Peripheral blood heterotypic lymphocyte AI automatic detection system

Abstract

The invention relates to the technical field of medical examination and artificial intelligence, in particular to a peripheral blood heterotypic lymphocyte AI automatic detection system which comprises a model training subsystem and an automatic detection subsystem. The training subsystem is used for extracting multi-modal morphological characteristics such as cell nucleuses, cytoplasm and spatial relations by constructing an expert annotation data set, and training a cell classification model of a CNN-Transform mixed architecture and an expert sub-model for processing difficult samples. After the detection subsystem performs same feature extraction on the blood smear to be detected, the cell classification model performs preliminary classification and outputs confidence, the expert sub-model is started for performing secondary classification on difficult cells with low confidence, and finally, all results are synthesized to generate a structured grading report. According to the method, the problems of false detection and high omission ratio caused by complex and changeable shapes of the heterotypic lymphocytes in the traditional method are solved, and high-precision and high-efficiency automatic detection is realized.

Description

Technical Field

[0001] This invention relates to the fields of medical testing and artificial intelligence technology, and in particular to an AI-automated detection system for atypical peripheral blood lymphocytes. Background Technology

[0002] In peripheral blood cell morphology examination, the identification and counting of atypical lymphocytes are crucial for the diagnosis of various viral infections and other diseases. Currently, this test primarily relies on laboratory physicians manually observing stained blood smears under an optical microscope, and the identification criteria are highly dependent on the physician's personal experience. Atypical lymphocytes are not a single morphology but rather a cell population with continuous lineage variations. Their morphology lies between normal lymphocytes and other mononuclear-macrophage system cells such as monocytes. This inherent morphological ambiguity and heterogeneity leads to the risk of subjective judgment differences and inconsistent results, even among experienced laboratory physicians.

[0003] However, existing automated blood analyzers or general cell identification algorithms still face significant technical bottlenecks when dealing with the specific task of atypical lymphocyte detection. The core problem is that existing technologies cannot accurately quantify and effectively integrate the subtle multidimensional morphological differences of atypical lymphocytes in the nucleus, cytoplasm, and their spatial relationships. This results in high false positive and false negative rates in complex real-world scenarios, which limits the reliable application of automated technology in atypical lymphocyte screening and fails to provide clinicians with highly specific and accurate test results comparable to those of senior experts. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides an AI-automated detection system for peripheral blood atypical lymphocytes, aiming to improve the problem in the existing technology that it is difficult to accurately quantify and effectively integrate the multidimensional morphological differences of atypical lymphocytes in the cell nucleus, cytoplasm and their spatial relationships, resulting in high false detection and false negative rates in complex real-world scenarios.

[0005] This invention provides the following technical solution: an AI-automated detection system for atypical lymphocytes in peripheral blood, the system comprising a model training subsystem and an automated detection subsystem: The model training subsystem includes the following modules: The training image acquisition and dataset construction module is used to acquire multiple stained first peripheral blood smear digital images and construct a training dataset containing expert annotations of atypical lymphocytes. The training feature extraction and annotation acquisition module is used to extract the first multimodal morphological features from the cell images in the training dataset and acquire the corresponding expert annotations. The cell classification model training module is used to train a cell classification model based on the extracted first multimodal morphological features and the corresponding expert annotations, so as to obtain a trained cell classification model. The expert sub-model training module is used to train an expert sub-model based on the difficult samples in the training dataset to obtain a trained expert sub-model. The automated detection subsystem includes the following modules: The detection image acquisition and feature extraction module is used to acquire the digital image of the stained second peripheral blood smear to be detected, and to extract the second multimodal morphological features of the candidate cells in the digital image of the second peripheral blood smear. The preliminary classification and confidence assessment module is used to input the extracted second multimodal morphological features into the trained cell classification model to obtain the preliminary classification results of candidate cells and the corresponding confidence scores. The module for identifying and classifying problematic cells is used to determine whether a candidate cell is a problematic cell based on the confidence level, and to activate the trained expert sub-model to perform secondary classification on the problematic cells in order to determine the final classification result of all candidate cells. The grading report generation module is used to synthesize the final classification results of all candidate cells and generate a grading report that includes the detection status of atypical lymphocytes.

[0006] Preferably, the process of constructing the training dataset includes: Multiple high-resolution digital images of peripheral blood smears stained with Wright-Gymsa were acquired using a microscope digital scanning system. At least two hematology experts independently identified and labeled the lymphocytes in the field of view based on the high-resolution digital images. Consistency arbitration is conducted on the expert annotation results. When the annotations are consistent, the cell image and its annotations are directly included in the training dataset. When the annotations are inconsistent, a third senior expert makes the final decision, and the decision is included in the training dataset. The arbitrated cell images are classified into normal lymphocytes, atypical lymphocytes, and other easily confused cell categories, ultimately forming the training dataset containing expert annotations for atypical lymphocytes.

[0007] Preferably, the process for extracting the first multimodal morphological features includes: For each cell image in the training dataset that has been arbitrated, instance segmentation of the cell nucleus, cytoplasm, and cell body is performed to obtain the cell nucleus mask, cytoplasm mask, and cell body mask; Based on the nuclear mask, nuclear morphological modal features are calculated, including nuclear shape category, nuclear misalignment degree, nuclear chromatin roughness and structure; Based on the cytoplasmic mask, cytoplasmic modal characteristics are calculated, including cytoplasmic staining characteristics, number and distribution of vacuoles, cytoplasmic transparency, and presence of granules. Based on the cell body mask, cell body morphological modal features are calculated, including cell body size, shape regularity, and presence or absence of pseudopodia. Based on the spatial relationship between the nuclear mask, cytoplasmic mask, and cell body mask, spatial relationship modal features are calculated, including the nucleocytoplasmic ratio and the relative positions of the nucleus and cytoplasm. The extracted nuclear morphological modal features, cytoplasmic modal features, cell body morphological modal features, and spatial relationship modal features are spliced ​​and normalized to form the first multimodal morphological feature vector of the cell. Among them, the nuclear morphological modal characteristics, cytoplasmic modal characteristics, cell body morphological modal characteristics, and spatial relationship modal characteristics are used to distinguish the morphological subtypes of type I, type II, and type III atypical lymphocytes.

[0008] Preferably, the process for obtaining the trained cell classification model includes: A cell classification model is constructed using a hybrid architecture with CNN as the backbone for feature extraction and Transformer encoder as the feature enhancer. The first multimodal morphological feature vector is used as the input feature, and the expert annotations bound to it are used as supervision labels to construct model training samples; The cell classification model is trained using the training samples to output a predicted category; During training, the difference between the cell classification model's predicted class and the expert-annotated supervision label is calculated to form the loss function, and the backpropagation algorithm is used to optimize the model parameters. When the performance of the cell classification model on the preset validation set reaches a stable state, the training is completed, the model parameters at this time are saved, and the trained cell classification model is obtained.

[0009] Preferably, the process for obtaining the trained expert sub-model includes: From the training dataset, cell samples whose prediction confidence of the trained cell classification model is lower than a preset threshold are selected to form a difficult sample set; An expert sub-model is constructed by taking the first multimodal morphological feature vector corresponding to the difficult sample set as input and the expert annotations bound to it as supervision labels, and then training the expert sub-model in a concentrated manner. When the classification accuracy of the expert sub-model on the difficult sample validation set surpasses that of the cell classification model, the training is completed, and the trained expert sub-model is obtained.

[0010] Preferably, the extraction process of the second multimodal morphological features includes: Digital images of the stained second peripheral blood smear to be tested are acquired using a microscope digital scanning system; Using a deep learning-based cell detection algorithm, all candidate lymphocytes were located and cropped from the digital image of the second peripheral blood smear. For each candidate lymphocyte image, the nucleus, cytoplasm, and cell body are segmented to obtain the nuclear mask, cytoplasmic mask, and cell body mask; Based on the nuclear mask, nuclear morphological modal features are calculated, including nuclear shape category, nuclear misalignment degree, nuclear chromatin roughness and structure; Based on the cytoplasmic mask, cytoplasmic modal characteristics are calculated, including cytoplasmic staining characteristics, number and distribution of vacuoles, cytoplasmic transparency, and presence of granules. Based on the cell body mask, cell body morphological modal features are calculated, including cell body size, shape regularity, and presence or absence of pseudopodia. Based on the spatial relationship between the nuclear mask, cytoplasmic mask, and cell body mask, spatial relationship modal features are calculated, including the nucleocytoplasmic ratio and the relative positions of the nucleus and cytoplasm. The nuclear morphological modal features, cytoplasmic modal features, cell body morphological modal features, and spatial relationship modal features extracted from the candidate cell are spliced ​​and normalized to form the second multimodal morphological feature vector of the cell.

[0011] Preferably, the process for obtaining the preliminary classification results and corresponding confidence scores of the candidate cells includes: Load the trained cell classification model and batch input the second multimodal morphological feature vectors of candidate cells into the trained cell classification model; The cell classification model extracts deep local features from the input features, then fuses the deep local features with global contextual information through a self-attention mechanism, and finally outputs a probability distribution vector. The category corresponding to the maximum probability value in the probability distribution vector is taken as the preliminary classification result of the candidate cell; The maximum probability value itself is used as the confidence level corresponding to the preliminary classification result.

[0012] Preferably, the process for determining whether a candidate cell is a problematic cell includes: A preset confidence threshold is used to distinguish between problematic and non-problematic cells; The confidence score of each candidate cell is compared with the confidence threshold. If the confidence level of a candidate cell is lower than the confidence threshold, the candidate cell is determined to be a difficult cell. If the confidence level of a candidate cell is not lower than the confidence threshold, then the candidate cell is determined to be a non-problematic cell.

[0013] Preferably, the process for obtaining the final classification result includes: For candidate cells identified as problematic cells, the trained expert sub-model is loaded. The second multimodal morphological feature vector corresponding to the problematic cells is input into the trained expert sub-model; The expert sub-model processes the input features and outputs the secondary classification result of the problematic cell; For candidate cells that are determined to be non-problematic cells, their preliminary classification results are directly used as their final classification results. The final classification results of all non-problematic cells and the secondary classification results of all problematic cells are combined to form a set of final classification results for all candidate cells.

[0014] Preferably, the generation process of the hierarchical report includes: The number of cells classified as atypical lymphocytes is counted in the final classification result set of all candidate cells; Based on the number of atypical lymphocytes and the total number of lymphocytes, calculate the percentage of atypical lymphocytes; Based on a preset clinical significance threshold, the percentage of atypical lymphocytes is mapped to the corresponding risk level; Generate a structured report containing the percentage of atypical lymphocytes, the risk level, and images of all cells identified as atypical lymphocytes and their location information in the original smear.

[0015] The present invention has the following beneficial effects: 1. In this invention, by introducing multimodal morphological feature extraction, the morphological features that rely on qualitative description are transformed into calculable quantitative indicators. Combined with the local and global feature capture capabilities of the CNN-Transformer hybrid model, the system can keenly identify subtle differences between atypical lymphocytes and easily confused cells such as normal lymphocytes and monocytes, thereby fundamentally reducing the false detection and false negative rates.

[0016] 2. In this invention, a confidence-based cascaded judgment strategy is adopted, which allows the main model to quickly process most simple cases, while low-confidence, difficult cells are automatically routed to more complex expert sub-models for secondary diagnosis. This design simulates the decision-making process of senior experts, and concentrates superior computing power to tackle the most difficult cells without significantly increasing the overall computing overhead, thus achieving the optimal balance between detection efficiency and ultimate accuracy.

[0017] 3. In this invention, a structured report containing the percentage of atypical lymphocytes, clinical risk level, and key cell image evidence can be generated. This not only provides doctors with intuitive and quantitative diagnostic basis, but also enhances the credibility and interpretability of the results through visual evidence. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the architecture of an AI-automated detection system for atypical peripheral blood lymphocytes proposed in this invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In a first embodiment of the present invention, the present invention provides an AI-automated detection system for atypical lymphocytes in peripheral blood, such as... Figure 1 As shown, the system includes a model training subsystem and an automated detection subsystem.

[0021] The model training subsystem includes the following modules: The training image acquisition and dataset construction module is used to acquire multiple stained digital images of the first peripheral blood smear and construct a training dataset containing expert annotations of atypical lymphocytes.

[0022] Furthermore, the process of constructing the training dataset includes: Multiple high-resolution digital images of peripheral blood smears stained with Wright-Gymsa were acquired using a microscope digital scanning system. At least two hematology experts independently identified and labeled the lymphocytes in the field of view based on high-resolution digital images; Consistency arbitration is conducted on the expert annotation results. When the annotations are consistent, the cell image and its annotations are directly included in the training dataset. When the annotations are inconsistent, a third senior expert makes the final decision, and the decision is included in the training dataset. The arbitrated cell images are classified into normal lymphocytes, atypical lymphocytes, and other easily confused cell categories, ultimately forming a training dataset containing expert annotations for atypical lymphocytes.

[0023] Specifically, a digital scanning system equipped with a high numerical aperture microscope is used to automatically scan peripheral blood smears processed with the standard Wright-Gymsa staining procedure, acquiring multiple high-resolution digital images with a resolution of no less than 0.1 micrometers per pixel. Subsequently, at least two clinically experienced hematology experts identify and label each lymphocyte in the field of view based on these high-resolution digital images in an independent diagnostic environment. The experts use specialized labeling software to accurately delineate the outlines of the cell nucleus and cytoplasm on the cell images and classify them into "normal lymphocytes," "atypical lymphocytes," or "other easily confused cells" based on their morphological characteristics. Among them, "other easily confused cells" mainly include monocytes and plasma cells. To ensure the authority and consistency of the annotation results, the system introduces an arbitration mechanism to compare the independent annotation results of all experts. When two experts have completely consistent annotation categories for the same cell, the cell image and its annotation are directly included in the training dataset. When the annotation opinions are inconsistent, a third, more experienced verification expert makes the final decision, and this decision is used as the final standard for inclusion in the training dataset. Finally, all cell images that have undergone the above arbitration process and their corresponding authoritative annotation information are integrated to form a large-scale, high-quality training dataset. This dataset is stored in a structured manner, in which each cell image precisely corresponds to its arbitrated category label, providing a reliable data foundation for subsequent model training.

[0024] The training feature extraction and annotation module is used to extract the first multimodal morphological features from cell images in the training dataset and obtain the corresponding expert annotations.

[0025] Furthermore, the process for extracting the first multimodal morphological features includes: For each cell image in the training dataset that has been arbitrated, instance segmentation of the cell nucleus, cytoplasm, and cell body is performed to obtain the cell nucleus mask, cytoplasm mask, and cell body mask; Based on the nuclear mask, nuclear morphological modal features are calculated, including nuclear shape category, nuclear misalignment degree, nuclear chromatin roughness and structure; Based on the cytoplasmic mask, cytoplasmic modal characteristics are calculated, including cytoplasmic staining characteristics, number and distribution of vacuoles, cytoplasmic transparency, and presence of granules. Based on the cell body mask, cell body morphological modal features are calculated, including cell body size, shape regularity, and presence or absence of pseudopodia. Based on the spatial relationship between the nuclear mask, cytoplasmic mask, and cell body mask, spatial relationship modal features are calculated, including the nucleocytoplasmic ratio and the relative positions of the nucleus and cytoplasm. The extracted nuclear morphological modal features, cytoplasmic modal features, cell body morphological modal features, and spatial relationship modal features are spliced ​​and normalized to form the first multimodal morphological feature vector of the cell. Among them, the nuclear morphological modal characteristics, cytoplasmic modal characteristics, cell body morphological modal characteristics, and spatial relationship modal characteristics are used to distinguish the morphological subtypes of type I, type II, and type III atypical lymphocytes.

[0026] Specifically, for each cell image in the training dataset that has been arbitrated, a deep learning-based instance segmentation model, such as Mask R-CNN or its improved model, which has been pre-trained on a large cell image dataset, is used for processing. This model outputs three precise pixel-level binary masks: a nucleus mask, which identifies the nucleus region with a foreground pixel value of 1 and a background value of 0; a cytoplasm mask, which identifies the cytoplasm region; and a cell body mask, which identifies the outer contour region of the entire lymphocyte. This is usually obtained by performing a morphological closing operation on the union of the nucleus mask and the cytoplasm mask to ensure the integrity of the contour. Based on the aforementioned mask and original color image, four major modal features are calculated. The calculation process for the nuclear morphology modal features is as follows: The nuclear shape category is calculated by extracting the outline of the nuclear mask and calculating its Hu invariant moments, Zernike moments, and other shape descriptors. Using a predefined classifier, such as a support vector machine, it is classified into categories like "circular," "elliptical," "kidney-shaped," and "lobed," and represented by a one-hot encoded vector. The degree of nuclear skewness is calculated by calculating the Euclidean distance between the centroid of the nuclear mask and the centroid of the cell body mask, and then normalizing this distance by dividing it by the equivalent radius of the cell body to obtain a quantized skewness coefficient. Finally, the nuclear chromatin roughness and structure are calculated by calculating the gray-level co-occurrence matrix in the gray-level region of the original image corresponding to the nuclear mask, and extracting the following texture parameters: roughness, quantified by contrast parameters; and structure, which is described by entropy and second-order angular moments as whether the chromatin is "rough mesh or small patchy" or "fine mesh." The calculation of cytoplasmic modal features involves the following steps: First, calculating cytoplasmic staining features. Within the cytoplasmic mask region, the average intensity value of all pixels in a specific color channel (focusing on the blue channel for Wright-Gymsa staining) is calculated to quantify the basophilicity of "deep blue," "light blue," or "grayish blue." Second, calculating the number and distribution of vacuoles. On the contrast-enhanced image of the cytoplasmic region, a local brightness extremum detection and region growing algorithm is used to identify "vacuolar" regions with significantly higher brightness than the surrounding cytoplasm. The number of vacuoles is counted, and the ratio of the total area of ​​vacuolar regions to the total area of ​​the cytoplasm is calculated as the distribution density. Third, calculating cytoplasmic transparency. This involves calculating the standard deviation of pixel intensity in the cytoplasmic region; a lower variance value may correspond to a more uniform and transparent cytoplasm (such as type II). Fourth, calculating the presence of granules. Within the cytoplasmic region, a multi-scale speckle detection algorithm is applied to identify fine granular structures such as azurophilic granules and their number is counted. The calculation of cell body morphology modal features specifically involves: calculating cell body size by directly calculating the area of ​​the cell body mask; calculating shape regularity by calculating the roundness of the cell body mask, i.e., (4π * area) / (circumference^2). The closer this value is to 1, the more regular (round) it is, and the lower it is, the more irregular it is. At the same time, its Fourier descriptor is calculated to capture complex contours. The cell body mask is extracted by skeletonization, and the number and length of skeleton branch points are analyzed. If there are long branches whose length exceeds a certain proportion of the equivalent radius of the cell body, it is determined that there are "pseudopodia", which is represented by Boolean values ​​or length parameters. The calculation of spatial relationship modal features specifically involves calculating the nucleocytoplasmic ratio, which is the ratio of the area of ​​the nuclear mask to the area of ​​the cytoplasmic mask; and the nucleocytoplasmic relative position, which is the direction and distance of the centroid of the nuclear mask relative to the boundary of the cell body mask, and can be represented by polar coordinates. All the feature parameters calculated above, including classification codes, quantization coefficients, intensity values, proportion values, quantity, Boolean values, etc., are concatenated into a one-dimensional numerical vector in a predetermined fixed order. This vector is then Z-score normalized to eliminate the influence of differences in the dimensions and value ranges of different features, ultimately forming the standardized first multimodal morphological feature vector of the cell. This feature vector is bound to the expert annotations determined by arbitration in the training dataset for this cell, constituting a complete training sample. The extracted multimodal morphological features are designed to effectively characterize and distinguish the three classic clinical atypical lymphocyte morphological subtypes. Key distinguishing features include: Type I (vacuolar / plasma cell type), characterized by high cytoplasmic staining characteristic (basophilic) value, significant vacuolar number and distribution density parameters, and potentially high nuclear eccentricity parameter; Type II (irregular / monocyte type), characterized by low cell shape regularity parameter (irregular), true pseudopodia, high cytoplasmic transparency parameter (low variance), and relatively large cell size parameter; and Type III (immature type), characterized by nuclear chromatin structure parameters (such as low entropy and high second moment) indicating fine and uniform chromatin, high cytoplasmic staining characteristic (basophilic) value, and nuclear shape category possibly pointing to "round" or "oval".

[0027] The cell classification model training module is used to train a cell classification model based on the extracted first multimodal morphological features and corresponding expert annotations, so as to obtain a trained cell classification model.

[0028] Furthermore, the process for obtaining a trained cell classification model includes: A cell classification model is constructed using a hybrid architecture with CNN as the backbone for feature extraction and Transformer encoder as the feature enhancer. The first multimodal morphological feature vector is used as the input feature, and the expert annotations bound to it are used as supervision labels to construct model training samples; The cell classification model is trained using training samples, and the predicted category is output. During training, the difference between the cell classification model's predicted class and the expert-annotated supervision label is calculated to form the loss function, and the backpropagation algorithm is used to optimize the model parameters. When the performance of the cell classification model on the preset validation set reaches a stable state, the training is completed, the model parameters at this time are saved, and the trained cell classification model is obtained.

[0029] Specifically, firstly, a cell classification model is constructed, which adopts a hybrid architecture combining a convolutional neural network (CNN) and a Transformer encoder. Specifically, the feature extraction backbone is a six-layer CNN network with an input dimension of the same length as the first multimodal morphological feature vector, used to extract deep local correlation features from the input features. The output feature map of this CNN backbone is then flattened into a sequence and input into a Transformer encoder, which contains two layers, each equipped with eight attention heads, to capture the global dependencies between different feature points in the sequence through its self-attention mechanism. Finally, a fully connected layer is connected as a classifier, and the number of its output nodes corresponds to the number of cell classification categories, including normal lymphocytes, atypical lymphocytes, and other easily confused cells. Next, training samples are constructed. The first multimodal morphological feature vector obtained earlier is used as the input feature X, and the expert annotations bound to it are encoded one-hot and used as the supervision label Y. Together, they form a training sample. The set of all training samples constitutes the training set of the model. Then, the cell classification model is trained using the training set. The training process includes forward propagation and backward propagation. In forward propagation, the input feature X is processed by the CNN backbone and the Transformer encoder. Finally, the classifier outputs a predicted probability distribution for each category. The difference between the model's predicted probability distribution and the true label Y is calculated using the cross-entropy loss function. The calculation formula is as follows: ; in, This represents the i-th component of the real label. The i-th component represents the model's predicted probability. During backpropagation, the stochastic gradient descent algorithm or its variants are used to update all parameters in the model based on the gradient of the loss function, including the weights of the CNN, the parameters of the Transformer, and the weights and biases of the fully connected layers. During training, a validation set, which is pre-defined from the total data and not used for training, is used to monitor model performance. When the model's classification accuracy on the validation set no longer improves by more than 0.1 percent within twenty consecutive training cycles, the model performance is considered to have reached a stable state, training is terminated, and all parameters of the trained model are saved to obtain the trained cell classification model for use by the automated detection subsystem.

[0030] The expert sub-model training module is used to train an expert sub-model based on difficult samples in the training dataset, and obtain a trained expert sub-model.

[0031] Furthermore, the process for obtaining the trained expert sub-model includes: From the training dataset, cell samples whose prediction confidence of the trained cell classification model is lower than a preset threshold are selected to form a difficult sample set; An expert sub-model is constructed by taking the first multimodal morphological feature vector corresponding to the difficult sample set as input and the expert annotations bound to it as supervision labels, and then training the expert sub-model in a concentrated manner. When the expert sub-model outperforms the cell classification model in classification accuracy on the difficult sample validation set, training is complete, and a well-trained expert sub-model is obtained.

[0032] Specifically, firstly, a difficult sample set is constructed. The first multimodal morphological feature vector from the entire training dataset is input into the previously trained cell classification model for forward propagation, and the prediction confidence of each cell sample is obtained. This confidence is the maximum value in the probability distribution vector output by the model. A confidence threshold of 0.85 is set, and all cell samples with prediction confidence below this threshold are filtered out to form a difficult sample set. The samples in this set are cases that the main model considers difficult to distinguish or has low confidence in. Subsequently, an expert sub-model is constructed. This model is a deeper and more complex neural network than the cell classification model. In one specific embodiment, the expert sub-model is a twelve-layer deep fully connected network. Its input layer dimension is the same as the length of the first multimodal morphological feature vector, and its output layer dimension is the same as the number of classification categories. The number of neurons in the intermediate hidden layers decreases layer by layer. The model uses the ReLU activation function to introduce nonlinearity. Then, the expert sub-model is trained intensively using a set of difficult samples. The first multimodal morphological feature vector corresponding to the difficult sample set is used as input, and the expert annotations bound to it are used as supervision labels. The training process is also optimized using the cross-entropy loss function and the stochastic gradient descent algorithm. Unlike the training of the main model, the goal of this training is to enable the expert sub-model to focus on learning and mastering the subtle morphological differences of these difficult samples. A separate portion of the difficult sample set is allocated as a difficult sample validation set. During training, the classification accuracy of the expert sub-model is periodically evaluated on this validation set. When the classification accuracy (e.g., accuracy or F1 score) of the expert sub-model on this validation set consistently and stably surpasses the performance of the cell classification model on the same validation set, the expert sub-model is considered to have completed training. The parameters of the expert sub-model at this point are saved, resulting in a well-trained expert sub-model, which is then used for secondary classification of difficult cells during the automated detection phase.

[0033] The detection image acquisition and feature extraction module is used to acquire the digital image of the stained second peripheral blood smear to be detected, and to extract the second multimodal morphological features of the candidate cells in the digital image of the second peripheral blood smear.

[0034] Furthermore, the extraction process for the second multimodal morphological feature includes: Digital images of the stained second peripheral blood smear to be tested are acquired using a microscope digital scanning system; Using a deep learning-based cell detection algorithm, all candidate lymphocytes were located and cropped from the digital image of the second peripheral blood smear. For each candidate lymphocyte image, the nucleus, cytoplasm, and cell body are segmented to obtain the nuclear mask, cytoplasmic mask, and cell body mask; Based on the nuclear mask, nuclear morphological modal features are calculated, including nuclear shape category, nuclear misalignment degree, nuclear chromatin roughness and structure; Based on the cytoplasmic mask, cytoplasmic modal characteristics are calculated, including cytoplasmic staining characteristics, number and distribution of vacuoles, cytoplasmic transparency, and presence of granules. Based on the cell body mask, cell body morphological modal features are calculated, including cell body size, shape regularity, and presence or absence of pseudopodia. Based on the spatial relationship between the nuclear mask, cytoplasmic mask, and cell body mask, spatial relationship modal features are calculated, including the nucleocytoplasmic ratio and the relative positions of the nucleus and cytoplasm. The nuclear morphological modal features, cytoplasmic modal features, cell body morphological modal features, and spatial relationship modal features extracted from the candidate cell are spliced ​​and normalized to form the second multimodal morphological feature vector of the cell.

[0035] Specifically, using a microscope digital scanning system of the same model or with comparable performance as the one used in the training phase, the peripheral blood smear to be detected, which has already undergone the standard Wright-Gymsa staining procedure, is automatically scanned to acquire a high-resolution second peripheral blood smear digital image. The resolution, color space, and other parameters of this image are consistent with the standard of the first peripheral blood smear digital image acquired in the training phase. A cell detection algorithm based on deep learning, such as YOLOv5, Faster R-CNN, or RetinaNet architecture, and a model pre-trained on a general blood cell image dataset containing lymphocytes, is used to analyze the entire second peripheral blood smear digital image. This model can identify and output the bounding box coordinates of all suspected lymphocytes in the image. Based on these bounding box coordinates, a local image of each candidate lymphocyte is precisely cropped from the original digital image as the input unit for subsequent processing. For each candidate lymphocyte image obtained from the above cropping, a deep learning-based instance segmentation model (e.g., the same version of Mask R-CNN) is used for processing, identical to that used in the training phase. This model performs forward inference on the candidate cell image, simultaneously outputting three precise pixel-level binary masks: a nuclear mask, a cytoplasmic mask, and a cell body mask. Based on the obtained masks and the original color image of the candidate cell, the model uses the same calculation formulas, algorithms, and parameters as in the feature extraction phase to calculate the four major modal features one by one. Specifically, the calculation of nuclear morphology modal features includes calculating the nuclear shape category, using the same shape classifier used in the training phase, classifying it into categories such as "circular" and "elliptical" based on the nuclear mask outline, and generating encoding vectors in the same format; calculating the nuclear offset degree, using the same normalization calculation formula as in the training phase; calculating the nuclear chromatin roughness and structure, calculating the gray-level co-occurrence matrix in the gray-level region corresponding to the nuclear mask, and extracting three texture parameters—contrast, entropy, and second angular moment—identical to those used in the training phase. The calculation of cytoplasmic modal features includes: calculating cytoplasmic coloring features, which calculates the average intensity value of the cytoplasmic mask region in the blue channel; calculating the number and distribution of vacuoles, which uses the same local extremum detection and region growing algorithm to identify vacuoles and calculate their number and area ratio; calculating cytoplasmic transparency, which calculates the standard deviation of pixel intensity in the cytoplasmic region; and calculating particle presence, which uses the same multi-scale spot detection algorithm to identify and count the number of azurite particles. The calculation of cell body morphological modal characteristics includes calculating cell body size, calculating cell body mask area, calculating shape regularity, calculating the same roundness and Fourier descriptor, and determining the presence or absence of pseudopodia using the same skeletonization and branching analysis methods. The calculation of spatial relationship modal characteristics includes calculating the nucleocytoplasmic ratio, which is the ratio of the nuclear area to the cytoplasmic area, and calculating the relative position of the nucleocytoplasm, which is the same polar coordinate parameter of the nuclear centroid relative to the cell boundary. All the feature parameters calculated above are concatenated in the same order preset during the training phase to form a one-dimensional original feature vector. Then, the global feature mean and standard deviation, which were pre-calculated and saved during the model training phase, are loaded, and the original feature vector is Z-score normalized. This operation ensures that the second multimodal morphological feature vector extracted during the detection phase and the first multimodal morphological feature vector obtained during the training phase are in the same numerical distribution space, so that they can be directly and correctly processed by the trained classification model. Finally, each candidate lymphocyte that has completed the above process generates a standardized second multimodal morphological feature vector, which will be used as the input for the subsequent "preliminary classification and confidence assessment module".

[0036] The preliminary classification and confidence assessment module is used to input the extracted second multimodal morphological features into the trained cell classification model to obtain the preliminary classification results of candidate cells and the corresponding confidence scores.

[0037] Furthermore, the process for obtaining the preliminary classification results and corresponding confidence scores of candidate cells includes: Load the trained cell classification model and batch input the second multimodal morphological feature vectors of candidate cells into the trained cell classification model; The cell classification model extracts deep local features from the input features, then fuses the deep local features with global contextual information through a self-attention mechanism, and finally outputs a probability distribution vector. The category corresponding to the highest probability value in the probability distribution vector is taken as the preliminary classification result of the candidate cell; The highest probability value itself is used as the confidence level corresponding to the preliminary classification result.

[0038] Specifically, the pre-trained cell classification model, saved during the previous training phase, is loaded from the storage medium and deployed in the memory of the computing device. This model contains optimized weights and bias parameters. Subsequently, the second multimodal morphological feature vectors of multiple candidate cells obtained after preprocessing are combined into a batch and input into the loaded cell classification model in batches. The batch processing method helps to improve computational efficiency. After receiving the input feature vectors, the model performs forward propagation calculations in sequence. The input features first pass through the CNN backbone network of the model for deep local feature extraction. This process is achieved through multi-layer convolution and pooling operations. Then, the Transformer encoder performs global context information fusion on the extracted deep local feature sequence. Its self-attention mechanism can calculate the correlation weights between different feature elements in the sequence and generate an enhanced feature representation that integrates global information. Finally, the enhanced feature is passed to the fully connected classifier at the end. The classifier outputs a probability distribution vector. The dimension of this vector is the same as the number of cell categories. Each element value in the vector represents the probability that the model determines that the cell belongs to the corresponding category. The sum of all element values ​​is 1. For each candidate cell, the highest probability value is identified from its output probability distribution vector. The category label corresponding to the highest probability value is then assigned. For example, if the highest value corresponds to the atypical lymphocyte category, the cell is initially classified as an atypical lymphocyte. This is the preliminary classification result for the cell. At the same time, the highest probability value itself is taken as a value between 0 and 1 and directly used as the confidence level for this preliminary classification result. This confidence level quantifies the model's confidence in its classification decision.

[0039] The module for identifying and classifying questionable cells is used to determine whether a candidate cell is a questionable cell based on confidence level, and to activate a pre-trained expert sub-model to perform secondary classification on questionable cells in order to determine the final classification result of all candidate cells.

[0040] Furthermore, the process for determining whether candidate cells are problematic cells includes: A preset confidence threshold is used to distinguish between problematic and non-problematic cells; The confidence score of each candidate cell is compared with the confidence threshold. If the confidence level of a candidate cell is lower than the confidence threshold, the candidate cell is determined to be a difficult cell. If the confidence level of a candidate cell is not lower than the confidence threshold, then the candidate cell is determined to be a non-problematic cell.

[0041] Furthermore, the process for obtaining the final classification results includes: For candidate cells identified as problematic cells, load the trained expert sub-model; The second multimodal morphological feature vector corresponding to the problematic cells is input into the trained expert sub-model; The expert sub-model processes the input features and outputs the secondary classification result of the problematic cell; For candidate cells that are determined to be non-problematic cells, their preliminary classification results are directly used as their final classification results. The final classification results of all non-problematic cells and the secondary classification results of all problematic cells are combined to form a set of final classification results for all candidate cells.

[0042] Specifically, the system presets a confidence threshold of 0.90 to distinguish between problematic and non-problematic cells. The initial classification confidence of each candidate cell is compared with this threshold. If the confidence of a candidate cell is lower than 0.90, the system determines that the cell is a problematic cell. If the confidence of a candidate cell is equal to or higher than 0.90, the system determines that the cell is a non-problematic cell. Then, based on the above judgment results, the final classification result of all candidate cells is determined. For candidate cells that are determined to be problematic cells, the system performs the following operations: loads the trained expert sub-model saved in the previous training phase from the storage medium, inputs the second multimodal morphological feature vector corresponding to the problematic cell into the loaded expert sub-model, the expert sub-model performs forward propagation processing on the input feature vector, and outputs a secondary classification result. This secondary classification result is the final classification result of the problematic cell. For candidate cells identified as non-problematic cells, the system no longer activates the expert sub-model. Instead, it directly uses the preliminary classification result obtained from the preliminary classification and confidence assessment module as the final classification result for that cell. Finally, the system creates a final classification result set, which summarizes the final classification results of all non-problematic cells and the secondary classification results of all problematic cells given by the expert sub-model. This final classification result set contains authoritative classification information for all identified candidate cells in the current peripheral blood smear digital image to be tested, providing a data foundation for generating the final structured report.

[0043] The grading report generation module is used to synthesize the final classification results of all candidate cells and generate a grading report that includes the detection status of atypical lymphocytes.

[0044] Furthermore, the process for generating tiered reports includes: The number of cells classified as atypical lymphocytes is counted in the final classification result set of all candidate cells; Calculate the percentage of atypical lymphocytes based on the number of atypical lymphocytes and the total number of lymphocytes; Based on a preset clinical significance threshold, the percentage of atypical lymphocytes is mapped to the corresponding risk level; Generate a structured report containing the percentage of atypical lymphocytes, risk level, and images of all cells identified as atypical lymphocytes and their location information in the original smear.

[0045] Specifically, the system first performs statistical analysis on the final classification result set. Specifically, it iterates through the set, counting the number of cells classified as atypical lymphocytes, denoted as . Typically, the sum of the number of cells classified as normal lymphocytes and atypical lymphocytes in the count set is denoted as . This is the total number of lymphocytes in the current field of view. The system then calculates the percentage of atypical lymphocytes using the formula: Atypical lymphocyte percentage = typical / *100%; Next, the system maps the calculated percentage of atypical lymphocytes to the corresponding risk level based on the preset clinical significance threshold. The clinical significance threshold is set as follows: when the percentage of atypical lymphocytes is less than 5%, it is mapped to a negative risk level; when the percentage is greater than or equal to 5% and less than 10%, it is mapped to a suspected positive risk level; and when the percentage is greater than or equal to 10%, it is mapped to a positive risk level. Finally, the system automatically generates a structured test report, which is an electronic document. The report includes at least the calculated percentage of atypical lymphocytes, the specific risk level determined according to the above threshold, snapshots of all cells identified as atypical lymphocytes, and the coordinate position information of each atypical lymphocyte image in the original digital image of the second peripheral blood smear. After this structured report is generated, it can be automatically transmitted to the hospital's laboratory information system or printed as a paper report for clinicians to review and use for diagnosis.

[0046] It should be noted that, compared with the inconsistent reporting methods of traditional manual microscopy, for example, some laboratories report morphologically similar cells as 'atypical lymphocytes' while others report them as 'abnormal lymphocytes' or 'abnormal cells' (see Table 1 below), the system of this invention strictly follows a standardized classification system. For cells identified as atypical lymphocytes, the report clearly classifies them as 'atypical lymphocytes' and provides their percentage, risk level, and image evidence. This provides clinicians with unified, objective, and directly comparable test results, effectively avoiding clinical interpretation confusion caused by inconsistent naming and standards. Table 1. Example of recommended reporting methods for blood smear results with atypical lymphocytosis: White blood cell classification Reporting Method 1 Reporting Method 2 Reporting Method 3 neutrophils 35% 35% 35% Monocytes 5％ 5% 5% lymphocytes 45% 45% 45% Atypical lymphocytes 15% - - Atypical lymphocytes - 15% - abnormal cells - - 15% Remark -35% - Most of the abnormal cells resemble atypical lymphocytes. Note: * indicates main morphological features and suggestions; - indicates no data.

[0047] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-automated detection system for atypical lymphocytes in peripheral blood, characterized in that, The system includes the following model training subsystem and automated detection subsystem: The model training subsystem includes the following modules: The training image acquisition and dataset construction module is used to acquire multiple stained first peripheral blood smear digital images and construct a training dataset containing expert annotations of atypical lymphocytes. The training feature extraction and annotation acquisition module is used to extract the first multimodal morphological features from the cell images in the training dataset and acquire the corresponding expert annotations. The cell classification model training module is used to train a cell classification model based on the extracted first multimodal morphological features and the corresponding expert annotations, so as to obtain a trained cell classification model. The expert sub-model training module is used to train an expert sub-model based on the difficult samples in the training dataset to obtain a trained expert sub-model. The automated detection subsystem includes the following modules: The detection image acquisition and feature extraction module is used to acquire the digital image of the stained second peripheral blood smear to be detected, and to extract the second multimodal morphological features of the candidate cells in the digital image of the second peripheral blood smear. The preliminary classification and confidence assessment module is used to input the extracted second multimodal morphological features into the trained cell classification model to obtain the preliminary classification results of candidate cells and the corresponding confidence scores. The module for identifying and classifying problematic cells is used to determine whether a candidate cell is a problematic cell based on the confidence level, and to activate the trained expert sub-model to perform secondary classification on the problematic cells in order to determine the final classification result of all candidate cells. The grading report generation module is used to synthesize the final classification results of all candidate cells and generate a grading report that includes the detection status of atypical lymphocytes.

2. The AI-automated detection system for atypical lymphocytes in peripheral blood according to claim 1, characterized in that, The process of constructing the training dataset includes: Multiple high-resolution digital images of peripheral blood smears stained with Wright-Gymsa were acquired using a microscope digital scanning system. At least two hematology experts independently identified and labeled the lymphocytes in the field of view based on the high-resolution digital images. Consistency arbitration is conducted on the expert annotation results. When the annotations are consistent, the cell image and its annotations are directly included in the training dataset. When the annotations are inconsistent, a third senior expert makes the final decision, and the decision is included in the training dataset. The arbitrated cell images are classified into normal lymphocytes, atypical lymphocytes, and other easily confused cell categories, ultimately forming the training dataset containing expert annotations for atypical lymphocytes.

3. The AI-automated detection system for atypical lymphocytes in peripheral blood according to claim 1, characterized in that, The process of the first multimodal morphological feature extraction includes: For each cell image in the training dataset that has been arbitrated, instance segmentation of the cell nucleus, cytoplasm, and cell body is performed to obtain the cell nucleus mask, cytoplasm mask, and cell body mask; Based on the nuclear mask, nuclear morphological modal features are calculated, including nuclear shape category, nuclear misalignment degree, nuclear chromatin roughness and structure; Based on the cytoplasmic mask, cytoplasmic modal characteristics are calculated, including cytoplasmic staining characteristics, number and distribution of vacuoles, cytoplasmic transparency, and presence of granules. Based on the cell body mask, cell body morphological modal features are calculated, including cell body size, shape regularity, and presence or absence of pseudopodia. Based on the spatial relationship between the nuclear mask, cytoplasmic mask, and cell body mask, spatial relationship modal features are calculated, including the nucleocytoplasmic ratio and the relative positions of the nucleus and cytoplasm. The extracted nuclear morphological modal features, cytoplasmic modal features, cell body morphological modal features, and spatial relationship modal features are spliced ​​and normalized to form the first multimodal morphological feature vector of the cell. Among them, the nuclear morphological modal characteristics, cytoplasmic modal characteristics, cell body morphological modal characteristics, and spatial relationship modal characteristics are used to distinguish the morphological subtypes of type I, type II, and type III atypical lymphocytes.

4. The AI-automated detection system for atypical lymphocytes in peripheral blood according to claim 3, characterized in that, The process for obtaining the trained cell classification model includes: A cell classification model is constructed using a hybrid architecture with CNN as the backbone for feature extraction and Transformer encoder as the feature enhancer. The first multimodal morphological feature vector is used as the input feature, and the expert annotations bound to it are used as supervision labels to construct model training samples; The cell classification model is trained using the training samples to output a predicted category; During training, the difference between the cell classification model's predicted class and the expert-annotated supervision label is calculated to form the loss function, and the backpropagation algorithm is used to optimize the model parameters. When the performance of the cell classification model on the preset validation set reaches a stable state, the training is completed, the model parameters at this time are saved, and the trained cell classification model is obtained.

5. The AI-automated detection system for atypical lymphocytes in peripheral blood according to claim 4, characterized in that, The process for obtaining the trained expert sub-model includes: From the training dataset, cell samples whose prediction confidence of the trained cell classification model is lower than a preset threshold are selected to form a difficult sample set; An expert sub-model is constructed by taking the first multimodal morphological feature vector corresponding to the difficult sample set as input and the expert annotations bound to it as supervision labels, and then training the expert sub-model in a concentrated manner. When the classification accuracy of the expert sub-model on the difficult sample validation set surpasses that of the cell classification model, the training is completed, and the trained expert sub-model is obtained.

6. The AI-automated detection system for atypical lymphocytes in peripheral blood according to claim 1, characterized in that, The extraction process for the second multimodal morphological features includes: Digital images of the stained second peripheral blood smear to be tested are acquired using a microscope digital scanning system; Using a deep learning-based cell detection algorithm, all candidate lymphocytes were located and cropped from the digital image of the second peripheral blood smear. For each candidate lymphocyte image, the nucleus, cytoplasm, and cell body are segmented to obtain the nuclear mask, cytoplasmic mask, and cell body mask; Based on the nuclear mask, nuclear morphological modal features are calculated, including nuclear shape category, nuclear misalignment degree, nuclear chromatin roughness and structure; Based on the cytoplasmic mask, cytoplasmic modal characteristics are calculated, including cytoplasmic staining characteristics, number and distribution of vacuoles, cytoplasmic transparency, and presence of granules. Based on the cell body mask, cell body morphological modal features are calculated, including cell body size, shape regularity, and presence or absence of pseudopodia. Based on the spatial relationship between the nuclear mask, cytoplasmic mask, and cell body mask, spatial relationship modal features are calculated, including the nucleocytoplasmic ratio and the relative positions of the nucleus and cytoplasm. The nuclear morphological modal features, cytoplasmic modal features, cell body morphological modal features, and spatial relationship modal features extracted from the candidate cell are spliced ​​and normalized to form the second multimodal morphological feature vector of the cell.

7. The AI-automated detection system for atypical lymphocytes in peripheral blood according to claim 6, characterized in that, The process for obtaining the preliminary classification results and corresponding confidence scores of the candidate cells includes: Load the trained cell classification model and batch input the second multimodal morphological feature vectors of candidate cells into the trained cell classification model; The cell classification model extracts deep local features from the input features, then fuses the deep local features with global contextual information through a self-attention mechanism, and finally outputs a probability distribution vector. The category corresponding to the maximum probability value in the probability distribution vector is taken as the preliminary classification result of the candidate cell; The maximum probability value itself is used as the confidence level corresponding to the preliminary classification result.

8. The AI-automated detection system for atypical lymphocytes in peripheral blood according to claim 7, characterized in that, The process for determining whether a candidate cell is a problematic cell includes: A preset confidence threshold is used to distinguish between problematic and non-problematic cells; The confidence score of each candidate cell is compared with the confidence threshold. If the confidence level of a candidate cell is lower than the confidence threshold, the candidate cell is determined to be a difficult cell. If the confidence level of a candidate cell is not lower than the confidence threshold, then the candidate cell is determined to be a non-problematic cell.

9. The AI-automated detection system for atypical lymphocytes in peripheral blood according to claim 8, characterized in that, The process for obtaining the final classification result includes: For candidate cells identified as problematic cells, the trained expert sub-model is loaded. The second multimodal morphological feature vector corresponding to the problematic cells is input into the trained expert sub-model; The expert sub-model processes the input features and outputs the secondary classification result of the problematic cell; For candidate cells that are identified as non-problematic cells, their preliminary classification results are directly used as their final classification results. The final classification results of all non-problematic cells and the secondary classification results of all problematic cells are combined to form a set of final classification results for all candidate cells.

10. The AI-automated detection system for atypical lymphocytes in peripheral blood according to claim 9, characterized in that, The process for generating a tiered report includes: The number of cells classified as atypical lymphocytes is counted in the final classification result set of all candidate cells; Based on the number of atypical lymphocytes and the total number of lymphocytes, calculate the percentage of atypical lymphocytes; Based on a preset clinical significance threshold, the percentage of atypical lymphocytes is mapped to the corresponding risk level; Generate a structured report containing the percentage of atypical lymphocytes, the risk level, and images of all cells identified as atypical lymphocytes and their location information in the original smear.

Images