Deep Learning-Based Blood Smear Whole-Slide Red Blood Cell Abnormality Morphology Assessment System
By integrating multi-level deep learning with multi-feature extraction algorithms, and combining the ResNet-50 model with cascaded feature selection, the problems of low efficiency in traditional manual microscopic examination and difficulty in accurately distinguishing subtle morphological differences by deep learning methods are solved, achieving efficient and accurate assessment and interpretable decision-making for abnormal morphology of red blood cells.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HANGZHOU FIRST PEOPLES HOSPITAL
- Filing Date
- 2026-01-20
- Publication Date
- 2026-05-26
AI Technical Summary
Traditional manual microscopic examination is inefficient and highly subjective. Existing deep learning methods struggle to accurately distinguish subtle local morphological differences and lack transparent interpretation, resulting in low accuracy in assessing abnormal red blood cell morphology and frequent missed or false detections.
This paper adopts a multi-level deep learning and multi-feature extraction algorithm, uses the ResNet-50 model and cascaded feature selection multi-feature extraction units for dual classification, and combines morphology, staining features and image quality for arbitration judgment to construct an interpretable decision evidence chain.
It improves the accuracy and efficiency of identifying abnormal red blood cell morphology, constructs an interpretable chain of decision-making evidence, and achieves accurate assessment at the whole slide scale.
Smart Images
Figure CN121544632B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical diagnostic technology, specifically relating to a deep learning-based system for evaluating abnormal morphology of red blood cells in a whole blood smear. Background Technology
[0002] In the diagnosis of hematological diseases, abnormal changes in red blood cell morphology are one of the important diagnostic criteria. Traditional manual microscopic examination suffers from problems such as low efficiency, high subjectivity, and high rate of missed diagnoses, which limits the development of cell morphology.
[0003] In recent years, deep learning-based technologies have been gradually applied to blood cell morphology analysis, but they still face technical bottlenecks. First, existing deep learning methods (such as ResNet and MobileNet) rely on end-to-end classification. Their advantage lies in distinguishing global features, such as overall color distribution and the presence of cell nuclei, but subtle local morphological differences are easily overlooked. For example, bite marks and vesicular cells, which are morphologically similar to mature red blood cells, are often difficult to distinguish accurately. Second, deep learning models suffer from "black box decision-making," lacking transparent explanations of the decision-making basis and failing to provide morphological quantitative evidence that meets clinical pathology standards, leading to reduced trust in AI results among physicians. Furthermore, quantifiable feature engineering and deep learning are often applied in isolation, only achieving a "1:1" cell morphology capture, making accurate cell morphology identification difficult. In addition, the large number of red blood cells in a full-slide blood smear makes it easy to miss or misdetect abnormal red blood cells, hindering effective full-data evaluation.
[0004] Therefore, this invention proposes a technique that integrates multi-level deep learning and multi-feature extraction algorithms to improve recognition accuracy while constructing an interpretable chain of decision evidence, thereby achieving accurate assessment of abnormal red blood cell classification at the whole slide scale. Summary of the Invention
[0005] To address the shortcomings of the existing technologies, this invention provides a deep learning-based system for evaluating abnormal morphology of red blood cells in a whole blood smear.
[0006] To solve the above problems, the present invention provides the following technical solution:
[0007] A deep learning-based system for evaluating abnormal morphology of red blood cells in whole blood smears includes a dual classification module, a fitting and comparison module, and an evaluation output module, wherein:
[0008] The dual classification module uses a trained deep convolutional neural network in conjunction with a multi-feature extraction unit based on cascaded feature filtering to perform dual classification on each red blood cell in the input blood smear whole slide image, and obtains the dual classification result;
[0009] The fitting and comparison module is used to fit and compare the dual classification results. If the dual classification results are consistent, it is the final classification. If the dual classification results are inconsistent, the confidence of the deep convolutional neural network classification result is calculated by combining morphological features, staining features, classification features and image quality. Arbitration judgment is carried out based on the preset confidence threshold, and the final red blood cell abnormality morphology classification result is output.
[0010] The evaluation output module is used to perform statistical analysis on the final classification of red blood cells in the entire blood smear and output an evaluation report on abnormal red blood cells in the entire smear.
[0011] In the above technical solution, the deep convolutional neural network adopts the ResNet-50 model and is trained and validated based on an abnormal red blood cell database;
[0012] The abnormal red blood cell database is constructed as follows: acquire full-slide images of blood smears, covering 11 types of red blood cells including: macrocytes, microcytes, nucleated red blood cells, polychromatic red blood cells, serrated red blood cells, bite marks, vesicular cells, basophilic stippling, hemoglobin crystal red blood cells, sickle cells, and irregular contracted red blood cells; classify and label each red blood cell obtained, and divide it into training and validation sets;
[0013] The network model is trained on the training set, using the cross-entropy loss function and the Adam optimizer. The model is iterated until it converges, and then evaluated on the validation set. The model with the highest accuracy is saved to obtain a trained deep convolutional neural network.
[0014] Furthermore, the multi-feature extraction unit based on cascaded feature screening is used to extract seven types of cascaded features of each red blood cell, including: color feature factor FC, roundness C, circle area S, shape factor Y, boundary roughness B, gray-scale distribution uniformity H, and central moment Z.
[0015] Furthermore, the color feature factor FC specifically refers to:
[0016] ,
[0017] Among them, R i G i Bi represents the red, green, and blue channel values of pixel i, and N represents the total number of pixels in the red cell region to be evaluated.
[0018] The specific value of roundness C is:
[0019] ,
[0020] Where P is the perimeter of the red blood cell region to be evaluated, and A is the area enclosed by the perimeter.
[0021] The area S of the circle is specifically:
[0022] ,
[0023] Where n is the number of edge points in the red blood cell region to be evaluated; the subscript i is the edge point identifier, and x i y i Let (x, y) be the coordinates of point i.
[0024] The shape factor Y is specifically:
[0025] ;
[0026] Boundary roughness B is specifically:
[0027] ,
[0028] Where, d i It is the difference between the straight-line distance from the i-th edge point to the center of the best-fit circle of the red blood cell to be evaluated and the radius of the fitted circle;
[0029] The grayscale distribution uniformity H is specifically as follows:
[0030] ,
[0031] Where L is the number of gray levels after quantization. Represents the normalized gray-level co-occurrence matrix;
[0032] The central moment Z is specifically:
[0033] ,
[0034] Where f(x,y) is a function describing whether the shape distribution of the object is uniform. , and q are the mean values of the distribution in the x and y directions, respectively, and p and q are the orders.
[0035] Furthermore, the method for classifying red blood cells in an input blood smear slide image using a multi-feature extraction unit based on cascaded feature filtering is as follows:
[0036] (1) If FC∈(0,100], then the output is polychromatic erythrocytes;
[0037] (2) If FC∈(100,200] ∩S∈(0,25], then the output is small red cells;
[0038] (3) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0,0.5] ∩H∈(0.5,0.7], then the output is irregularly contracted red blood cells;
[0039] (4) If FC∈(100,200]∩S∈(25,60] ∩C∈(0,0.5] ∩H∈(0.7,0.9] ∩Z∈(15,18], then the output is a serrated red blood cell;
[0040] (5) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0,0.5] ∩H∈(0.7,0.9] ∩Z∈(18,+∞], then the output is sickle cell;
[0041] (6) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0.5,0.7], then the output is bite mark cells;
[0042] (7) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0.7,1.0] ∩B∈(0,0.5], then the output is a basophilic dot pattern;
[0043] (8) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0.7,1.0] ∩B∈(0.5,1], then the output is a vesicular cell;
[0044] (9) If FC∈(100,200] ∩S∈(60,+∞], then the output is a large red blood cell;
[0045] (10) If FC∈(200,+∞] ∩Y∈(1.3,1.8], then the output is nucleated red blood cells;
[0046] (11) If FC∈(200,+∞] ∩Y∈(2.5,+∞], then the output is hemoglobin crystal red blood cells.
[0047] Furthermore, the determination of the confidence level (CF) based on morphological, staining, and classification features, combined with image quality, specifically involves:
[0048]
[0049] Among them, F morph F is the morphological confidence factor, taking values of (0, 1). color D is the confidence factor for staining characteristics, with values ranging from (0, 1). quality F represents the image quality degradation coefficient, with values ranging from (0.8 to 1.0), all generated by the ResNet-50 model; model α represents the probability value of the corresponding classification label output by the Softmax layer of the ResNet-50 model, with a value range of (0,1); α, β, and γ are weight coefficients used to balance F. morph F color Fmodel The contribution of the confidence level CF satisfies α≥0, β≥0, γ≥0 and α+β+γ=1.
[0050] Furthermore, if the results of the double classification are inconsistent, arbitration is conducted based on the confidence level (CF) and the confidence threshold.
[0051] When CF > 90%, the classification result of the deep convolutional neural network is used as the final classification.
[0052] When CF ≤ 90%, the classification result of the multi-feature extraction unit is used as the final classification.
[0053] An electronic device, comprising:
[0054] One or more processors;
[0055] Memory, used to store one or more programs;
[0056] When the one or more programs are executed by the one or more processors, the one or more processors perform the functions of the system as described in any of the preceding claims.
[0057] A computer-readable storage medium storing computer-executable instructions, which, when executed, are used to perform the functions of the system as described in any of the preceding claims.
[0058] The beneficial effects of the present invention are at least as follows:
[0059] This invention integrates multi-layered deep learning technology with multi-feature extraction algorithms. Deep learning addresses the poor generalization ability of multi-feature extraction algorithms, while multi-feature extraction algorithms overcome the limitations of "black box decision-making" in artificial intelligence, which leads to unclear rules and limited accuracy. Furthermore, the cascaded feature selection technique employed in the integrated multi-feature extraction algorithm technology significantly optimizes the detection process compared to parallel feature selection, doubling the detection efficiency and saving computational resources. The system provided by this invention improves recognition accuracy while constructing an interpretable chain of decision-making evidence, enabling precise assessment of abnormal red blood cell classification at the whole slide scale. Attached Figure Description
[0060] The present invention will be further described below with reference to the accompanying drawings and embodiments:
[0061] Figure 1 This is a block diagram of the deep learning-based blood smear whole-slide red blood cell abnormality assessment system of the present invention;
[0062] Figure 2 This is a flowchart of the deep learning-based method for evaluating abnormal morphology of red blood cells in a whole blood smear, as described in this invention.
[0063] Figure 3 This is a logic diagram of a multi-feature extraction algorithm for classifying red blood cells. Detailed Implementation
[0064] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention will be further described below with reference to the accompanying drawings and embodiments:
[0065] According to a specific embodiment of the present invention, a deep learning-based system for evaluating abnormal morphology of red blood cells in a whole blood smear is provided, such as... Figure 1 As shown, it includes a dual classification module, a fitting and comparison module, and an evaluation output module, wherein:
[0066] The dual classification module uses a trained deep convolutional neural network in conjunction with a multi-feature extraction unit (with built-in multi-feature extraction algorithm) based on cascaded feature filtering to perform dual classification on each red blood cell in the input blood smear whole slide image, and obtains the dual classification result;
[0067] The fitting and comparison module is used to fit and compare the dual classification results. If the dual classification results are consistent, it is the final classification. If the dual classification results are inconsistent, the confidence level of the classification result of the deep convolutional neural network is determined based on morphology, staining features, and classification features, combined with image quality. Arbitration is performed based on the confidence level threshold to obtain the final classification.
[0068] The evaluation output module is used to perform statistical analysis on the final classification of red blood cells in the entire blood smear and output an evaluation report on abnormal red blood cells in the entire smear.
[0069] like Figure 2 The diagram shown is a flowchart of the method upon which the above system is based, including the following steps:
[0070] S1. Obtain whole-slide images of blood smears covering 11 types of abnormal red blood cell morphologies, and have them labeled by a professional physician. Divide the image set into a training set, a validation set, and a test set in a 7:2:1 ratio to construct an abnormal red blood cell database. The 11 types of abnormal red blood cell morphologies include: macrocytes, microcytes, nucleated red blood cells, polychromatic erythrocytes, serrated red blood cells, bite marks, vesicular cells, basophilic stippling, hemoglobin crystallized red blood cells, sickle cells, and irregularly contracted red blood cells.
[0071] S2. The ResNet-50 model is used as the multi-class classification model, and the training set obtained in S1 is used to train the model. During training, the cross-entropy loss function and Adam optimizer are used, and the model is converged through multiple rounds of iteration. The model is evaluated by inputting the validation set, and the model with the highest accuracy is saved. The test set of red blood cell images is input into the trained ResNet-50 model, and the classification label and confidence score of each sample are output. At the same time, the multi-feature extraction algorithm extracts seven core features of red blood cells (color feature factor (FC), roundness (C), circle area (S), shape factor (Y), boundary roughness (B), gray-level distribution uniformity (H), and central moment (Z)), and judges the category based on preset logic, as follows:
[0072] The characteristic factors include:
[0073] The color feature factor (FC) is calculated by function (1): (1)
[0074] Among them, R i G i Bi represents the red, green, and blue channel values of pixel i. N represents the total number of pixels in the region.
[0075] Circularity (C) is calculated by function (2): (2)
[0076] Where P is the perimeter of the figure and A is the area enclosed by the perimeter.
[0077] The area (S) of the circle is calculated using function (3): (3)
[0078] Where n is the number of edge points; x i y i Let (x, y) be the coordinates of point i; the subscript indicates the edge point.
[0079] The shape factor (Y) is calculated using functions (3) and (4): (4)
[0080] Where n is the number of edge points; x i y i Let (x, y) be the coordinates of point i; the subscript indicates the edge point; S is the area of the red blood cell; π is the value of pi.
[0081] Boundary roughness (B) is calculated by function (5): (5)
[0082] Where n is the total number of red blood cell edge points, d iIt is the difference between the straight-line distance from the i-th boundary point to the center of the best-fit circle of the red blood cell to be evaluated and the radius of the fitted circle.
[0083] The gray-scale distribution uniformity (H) is calculated by function (6): (6)
[0084] Where L is the number of gray levels after quantization. This represents the normalized gray-level co-occurrence matrix.
[0085] The central moment (Z) is calculated from the function (7): (7)
[0086] Here, f(x,y) is a function describing whether the shape distribution of the object is uniform. In this example, f(x,y) is the distribution density function of red blood cells on the two-dimensional plane of the blood smear. , Let p and q be the mean values of the distribution in the x and y directions, respectively, and p and q be the orders, both of which are non-negative integers (0, 1, 2, 3, ...). Together they form the p+q order central moments, which are used here to determine whether the morphology of red blood cells is uniform and how dispersed they are. Orders 0 to 2 are sufficient. In this example, the second order in the x direction is used, with p=2 and q=0. The higher the order, the more complex the morphological features described.
[0087] S3. Based on the seven extracted features, such as Figure 3 As shown, the red blood cell categories are determined sequentially according to the following logic:
[0088] (1) If FC∈(0,100], then the output is polychromatic erythrocytes;
[0089] (2) If FC∈(100,200] ∩S∈(0,25], then the output is small red blood cells; “∩” means that the corresponding red blood cell category can only be output if all the given conditions are met.
[0090] (3) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0,0.5] ∩H∈(0.5,0.7], then the output is irregularly contracted red blood cells;
[0091] (4) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0,0.5] ∩H∈(0.7,0.9] ∩Z∈(15,18], then the output is a serrated red blood cell;
[0092] (5) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0,0.5] ∩H∈(0.7,0.9] ∩Z∈(18,+∞], then the output is sickle cell;
[0093] (6) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0.5,0.7], then the output is bite mark cells;
[0094] (7) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0.7,1.0] ∩B∈(0,0.5], then the output is a basophilic dot pattern;
[0095] (8) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0.7,1.0] ∩B(∈0.5,1]), then the output is a vesicular cell;
[0096] (9) If FC∈(100,200] ∩S∈(60,+∞], then the output is a large red blood cell;
[0097] (10) If FC∈(200,+∞] ∩Y∈(1.3,1.8], then the output is nucleated red blood cells;
[0098] (11) If FC∈(200,+∞] ∩Y∈(2.5,+∞], then the output is hemoglobin crystal red blood cells.
[0099] S4. Fit and compare the classification results of the ResNet-50 model and the multi-feature extraction algorithm, and arbitrate according to the following rules: If the initial classification and the secondary classification results are consistent, output the classification directly; if the two results are inconsistent, calculate the confidence score (CF) of the ResNet-50 model and arbitrate based on the threshold: (a) when the initial classification confidence score is > 90%, adopt the classification result of the ResNet-50 model; (b) when the initial classification confidence score is ≤ 90%, adopt the classification result of the multi-feature extraction algorithm.
[0100] The confidence score (CF) reflects the reliability of the classification results of the ResNet-50 model and is calculated by function (8):
[0101] (8)
[0102] Among them, F morph For the morphological confidence factor (with values of (0, 1]), F color D is the confidence factor for staining characteristics (with values of (0, 1)); quality The image quality attenuation coefficients (with values of (0.8, 1.0)) are all generated by the trained ResNet-50 model. During model training, the training set is manually labeled, and expert annotations are obtained through a comprehensive evaluation of image quality indicators such as the morphological features of the target red blood cell region, the feature detection and evaluation of red blood cell staining intensity, and the image quality indicators of the target analysis region. modelα represents the probability value of the corresponding classification label output by the Softmax layer of the ResNet-50 model, with a value range of (0,1); α, β, and γ are weight coefficients used to balance F. morph F color F model The contribution of the confidence level CF satisfies α≥0, β≥0, γ≥0 and α+β+γ=1.
[0103] Statistical analysis of the red blood cell classification results of the entire blood smear was performed to generate an evaluation report, which included:
[0104] Abnormal Red Blood Cell Types: Lists the types present in the 11 types of red blood cells detected;
[0105] Quantity and proportion: The count (number) of each type of abnormal red blood cell and its proportion (%) of the total red blood cells;
[0106] Image display: Typical images of each type of abnormal red blood cell (with bounding boxes and categories marked) to visually demonstrate morphological characteristics.
[0107] Example:
[0108] 10,000 clinical blood smear images were collected, covering 11 types of abnormal red blood cell morphology. The sample numbers for each type were as follows: 950 macrocytes, 980 microcytes, 920 nucleated red blood cells, 900 polychromatic erythrocytes, 960 serrated erythrocytes, 930 bite marks, 910 vesicular cells, 940 basophilic stippling cells, 970 hemoglobin crystallized red blood cells, 990 sickle cells, and 1550 irregularly contracted red blood cells. 1300 normal red blood cell samples were also included. All blood smears were stained using the Wright-Giemsa staining method for 15 minutes with a buffer pH of 6.4-6.8 to ensure staining quality met clinical testing standards. A labeling team of three physicians with over 10 years of experience in hematology and pathology diagnosis meticulously reviewed and categorized the red blood cells in the images. The annotations included: red blood cell bounding boxes (coordinates accurate to pixels) and category labels (11 categories of abnormal or normal morphologies). After annotation, a cross-validation mechanism was used: if the annotation results from three physicians were consistent, they were directly used as the "standard result"; if there were discrepancies, a fourth senior physician (with over 20 years of experience) reviewed and determined the final annotation result, ensuring an accuracy rate of ≥99.5%. The 10,000 annotated images were randomly divided into training, validation, and test sets in a 7:2:1 ratio, with 7,000 images in the training set (including samples of various abnormal morphologies and normal red blood cells, in the same proportion as the total sample), 2,000 images in the validation set, and 1,000 images in the test set. Stratified sampling was used during the dataset partitioning process to avoid the absence of any particular type of sample in a single set, ensuring the representativeness of the dataset.
[0109] A ResNet-50 model was used as the multi-class classification model, trained with a training set of 7000 examples. During training, the cross-entropy loss function and Adam optimizer were used, and the model converged through multiple iterations. A validation set of 2000 examples was used to evaluate the model, and the model with the highest accuracy was saved. 1000 blood smear images from the test set were input into the trained ResNet-50 model to obtain the initial classification result (model classification label) for each red blood cell. Simultaneously, the ROI of the same red blood cell was input into a multi-feature extraction unit to obtain the binary classification result (feature classification label). Specifically, the U-Net semantic segmentation model was used to segment the red blood cell region of the blood smear slide image, obtaining the mask of each red blood cell. Then, the ROI region of each red blood cell was extracted by multiplying the mask by the pixels of the original image (excluding background and other cell interference). Specific calculation of 7 types of features (1) Color feature factor Fc calculation example: A red blood cell ROI contains N=1200 pixels, and the sum of Ri+Gi+Bi of all pixels is 98400, then Fc=98400 / 1200=82, which meets the classification rule (1) FC∈(0,100], and is initially determined to be a polychromatic red blood cell; (2) Circular area S calculation example: A red blood cell edge point n=120, some coordinates are (50,50), (51,52), ..., (50,50), and the S is calculated to be 42.3 pixels when substituted into the formula. 2 (3) Example of roundness C calculation: The perimeter of the above red blood cell is P = 23.5 pixels, and the area is A = 42.3 pixels. 2 Therefore, C = 4 × 3.1416 × 42.3 / (23.5) 2 ≈1.90. (4) Example of shape factor Y calculation: For the above red blood cell C≈1.90, then Y=1 / 1.90≈0.526, reflecting the degree to which the shape of the red blood cell deviates from a circle. (5) Calculation of boundary roughness B: Example: For a certain vesicle cell edge point n=150, the fitted circle radius R=8.5 pixels, all di 2 The sum of the values is 37.5, so B = 37.5 / 150 = 0.25, which meets the classification rule (7) B∈(0,0.5). (6) Example of calculating the uniformity of gray distribution H: In the normalized gray co-occurrence matrix of a certain basophilic stippled red blood cell, most elements are concentrated near the main diagonal. The value is 0.89, i.e. H=0.89. (7) Selection of the order of the central moment Z: In this embodiment, p=2 and q=0 (2nd order central moment in the x direction) are selected to reflect the dispersion of the gray distribution in the x direction; Calculation example: For a certain serrated red blood cell, x̄=112 and ȳ=97, then Z=16.8, which meets the classification rule (4) Z∈(15,18).
[0110] The initial and secondary classification results are fitted and compared. A certain red blood cell model classifies it as "sickle cell," and the feature classification is "serrated red blood cell." The confidence score (CF) is then calculated, determined through cross-validation on the validation set and grid search. The optimal combination is α=0.3, β=0.2, and γ=0.5. F morph=0.85、 F color=0.92、 F model=0.88 D quality =0.95, then CF=(0.3×0.85+0.2×0.92+0.5×0.88)×0.95=(0.255+0.184+0.44)×0.95=0.879×0.95≈0.835 (83.5%). The final feature classification result "serrated red blood cells" was adopted, and the classification was verified to be correct by comparing it with the "standard result".
[0111] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0112] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0113] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0114] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0115] The embodiments described above are merely some preferred embodiments of the present invention, and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.
Claims
1. A deep learning-based system for evaluating abnormal morphology of red blood cells in a whole blood smear, characterized in that, It includes a dual classification module, a fitting and comparison module, and an evaluation output module, among which: The dual classification module uses a trained deep convolutional neural network in conjunction with a multi-feature extraction unit based on cascaded feature filtering to perform dual classification on each red blood cell in the input blood smear whole slide image, and obtains the dual classification result; The fitting and comparison module is used to fit and compare the dual classification results. If the dual classification results are consistent, it is the final classification. If the dual classification results are inconsistent, the confidence of the deep convolutional neural network classification result is calculated by combining morphological features, staining features, classification features and image quality. Arbitration judgment is carried out based on the preset confidence threshold, and the final red blood cell abnormality morphology classification result is output. The evaluation output module is used to perform statistical analysis on the final classification of red blood cells in the entire blood smear and output an evaluation report on abnormal red blood cells in the entire smear. The multi-feature extraction unit based on cascaded feature screening is used to extract seven types of cascaded features of each red blood cell, including: color feature factor FC, roundness C, circle area S, shape factor Y, boundary roughness B, gray level distribution uniformity H, and central moment Z. The method for classifying red blood cells in an input blood smear slide image using a multi-feature extraction unit based on cascaded feature filtering is as follows: (1) If FC∈(0,100], then the output is polychromatic erythrocytes; (2) If FC∈(100,200] ∩S∈(0,25], then the output is small red cells; (3) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0,0.5] ∩H∈(0.5,0.7], then the output is irregularly contracted red blood cells; (4) If FC∈(100,200]∩S∈(25,60] ∩C∈(0,0.5] ∩H∈(0.7,0.9] ∩Z∈(15,18], then the output is a serrated red blood cell; (5) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0,0.5] ∩H∈(0.7,0.9] ∩Z∈(18,+∞], then the output is sickle cell; (6) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0.5,0.7], then the output is bite mark cells; (7) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0.7,1.0] ∩B∈(0,0.5], then the output is a basophilic dot pattern; (8) If FC∈(100,200] ∩S∈(25,60] ∩C∈(0.7,1.0] ∩B∈(0.5,1], then the output is a vesicular cell; (9) If FC∈(100,200] ∩S∈(60,+∞], then the output is a large red blood cell; (10) If FC∈(200,+∞] ∩Y∈(1.3,1.8], then the output is nucleated red blood cells; (11) If FC∈(200,+∞] ∩Y∈(2.5,+∞], then the output is hemoglobin crystal red blood cells.
2. The deep learning-based blood smear whole-slide erythrocyte abnormality assessment system according to claim 1, characterized in that, The deep convolutional neural network uses the ResNet-50 model and is trained and validated based on an abnormal red blood cell database. The abnormal red blood cell database is constructed as follows: acquire full-slide images of blood smears, covering 11 types of red blood cells including: macrocytes, microcytes, nucleated red blood cells, polychromatic red blood cells, serrated red blood cells, bite marks, vesicular cells, basophilic stippling, hemoglobin crystal red blood cells, sickle cells, and irregular contracted red blood cells; classify and label each red blood cell obtained, and divide it into training and validation sets; The network model is trained on the training set, using the cross-entropy loss function and the Adam optimizer. The model is iterated until it converges, and then evaluated on the validation set. The model with the highest accuracy is saved to obtain a trained deep convolutional neural network.
3. The deep learning-based blood smear whole-slice red blood cell abnormality assessment system according to claim 1, characterized in that, The color feature factor FC is specifically: ， Among them, R i G i Bi represents the red, green, and blue channel values of pixel i, and N represents the total number of pixels in the red cell region to be evaluated. The specific value of roundness C is: ， Where P is the perimeter of the red blood cell region to be evaluated, and A is the area enclosed by the perimeter. The area S of the circle is specifically: ， Where n is the number of edge points in the red blood cell region to be evaluated; the subscript i is the edge point identifier, and x i y i Let (x, y) be the coordinates of point i. The shape factor Y is specifically: ； Boundary roughness B is specifically: ， Where, d i It is the difference between the straight-line distance from the i-th edge point to the center of the best-fit circle of the red blood cell to be evaluated and the radius of the fitted circle; The grayscale distribution uniformity H is specifically as follows: ， Where L is the number of gray levels after quantization. Represents the normalized gray-level co-occurrence matrix; The central moment Z is specifically: ， Where f(x,y) is a function describing whether the shape distribution of the object is uniform. , and q are the mean values of the distribution in the x and y directions, respectively, and p and q are the orders.
4. The deep learning-based blood smear whole-slide erythrocyte abnormality assessment system according to claim 2, characterized in that, The confidence level (CF) is determined based on morphological, staining, and classification features, combined with image quality, as follows: Among them, F morph F is the morphological confidence factor, taking values of (0, 1). color D is the confidence factor for staining characteristics, with values ranging from (0, 1). quality F represents the image quality degradation coefficient, with values ranging from (0.8 to 1.0), all generated by the ResNet-50 model; model α represents the probability value of the corresponding classification label output by the Softmax layer of the ResNet-50 model, with a value range of (0,1); α, β, and γ are weight coefficients used to balance F. morph F color F model The contribution of the confidence level CF satisfies α≥0, β≥0, γ≥0 and α+β+γ=1.
5. The deep learning-based blood smear whole-slice red blood cell abnormality assessment system according to claim 4, characterized in that, If the results of the double classification are inconsistent, arbitration will be conducted based on the confidence level (CF) and the confidence threshold. When CF > 90%, the classification result of the deep convolutional neural network is used as the final classification. When CF ≤ 90%, the classification result of the multi-feature extraction unit is used as the final classification.
6. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the functions of the system as described in any one of claims 1-5.
7. A computer-readable storage medium storing computer-executable instructions, which, when executed, perform the functions of the system according to any one of claims 1-5.