A cervical cell seven-classification method and system based on hierarchical routing and boundary expert fusion

Abstract

The application discloses a cervical cell seven-classification method and system based on hierarchical routing and boundary expert fusion, and belongs to the technical field of image processing. In order to solve the technical problem that the existing technology directly adopts a single multi-classification model for cell classification, resulting in unreliable gray area boundary discrimination and finally leading to unstable cell classification result discrimination, the application utilizes multi-dimensional morphological indexes and detection confidence to calculate quality points; single cells and cell groups and halo cells are shunted for processing, the cell groups and the halo cells are directly outputted in types, the single cells enter subsequent processes, different morphological cells are avoided from being mixed into the same classifier, the single model training target inconsistency and feature deviation problem is solved; the single cell five-classification basic probability is acquired firstly, target areas of ASC-US and LSIL and ASC-H and HSIL are screened out through boundary uncertainty comprehensive points and quality points, finally probability is obtained through fusion of morphological indexes, bias probability and fusion strength, and the class with the maximum probability is selected as the seven-classification result. The application is used for cervical cell classification.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method and system for classifying cervical cell images. Background Technology

[0002] Cervical cytology screening is an important means of early detection and risk stratification of cervical cancer and precancerous lesions. Clinically, the Bethesda system is often used for cell morphology interpretation. In addition to normal squamous epithelial cells (NILM), actual samples may also contain atypical squamous cells of undetermined significance (ASC-US), low-grade squamous intraepithelial lesions (LSIL), atypical squamous cells that cannot rule out high-grade lesions (ASC-H), and high-grade squamous intraepithelial lesions (HSIL). Various abnormal cells gradually transition in multiple dimensions such as nuclear-cytoplasmic ratio, nuclear hyperchromatography, atypia, nuclear membrane irregularity, and staining homogeneity, without clear boundaries, making the differentiation between ASC-US and LSIL, and ASC-H and HSIL, significantly more difficult than other categories.

[0003] Furthermore, cell clumps are commonly found in actual slide preparation, and cell overlap, adhesion, and occlusion lead to poor stability in segmentation and single-cell feature calculation. Meanwhile, hollow cells (cells with hollowed-out structures) associated with HPV infection have characteristic perinuclear vacuolar structures, whose morphology easily interferes with conventional cell classification models, making them prone to misclassification as abnormally elevated or decreased levels. They also frequently coexist with lesions such as LSIL, further exacerbating class boundary confusion and affecting classification accuracy.

[0004] Existing automatic classification methods or systems often directly employ a single multi-classification model to output the probability distribution of the input cell image. In real clinical data, the gray area boundaries between ASC-US and LSIL, and between ASC-H and HSIL, often exhibit uncertainties such as "close probabilities, insufficient confidence, and fluctuating ranking," leading to unstable classification results for cell images by existing automatic classification methods. Summary of the Invention

[0005] To overcome the technical problem that existing technologies directly use a single multi-classification model for cell classification, resulting in unreliable gray area boundary discrimination and ultimately unstable cell classification results, this invention provides a cervical cell seven-classification method, system, and medium based on hierarchical routing and boundary expert fusion.

[0006] This invention is achieved through the following technical solution:

[0007] This invention provides a seven-classification method for cervical cells based on hierarchical routing and boundary expert fusion, comprising:

[0008] S1: Acquire cervical cytology images and preprocess them to obtain standardized images;

[0009] S2: Perform target detection on the standardized image to obtain the target region and the corresponding detection confidence level. ;

[0010] S3: Morphological index scoring and detection confidence based on the target region Determine the quality score of the current target area Based on quality score Set fusion strength ;

[0011] The morphological indicators are adopted using any one of the following two combinations:

[0012] Combination 1 includes at least nucleocytoplasmic ratio, nuclear dark staining ratio, degree of heteromorphism, area ratio, nuclear irregularity, and nuclear clarity;

[0013] Combination 2 includes at least nucleocytoplasmic ratio, nuclear deep staining, degree of atypia, area ratio, nuclear irregularity, and nuclear clarity;

[0014] S4: Determine the type of the target region, including single cells, cell clusters, and empty halo cells; if it is a single cell, execute S5; otherwise, directly output the current type as the cervical cell seven-class classification result.

[0015] S5: Obtain the basic probability distribution of the target region for the five categories, including NILM, ASC-US, LSIL, ASC-H, and HSIL. ;

[0016] When the two categories with the highest probabilities in the five-category basic probability distribution are ASC-US and LSIL, or ASC-H and HSIL, a boundary uncertainty composite score is constructed. If the boundary uncertainty composite score and the quality score... If the preset conditions are met simultaneously, the target area is determined to be a region with uncertain boundaries and S6 is executed.

[0017] Otherwise, the five-category basic probability distribution will be used. The cell type with the highest probability in the classification is used as the result of the seven-category cervical cell classification.

[0018] The order of the two highest categories is not specified.

[0019] S6: Morphological index scoring, morphological index and quality score based on the current boundary uncertainty region. Calculate the probability of bias towards LSIL or HSIL; based on fusion strength Five-category basic probability distribution The final probability distribution of the five-class classification is obtained by considering the probability of leaning towards LSIL or HSIL. The highest probability value among them is used as the classification result of cervical cells.

[0020] Furthermore, based on the detection confidence level The quality score of the candidate target region is calculated using the kernel clarity score. The specific process includes: setting the detection confidence level separately. By normalizing the kernel clarity score, the normalized detection confidence is obtained. Compared with normalized kernel clarity score Then, the normalized detection confidence level is used. Compared with normalized kernel clarity score The quality score is obtained by product or weighted summation. .

[0021] Furthermore, the quality-based The rules for setting the fusion strength include:

[0022] When the quality is divided When the mass score is less than the first preset mass score threshold, the fusion strength is set to the first fusion strength value;

[0023] When the quality is divided When the value of the fusion strength is greater than or equal to the first preset quality score threshold and less than the second preset quality score threshold, the value of the fusion strength is set to the second fusion strength value.

[0024] When the quality is divided When the value of the fusion strength is greater than or equal to the second preset quality score threshold, the value of the fusion strength is set to the third fusion strength value;

[0025] The first preset quality score threshold is less than the second preset quality score threshold, and the second fusion strength value is less than the third fusion strength value.

[0026] Furthermore, the method for determining the halo cells includes: calculating the halo evidence score for the target region; if the score exceeds a threshold... Furthermore, if the nuclear features meet the morphological constraints of HPV-related changes, the current target region will be identified as an empty halo cell.

[0027] The halo evidence score is calculated using the following formula:

[0028] ,

[0029] In the formula, Indicates evidence of vertigo; Indicates the weighting coefficients, all of which are greater than 0 and ; Represents the mean operator; Represents the variance operator; Represents the set of pixel intensities in the kernel region; This represents the set of pixel intensities within the annular halo region; Indicates the ring width; express The result after normalization; express The result after normalization; express The result after normalization;

[0030] The morphological constraints for HPV-related changes include: area ratio score or nucleocytoplasmic ratio score, nuclear dark staining ratio score, degree of irregularity score, nuclear irregularity score, and mass score. When all preset conditions are met, the constraints on HPV-related morphological changes are confirmed to be satisfied.

[0031] Furthermore, the method for determining the cell cluster includes:

[0032] If the number of cell instances in the current target area is greater than or equal to a preset threshold, it is determined to be a cluster of cells;

[0033] If there are no cell instances in the current target area, the number of cell instances is approximated by the nuclear peak count, connected component count, or the ratio of the region area to the median area of ​​a single cell. If the approximation is greater than or equal to a preset threshold, it is determined to be a cluster of cells.

[0034] Furthermore, the boundary uncertainty comprehensive analysis described in S5 The calculation formula is:

[0035] ,

[0036] In the formula, , , Indicates weight, + + =1; Represents the probability of five categories The marginal difference in probability between the top 1 and top 2 probabilities. , Represents the probability of five categories The highest probability value in, Represents the probability of five categories The second largest probability value; Represents the normalized five-class probability The entropy value, Represents the probability of five categories The entropy value; ,in Represents the probability of five categories Cell category index Indicates the first The probability value corresponding to the cell type.

[0037] Furthermore, when the two classes with the highest probabilities in the five-category basic probability distribution are ASC-US and LSIL, the scores are based on the area ratio score, heterogeneity score, nuclear staining ratio score, nuclear irregularity, and quality score of the current target region. Calculate the probability of bias towards LSIL. The expression is:

[0038] ,

[0039] ,

[0040] ,

[0041] In the formula, This represents the input feature vector of the left branch; This represents the learnable weight vector of the left branch; Represents the learnable bias vector of the left branch; This indicates a linear score for the unactivated left branch; Indicates the activation function;

[0042] Furthermore, when the two categories with the highest probabilities in the five-category basic probability distribution are ASC-H and HSIL, respectively, the nucleocytoplasmic ratio score, heteromorphism score, nuclear dark staining ratio score, nuclear irregularity, and mass score of the current target region are used as the basis for further analysis. Calculate the probability of bias towards HSIL. :

[0043] ,

[0044] ,

[0045] ,

[0046] In the formula, This represents the input feature vector of the right branch; This represents the learnable weight vector of the right branch; Represents the learnable bias vector of the right branch; This represents the linear score of the right branch that has not undergone an activation function; This represents the activation function.

[0047] Furthermore, the final probability distribution of the five categories The method for determining it includes the following steps:

[0048] Let the cervical cells in the current uncertain region be type A and type B, with base probabilities of respectively... , The probability of being biased towards type B is Using the sum of the basic probabilities of type A and type B as a conserved quantity, and utilizing the probability biased towards type B... and the fusion strength The basic probabilities of type A and type B respectively , Adjustments are made to obtain the final probabilities of type A and type B. , In the five-part differential autoimmune classification of single cells, the probability of type C, excluding types A and B, remains consistent with the baseline probability; the formula is expressed as:

[0049] ,

[0050]

[0051] ,

[0052] .

[0053] The second aspect of the present invention provides a cervical cell seven-classification system based on hierarchical routing and boundary expert fusion, comprising: an image preprocessing module, a target detection module, a quality gating module, a hierarchical routing module, a boundary expert triggering module, and a boundary expert classification module;

[0054] The image preprocessing module is used to acquire cervical cytology images and perform preprocessing to obtain standardized images;

[0055] The target detection module is used to perform target detection on the standardized image, obtaining the target region and the corresponding detection confidence level. ;

[0056] The quality gating module is used to score and determine the detection confidence level based on the morphological indicators of the target region. Determine the quality score of the current candidate target region Based on quality score Set fusion strength The morphological indicators include at least the nucleocytoplasmic ratio, nuclear hyperchromaticity or hyperchromaticity ratio, degree of heteromorphism, area ratio, nuclear irregularity, and nuclear clarity.

[0057] The hierarchical routing module is used to classify candidate target regions, including single cells, cell clusters, and halo cells; if it is a single cell, it enters the boundary expert triggering module; otherwise, it directly outputs the current type as the cervical cell seven-class classification result.

[0058] The boundary expert triggering module is used to obtain the five-category basic probability distribution of the target area, including NILM, ASC-US, LSIL, ASC-H, and HSIL. When the two categories with the highest probabilities in the five-category basic probability distribution are ASC-US and LSIL, or ASC-H and HSIL, a boundary uncertainty comprehensive score is constructed. If the boundary uncertainty comprehensive score and the quality score simultaneously meet preset conditions, the target area is determined to belong to the boundary uncertainty area and enters the boundary expert module; otherwise, the cell type corresponding to the highest probability in the five-category basic probability distribution is taken as the cervical cell seven-category result.

[0059] The boundary expert module is used to score morphological indicators, morphological indicators and quality scores based on the current boundary uncertainty region. Calculate the probability of bias towards LSIL or HSIL; based on fusion strength Five-category basic probability distribution The final probability distribution of the five-class classification is obtained by considering the probability of leaning towards LSIL or HSIL. The highest probability value among them is used as the result of the seven-category cervical cell classification.

[0060] The beneficial effects of this invention are:

[0061] This invention calculates a quality score using multi-dimensional morphological indicators and detection confidence, and sets the fusion intensity accordingly. It separates single cells from cell clusters and halo cells, with cell clusters and halo cells directly outputting their types, while single cells enter subsequent processes. This avoids different morphological cells being mixed into the same classifier, solving the problems of inconsistent training targets and feature shifts in single-model training. First, it obtains the basic five-class probabilities of single cells, then uses the boundary uncertainty comprehensive score combined with the quality score to screen out the target regions for ASC-US and LSIL, and ASC-H and HSIL. Finally, it fuses morphological indicators, bias probabilities, and fusion intensity to obtain the final probability, solving the problems of similar probabilities and unstable discrimination in gray area samples. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0063] Figure 1 This is a flowchart illustrating a seven-classification method for cervical cells based on hierarchical routing and boundary expert fusion, as one embodiment of the method of the present invention.

[0064] Figure 2This is a schematic diagram of the structure of a cervical cell seven-classification system based on hierarchical routing and boundary expert fusion in one embodiment of the method of the present invention;

[0065] Figure 3 This is a schematic diagram of morphological feature extraction in one embodiment of the method of the present invention;

[0066] Figure 4 This is a schematic diagram comparing the five-class probability distribution before and after the boundary expert fusion mechanism is applied in one embodiment of the method of the present invention. Detailed Implementation

[0067] The technical solutions of 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. Specific Implementation Method 1

[0069] Please refer to Figure 1 This embodiment provides a seven-classification method for cervical cells based on hierarchical routing and boundary expert fusion, including the following steps:

[0070] Step S1: Obtain cervical cytology images and perform preprocessing to obtain standardized images. Specifically, this includes steps S101 to S102:

[0071] Step S11: Obtain the cervical cytology image to be processed and perform basic legality verification on it.

[0072] The verification includes, but is not limited to: verifying whether the pixel size, physical resolution, and number of image channels of the image meet preset standards.

[0073] Step S12: Perform standardized preprocessing on the cervical cytology images.

[0074] The standardized preprocessing includes at least one of the following: color normalization, contrast stretching, and lightweight noise reduction.

[0075] Step S2: Perform target detection on the standardized image to obtain the target region and the corresponding detection confidence level. Specifically, target detection or instance segmentation is performed on the cervical cytology image, and for each candidate target region, a corresponding detection confidence score is generated and output. The candidate target region is processed using a non-maximum suppression algorithm to remove overlapping and redundant regions and merge duplicate targets.

[0076] Step S3: Calculate the morphological indicators and morphological indicator scores for each candidate target region, and then use the morphological indicator scores and detection confidence levels as a basis for the calculation. The quality score of the candidate target region is obtained. Based on quality score Set the blending strength. Specifically:

[0077] The morphological indicators are adopted using any one of the following two combinations:

[0078] Combination 1 includes at least: nucleocytoplasmic ratio Area ratio nuclear deep staining ratio Nuclear irregularity Degree of heterogeneity and nuclear clarity .

[0079] Combination 2 includes at least the nucleo-mass ratio. nuclear deep staining Degree of heterogeneity Area ratio Nuclear irregularity and nuclear clarity ;

[0080] Please refer to Figure 3 , Figure 3 This is a schematic diagram of cervical cell morphological feature extraction provided in this embodiment.

[0081] The morphological index score is mapped to a scoring item with a uniform dimension through monotonic mapping or piecewise mapping based on the quantiles of the training set. This embodiment preferably uses piecewise linear monotonic mapping based on the quantile thresholds of the training set to obtain the nucleocytoplasmic ratio score. Area ratio score Nuclear deep staining score Nuclear deep staining ratio score Nuclear irregularity score Degree of heterogeneity score Nuclear clarity score The range of each of the scoring items is The rating mapping satisfies monotonicity, meaning that the higher the score of a rating item, the more the morphological indicator is biased towards an "abnormal direction".

[0082] (1) The nucleus-mass ratio , where is the ratio of nuclear area to cytoplasmic area, where cytoplasmic area = cell area - nuclear area. In this embodiment, the nucleocytoplasmic ratio is... The calculation formula is:

[0083]

[0084] In the formula, Indicates the core area; Indicates cell area; This indicates a term to prevent division by zero.

[0085] (2) The area ratio The ratio of the nuclear area to the cell area is used in this embodiment. The calculation formula is:

[0086]

[0087] In the formula, This indicates a term to prevent division by zero.

[0088] (3) The nuclear deep staining ratio is This is the ratio of the average intensity (or H channel intensity) of the nuclear region to the cytoplasmic region. In this embodiment, the nuclear staining intensity ratio is calculated using the average intensity of the nuclear and cytoplasmic regions. The formula is:

[0089]

[0090] In the formula, This represents the set of pixel intensities in the kernel region, i.e., the pixel values ​​sampled within the kernel mask; This represents the set of pixel intensities in the cytoplasmic region; This is the mean operator, which means taking the average intensity of all pixels in the region; This indicates a term to prevent division by zero.

[0091] (4) The aforementioned nuclear irregularity (roundness inverse measure) The calculation formula is:

[0092]

[0093] In the formula, Indicates the core area. Indicates the perimeter of the core. This indicates a term to prevent division by zero.

[0094] (5) The degree of heterogeneity The calculation is performed using a weighted approach with multiple features, and the formula is as follows:

[0095]

[0096] In the formula, Indicates configurable or learnable parameters; Represents the normalized kernel area ; This represents the average pixel intensity of the kernel region after normalization. This represents the standard deviation of pixel intensity in the kernel region after normalization.

[0097] (6) The nuclear clarity The calculation method is as follows:

[0098] Let the grayscale image of the cell nucleus region be... The kernel mask (the set of pixels within the kernel region) is The Laplace operator is ,but:

[0099]

[0100]

[0101] The kernel clarity can be expanded into the form of "mean + variance" as follows:

[0102]

[0103]

[0104] in, This represents the mean of the Laplace response within the core region; Represents pixels within the kernel region The Laplacian response value at that location is used to characterize the intensity of local high-frequency changes at that pixel position;

[0105] Based on the detection confidence level With nuclear clarity Calculate the quality score of the candidate target region The specific process includes:

[0106] First, let's analyze the detection confidence level. With nuclear clarity score Normalization mapping is performed to obtain the normalized detection confidence and normalized kernel clarity score:

[0107]

[0108]

[0109] in, This represents the normalized detection confidence level; This represents the normalized kernel clarity score. This represents the truncation function. ; To prevent division by zero of the stable term; , These represent the lower and upper bounds of the detection confidence mapping threshold, respectively, used for linear scaling; , These represent the lower and upper bounds of the kernel clarity score mapping threshold, respectively, used to map the original kernel clarity score. Linear scaling to Interval. In this embodiment, , , , , .

[0110] Then the normalized detection confidence level is used. Compared with normalized kernel clarity score The mass score is obtained by multiplication or weighted summation. .

[0111] The mass score is calculated using a product method. The formula is: .

[0112] The quality score is calculated using a weighted summation method. The formula is: ,in , Indicates the weighting coefficient. .

[0113] This embodiment uses a weighted summation method to calculate the quality score. , , Using a weighted approach can control the overall quality from decreasing synchronously when any quality item is low.

[0114] The quality score The rules for setting the fusion strength include:

[0115] When the quality is divided When the score is less than the first preset quality score threshold, set the first fusion strength value and mark it as needing review. This can be done by setting need_review=1.

[0116] When the quality is divided When the value of the fusion strength is greater than or equal to the first preset quality score threshold and less than the second preset quality score threshold, the value of the fusion strength is set to the second fusion strength value.

[0117] When the quality is divided When the value of the fusion strength is greater than or equal to the second preset quality score threshold, the value of the fusion strength is set to the third fusion strength value;

[0118] The first preset quality score threshold is less than the second preset quality score threshold, and the second fusion strength value is less than the third fusion strength value.

[0119] In this embodiment, the first preset quality score threshold is 0.35, the second preset quality score threshold is 0.65, the first fusion strength value is 0, the second fusion strength value is 0.3, and the third fusion strength value is 0.8.

[0120] Quality score Based on the detection confidence and kernel clarity scores, the algorithm is used to suppress low-quality samples from triggering error correction and output review suggestions, thereby reducing the risk of "making more mistakes with each correction".

[0121] Step S4: Determine the type of the target region, including single cells, cell clusters, and empty halo cells; if it is a single cell, execute S500; otherwise, directly output the current type as the classification result of cervical cells.

[0122] (1) Method for determining empty halo cells:

[0123] Calculate the halo evidence score for candidate targets. If it exceeds the evidence score threshold... Furthermore, if the nuclear features satisfy the morphological constraints of HPV-related changes, the current candidate target is identified as a halo cell. The halo evidence score is calculated using the following formula:

[0124] (12)

[0125] In the formula, The vacuolar evidence score is used to measure the significance of the "peripheral vacuolar / ring-shaped halo" structure. This represents the weighting coefficients used to balance the three types of evidence: brightness difference, texture sparsity, and ring width. Contributions in and ; This represents the mean operator, which calculates the average intensity of pixels within a region. This represents the variance operator, which calculates the variance of pixel intensity within a region (depicting texture / intensity fluctuations). This represents the set of pixel intensities in the kernel region, i.e., within the kernel mask. Internally sampled pixel values; This represents the set of pixel intensities in the annular halo region, i.e., the region masked within the annular region. Internally sampled pixel values; It is constructed by "expanding the nuclear mask to obtain a ring-shaped region"; Indicates the ring width (thickness of the halo band); express The result after normalization; express The result after normalization; express The result after normalization.

[0126] In this embodiment, the formula for calculating the ring width is: ,in This represents the area of ​​the annular region (in pixels). The kernel perimeter (pixel length). To prevent division by zero of the stable term, the value is taken as... In this embodiment, the annular region mask ,in, This represents the morphological expansion radius corresponding to the outer boundary of the generated annular region. This represents the morphological dilation radius corresponding to the boundary of the generated annular region. Both are in pixels and satisfy the following conditions: .

[0127] The evidence score threshold can be calculated by taking the empty halo evidence score of all samples on the training set or validation set, and then statistically analyzing the empty halo evidence scores of non-empty halo samples into a distribution. Then from the distribution The highest quantile was used as the threshold for the evidence score.

[0128] In this embodiment, the evidence threshold is used. The range of values ​​is , preferably 0.70.

[0129] The criteria for determining whether the nuclear features meet the morphological constraints of HPV-related changes are as follows: The area ratio score (or nucleocytoplasmic ratio score), dark staining ratio score, degree of irregularity score, nuclear irregularity score, and mass score all meet preset conditions, confirming that the constraints of HPV-related morphological changes are satisfied. In this implementation, the standard is set to 0.20. Area ratio score 0.75 (or 0.35) Nucleus-to-mass ratio score 0.80), Darker shade score 0.55, Degree of Abnormality Score 0.20, Degree of Abnormality Score 0.85, mass fraction 0.35.

[0130] (2) Methods for dividing clumped cells:

[0131] If the number of cell instances in the current target area is greater than or equal to a preset threshold, it is determined to be a cluster of cells; if there are no cell instances in the current target area, the number of cell instances can be approximated using nuclear peak count / connected region count / or the ratio of region area to the median area of ​​a single cell. , Indicates the area of ​​the candidate region. The median area of ​​a single cell. This represents the set of single-cell area samples.

[0132] This invention uses object type routing to treat cell clusters and koilocytes / hollows as pre-classification categories, reducing their systematic interference on single-cell branches (main model five-class classification and boundary error correction). It is particularly suitable for complex scenarios with occlusion and overlap, difficult boundary segmentation, and koilocyte / hollow structure interference, thereby improving overall classification consistency and robustness.

[0133] Step S5: Obtain the basic probability distribution of the target region, including NILM, ASC-US, LSIL, ASC-H, and HSIL. When the two categories with the highest probabilities in the five-category basic probability distribution are ASC-US and LSIL, or ASC-H and HSIL, a boundary uncertainty composite score is constructed. If the boundary uncertainty composite score and the quality score are... If both preset conditions are met, execute S6; otherwise, apply the five-category basic probability distribution. The cell type with the highest probability in the classification is used as the classification result for cervical cells.

[0134] Step S501: Input the single-cell image into the convolutional neural network and output the five-class classification probability. .

[0135] In this embodiment, the convolutional neural network is a ResNet-50 network, and its original final classification layer is replaced with a fully connected layer with an output dimension of 5. The fully connected layer directly outputs the raw logits of each category, which are then converted into probability distributions corresponding to the NILM, ASC-US, LSIL, ASC-H, and HSIL categories by the Softmax function. The maximum probability is used as the confidence score for the final prediction. To improve the interpretability of the confidence level, a temperature calibration operation was performed on the logits, so that the calibrated confidence level is closer to the actual prediction accuracy.

[0136] The temperature calibration can be expressed as:

[0137] (13)

[0138] in, For the unlabeled first Primitive logit; This represents the first value obtained after temperature calibration and Softmax normalization. Class prediction probability; Indicates cell category index, These correspond to NILM, ASC-US, LSIL, ASC-H, and HSIL, respectively. This represents the temperature calibration parameter, a scalar value greater than 0, used to adjust the smoothness of the output probability distribution.

[0139] Step S502: When the two classes with the highest probabilities in the five-category probability distribution are ASC-US and LSIL, or ASC-H and HSIL, respectively, construct the boundary uncertainty composite score. If the boundary uncertainty composite score and the quality score are... When the preset conditions are met, the corresponding expert is triggered; otherwise, the cell category corresponding to the maximum probability in the five-category probability distribution is directly used as the cervical cell classification result. The order of the Top 1 and Top 2 probability categories is not limited, only requiring a combination of ASC-US and LSIL or ASC-H and HSIL.

[0140] The boundary uncertainty comprehensive analysis The calculation formula is:

[0141] (14)

[0142] In the formula, , , Indicates weight, + + =1; Represents the probability of five categories The marginal difference in probability between the top 1 and top 2 probabilities. , Represents the probability of five categories The highest probability value in, Represents the probability of five categories The second largest probability value; Represents the normalized five-class probability The entropy value, Represents the probability of five categories The entropy value, .

[0143] When the boundary is uncertain, the comprehensive analysis Greater than or equal to the preset trigger threshold and quality score If the value is greater than or equal to the quality gating threshold, the current target area is determined to be a boundary uncertain area, and step S600 is executed; otherwise, the cell type corresponding to the highest probability in the five-category basic probability distribution is taken as the cervical cell seven-category result.

[0144] Step S6: Based on the morphological index score, morphological index and quality score of the current boundary uncertain region. Calculate the probability of bias towards LSIL or HSIL; based on fusion strength Five-category basic probability distribution The final probability distribution of the five-class classification is obtained by considering the probability of leaning towards LSIL or HSIL. The classification result of cervical cells is then determined based on the highest probability value among the classifications. This includes the following steps:

[0145] Step S61: Output the probability that the region is biased towards LSIL or HSIL using a two-branch calculation process. The two-branch calculation process can be implemented based on logistic regression, lightweight multilayer perceptron (MLP), or gradient boosting decision tree (GBDT). In this embodiment, logistic regression is used to complete the probability calculation. The two-branch calculation process includes an Expert-L branch and an Expert-H branch.

[0146] Specifically, this includes the following two scenarios:

[0147] Scenario 1: Five-class classification probability of the current boundary uncertain region When the two categories with the highest probabilities of Top 1 and Top 2 are ASC-US and LSIL respectively, the probability of bias towards LSIL is calculated using the Expert-L branch.

[0148] Specifically, based on the morphological index score (area ratio score) of the current target region (single cell), Degree of heterogeneity score Nuclear deep staining ratio score ), morphological indicators (nuclear irregularity) ) and quality score Constructing the left branch input feature vector Given the Expert-L branch, output the probability that the value is more biased towards LSIL in the {ASC-US,LSIL}. :

[0149] (15)

[0150] (16)

[0151] (17)

[0152] In the formula, This represents the input feature vector of the Expert-L branch; This represents the learnable weight vector of the Expert-L branch; This represents the learnable bias vector for the Expert-L branch; This indicates a linear score without going through the Expert-L branch of the Sigmoid function; This represents the Sigmoid activation function, which is defined as follows: , This indicates the object being processed by the current activation function.

[0153] Scenario 2: When the boundary region has an uncertain five-class classification probability When the two categories of the top 1 and top 2 probabilities are ASC-H and HSIL respectively, the probability of bias towards HSIL is calculated using the Expert-H branch.

[0154] Specifically, based on the morphological index score (nucleocytoplasmic ratio score) of the current target region (single cell), Degree of heterogeneity score Nuclear deep staining ratio score ), morphological indicators (nuclear irregularity) ) and quality score Constructing input feature vectors from branches Given the Expert-H branch, output the probability that the value is more biased towards HSIL in the {ASC-H, HSIL}. :

[0155] (18)

[0156] (19)

[0157] (20)

[0158] In the formula, This represents the input feature vector of the right branch; This represents the learnable weight vector of the Expert-H branch; This represents the learnable bias vector for the Expert-H branch; This indicates a linear score without going through the Expert-H branch of the Sigmoid function; This represents the Sigmoid activation function, which is the same as the activation function in the Expert-L branch.

[0159] During the execution of scenarios one and two above, to avoid over-correction of low-quality samples, a quality score can be introduced to optimize the linear scoring that has not undergone sigmoid, which can be expressed as: or , This represents the optimized linear scoring result. express or , This indicates the threshold for triggering quality gating. Indicates the quality modulation intensity coefficient. When the sample quality is high, the inhibitory effect on the scoring is weakened; when the sample quality is low, the scoring is further suppressed.

[0160] Step S62: Based on the probability that the current boundary uncertainty region is biased towards LSIL or HSIL, determine the final probability values ​​of the two classes in the current boundary uncertainty region, thereby determining the final five-class probability distribution. .

[0161] Let the cervical cells in the current uncertain region be type A and type B, with initial probabilities of respectively. , The probability of being biased towards type B is Using the sum of the initial probabilities of type A and type B as a conserved quantity, and utilizing the probability biased towards type B... and the fusion strength Initial probabilities for type A and type B respectively , Adjustments are made to obtain the final probabilities of type A and type B. , In the five-class differential autoimmune disease of single cells, all types except A and B, type C maintains the same probability as the initial probability. This can be represented as:

[0162] (twenty one)

[0163] (twenty two)

[0164] (twenty three)

[0165] (twenty four)

[0166] Local probability fusion only redistributes probabilities between the two classes at the boundary, while keeping the probabilities of other classes unchanged, thereby avoiding unnecessary perturbations to NILM and non-boundary classes and ensuring the stability of the overall output.

[0167] This invention decouples the two most easily confused gray-zone boundaries in clinical practice (ASC-US vs LSIL, ASC-H vs HSIL) from the unified seven-class / five-class classification problem: only when the main model presents uncertainty regarding this boundary is the corresponding boundary expert triggered to perform a two-choice refinement judgment, thereby specifically improving the consistency and accuracy of gray-zone boundary discrimination without increasing the overall model complexity. Please refer to [reference needed]. Figure 4 , Figure 4 This is a schematic diagram comparing the five-class probability obtained in step S5 and the final five-class probability distribution obtained after processing in step S6 in this embodiment.

[0168] Step S700: Output the seven-category label for cervical cells.

[0169] If the current target area is determined to be an empty halo cell, then the empty halo cell label is output directly;

[0170] If the current target region is determined to be a cell cluster, the corresponding cell cluster label is output.

[0171] If the current target region is determined to be a single cell, then output the final five-class classification probability of the single cell. The cell classification corresponding to the class with the highest probability is used as the final label.

[0172] In addition to the cell classification labels mentioned above, the method of this invention can also output whether a boundary expert is triggered, whether a review is recommended, and cell cluster subclass labels, so that the decision path of each sample can be audited and traced, which facilitates clinical quality control and human-machine collaborative review.

[0173] During the training phase of the method of this invention, for the sample set corresponding to each parent class label, one, two, or all three indicators of comprehensive severity score, quality score, and boundary uncertainty comprehensive score are selected for binning. For each selected indicator, the sample set is divided into at least two corresponding bins based on its quantile threshold, namely the severity bin, the quality-limited bin, and the boundary bin. The parent class label is cross-combined with the labels of each bin to obtain the fission subclass label. The fission subclass label is used for auxiliary classification supervision, sample reweighting, hard example sampling, or course learning.

[0174] The parent class labels include NILM, ASC-US, LSIL, ASC-H, HSIL, halo cells, and cell clusters;

[0175] The comprehensive severity score is obtained by weighting and summing the nucleocytoplasmic ratio score, atypicality score, nuclear deep staining ratio score, area ratio score, and nuclear irregularity score using a preset weighting rule.

[0176] Using a comprehensive severity score to construct fission subclass labels for training samples can enhance the main classification model's ability to represent fine-grained differences in lesion severity. The construction steps include:

[0177] (1) Extract preset morphological indicators for each training sample and map the morphological indicators to corresponding scoring items;

[0178] (2) The scoring items are weighted and fused according to the preset weighting rules to obtain the comprehensive severity score of each training sample. ;

[0179] (3) Within the training sample set corresponding to each parent class label, the statistics are... quantile threshold;

[0180] (4) Based on the quantile threshold, the corresponding parent class samples are divided into three severity buckets: low severity, medium severity, and high severity;

[0181] (5) Combine the parent class tag with the severity bucket tag to obtain the fission subclass tag;

[0182] (6) During the training phase, auxiliary classification supervision, sample reweighting, hard example sampling or course learning are performed based on the fission subclass labels; during the inference phase, only the classification results corresponding to the original parent class labels are output, and the fission subclass labels are not output.

[0183] The comprehensive severity score The calculation equation is as follows:

[0184] (25)

[0185] In the formula, This represents a learnable parameter, whose value combinations can be configured specifically for different boundary scenarios.

[0186] In this embodiment, the comprehensive severity score The quantile binning rule is as follows:

[0187] (26)

[0188] (27)

[0189] (28)

[0190] In the formula, Indicates low severity. Indicates moderate to severe severity. Pxx indicates high severity, P33 indicates the xxth percentile, and P33 indicates the threshold that is only greater than 33% of the training samples. This threshold is automatically generated by statistical analysis of the training set data and does not require manual intervention.

[0191] The training process of the method of the present invention includes the following steps:

[0192] First, training samples with parent class labels are obtained. The training samples are preprocessed, region detection and feature extraction are performed to form sample input for model training. For the sample set corresponding to each parent class label, the samples are binned according to one, two or all three indicators mentioned above: comprehensive severity score, quality score and boundary uncertainty comprehensive score. The parent class label and the bin label are cross-combined to generate fission subclass labels.

[0193] During training, the fission subclass labels can be used as auxiliary supervision information during the training phase. Without changing the final seven-class classification output system, based on the shared feature extraction backbone network, while the main classification branch outputs the classification results corresponding to the original parent class labels, a fission subclass auxiliary branch is added to predict the fission subclass labels, and the fission subclass auxiliary supervision loss is calculated based on the prediction results. The auxiliary supervision loss is jointly optimized with the main classification loss to enhance the model's ability to represent fine-grained morphological differences, boundary samples, and difficult samples.

[0194] The fission subclass labels can also be used for sample reweighting. Specifically, different training weights can be assigned to different samples based on the bucket attributes corresponding to the fission subclass label to which the sample belongs; among them, samples located in high severity buckets, boundary buckets, or quality-constrained buckets can be assigned higher training weights than ordinary samples, so as to improve the model's attention to key boundary samples, difficult samples, and low-quality samples.

[0195] The fission subclass labels can also be used for hard example sampling. Specifically, when constructing training batches, the sampling probability of boundary bucket samples, high severity bucket samples, or quality-constrained bucket samples can be increased, so that the training process pays more attention to easily confused samples and hard samples, thereby improving the model's ability to distinguish between boundary categories.

[0196] The fission subclass labels can also be used for course learning. Specifically, training samples can be trained in a stepwise manner from easy to difficult according to the sample difficulty reflected by the fission subclass labels; for example, high-quality samples with low boundary uncertainty can be used for initial training first, and then boundary bucket samples, quality-constrained bucket samples and high-severity difficult samples can be gradually introduced to improve training stability and model generalization ability.

[0197] In a specific embodiment of the present invention, the application of the fission subclass labels can be one or a combination of several of the following: assisted classification supervision, sample reweighting, hard example sampling, or course learning. After training, an optimized target classification model is obtained; during the inference phase, only the seven-class output corresponding to the original parent class labels is retained, and the fission subclass labels are not output. This improves the model's ability to distinguish boundary samples and hard samples, as well as its overall generalization performance, without requiring additional manual annotation. Specific Implementation Method Two

[0199] This embodiment provides a cervical cell seven-classification system based on hierarchical routing and boundary expert fusion. Please refer to [link / reference]. Figure 2 It includes: image preprocessing module, target detection module, quality gating module, hierarchical routing module, boundary expert triggering module, and boundary expert module;

[0200] The image preprocessing module is used to acquire cervical cytology images and perform preprocessing to obtain standardized images;

[0201] The target detection module is used to perform target detection on the standardized image, obtaining the target region and the corresponding detection confidence level. .

[0202] The quality gating module is used to score and determine the detection confidence level based on the morphological indicators of the target region. Determine the quality score of the current candidate target region Based on quality score Set fusion strength The morphological indicators are adopted using any one of the following two combinations:

[0203] Combination 1 includes at least nucleocytoplasmic ratio, nuclear dark staining ratio, degree of heteromorphism, area ratio, nuclear irregularity, and nuclear clarity;

[0204] Combination 2 includes at least nucleocytoplasmic ratio, nuclear deep staining, degree of atypia, area ratio, nuclear irregularity, and nuclear clarity;

[0205] Quality gating mechanisms enable the system to maintain a conservative strategy under low-quality input, thereby improving the safety and controllability of clinical implementation.

[0206] The hierarchical routing module is used to classify candidate target regions by type, including single cells, cell clusters, and empty halo cells; if it is a single cell, S5 is executed; otherwise, the current type is directly output as the classification result of cervical cells.

[0207] This mechanism prevents "special objects" from being forcibly mixed with single-cell abnormality spectrum classes for training / judgment, reducing the accumulation of systematic errors and serving as a key foundational module for overall stability.

[0208] The boundary expert triggering module is used to obtain the five-class basic probability distribution of the target region, including NILM, ASC-US, LSIL, ASC-H, and HSIL. When the two categories with the highest probabilities in the five-category basic probability distribution are ASC-US and LSIL, or ASC-H and HSIL, a boundary uncertainty composite score is constructed. If the boundary uncertainty composite score and the quality score are... If both preset conditions are met, execute S6; otherwise, apply the five-category basic probability distribution. The cell type with the highest probability in the gray area is used as the classification result of cervical cells; this design conforms to the clinical interpretation rule that "boundaries are more dependent on fine-grained morphological evidence" and can significantly improve the consistency of gray area discrimination.

[0209] The boundary expert module is used to score morphological indicators, morphological indicators and quality scores based on the current boundary uncertainty region. Calculate the probability of bias towards LSIL or HSIL; based on fusion strength Five-category basic probability distribution The final probability distribution of the five-class classification is obtained by considering the probability of leaning towards LSIL or HSIL. The highest probability value among them is used as the classification result of cervical cells.

[0210] The boundary expert module includes an Expert-L submodule and an Expert-H submodule. The Expert-L submodule is used to calculate the probability of bias towards LSIL based on the area ratio score, heterogeneity score, nuclear staining ratio score, nuclear irregularity, and mass score of the current target region when the two most probable categories in the five-category basic probability distribution are ASC-US and LSIL, respectively. The Expert-H submodule is used to calculate the probability of bias towards HSIL based on the nuclear-mass ratio score, heterogeneity score, nuclear staining ratio score, nuclear irregularity, and mass score of the current target region when the two most probable categories in the five-category basic probability distribution are ASC-H and HSIL, respectively.

[0211] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0212] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A seven-classification method for cervical cells based on hierarchical routing and boundary expert fusion, characterized in that, include: S1: Acquire cervical cytology images and preprocess them to obtain standardized images; S2: Perform target detection on the standardized image to obtain the target region and the corresponding detection confidence level. ; S3: Morphological index scoring and detection confidence based on the target region Determine the quality score of the current target area Based on quality score Set fusion strength ; The morphological indicators are adopted using any one of the following two combinations: Combination 1 includes at least nucleocytoplasmic ratio, nuclear dark staining ratio, degree of heteromorphism, area ratio, nuclear irregularity, and nuclear clarity; Combination 2 includes at least nucleocytoplasmic ratio, nuclear deep staining, degree of atypia, area ratio, nuclear irregularity, and nuclear clarity; S4: Determine the type of the target region, including single cells, cell clusters, and empty halo cells; if it is a single cell, execute S5; otherwise, directly output the current type as the cervical cell seven-class classification result. S5: Obtain the basic probability distribution of the target region for the five categories, including NILM, ASC-US, LSIL, ASC-H, and HSIL. ; When the two categories with the highest probabilities in the five-category basic probability distribution are ASC-US and LSIL, or ASC-H and HSIL, a boundary uncertainty composite score is constructed. If the boundary uncertainty composite score and the quality score... If the preset conditions are met simultaneously, the target area is determined to be an area with uncertain boundaries and S6 is executed; Otherwise, the five-category basic probability distribution will be used. The cell type with the highest probability in the classification is used as the result of the seven-category cervical cell classification. The order of the two highest categories is not specified. S6: Morphological index scoring, morphological index and quality score based on the current boundary uncertainty region. Calculate the probability of bias towards LSIL or HSIL; based on fusion strength Five-category basic probability distribution The final probability distribution of the five-class classification is obtained by considering the probability of leaning towards LSIL or HSIL. The highest probability value among them is used as the classification result of cervical cells.

2. The cervical cell seven-classification method based on hierarchical routing and boundary expert fusion according to claim 1, characterized in that, Based on the detection confidence level The quality score of the candidate target region is calculated using the kernel clarity score. The specific process includes: setting the detection confidence level separately. By normalizing the kernel clarity score, the normalized detection confidence is obtained. Compared with normalized kernel clarity score Then, the normalized detection confidence level is used. Compared with normalized kernel clarity score The quality score is obtained by product or weighted summation. .

3. The cervical cell seven-classification method based on hierarchical routing and boundary expert fusion according to claim 2, characterized in that, The quality score The rules for setting the fusion strength include: When the quality is divided When the mass score is less than the first preset mass score threshold, the fusion strength is set to the first fusion strength value; When the quality is divided When the value of the fusion strength is greater than or equal to the first preset quality score threshold and less than the second preset quality score threshold, the value of the fusion strength is set to the second fusion strength value. When the quality is divided When the value of the fusion strength is greater than or equal to the second preset quality score threshold, the value of the fusion strength is set to the third fusion strength value; The first preset quality score threshold is less than the second preset quality score threshold, and the second fusion strength value is less than the third fusion strength value.

4. The cervical cell seven-classification method based on hierarchical routing and boundary expert fusion according to claim 3, characterized in that, The method for determining halo cells includes: calculating the halo evidence score for the target region; if the score exceeds a threshold... Furthermore, if the nuclear features meet the morphological constraints of HPV-related changes, the current target region will be identified as an empty halo cell. The halo evidence score is calculated using the following formula: ， In the formula, Indicates evidence of vertigo; Indicates the weighting coefficients, all of which are greater than 0 and ; Represents the mean operator; Represents the variance operator; Represents the set of pixel intensities in the kernel region; This represents the set of pixel intensities within the annular halo region; Indicates the ring width; express The result after normalization; express The result after normalization; express The result after normalization; The morphological constraints for HPV-related changes include: area ratio score or nucleocytoplasmic ratio score, dark staining ratio score, degree of irregularity score, nuclear irregularity score, and mass score. When all preset conditions are met, the constraints on HPV-related morphological changes are confirmed to be satisfied.

5. The cervical cell seven-classification method based on hierarchical routing and boundary expert fusion according to claim 4, characterized in that, The method for determining the cell cluster includes: If the number of cell instances in the current target area is greater than or equal to a preset threshold, it is determined to be a cluster of cells; If there are no cell instances in the current target area, the number of cell instances is approximated by the nuclear peak count, connected component count, or the ratio of the region area to the median area of ​​a single cell. If the approximation is greater than or equal to a preset threshold, it is determined to be a cluster of cells.

6. The cervical cell seven-classification method based on hierarchical routing and boundary expert fusion according to claim 5, characterized in that, S5 describes the boundary uncertainty comprehensive analysis The calculation formula is: ， In the formula, , , Indicates weight, + + =1; Represents the probability of five categories The marginal difference in probability between the top 1 and top 2 probabilities. , Represents the probability of five categories The highest probability value in, Represents the probability of five categories The second largest probability value; Represents the normalized five-class probability The entropy value, Represents the probability of five categories The entropy value; ,in Represents the probability of five categories Cell category index Indicates the first The probability value corresponding to the cell type.

7. The cervical cell seven-classification method based on hierarchical routing and boundary expert fusion according to claim 6, characterized in that, A two-branch calculation process is used to output the probability that the region is biased towards LSIL or HSIL; the two branches include the Expert-L branch and the Expert-H branch; When the two classes with the highest probabilities in the five-category basic probability distribution are ASC-US and LSIL, the score is based on the area ratio score, aberration score, nuclear staining ratio score, nuclear irregularity, and quality score of the current target region. Constructing the left branch input feature vector Given the Expert-L branch, output the probability that the value is more biased towards LSIL in the {ASC-US,LSIL}. ; When the two categories with the highest probabilities in the five-category basic probability distribution are ASC-H and HSIL, the nucleocytoplasmic ratio score, the degree of heteromorphism score, the nuclear dark staining ratio score, the nuclear irregularity, and the mass score of the current target region are used as the basis for the calculation. Construct a feature vector input from a branch, input to the Expert-H branch, and output the probability biased towards HSIL. .

8. The cervical cell seven-classification method based on hierarchical routing and boundary expert fusion according to claim 7, characterized in that, Five-class final probability distribution The method for determining it includes the following steps: Let the cervical cells in the current uncertain region be type A and type B, with base probabilities of respectively... , The probability of being biased towards type B is Using the sum of the basic probabilities of type A and type B as a conserved quantity, and utilizing the probability biased towards type B... and the fusion strength The basic probabilities of type A and type B respectively , Adjustments are made to obtain the final probabilities of type A and type B. , In the five-part differential autoimmune classification of single cells, the probability of type C, excluding types A and B, remains consistent with the baseline probability; the formula is expressed as: ， ， ， 。 9. The cervical cell seven-classification method based on hierarchical routing and boundary expert fusion according to claim 1, characterized in that, During the training phase of the method of this invention, for the sample set corresponding to each parent label, one, two or all three indicators of comprehensive severity score, quality score and boundary uncertainty comprehensive score are selected for binning. For each selected indicator, the sample set is divided into at least two corresponding buckets based on its quantile threshold: severity bucket, quality-limited bucket, and boundary bucket. The parent class label is cross-combined with the labels of each bucket to obtain the fission subclass label. The parent class labels include NILM, ASC-US, LSIL, ASC-H, HSIL, halo cells, and cell clusters; The comprehensive severity score is obtained by weighting and summing the nucleocytoplasmic ratio score, atypicality score, nuclear deep staining ratio score, area ratio score, and nuclear irregularity score using a preset weighting rule.

10. A cervical cell seven-classification system based on hierarchical routing and boundary expert fusion, characterized in that, include: Image preprocessing module, target detection module, quality gating module, hierarchical routing module, boundary expert triggering module, and boundary expert classification module; The image preprocessing module is used to acquire cervical cytology images and perform preprocessing to obtain standardized images; The target detection module is used to perform target detection on the standardized image, obtaining the target region and the corresponding detection confidence level. ; The quality gating module is used to score and determine the detection confidence level based on the morphological indicators of the target region. Determine the quality score of the current candidate target region Based on quality score Set fusion strength The morphological indicators are adopted using any one of the following two combinations: Combination 1 includes at least nucleocytoplasmic ratio, nuclear dark staining ratio, degree of heteromorphism, area ratio, nuclear irregularity, and nuclear clarity; Combination 2 includes at least nucleocytoplasmic ratio, nuclear deep staining, degree of atypia, area ratio, nuclear irregularity, and nuclear clarity; The hierarchical routing module is used to identify the type of candidate target regions, including single cells, cell clusters, and halo cells. If it is a single cell, proceed to the boundary expert trigger module; otherwise, directly output the current type as the cervical cell seven-class classification result. The boundary expert triggering module is used to obtain the five-category basic probability distribution of the target area, including NILM, ASC-US, LSIL, ASC-H, and HSIL. When the two categories with the highest probabilities in the five-category basic probability distribution are ASC-US and LSIL, or ASC-H and HSIL, a boundary uncertainty comprehensive score is constructed. If the boundary uncertainty comprehensive score and the quality score simultaneously meet preset conditions, the target area is determined to belong to the boundary uncertainty area and enters the boundary expert module; otherwise, the cell type corresponding to the highest probability in the five-category basic probability distribution is taken as the cervical cell seven-category result. The boundary expert module is used to score morphological indicators, morphological indicators and quality scores based on the current boundary uncertainty region. Calculate the probability of bias towards LSIL or HSIL; based on fusion strength Five-category basic probability distribution The final probability distribution of the five-class classification is obtained by considering the probability of leaning towards LSIL or HSIL. The highest probability value among them is used as the result of the seven-category cervical cell classification.

Images