset (incorporating red blood cell images under different disease types and staining conditions), optimizing the parameters of the dual-task deep learning model (enhancing the ability to identify rare target red blood cell variants), and refining the weight settings of morphological integration indicators (calibrating parameter weights by combining more clinical case data), thereby providing stronger technical support for the early diagnosis, disease assessment, and treatment monitoring of clinical diseases.

[0073] In summary, this invention utilizes a dual-task deep learning model for simultaneous analysis of red blood cell localization and target structure, effectively addressing the pain points of traditional manual testing, such as long processing times and significant subjective variations. It significantly improves counting accuracy and detection efficiency, avoiding false positives and human error. It can accurately provide quantitative diagnostic evidence for anemia diseases such as thalassemia, verifying the applicability and reliability of the technology in clinical scenarios. This promotes the transformation of blood morphology testing from subjective experience-based judgment to objective, quantitative, and standardized methods, helping to shorten the diagnostic cycle, improve diagnostic accuracy, and reduce the risk of missed or misdiagnosed cases.

[0074] Figure 5 This is a block diagram of the target red blood cell image recognition, counting, and proportion system according to an embodiment of this application.

[0075] like Figure 5 As shown, the target-shaped red blood cell image recognition, counting, and proportion system 10 includes: an image acquisition module 100, a dual-task deep learning detection module 200, a counting and proportion analysis module 300, and a diagnostic report generation module 400.

[0076] The system includes: an image acquisition module 100 for acquiring high-resolution digital images of bone marrow or peripheral blood smears under an oil immersion lens; a dual-task deep learning detection module 200 for inputting the high-resolution digital images into a dual-task deep learning detection model, which simultaneously performs pixel-level semantic segmentation of individual red blood cells and their cytoplasmic target structures; a counting and proportion analysis module 300 for determining the existence and classifying each red blood cell based on its location and pixel-level semantic segmentation, counting the number of target red blood cells and the total number of red blood cells, and determining the proportion of target red blood cells using a proportion algorithm; and a diagnostic report generation module 400 for generating morphological integrated indicators based on the proportion of target red blood cells, fusing red blood cell parameters, and outputting a diagnostic report.

[0077] It should be noted that the foregoing explanation of the target red blood cell image recognition, counting and proportion method embodiment also applies to the target red blood cell image recognition, counting and proportion system of this embodiment, and will not be repeated here.

[0078] The target-shaped red blood cell image recognition, counting, and proportion system proposed in this application, through precise analysis of high-resolution digital images of bone marrow or peripheral blood smears under oil immersion, utilizes a dual-task deep learning detection model to simultaneously perform individual red blood cell localization and pixel-level semantic segmentation of the target structure. This effectively overcomes the shortcomings of traditional manual microscopic counting, such as reliance on laboratory personnel experience, large subjective errors, poor repeatability, and low efficiency. The simultaneous localization and segmentation operations also avoid the error accumulation problem caused by step-by-step processing, significantly improving the accuracy of target-shaped red blood cell recognition and the reliability of counting results. The automated classification, statistics, and proportion calculation process greatly shortens the testing time for a single sample, reduces labor costs, and has the ability to process clinical samples in batches, meeting the application needs of large-scale screening. For target-shaped red blood cells with atypical morphology, pixel-level semantic segmentation can accurately capture the characteristic details of the target structure in the cytoplasm, further reducing the probability of missed detection and false detection. Furthermore, the morphological integrated indicators and diagnostic reports generated by fusing key parameters such as mean corpuscular volume (MCV) and hemoglobin concentration can provide objective and quantitative references for the differential diagnosis of related blood diseases such as thalassemia and iron deficiency anemia, thus facilitating efficient decision-making in diagnosis and treatment. This solves the problems of existing technologies, such as the lack of standardized image acquisition procedures and the difficulty in accurately distinguishing between target and non-target red blood cells.

[0079] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: The memory 601, the processor 602, and the computer program stored on the memory 601 and capable of running on the processor 602.

[0080] When the processor 602 executes the program, it implements the target red blood cell image recognition, counting, and proportion method provided in the above embodiments.

[0081] Furthermore, electronic devices also include: Communication interface 603 is used for communication between memory 601 and processor 602.

[0082] The memory 601 is used to store computer programs that can run on the processor 602.

[0083] The memory 601 may include high-speed RAM (Random Access Memory) memory, and may also include non-volatile memory, such as at least one disk storage.

[0084] If the memory 601, processor 602, and communication interface 603 are implemented independently, then the communication interface 603, memory 601, and processor 602 can be interconnected via a bus to complete communication between them. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0085] Optionally, in a specific implementation, if the memory 601, processor 602, and communication interface 603 are integrated on a single chip, then the memory 601, processor 602, and communication interface 603 can communicate with each other through an internal interface.

[0086] The processor 602 may be a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of this application.

[0087] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for identifying, counting, and determining the proportion of target red blood cells in an image.

[0088] Furthermore, this application also provides a computer program product, including a computer program or instructions, which, when executed, implement the above-described method for identifying, counting, and determining the proportion of target red blood cells in an image.

[0089] In the description of this specification, the references to "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0090] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0091] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0092] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by suitable instructions. For example, if implemented in hardware as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0093] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0094] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for identifying, counting, and determining the proportion of target-shaped red blood cells in an image, characterized in that, include: Acquire high-resolution digital images of bone marrow or peripheral blood smears under an oil immersion lens; The high-resolution digital image is input into a dual-task deep learning detection model, and the localization of individual red blood cells and pixel-level semantic segmentation of their intracytoplasmic target structures are performed simultaneously through the dual-task deep learning detection model. Based on the localization and pixel-level semantic segmentation, the existence of each individual red blood cell is determined and classified, the number of target red blood cells and the total number of red blood cells are counted, and the proportion of target red blood cells is determined by the proportion algorithm. Based on the percentage of target-shaped red blood cells, red blood cell parameters are fused to generate integrated morphological indicators, and a diagnostic report is output.

2. The method for identifying, counting, and determining the proportion of target-shaped red blood cells according to claim 1, characterized in that, Acquire high-resolution digital images of bone marrow or peripheral blood smears under an oil immersion lens, including: Construct a dedicated platform for acquiring oil immersion field-of-view images; Based on the dedicated oil immersion field-of-view image acquisition platform, the oil immersion objective focal length is calibrated, the illumination intensity is standardized, and the specimen smear quality is verified to determine the image acquisition adaptation parameters. Based on the image acquisition adaptation parameters, the original image is acquired through a high-magnification microscopic imaging device, and after impurity interference removal and imaging noise suppression processing, a high-resolution digital image that meets the analysis standards is generated.

3. The method for identifying, counting, and determining the proportion of target-shaped red blood cells according to claim 1, characterized in that, The dual-task deep learning detection model simultaneously performs pixel-level semantic segmentation of individual red blood cells and their intracytoplasmic target structures, including: Construct a dual-task deep learning detection model framework based on the fusion of object detection and semantic segmentation; Based on the dual-task deep learning detection model framework, the localization branch learns the morphological contour and boundary features of individual red blood cells through feature pyramid and adaptive anchor box matching, and outputs the bounding box coordinates, confidence level and individual identifier of red blood cells in real time. The segmentation branch focuses on the gray-scale distribution, morphological features and pixel differences of the target structure in the cytoplasm through multi-scale feature extraction and attention mechanism, and performs pixel-level semantic segmentation of the target structure simultaneously to generate an accurate target structure segmentation mask. Through joint training of the localization branch and the segmentation branch and optimization of the loss function, a synchronous analysis result containing individual red blood cell localization information and intracytoplasmic target structure segmentation information is output.

4. The method for identifying, counting, and determining the proportion of target-shaped red blood cells according to claim 1, characterized in that, The presence and classification of each individual red blood cell include: Construct a model for determining and classifying the existence of individual red blood cells; Based on the red blood cell individual existence determination and classification model, combined with the morphological contour, boundary features and pixel distribution patterns in the red blood cell location information, it is determined whether the area contains a valid red blood cell individual, and suspected areas such as non-cellular impurities, severely broken and excessively overlapping areas are excluded to obtain valid red blood cell individuals. Based on the morphological characteristics, grayscale differences, and spatial distribution of the target structure within the cytoplasm of the effective red blood cells, they are classified as target red blood cells or non-target red blood cells.

5. The method for identifying, counting, and determining the proportion of target-shaped red blood cells according to claim 1, characterized in that, The number of target-shaped red blood cells and the total number of red blood cells were counted, including: Construct a dedicated model for red blood cell count; Based on the red blood cell count model, valid red blood cell individuals after existence determination and classification are included, and the total number of red blood cells that meet the criteria within the field of view is counted to obtain the statistical classification results. The number of individuals marked as target red blood cells in the statistical classification results is synchronized to form a clear and corresponding statistical data on the total number of red blood cells and the number of target red blood cells.

6. The method for identifying, counting, and determining the proportion of target-shaped red blood cells according to claim 1, characterized in that, The formula for the dual-task deep learning detection model is as follows: ； ； ； ； ； ； ； ； ； ； in, For the first Layer fusion features; For upsampling operators; for Deep and strong semantic features; This is an element-wise summation operation; The first output for the backbone network High-resolution features in shallow layers; For intersection, union, and comparison; For anchor frame; The true bounding box; For intersection operations; This is a union operation; For classification loss; This represents the number of positive sample anchor frames; This represents the total number of anchor frames; For indicator functions; For anchor frame serial number; The set of positive samples; For anchor frame The probability of being predicted as a red blood cell; For focusing parameters; For bounding box regression loss; To predict bounding box parameters; These are the parameters of the actual bounding box; For smoothing L1 loss function; The difference between the predicted value and the actual value; To determine the total loss for the branch; These are the classification loss weight coefficients; These are the regression loss weight coefficients; This is an attention weight map; Use the Sigmoid activation function; This is a 1×1 convolution operation; This is a global average pooling operation; This is a feature map of the target core structure after multi-scale feature extraction; This is a global max pooling operation; The feature map is after attention weighting; For branching loss; The feature map height; The width of the feature map; For pixels The probability of predicting a target structure; For pixels Authentic labeling; It is the minimum value; For the first The probability value of each pixel being predicted as a bullseye structure; The total loss due to the combined efforts of both missions; To determine the weighting coefficients for the branch loss; The weighting coefficients for the branch loss are used.

7. A target-shaped red blood cell image recognition, counting, and proportion system, characterized in that, include: The image acquisition module is used to acquire high-resolution digital images of bone marrow or peripheral blood smears under an oil immersion lens. The dual-task deep learning detection module is used to input the high-resolution digital image into the dual-task deep learning detection model, and simultaneously perform pixel-level semantic segmentation of individual red blood cells and their intracytoplasmic target structures through the dual-task deep learning detection model. The counting and percentage analysis module is used to determine the existence and classify each individual red blood cell based on the location and pixel-level semantic segmentation, count the number of target red blood cells and the total number of red blood cells, and determine the percentage of target red blood cells through the percentage algorithm. The diagnostic report generation module is used to generate morphological integrated indicators based on the target red blood cell percentage, integrate red blood cell parameters, and output a diagnostic report.

8. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the program to implement the target red blood cell image recognition, counting, and proportion method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When a computer program or instruction is executed, it implements the target red blood cell image recognition, counting, and proportion method according to any one of claims 1-6.

10. A computer program product, comprising a computer program or instructions, characterized in that, When a computer program or instruction is executed, it implements the target red blood cell image recognition, counting, and proportion method according to any one of claims 1-6.

Images