Blood cell morphology AI quality control system based on transfer learning

Abstract

The invention relates to the technical field of medical image analysis, in particular to a blood cell morphology AI quality control system based on transfer learning. Comprising an image acquisition module for receiving a full-slide image from a digital slide scanner; the quality control module is used for carrying out multi-parameter quality analysis on a single-layer area in the full-slide image and outputting a composite quality score; the diagnosis and classification module is used for diagnosing the full-slide image according to the composite mass score, excluding unqualified smears and classifying the passing smears; the data management module is used for storing a quality analysis result and a quality analysis score of the full-slide image; and the user interface module is used for displaying the diagnosis result and the classification result of the diagnosis classification module through a user interface. According to the method, automatic image quality evaluation is introduced before diagnostic analysis, so that the problem that the performance of an AI classifier is affected by the quality of an original image in the prior art is solved, and the reliability and accuracy of automatic blood morphological analysis are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image analysis, in particular to a blood cell morphology AI quality control system based on transfer learning. BACKGROUND

[0002] Manual microscopic review of peripheral blood smears is the cornerstone of hematology diagnosis, used to confirm results from automated blood cell counters and provide critical cell morphological details. However, the manual microscopic review process is time-consuming, labor-intensive, and highly subjective, leading to significant variability in results between observers. Its accuracy depends largely on the experience and training level of the pathologist, which makes results difficult to standardize.

[0003] The advent of automated blood cell analyzers and digital morphological analysis systems has changed laboratory workflow by improving speed, efficiency, and throughput. These systems are capable of performing cell counting and pre-classification automatically using artificial intelligence (AI). However, the performance of these advanced systems heavily depends on the quality of their input, i.e., physical blood smears and their subsequent digitized images. Poor smear preparation (e.g., uneven distribution, improper thickness) and inconsistent staining are major sources of error. Existing DM analyzers are very sensitive to these variations, which can lead to false classification and unreliable results. For example, staining precipitates can be mistaken for bacteria or platelets, while water artifacts can be mistaken for blood parasites. This indicates that it is not sufficient to simply apply a powerful AI classifier; the quality of the input data must first be ensured.

[0004] Existing hematology quality control schemes focus mainly on the analysis system itself, including the analyzer hardware, reagents, and calibrators. These schemes aim to ensure the proper functioning of the machine. External quality control or proficiency testing evaluates long-term overall performance by comparing laboratory results with peer groups. However, there is a lack of an automated quality control scheme for digital blood smear images themselves, on a sample-by-sample basis. SUMMARY

[0005] The present application aims to overcome the shortcomings of the prior art and provide an automated, objective, and reliable blood smear image quality control method and system, thereby improving the accuracy of subsequent AI-based morphological analysis.

[0006] The blood cell morphology AI quality control system based on transfer learning comprises an image acquisition module, a quality control module, a diagnostic classification module, a data management module, and a user interface module.

[0007] The image acquisition module is connected with a digital slide scanner and is used to receive whole slide images from the digital slide scanner.

[0008] The quality control module is configured to perform multi-parameter quality analysis on the single-layer region in the whole slide image.

[0009] The quality control module comprises an artifact analysis unit, a staining quality analysis unit, and a cell distribution uniformity analysis unit, which are respectively configured to perform artifact analysis, staining quality analysis, and cell distribution uniformity analysis on the single-layer region in the whole slide image.

[0010] The diagnostic classification module is configured to perform diagnosis on the whole slide image according to the multi-parameter quality analysis results of the single-layer region in the whole slide image by the quality control module, to exclude unqualified smears and to classify the passed smears to determine the blood cell categories corresponding to the smears.

[0011] The data management module is configured to store the output results of the quality control module and the diagnostic classification module, and to connect with a laboratory information system to synchronize the data with the laboratory information system.

[0012] The user interface module is configured to display the data stored by the data management module through a user interface.

[0013] Preferably, the artifact analysis is configured to detect and quantify physical artifacts in the single-layer region in the whole slide image, wherein the physical artifacts include tissue fold, bubble, dust, scratch, ink, or fibrin filament.

[0014] Preferably, the artifact analysis employs a pre-trained image segmentation neural network model to perform pixel-level segmentation on the artifacts in the single-layer region in the whole slide image, and determines the percentage of the area occupied by the artifacts as the result of the artifact analysis.

[0015] Preferably, the staining quality analysis specifically comprises:

[0016] Converting the whole slide image from an RGB color space to a color space capable of separating luminance and chrominance information, such as HSD or CIELAB;

[0017] Identifying pixel groups corresponding to red blood cell cytoplasm and white blood cell nuclei by color thresholding;

[0018] Calculating and comparing the color histogram statistics such as mean, standard deviation, and peak position of these pixel groups with a pre-set gold standard profile established by a group of expert-verified best staining smears, to quantitatively evaluate staining intensity, color balance of eosin and hematoxylin, and the presence of staining precipitate.

[0019] Preferably, the cell distribution uniformity analysis specifically comprises:

[0020] Locating the centroids of red blood cells in the single-layer region in the whole slide image by an object detection algorithm;

[0021] dividing the single-layer region in the whole slide image into a uniform two-dimensional grid;

[0022] counting the number of cell centers within each grid cell;

[0023] calculating the coefficient of variation of the cell counts across all grid cells as a quantitative measure of uniformity, where a lower coefficient of variation value indicates higher uniformity.

[0024] Preferably, the output of the quality control module is a composite quality score, which is calculated by a weighted scoring model, specifically including:

[0025] normalizing the analysis results of each parameter, such as the artifact area percentage, staining deviation, and distribution coefficient of variation, to a common numerical scale, for example, 0 to 1, using the min-max scaling method;

[0026] assigning each normalized result a predetermined weight according to its expected clinical impact on the accuracy of the final diagnosis;

[0027] calculating the weighted sum of all weighted results to obtain the composite quality score;

[0028] The formula for calculating the composite quality score is:

[0029]

[0030] where CQI is the composite quality score, w i is the weight of the normalized result of the i-th parameter, S i is the score of the i-th parameter after normalization.

[0031] Preferably, the step of the diagnostic classification module performing diagnostic classification on the whole slide image includes:

[0032] comparing the composite quality score output by the quality control module with pre-set rejection threshold and review threshold;

[0033] According to the comparison result, the following operations are performed: if the composite quality score is lower than the rejection threshold, the analysis is aborted and a quality failure is reported to the laboratory information system; if the composite quality score is between the rejection threshold and the review threshold, the classification is continued but a manual review is required for the result; if the composite quality score is higher than the review threshold, the result is accepted as automatically reviewed and passed;

[0034] The whole slide image that passes the review is classified using a deep learning model obtained through transfer learning, and the corresponding confidence score is output.

[0035] Preferably, the training method of the deep learning model is:

[0036] Selecting a convolutional neural network pre-trained on a large non-medical image dataset as a base model;

[0037] Modifying the architecture of the base model, removing its original top classification layer and replacing it with a new set of fully connected layers tailored for the blood cell classification task;

[0038] Fine-tuning the modified model, i.e. retraining the newly added layers on a large, high-quality and expert-annotated blood cell image dataset, so that the model is specialized for blood cell classification.

[0039] Preferably, the diagnostic classification module also uses the composite quality score as a modulation factor when performing classification, adjusting the confidence score output by the deep learning model for each cell classification;

[0040] The formula for adjusting the confidence score is:

[0041] C' = C x f(CQI);

[0042] Where C is the confidence score output by the classification model, C' is the adjusted confidence score, and f(CQI) is a monotonically increasing function of the composite quality score CQI.

[0043] The blood cell morphology AI quality control method based on transfer learning includes the following steps:

[0044] Obtaining a digitized whole slide image of a blood smear from a digital slide scanner;

[0045] Performing multi-parameter quality analysis on the single layer area in the whole slide image, which includes:

[0046] Artifact analysis for detecting and quantifying physical artifacts within the single layer area;

[0047] Staining quality analysis for evaluating the staining characteristics of cells within the single layer area;

[0048] Cell distribution uniformity analysis for measuring the spatial distribution of cells within the single layer area;

[0049] Based on the results of the multi-parameter quality analysis, a composite quality index is generated;

[0050] Comparing the composite quality index with pre-set rejection threshold and review threshold, excluding unqualified smears, and using a deep learning model obtained through transfer learning to classify the passed smears to determine their corresponding blood cell categories.

[0051] Compared with the prior art, the application has the advantages that:

[0052] The quality control module of the application can systematically analyze and quantify key factors affecting diagnostic quality, including physical artifacts (such as air bubbles, folds), staining quality (such as intensity, balance, precipitation), and uniformity of cell distribution. By integrating these independent metrics into a single, easily understood composite quality index (CQI), an objective and repeatable quality score can be provided for each digital smear.

[0053] By automatically rejecting images of substandard quality before analysis begins, the application ensures that only high-quality, reliable data enters the downstream diagnostic module. By preventing classification errors due to image quality issues from the source, the reliability and accuracy of the final diagnostic result are significantly improved.

[0054] By using CQI to dynamically adjust the confidence score of the final diagnostic classification, the application provides a more transparent, interpretable and trustworthy diagnostic result for clinicians, enabling them to combine the objective score of image quality to judge the reliability of the diagnostic conclusion, thereby enhancing the credibility of the system. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 System architecture diagram of the blood cell morphology AI quality control system based on transfer learning provided by the application. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solutions and advantages of the application clearer, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application.

[0057] Example 1:

[0058] Referring to Figure 1 , the embodiment of the application provides a blood cell morphology AI quality control system based on transfer learning. The system includes an image acquisition module, a quality control module, a diagnostic classification module, a data management module and a user interface module in terms of software;

[0059] The image acquisition module is connected with a digital slide scanner and is used to receive full slide images from the digital slide scanner;

[0060] The quality control module is used to perform multi-parameter quality analysis on the single layer area in the full slide image;

[0061] The quality control module includes an artifact analysis unit, a staining quality analysis unit, and a cell distribution uniformity analysis unit, which are respectively used for artifact analysis, staining quality analysis, and cell distribution uniformity analysis on the single-layer area in the whole slide image.

[0062] The diagnostic classification module performs diagnosis on the whole slide image according to the multi-parameter quality analysis result of the single-layer area in the whole slide image by the quality control module, excludes unqualified smears, and classifies the passed smears to determine the corresponding blood cell category;

[0063] The data management module is used for storing the output results of the quality control module and the diagnostic classification module, and is connected with a laboratory information system to synchronize the data with the laboratory information system;

[0064] The user interface module is used for displaying the data stored by the data management module through a user interface.

[0065] In terms of hardware, it includes a digital slide scanner for digitizing glass smears into high-resolution (e.g., 40x to 100x objective lens) whole slide images, and a processing server containing one or more central processing units (CPUs), large-capacity random access memories (RAMs), and at least one high-performance graphics processing unit (GPU) for accelerating deep learning calculations.

[0066] Embodiment 2:

[0067] The embodiment of the present application provides a blood cell morphology AI quality control method based on transfer learning, which specifically comprises:

[0068] S1, image acquisition and preprocessing:

[0069] The image acquisition module acquires the whole slide image from the digital slide scanner. Then, the system identifies the "single-layer area" in the smear, and the cell distribution in the area is moderate, neither too dense nor too sparse, which is an ideal area for morphology evaluation.

[0070] It should be noted that the identified single-layer area is divided into multiple smaller image blocks for parallel processing.

[0071] S2, multi-parameter quality feature extraction:

[0072] The quality control module performs multiple quality analyses on the image blocks of the single-layer area in parallel.

[0073] Artifact detection:

[0074] Model: A deep learning segmentation model is employed, preferably U-Net or its variants, such as DoubleUNet. The model is pre-trained on a large blood smear image dataset, which is expert-annotated and contains various common artifacts such as tissue folds, bubbles, dust, scratches, ink / pen marks, and large fibrin filaments.

[0075] Output: The model outputs a pixel-level segmentation mask for the entire single-layer region, accurately marking the position and shape of all artifacts.

[0076] Metric (Artifact Score): Calculates the percentage of the total area of ​​a single layer covered by detected artifacts. This percentage is the quantitative measure of the artifacts.

[0077] Staining quality assessment:

[0078] Color space conversion: Converting image blocks from the RGB color space to a color space that can separate luminance and chrominance information, such as HSD or CIELAB.

[0079] Component recognition: The system identifies pixels corresponding to red blood cell cytoplasm (pink / red) and white blood cell nucleus (purple / blue).

[0080] Statistical analysis: Calculate the color histograms and statistical characteristics (such as mean, standard deviation, peak position, and distribution width) of these components.

[0081] Metrics (Staining Score): The calculated statistics are compared to a predefined "gold standard" profile derived from a set of expert-validated smears representing the best staining results. The deviation between the actual image and the gold standard is quantified, generating multiple scores, including:

[0082] Staining intensity: Assess whether the staining is too light or too dark.

[0083] Color balance: Assess whether the balance of eosin and hematoxylin is correct.

[0084] Staining precipitate: Detects tiny, dark-colored, irregularly shaped objects that do not conform to cell morphology by using its unique color and texture characteristics. A fraction is generated based on the density of the precipitate.

[0085] Cell distribution uniformity assessment:

[0086] Cell localization: The system performs preliminary cell detection (e.g., using a simple thresholding method or a fast object detection model) to locate the centroid of red blood cells within a monolayer region.

[0087] Mesh generation: Divide a single layer into a uniform grid (e.g., a 10x10 grid).

[0088] Density computation: Count the number of cell centers within each grid cell.

[0089] Metric (Distribution CV): Compute the coefficient of variation of cell counts across all grid cells. A low CV value indicates an even smear, while a high CV value indicates cell clustering and blank areas.

[0090] S3, Composite Quality Index (CQI) generation:

[0091] Combine these heterogeneous metrics (artifact area percentage, staining bias score, distribution CV) into a composite quality index.

[0092] In this embodiment, a weighted scoring model is chosen, specifically including:

[0093] Normalization: Normalize each individual metric (e.g. artifact score, staining intensity score, distribution score) to a common scale (e.g. 0 to 1) using min-max scaling.

[0094] Weight assignment: Assign weights to each metric based on its degree of influence on clinical diagnosis through expert consultation. For example, a severe artifact might receive a higher weight than a mild distribution unevenness.

[0095] CQI computation: Compute the weighted sum of all normalized scores with their corresponding weights, as follows:

[0096]

[0097] where w i is the weight of the i-th parameter normalized result, S i is the normalized score of the i-th parameter.

[0098] In another preferred embodiment, a fuzzy logic inference system can also be used to compute the composite quality index, specifically including:

[0099] Fuzzification: Map each crisp input metric (e.g. artifact area of 8%) to a fuzzy set through a membership function. For example, an artifact area of 8% might be defined as 70% "low" and 30% "medium". A staining bias of 0.4 might be defined as 50% "good" and 50% "acceptable".

[0100] Fuzzy rule base: Define a set of human-readable rules, for example:

[0101] IF (artifact is high) OR (staining is poor) THEN (quality is unacceptable)

[0102] IF (artifact is low) AND (staining is good) AND (distribution is even) THEN (quality is excellent)

[0103] Inference and De-Fuzzification: Evaluate all rules in parallel using a fuzzy inference engine (e.g., Mamdani type). The resulting fuzzy output is converted back to a single crisp value (i.e., CQI, range 0-100) using a de-fuzzification method (e.g., Centroid method).

[0104] S4. CQI-based gating and result modulation:

[0105] Compare the generated CQI with two configurable thresholds: T reiject (reject threshold) and T review (review threshold).

[0106] If CQI < T reiject : Analysis aborted. Output "quality failed" status to the laboratory information system and suggest rescan or make a new smear.

[0107] If T reject ≤ CQI < T review : Continue with classification but attach a "human review needed" flag to the result sent to the laboratory information system. The user interface will display the CQI score and highlight the specific quality defect items.

[0108] If CQI ≥ T review : Continue with classification, result labeled as "automatically reviewed passed".

[0109] Confidence modulation: The final confidence score of a reviewed cell classification is modulated by CQI, with the modulation formula:

[0110] C' = C x f(CQI);

[0111] where C is the confidence score output by the classification model, C' is the adjusted confidence score, and f(CQI) is a monotonically increasing function of the composite quality score CQI.

[0112] The confidence modulation ensures that even a high-confidence prediction on a borderline quality smear is presented to the user with appropriate caution.

[0113] S5. Cell classification based on transfer learning:

[0114] Base model selection: Choose a CNN pre-trained on a large general-purpose image dataset (e.g., ImageNet) as the base architecture, such as EfficientNet, ResNet, or VGG.

[0115] Model adaptation: The initial convolutional layers of the base model are "frozen" to preserve their ability to learn low-level features such as detecting edges, textures, and shapes. The classification layers at the end of the model, specific to the original task, are removed.

[0116] New classifier head: A new set of fully connected layers is added at the end of the model, with the number of neurons in the output layer corresponding to the desired blood cell classes (e.g., neutrophils, eosinophils, basophils, lymphocytes, monocytes, blasts, nucleated red blood cells, etc.).

[0117] Fine-tuning: The modified model is trained on a large, curated, and expert-labeled high-quality blood cell image dataset. This fine-tuning process adjusts the weights of the newly added layers and possibly the entire network (at a lower learning rate), allowing the model to specialize in the blood cell classification task.

[0118] In the description of the present specification, the description of the terms "one embodiment", "an example", "a specific example", and the like means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In the present specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Also, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0119] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details of the present invention, nor limit the present invention to the specific embodiments described. It is obvious that many modifications and variations can be made according to the content of the present specification. The present specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can well understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A blood cell morphology AI quality control system based on transfer learning, characterized in that, The system comprises an image acquisition module, a quality control module, a diagnostic classification module, a data management module and a user interface module. The image acquisition module is connected with a digital slide scanner and is configured to receive whole slide images from the digital slide scanner. The quality control module is configured to perform multi-parameter quality analysis on the single-layer area in the whole slide images. The quality control module comprises an artifact analysis unit, a staining quality analysis unit and a cell distribution uniformity analysis unit, which are respectively configured to perform artifact analysis, staining quality analysis and cell distribution uniformity analysis on the single-layer area in the whole slide images. The diagnostic classification module is configured to perform diagnosis on the whole slide images according to the multi-parameter quality analysis results of the single-layer area in the whole slide images by the quality control module, to exclude unqualified smears and to classify the qualified smears to determine the blood cell categories corresponding to the smears. The data management module is configured to store the output results of the quality control module and the diagnostic classification module, and is connected with a laboratory information system to synchronize the data with the laboratory information system. The user interface module is configured to display the data stored in the data management module through a user interface.

2. The migration learning based blood cell morphology AI quality control system according to claim 1, wherein, The artifact analysis is configured to detect and quantify physical artifacts in the single-layer area in the whole slide images, wherein the physical artifacts include tissue fold, bubble, dust, scratch, ink or fibrin filament.

3. The migration learning based blood cell morphology AI quality control system according to claim 1, wherein, The artifact analysis uses a pre-trained image segmentation neural network model to perform pixel-level segmentation on the artifacts in the single-layer area in the whole slide images, and determines the percentage of the area occupied by the artifacts as the result of the artifact analysis.

4. The transfer learning-based blood cell morphology AI quality control system according to claim 1, wherein, The staining quality analysis specifically comprises: Converting the whole slide images from an RGB color space to a color space capable of separating brightness and chrominance information, such as HSD or CIELAB; Identifying pixel groups corresponding to red blood cell cytoplasm and white blood cell nucleus by color thresholding method; Calculating and comparing the color histogram statistics such as mean, standard deviation and peak position of these pixel groups with a pre-set gold standard profile established by a group of expert-verified best staining smears, to quantitatively evaluate the staining intensity, the color balance of eosin and hematoxylin, and the presence of staining precipitate.

5. The transfer learning-based blood cell morphology AI quality control system according to claim 1, wherein, The cell distribution uniformity analysis specifically comprises: Locating the centroids of red blood cells in the single-layer area in the whole slide images by object detection algorithm; Dividing the single-layer area in the whole slide images into a uniform two-dimensional grid; Counting the number of cell centroids in each grid cell; Calculating the coefficient of variation of cell counts in all grid cells as a quantitative measure of distribution uniformity, wherein a lower coefficient of variation value indicates a higher uniformity.

6. The transfer learning-based blood cell morphology AI quality control system according to claim 1, wherein, The output result of the quality control module is a composite quality score, which is calculated by a weighted scoring model, specifically comprising: Normalizing the analysis results of each parameter, such as artifact area percentage, staining deviation and distribution coefficient of variation, to a common numerical scale, for example 0 to 1, by using the min-max scaling method; Assigning each normalized result a predetermined weight according to its expected clinical impact on the accuracy of the final diagnosis; Calculating the weighted sum of all weighted results to obtain the composite quality score. The formula for calculating the composite quality score is: Wherein, CQI is a composite quality score, w i is the weight of the i-th parameter normalization result, S i is the score of the i-th parameter after normalization.

7. The transfer learning-based blood cell morphology AI quality control system according to claim 1, wherein, The step of diagnosing and classifying the whole slide image by the diagnostic classification module comprises: Comparing the composite quality score output by the quality control module with a pre-set rejection threshold and review threshold; According to the comparison result, the following operations are performed: if the composite quality score is lower than the rejection threshold, the analysis is stopped and the laboratory information system is reported that the quality is unqualified; if the composite quality score is between the rejection threshold and the review threshold, the classification is continued but a mark indicating that manual review is needed is added to the result; if the composite quality score is higher than the review threshold, the result is accepted as automatically reviewed and passed; The whole slide image that passes the review is classified by using a deep learning model obtained through transfer learning, and the corresponding confidence score is output.

8. The transfer learning-based blood cell morphology AI quality control system according to claim 7, wherein, The training method of the deep learning model is: A convolutional neural network pre-trained on a large non-medical image dataset is selected as a base model; The architecture of the base model is modified, the original top classification layer is removed, and a new set of fully connected layers customized for the blood cell classification task is replaced; The modified model is fine-tuned, i.e. the newly added layers are retrained on a large, high-quality and expert-annotated blood cell image dataset, so that the model is specifically used for blood cell classification.

9. The transfer learning-based blood cell morphology AI quality control system according to claim 7, wherein, The diagnostic classification module also uses the composite quality score as a modulation factor when classifying, adjusting the confidence score output by the deep learning model for each cell classification; The formula for adjusting the confidence score is: C' = C x f(CQI); Where C is the confidence score output by the classification model, C' is the adjusted confidence score, and f(CQI) is a monotonically increasing function of the composite quality score CQI.

10. A blood cell morphology AI quality control method based on transfer learning, characterized in that, The steps include: Obtaining digitized whole slide images of blood smears from a digital slide scanner; Performing multi-parameter quality analysis on the single-layer area in the whole slide image, which includes: Artifact analysis for detecting and quantifying physical artifacts in the single-layer area; Staining quality analysis for evaluating the staining characteristics of cells in the single-layer area; Cell distribution uniformity analysis for measuring the spatial distribution of cells in the single-layer area; Based on the results of the multi-parameter quality analysis, a composite quality index is generated; Comparing the composite quality index with a pre-set rejection threshold and review threshold to exclude unqualified smears, and classifying the passed smears by using a deep learning model obtained through transfer learning to determine their corresponding blood cell categories.

Images