A standardized cervical cancer pathological image database construction method

Abstract

The application relates to the technical field of medical image database, and discloses a standardized cervical cancer pathological image database construction method, which comprises the following steps: step 1, specimen number is hashed to obtain pseudo-anonymous specimen number, unique slice identification and cell-level unique identification are generated; step 2, a unified reference system is established, the reference layer pixel spacing, the down-sampling multiple and the coordinate mapping are determined; step 3, the image is converted into an optical density space and normalized according to a target staining base to obtain a standardized image; step 4, cross-modal anchor point lists are established by manually comparing cytological and histological images; step 5, a hierarchical label set is generated by synthesizing multi-level unique labels according to the severity full sequence; step 6, the intersection-over-union index and the kappa consistency value are calculated according to the combination key, and a review to-do list is generated; step 7, point review and rule revision are carried out, and a data dictionary, a consistency report and an anchor point snapshot are output. The application realizes the standardized database construction and traceable quality control of cervical cancer cytological and histological pathological images.

Description

Technical Field

[0001] This invention belongs to the field of medical image database technology, specifically relating to a standardized method for constructing a cervical cancer pathological image database. Background Technology

[0002] Cervical cancer screening and diagnosis have long relied on two types of pathological data: cytological smears and histological biopsies. With the widespread adoption of digital slide scanning equipment and pathology information systems, a large number of cervical cytological and histological whole-slide images have accumulated, providing a data foundation for computer-aided diagnosis and related research. However, existing data is mostly scattered across different medical institutions or systems, with varying naming rules, specimen numbering, and slide management methods, lacking a unified identification system, making cross-institutional or cross-time period retrieval and tracing difficult.

[0003] At the image level, different scanning equipment and staining conditions result in significant differences in image resolution, color style, and noise characteristics, making it difficult to directly unify processing or jointly model. Furthermore, cytological images and histological whole-slice images are often generated by different departments at different times, lacking a clear spatial correspondence and structured anchor point records. Multi-level annotation at the cell, cell cluster, lesion region, and whole-slice levels also relies heavily on manual experience, lacking a unified severity system and consistency measurement mechanism. These factors result in significant deficiencies in existing cervical cancer pathology image data regarding cross-modal comparison, hierarchical annotation integration, and quality control, hindering the construction of a reproducible, scalable, and standardized database suitable for multi-center collaborative research. Summary of the Invention

[0004] This invention provides a standardized method for constructing a cervical cancer pathology image database, which solves the technical problem in related technologies that the lack of unified identification, spatial and staining standards for cross-institutional and cross-modal cervical cancer pathology image data leads to difficulties in consistent annotation and fusion, and the inability to quantify and control the quality.

[0005] This invention provides a method for constructing a standardized cervical cancer pathological image database, comprising the following steps:

[0006] Step 1: Hash the original specimen number corresponding to the cervical cancer pathology image to obtain a pseudo-anonymous specimen number, and combine it with the slide acquisition identification data to generate a unique slide identifier. Combine the unique slide identifier with the cell frame coordinates to generate a cell-level unique identifier.

[0007] Step 2: Establish a unified reference system, determine the pixel spacing of the reference layer, the fixed downsampling factor, and the coordinate mapping relationship of scaling, rotation, and translation from any layer to the reference layer;

[0008] Step 3: Convert the sliced ​​image from color space to optical density space and normalize it according to the target staining base to obtain a standardized image;

[0009] Step 4: Manually compare the cytological images and histological whole-section images of the same subject, and establish cross-modal anchor points by pairing the cell-level unique identifier with the histological lesion area identifier and its coordinates under the reference layer, forming a list of cross-modal anchor points;

[0010] Step 5: Define the total severity order of lesion categories, and combine multiple annotations of the same object at the cell, cell cluster, lesion area and slice level into a unique label to generate a hierarchical label set;

[0011] Step 6: Perform consistency measurement on the hierarchical label set according to the combination granularity of image level, lesion category, and equipment configuration, calculate the cross-union index and kappa consistency value, and generate a review pending list based on the cross-modal anchor point list to locate the objects.

[0012] Step 7: Perform point-to-point verification of the objects in the review to-do list between cytological images and histological whole-section images, make rule-based revisions, and output a data dictionary, consistency report, and cross-modal anchor point snapshot.

[0013] Furthermore, the original specimen numbers corresponding to the cervical cancer pathology images are hashed to obtain pseudo-anonymized specimen numbers, and unique slice identifiers and cell-level unique identifiers are generated, including:

[0014] Step 11: Calculate the pseudo-anonymous specimen number using a keyed message authentication hash algorithm on the original specimen number;

[0015] Step 12: Concatenate the pseudo-anonymous specimen number and the slice acquisition identifier data into a byte sequence in a fixed order and calculate using a message authentication hash algorithm to generate a unique slice identifier; wherein, the slice acquisition identifier data includes: site number, slice number, scan timestamp, and magnification;

[0016] Step 13: Represent the cell frame coordinates in pixel coordinates, with the origin at the top left corner. The cell frame coordinates include the horizontal start point, vertical start point, width, and height. Concatenate the slice unique identifier with the horizontal start point, vertical start point, width, and height in a fixed order to form a byte sequence and calculate it using a message authentication hash algorithm to obtain the cell-level unique identifier.

[0017] Furthermore, the pixel spacing of the reference layer, the fixed downsampling factor, and the coordinate mapping relationships from any layer to the reference layer for scaling, rotation, and translation are determined, including:

[0018] Step 21: Select the highest resolution layer as the reference layer using the unique identifier of the slice as the index, determine the pixel spacing of the reference layer, set a fixed downsampling factor, and calculate the pixel spacing of each layer by multiplying the pixel spacing of the reference layer by the corresponding power of the fixed downsampling factor.

[0019] Step 22: Obtain the image orientation based on the scan metadata and discretize it into a preset set of angles; determine the rotation parameters; determine the scaling parameters of each layer according to a fixed downsampling factor and layer number; determine the translation parameters with the upper left corner of the full-frame image of the reference layer as the origin and in pixels; and freeze the scaling parameters, rotation parameters, and translation parameters by associating them with the unique identifier of the slice using a read-only index table.

[0020] Step 23: Generate the coordinate mapping relationship from any layer to the reference layer in the order of scaling, rotation and translation. Convert the pixel coordinates and cell box coordinates to the reference layer according to the coordinate mapping relationship. The cell box coordinates are converted and reconstructed into the minimum bounding rectangle according to the four vertices. The coordinates are rounded down. The conversion from pixel length to physical length is completed according to the pixel spacing of the reference layer.

[0021] Furthermore, the sliced ​​image is converted from color space to optical density space and normalized according to the target staining base to obtain a normalized image, including:

[0022] Step 31: Locate the slice image using the slice unique identifier, extract the high quantile of the three-channel intensity of the blank background pixels as the reference white field, fix the numerical stability constant, and obtain the reference white field and the numerical stability constant.

[0023] Step 32: Convert the sliced ​​image from color space to optical density space, divide the three-channel intensity of each pixel by the three-channel intensity of the reference white field element by element, and take the negative of the natural logarithm to obtain the optical density matrix.

[0024] Step 33: Within the tissue region, singular value decomposition is used to obtain the source chromatin matrix composed of two principal directions, and non-negative least squares is used to calculate the chromatin coefficient matrix.

[0025] Step 34: Normalize the amplitude of the staining coefficient matrix using the target staining base, calculate the quantile ratio of each channel and scale the staining coefficient matrix channel by channel, and reconstruct the normalized optical density matrix under the target staining base.

[0026] Step 35: The normalized optical density matrix is ​​inversely transformed by the channel exponent and multiplied element-wise with the reference white field, cropped to the effective numerical domain, and the standardized image is obtained and stored in association with the unique identifier of the slice.

[0027] Furthermore, the disadvantages of the target chromophore described in step 34 include:

[0028] The device model, magnification, scanning timestamp, white balance number and staining reagent batch number are encoded in a fixed order and the device configuration fingerprint is calculated using a message authentication hash algorithm with a key. The device configuration fingerprint is used as the key to aggregate the slice set.

[0029] For each slice in the slice set, calculate the channel quantile vector using a fixed quantile for the coloring coefficient matrix, and then calculate the median quantile vector for each channel within the set.

[0030] The sum of the absolute differences between the channel quantile vector and the median quantile vector of each slice is used as the distance metric. The slice with the smallest distance metric is selected as the representative slice, and the source chromatogram matrix of the representative slice is used as the target chromatogram corresponding to the device configuration fingerprint.

[0031] Furthermore, establish cross-modal anchor points and create a list of cross-modal anchor points, including:

[0032] Step 41: Use the pseudo-anonymous specimen number to retrieve the unique identifier of the slices on both sides; when the pseudo-anonymous specimen numbers are the same, the dissection sources are the same, and the time difference does not exceed the upper limit of the preset time window, generate a candidate pairing set;

[0033] Step 42: Based on the coordinate mapping relationship, convert the pixel coordinates and cell box coordinates from any layer to their respective reference layers; convert the cell box coordinates according to the four vertices and reconstruct them into the minimum bounding rectangle in the reference layer, round the coordinates down and label the coordinate system.

[0034] Step 43: Manually verify the cell-level unique identifier and histological lesion region identifier within the candidate pairing set. Concatenate the pseudo-anonymous specimen number, bilateral slice unique identifier, paired coordinates, and source rules into a byte sequence according to a fixed field order, and calculate the cross-modal anchor record identifier using a message authentication hash algorithm with a key. Store the record after deduplication according to the identifier, and establish a foreign key association with the bilateral slice unique identifier to establish a cross-modal anchor, forming a cross-modal anchor list, and generating a version number and timestamp.

[0035] Furthermore, the severity sequence of lesion categories is defined, and multiple annotations of the same object at the cell, cell cluster, lesion region, and slide level are combined into a unique label, generating a hierarchical label set, including:

[0036] Step 51: Freeze the lesion category set and severity sequence, set category codes and evidence priorities, with histological evidence taking precedence over cytological evidence;

[0037] Step 52: Define a cell cluster object using a polygon represented by the reference layer coordinates. When all four vertices of the cell frame coordinates are within the polygon, add the corresponding cell-level unique identifier to the member set of the cell cluster object. Merge the candidate labels of the member set with the direct labels of the cell cluster object to form a candidate label set for the cell cluster object.

[0038] Step 53: Perform a total severity check of the parent-child hierarchy for each histological lesion region identifier and each slide unique identifier. When the total severity of the parent-level unique label is lower than that of any child-level unique label, generate a hierarchy conflict item and record the histological lesion region identifier, slide unique identifier, parent-level unique label, child-level unique label and timestamp.

[0039] Furthermore, consistency measures are applied to the hierarchical label set, calculating the intersection-union ratio (IU) and kappa consistency values. Based on the cross-modal anchor list, a review task list is generated, including:

[0040] Step 61: Combine image level, lesion category, and device configuration to form a key, and retrieve the object set from the hierarchical label set and multi-expert annotation based on the key; wherein, the image level is fixed as cell level, cell cluster, lesion area, and slice;

[0041] Step 62: Based on the coordinate mapping relationship, convert the geometric annotations in the object set to the reference layer, and generate a uniform resolution binary mask according to the pixel spacing of the reference layer.

[0042] Step 63: For each object, take the intersection pixel count of all expert binary masks as the numerator and the union pixel count as the denominator to obtain the object-level intersection-union ratio index; calculate the arithmetic mean of the object-level intersection-union ratio indices of all objects to obtain the combinatorial bond-level intersection-union ratio index; for each object in the object set, count the expert counts of each category to obtain the object-level observation consistency rate, and calculate the arithmetic mean of the object-level observation consistency rates of all objects as the combinatorial bond-level observation consistency rate; summarize the category counts of all objects and all experts into a marginal distribution, and sum the squares of the marginal proportions of each category as the combinatorial bond-level coupling consistency rate; subtract the combinatorial bond-level coupling consistency rate from the combinatorial bond-level observation consistency rate to obtain the numerator, and use the difference between the numerator and the combinatorial bond-level coupling consistency rate as the denominator to obtain the combinatorial bond-level kappa consistency value;

[0043] Step 64: When the cross-intersection over union (CUI) index at the composite key level is lower than the preset CUI threshold, select a preset number of objects from the object set according to the object-level CUI index from low to high to generate the first review set; when the cross-intersection over union (CUI) index at the composite key level is lower than the preset consistency threshold, generate the second review set according to the same rule; merge the two sets to remove duplicates, and generate a review pending list based on the cross-modal anchor list.

[0044] Furthermore, after establishing cross-modal anchor points, quality grading is performed, including:

[0045] The cell frame coordinates are converted from any layer to the cytological reference layer according to the coordinate mapping relationship. The corresponding point pairs associated with the cross-modal anchor point record are identified and the homography mapping is solved in the local neighborhood. The cell frame coordinates are then converted to the histological reference layer according to the homography mapping. The cross-union index is calculated according to the cross-union index in step 63. The coverage cross-union index is calculated for the two masks and used as the coverage.

[0046] The distance ratio is calculated as the ratio of the geometric centroid distance between the cell frame and the histological lesion region in the coordinates of the histological reference layer to the square root of the area of ​​the histological lesion region, and the time ratio is calculated as the ratio of the scan time difference to the upper limit of the preset time window.

[0047] The anchor quality score is obtained by weighted summation of coverage, distance ratio and time ratio. When the anchor quality score is lower than the preset score threshold, the object corresponding to the cross-modal anchor is added to the priority queue of the review task list and the anchor quality score is recorded.

[0048] Furthermore, the system performs point-to-point verification of objects on the review task list between cytological images and histological whole-section images, revises coordinates and categories according to rules, and outputs a data dictionary, consistency report, and cross-modal anchor snapshots, including:

[0049] Step 71: Read the object key, slide unique identifier, baseline layer coordinates and cross-modal anchor point record identifier from the review to-do list; obtain the cell-level unique identifier and histological lesion area identifier from the cross-modal anchor point list; call the coordinate mapping relationship to convert the coordinates of any layer to the baseline layer, and complete the synchronous positioning of the cytological image and the histological whole slide image.

[0050] Step 72: Verify the geometry of cell-level unique identifiers and histological lesion region identifiers in the baseline layer; when the geometry needs to be revised, record the new polygon and reconstruct it into the smallest bounding rectangle, round the coordinates down, generate a binary mask of the old and new geometry, calculate the area difference and the intersection-union ratio of the old and new geometry, and write the revised coordinates back to the baseline layer.

[0051] Step 73: Based on the fixed decision-making rules of total severity, evidence priority, and category coding, candidate labels are selected to determine a unique label; when there is a cross-modal anchor point record that is associated with the histological lesion area, the evidence priority is marked and a category revision record is generated.

[0052] Step 74: Write the slice unique identifier, object key, baseline coordinates, unique label, category code, severity, evidence priority, cross-modal anchor record identifier, reviser and Coordinated Universal Time timestamp into a read-only index table in a fixed field order, generate an integrity summary and associate it with the slice unique identifier foreign key; update the version number according to the three-stage rule of major version, minor version and patch version, and output the data dictionary, consistency report and cross-modal anchor snapshot.

[0053] The beneficial effects of this invention are as follows: By unifying the reference system and coordinate mapping relationship, this invention converts arbitrary layer images and geometric annotations to the baseline layer, ensuring consistency in spatial scale and physical quantity; by deterministic transformation from color space to optical density space and normalization under the target chromatin, it reduces the differences between equipment and staining batches, resulting in standardized images with strong comparability; by establishing cross-modal anchor points and quality grading in cytology and histology, it achieves structured management of cross-modal controls; by using total severity sequence, evidence priority, and parent-child hierarchical verification, it synthesizes multi-level and multi-source annotations into a unique label and forms a hierarchical label set; by combining the intersection-union index and kappa consistency value and threshold-triggered review, it forms a traceable review closed loop and outputs a data dictionary, consistency report, and anchor point snapshot, thereby improving the standardization, reproducibility, and long-term maintenance capability of the database. Attached Figure Description

[0054] Figure 1 This is a flowchart of a standardized method for constructing a cervical cancer pathological image database according to the present invention. Detailed Implementation

[0055] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0056] like Figure 1 As shown, a method for constructing a standardized cervical cancer pathology image database includes the following steps:

[0057] Step 1: Hash the original specimen number corresponding to the cervical cancer pathology image to obtain a pseudo-anonymous specimen number, and combine it with the slide acquisition identification data to generate a unique slide identifier. Combine the unique slide identifier with the cell frame coordinates to generate a cell-level unique identifier.

[0058] Step 2: Establish a unified reference system, determine the pixel spacing of the reference layer, the fixed downsampling factor, and the coordinate mapping relationship of scaling, rotation, and translation from any layer to the reference layer;

[0059] Step 3: Convert the sliced ​​image from color space to optical density space and normalize it according to the target staining base to obtain a standardized image;

[0060] Step 4: Manually compare the cytological images and histological whole-section images of the same subject, and establish cross-modal anchor points by pairing the cell-level unique identifier with the histological lesion area identifier and its coordinates under the reference layer, forming a list of cross-modal anchor points;

[0061] Step 5: Define the total severity order of lesion categories, and combine multiple annotations of the same object at the cell, cell cluster, lesion area and slice level into a unique label to generate a hierarchical label set;

[0062] Step 6: Perform consistency measurement on the hierarchical label set according to the combination granularity of image level, lesion category, and equipment configuration, calculate the cross-union index and kappa consistency value, and generate a review pending list based on the cross-modal anchor point list to locate the objects.

[0063] Step 7: Perform point-to-point verification of the objects in the review to-do list between cytological images and histological whole-section images, make rule-based revisions, and output a data dictionary, consistency report, and cross-modal anchor point snapshot.

[0064] In one embodiment of the present invention, the original specimen number corresponding to the cervical cancer pathological image is hashed to obtain a pseudo-anonymous specimen number, which is then combined with the slide acquisition identification data to generate a unique slide identifier. The unique slide identifier is then combined with the cell frame coordinates to generate a cell-level unique identifier, including:

[0065] Step 11: Calculate the pseudo-anonymous sample number using a keyed message authentication hash algorithm on the original sample number. The original sample number represents the internal identifier used to refer to a single sample. In this embodiment, it is preferable to remove leading and trailing whitespace and unify capitalization before calculating the pseudo-anonymous sample number. The keyed message authentication hash algorithm refers to a class of algorithms that, based on a conventional hash function, introduce a key stored only in a controlled environment to perform message authentication hashing on the input data, such as a key-based hash message authentication structure. By introducing a key, on the one hand, it can be ensured that the same original sample number is mapped to a specific pseudo-anonymous sample number within the system, facilitating subsequent data association; on the other hand, even if the hash result is intercepted externally, the original sample number cannot be recovered through exhaustive search or simple comparison due to the lack of a key, thereby achieving pseudo-anonymization and privacy protection of the examinee's identity information at the sample level.

[0066] Step 12: Concatenate the pseudo-anonymous specimen number and the slide acquisition identification data into a byte sequence in a fixed order and calculate using a message authentication hash algorithm to generate a unique slide identifier. The slide acquisition identification data includes: site number, slide number, scan timestamp, and magnification. The site number is used to identify a specific medical institution or acquisition node, the slide number is used to identify different slides under the same subject, and the magnification is used to distinguish image data at different scan magnifications. The unique slide identifier is used to eliminate the risk of confusion caused by duplicate naming at the slide level.

[0067] Step 13: Represent the cell bounding box coordinates in pixel coordinates. Pixel coordinates refer to a two-dimensional discrete coordinate system with the top-left corner of the image as the origin. The cell bounding box coordinates include the horizontal start point, vertical start point, width, and height. The horizontal and vertical start points are used to mark the position of the top-left corner of the cell's bounding rectangle in the image, and the width and height are used to mark the pixel dimensions of the bounding rectangle in the horizontal and vertical directions. The unique slice identifier is concatenated with the horizontal and vertical start points, width, and height in a fixed order to form a byte sequence, and a message authentication hash algorithm is used to calculate the cell-level unique identifier. The pseudo-anonymous specimen number, the unique slice identifier, the cell-level unique identifier, and the cell bounding box coordinates are merged and stored to form a unique identifier set, ensuring stable and accurate referencing to the same cell object and its context slice information.

[0068] This embodiment achieves traceable pseudo-anonymity by performing keyed message authentication hashing on the original specimen number; it achieves globally unique identifiers at the slice and cell levels by hashing the pseudo-anonymous specimen number with the collection identifier and cell frame coordinates in a fixed order; and finally, it merges them into a unique identifier set in a read-only index table to ensure stable cross-institutional data association and privacy security.

[0069] In one embodiment of the present invention, a unified reference system is established, and the coordinate mapping relationships of the reference layer pixel spacing, fixed downsampling factor, and scaling, rotation, and translation from any layer to the reference layer are determined, including:

[0070] Step 21: Using the unique identifier of the slice as the retrieval key, read the layer list and pixel spacing information of the slice, and select the layer with the smallest pixel spacing as the reference layer; pixel spacing refers to the spatial sampling step size given by the scanning device during imaging; when there are multiple candidate layers, the layer with the smallest sequence number is preferred as the reference layer. The pixel spacing of the reference layer is used to represent the physical length corresponding to each pixel in the image of that layer, and the unit is preferably micrometers / pixel. The fixed downsampling factor is unique within the same slice and constant for all layers, used to characterize the resolution ratio relationship between adjacent layers. Based on the pixel spacing of the reference layer and the fixed downsampling factor, calculate and record the pixel spacing of each layer according to the layer sequence number, forming a lookup table of layer sequence number, pixel spacing, and resolution factor, and associate it with the unique identifier of the slice.

[0071] Step 22: Based on the scan metadata, the image orientation is read and discretely mapped to one of the preset angle sets to determine the rotation parameters, used to correct the consistency of the scan direction. The preset angles may include: 0 degrees, 90 degrees, 180 degrees, and 270 degrees. Scaling parameters are calculated based on the ratio of the pixel spacing of each layer to the pixel spacing of the reference layer, used to unify the spatial scale of layers with different resolutions. Translation parameters are determined using the top-left corner of the full-frame image of the reference layer as the origin, in pixels, to eliminate the offset between the image coordinate origin and the display window origin. The scaling, rotation, and translation parameters are associated with the unique slice identifier in the form of a read-only index table and frozen within the data version to ensure the determinism and traceability of subsequent coordinate conversions.

[0072] Step 23: For pixel coordinates or geometric annotations of any layer, generate coordinate mapping relationships to the base layer in a fixed order of scaling, rotation, and translation. For cell bounding box coordinates, it is preferable to convert them to four-vertex coordinates and reconstruct them into the smallest bounding rectangle under the base layer; the coordinates are discretized using a floor function to eliminate uncertainties introduced by floating-point errors. Physical quantities such as length and area are converted to units based on the pixel spacing of the base layer, preferably using micrometers and square micrometers as units of measurement, and the mapping relationship between the conversion results and the original coordinates is recorded to ensure that there is no semantic ambiguity when calling them in subsequent steps.

[0073] This embodiment ensures consistency in coordinates and length across layers by using a unified reference system and a fixed mapping order; parameters are associated with and frozen using a read-only index table and a unique slice identifier, improving the reproducibility and traceability of data across devices and layers, and providing stable input for anchor point establishment, label synthesis, and consistency measurement.

[0074] In one embodiment of the present invention, converting the sliced ​​image from a color space to a density space and normalizing it according to a target chromatin base to obtain a standardized image includes:

[0075] Step 31: Locate the slice image using its unique identifier, retrieve its blank background region, and extract the high quantile of the three-channel intensity as the reference white field. The high quantile is preferably 0.95 to suppress the influence of isolated bright pixels on the estimation. The numerical stability constant is a fixed positive number used to avoid numerical overflow or zeroing during successive division and logarithmic transformation; the numerical stability constant is fixed within the data version. The reference white field refers to the upper bound intensity vector of the three channels used for normalization under the slice, providing the upper bound of the channels; the numerical stability constant is used to maintain computational repeatability for pixels with intensity close to zero.

[0076] Step 32: Divide the intensity of each pixel's three channels by the intensity of the reference white field's three channels one by one, and take the negative of the natural logarithm to obtain the optical density matrix. This transformation is equivalent to expressing pixel values ​​in terms of absorbance under the Lambert-Beer law, thus converting the nonlinear coupling between channels into a linearly decomposable representation. The reference white field and the numerical stability constant are included in the denominator and logarithmic terms of this transformation to ensure consistency across image apertures.

[0077] Step 33: Decompose the optical density matrix within the tissue region. First, eliminate low-intensity pixels to remove background interference. Then, normalize the remaining pixels column-wise. Use singular value decomposition to obtain two principal directions and orthogonalize them to form the source chromatin base matrix. With the source chromatin base matrix fixed, calculate the chromatin coefficient matrix using non-negative least squares, and ensure physical interpretability with non-negativity constraints. The source chromatin base matrix describes the orientation of the two main chromatin channels in RGB coordinates; the chromatin coefficient matrix describes the concentration weight of each pixel in the two channels.

[0078] Step 34: Preset the target colorimetric base and freeze it within the data version. Calculate the high quantiles of the colorimetric coefficient matrix channel by channel to obtain channel statistics, and compare them with the target channel statistics to obtain the quantile ratio. The target colorimetric base refers to the two unified basis vectors across data versions. The quantile ratio refers to the channel-by-channel scaling coefficient when aligning the coefficient distribution to the target statistics. Scale the colorimetric coefficient matrix channel by channel according to the quantile ratio to align its amplitude to the target colorimetric base. Perform linear reconstruction on the scaled coefficients under the target colorimetric base to obtain the normalized optical density matrix. This step only changes the coefficient amplitude and the basis vector coordinate system, without changing the spatial position of the pixels.

[0079] Step 35: Perform an inverse exponential transformation on the normalized optical density matrix by channel and multiply it element-wise with the reference white field to obtain a three-channel intensity matrix; crop the intensity matrix to the effective numerical domain to obtain a standardized image; the effective numerical domain is preferably 0 to 255; associate the standardized image with the unique identifier of the slice and store it, while recording the reference white field, numerical stability constant, source staining matrix, target staining matrix, quantile parameter and unit aperture, and write it into a read-only index table to support subsequent retrieval and traceability.

[0080] This embodiment achieves repeatable mapping in optical density space through quantile white field and numerical stability constant; it achieves color caliber uniformity across devices and batches through source chromatin decomposition and target chromatin amplitude normalization; and finally outputs a standardized image associated with a unique slice identifier.

[0081] In one embodiment of the present invention, the disadvantages of the target chromophore described in step 34 include:

[0082] The device model, magnification, scanning timestamp, white balance number, and staining reagent batch number are encoded in a fixed order and the device configuration fingerprint is calculated using a message authentication hash algorithm with a key. The device configuration fingerprint is used as the key to aggregate the slice set. The device configuration fingerprint is a fixed-length identifier used to uniquely identify the combination of device and staining batch.

[0083] For each slice in the slice set, the channel quantile vector is calculated using a fixed quantile for the staining coefficient matrix, and the median quantile vector is obtained by calculating the median of each channel within the set; the fixed quantile is preferably 0.95; the median quantile vector is used as a robust scaling reference for fingerprints;

[0084] The sum of the absolute differences between the channel quantile vector and the median quantile vector of each slice is used as the distance metric. The slice with the smallest distance metric is selected as the representative slice, and the source chromatogram matrix of the representative slice is used as the target chromatogram corresponding to the device configuration fingerprint.

[0085] This embodiment uses device configuration fingerprint grouping and distance measurement to deterministically select target staining bases, eliminating the impact of device and batch differences on normalization caliber; it registers the data in a read-only index table and binds version information to ensure that the color standard of the same fingerprint slice is consistent and reproducible across versions.

[0086] In one embodiment of the present invention, cytological images and histological whole-section images of the same subject are manually compared. Cell-level unique identifiers and histological lesion region identifiers, along with their coordinates in the reference layer, are paired to establish cross-modal anchor points, forming a list of cross-modal anchor points, including:

[0087] Step 41: Using the pseudo-anonymous specimen number as the search key, obtain the unique identifier set of the cytology image and the histology whole slide image respectively; unify the scan timestamps of the cytology image and the histology whole slide image to Coordinated Universal Time, and read the anatomical source, which is a standardized enumeration field used to characterize the collection site; when the pseudo-anonymous specimen number is consistent, the anatomical source is consistent, and the absolute value of the scan time difference does not exceed the upper limit of the preset time window, record the binary slide pair according to the fixed field order to generate a candidate pairing set; at the same time, freeze the source rules to ensure the consistency and auditability of subsequent citations. The source rules include the pseudo-anonymous specimen number, anatomical source, upper limit of the time window, and timestamp caliber.

[0088] Step 42: For each pair of slices in the candidate pairing set, invoke the coordinate mapping relationship and convert the pixel coordinates of any layer on both sides to their respective reference layers in a fixed order of scaling, rotation, and translation. For the cell box coordinates on the cytology side, convert them to four-vertex coordinates and reconstruct them into the minimum bounding rectangle in the reference layer, rounding down the coordinates. While recording the coordinates, label the cytology reference layer coordinates and the histology reference layer coordinates to avoid coordinate system confusion when referencing across tables. This yields paired coordinates from the cell-level unique identifier to the cytology reference layer coordinates and from the histology lesion region identifier to the histology reference layer coordinates, which serve as direct input for the next step of anchor point establishment.

[0089] Step 43: Within the candidate pairing set, authorized personnel manually verify the pairing relationship between the cell-level unique identifier and the histological lesion region identifier, as well as the correctness of their paired coordinates. Subsequently, the data is encoded into a byte sequence according to fixed fields and a fixed order. A message authentication hash algorithm with a key is used to calculate the cross-modal anchor record identifier. This identifier is used as the primary key for deduplication and storage, and a foreign key association is established with the unique identifiers of the adjacent slices to create a cross-modal anchor and write it into the cross-modal anchor list. Simultaneously, the Coordinated Universal Time (UTC) timestamp and the three-segment data version number are recorded and fixed in a read-only index table. The above field order, hash algorithm, and key identifier are frozen within the data version to ensure the idempotency and cross-system consistency of the anchor identifier.

[0090] This embodiment limits candidate pairing by using a unified time caliber and dissection source, ensures spatial comparability by using a fixed mapping order and baseline layer coordinate reconstruction, generates anchor record identifiers by using a keyed message authentication hash algorithm and deduplicates them into the database, and finally forms a traceable and versionable cross-modal anchor list to support subsequent consistency measurement and fixed-point verification.

[0091] In one embodiment of the present invention, a total severity sequence of lesion categories is defined, and multiple annotations of the same object at the cell, cell cluster, lesion region, and slide level are combined into a unique label to generate a hierarchical label set, including:

[0092] Step 51: Freeze the lesion category set and severity total order. The lesion category is used to classify the pathological state, and the severity total order is a strict total order relationship with the lesion category. The larger the value, the higher the severity. Assign a category code to each lesion category. The category code is a unique positive integer. Set the evidence priority. The evidence priority is histological evidence over cytological evidence, which is used as the secondary adjudication criterion when the severity is equal. The determination of the unique label is carried out in a fixed order of severity total order first, evidence priority second, and category code last. The three rules are frozen within the data version.

[0093] Step 52: Define a cell cluster object using polygons represented by the reference layer coordinates. The polygons are drawn by authorized personnel on the reference layer. When all four vertices of the cell frame coordinates are within the polygon, add the corresponding cell-level unique identifier to the member set of the cell cluster object. Merge the candidate labels of each cell in the member set with the direct label of the cell cluster object to form a candidate label set of the cell cluster object, which is then called by the unique label determination rules.

[0094] Step 53: Perform total severity check (TLS) on each histological lesion region identifier and each slide unique identifier, requiring that the total severity of the parent-level unique label is not lower than the total severity of any child-level unique label. If this requirement is not met, generate a hierarchical conflict item and record the histological lesion region identifier, slide unique identifier, parent-level unique label, child-level unique label, and timestamp. The timestamp is uniformly set to Coordinated Universal Time (UTC). Store the hierarchical conflict item in a read-only index table and establish a foreign key association with the slide unique identifier for subsequent review and revision.

[0095] This embodiment achieves deterministic synthesis of unique labels by freezing the total severity sequence, category coding, and evidence priority; it achieves consistent standards for multi-granularity objects by defining cell clusters through geometric inclusion and aggregating candidate labels; and finally, it ensures the hierarchical consistency and traceability of the hierarchical label set by parent-child level verification and conflict recording.

[0096] In one embodiment of the present invention, the consistency of the hierarchical label set is measured at a combined granularity of image level, lesion category, and device configuration. The cross-union ratio (CUI) and kappa consistency values ​​are calculated, and a review task list is generated based on the cross-modal anchor point list to locate objects. This includes:

[0097] Step 61: Combine the image level, lesion category, and equipment configuration into a key. The image level is fixed as cell level, cell cluster, lesion region, and section. The lesion category is taken from the frozen lesion category set. The equipment configuration is a binary tuple of equipment model and staining configuration fingerprint. Using the key as the search condition, retrieve the object set from the hierarchical label set and multi-expert annotation. Each object in the object set has geometric annotation and classification annotation from no less than two experts. The object key value is a unique identifier for cell level, cell cluster object, histological lesion region identifier, or unique identifier for section.

[0098] Step 62: Based on the coordinate mapping relationship, convert the geometric annotations of each expert in the object set to the reference layer of the slice to which the object belongs in the order of scaling, rotation, and translation. Round the coordinates down and label the reference layer coordinates. Generate a binary mask with the same size as the slice according to the pixel spacing in the reference layer. The mask value is zero and one, and the even-odd rule is used for filling. The even-odd rule means that in the pixel coordinate system of the reference layer, with the pixel center as the judgment point, draw a semi-infinite ray along the positive horizontal direction and count the number of intersections between the ray and the polygon boundary. Odd numbers are inside and even numbers are outside. Ignore the horizontal edge and only count the lower vertex of the intersection vertex. The boundary point is processed according to the left closed, top closed, right open and bottom opening diameter. Generate a binary mask with a value of zero and one, and do not enable anti-aliasing.

[0099] Step 63: For each object, take the intersection pixel count of all expert masks as the numerator and the union pixel count as the denominator to calculate the object-level intersection-union ratio (IU). The IU of all object-level IU is arithmetically averaged to obtain the combinatorial-level IU, which measures the spatial overlap of multi-expert geometric annotations of the same object at the baseline layer, reflecting boundary consistency and used to locate objects with low consistency. For each object, the expert counts for each lesion category are counted to calculate the object-level observation consistency rate, where the object-level observation consistency rate is the probability of randomly selecting two experts within the object without replacement to give the same category. The arithmetic mean of the object-level observation consistency rates is used as the combinatorial-level observation consistency rate. Within the combinatorial key range, the category counts of all objects and all experts are summarized to obtain the marginal distribution. The sum of the squares of the marginal proportions of each category is used as the combinatorial-level coupling consistency rate. The numerator is obtained by subtracting the combinatorial-level coupling consistency rate from the combinatorial-level observation consistency rate, and the denominator is obtained by subtracting the combinatorial-level coupling consistency rate from one. The combinatorial-level kappa consistency value is calculated and used to quantify the consistency of multi-expert classification annotations after chance correction within a given combinatorial key.

[0100] Step 64: When the cross-intersection over union (CUI) index at the composite key level is lower than the preset CUI threshold, select a preset number of objects from the object set according to the object-level CUI index from low to high to generate the first review set; when the cross-intersection over union (CUI) index at the composite key level is lower than the preset consistency threshold, generate the second review set according to the same rule; merge the two sets to remove duplicates, and locate the review objects according to the cross-modal anchor list to generate a review pending list; store the review pending list in a read-only index table, and establish a foreign key association with the slice unique identifier, while recording the version number and the Coordinated Universal Time (UTC) timestamp; the preset CUI threshold, preset consistency threshold, and preset number are fixed as constants within the data version.

[0101] This embodiment defines the statistical scope by using a combination key, unifies the rasterization at the baseline layer, and calculates the cross-union ratio and kappa consistency value to form a reproducible closed loop for consistency measurement. By triggering with dual thresholds and combining a cross-modal anchor list, the review objects are accurately located, and finally a traceable review task list and versioned records are output.

[0102] In one embodiment of the present invention, quality grading is performed after the cross-modal anchor point is established, including:

[0103] The cell frame coordinates and histological lesion region identifiers are located using cross-modal anchor point record identifiers. The coordinate mapping relationship from step 23 is invoked to convert the cell frame coordinates from any layer to the cytological reference layer. Within a local neighborhood centered on the geometric centroid of the cell frame and with a fixed pixel size, at least four pairs of corresponding points associated with the cross-modal anchor point record identifier are collected. The homography mapping is solved, and the cell frame coordinates are converted to the histological reference layer. Binary masks are generated for the converted cell frames and histological lesion regions in the histological reference layer according to the binary mask generation rules of step 62. The coverage is obtained by using the cross-union ratio (CUN) index calculation method from step 63, where the ratio of the intersection pixels to the union pixels of the two masks is used. Homography mapping refers to a deterministic mapping that determines the local coordinates of the cytological reference layer as the local coordinates of the histological reference layer through projective transformation. The window size and sampling step size of the local neighborhood are fixed constants within the data version.

[0104] The geometric centroid distance between the converted cell frame and the histological lesion region is calculated in the coordinates of the histological reference layer, and the ratio of this distance to the square root of the area of ​​the histological lesion region is used as the distance ratio. The scanning timestamps of the cytological images and the histological whole-section images are unified to Coordinated Universal Time (UTC), and the ratio of the absolute value of the time difference to the upper limit of the preset time window is used as the time ratio. The geometric centroid is calculated using an area-weighted method based on vertex coordinates, and the units of area and distance are consistent with the pixel spacing of the reference layer. The upper limit of the preset time window is a fixed constant within the data version.

[0105] The anchor quality score is obtained by multiplying and summing the coverage, distance ratio, and time ratio by their respective preset weights. When the anchor quality score is lower than the preset score threshold, the object corresponding to the cross-modal anchor is added to the priority queue of the review task list. It is written to a read-only index table with fixed fields and a foreign key association is established with the unique identifiers of the slices on both sides. The cross-modal anchor record identifier, coverage, distance ratio, time ratio, anchor quality score, UTC timestamp, and data version number are recorded. The preset score threshold is a fixed constant within the data version.

[0106] This embodiment utilizes homography mapping under a unified reference system to achieve cross-modal alignment. Anchor point quality scores are obtained by weighted synthesis of coverage, distance ratio, and time ratio. A priority review queue is generated by triggering a threshold, ensuring that review resources are focused on low-quality anchor points, thereby improving the controllability and traceability of database annotation consistency.

[0107] In one embodiment of the present invention, the objects in the review task list are reviewed at specific points between cytological images and histological whole-section images, and regular revisions are performed. A data dictionary, a consistency report, and a cross-modal anchor point snapshot are then output, including:

[0108] Step 71: Read the object key, slide unique identifier, reference layer coordinates, and cross-modal anchor point record identifier from the review task list item by item. The object key is an index field used to uniquely identify the review target, and its value can be a cell-level unique identifier, a cell cluster object, a histological lesion region identifier, or a slide unique identifier. In the cross-modal anchor point list, obtain the cell-level unique identifier and histological lesion region identifier corresponding to the object according to the cross-modal anchor point record identifier. Call the coordinate mapping relationship, and convert the coordinates of any layer to the reference layer in a fixed order of scaling, then rotation, and then translation, and round down. The reference layer is the layer with the smallest pixel spacing of the slide. In this way, the synchronous positioning of the cytological image and the histological whole slide image under their respective reference layers is completed, forming a dual-modal side-by-side review view of the same object.

[0109] Step 72: Verify the geometry of the cell-level unique identifier and the histological lesion region identifier in the baseline layer. The geometry is a polygon or its smallest bounding rectangle in the coordinate system of the baseline layer. When the geometry needs to be revised, record the revised new polygon and reconstruct it as the smallest bounding rectangle in the baseline layer. The coordinates are discretized by rounding down. Generate binary masks of the geometry before and after the revision in the baseline layer according to the mask generation rules in step 62. Calculate the difference in mask area before and after the revision. Calculate the intersection-union ratio (IU) of the old and new IU according to the IU index calculation method in step 63. Write the revised baseline layer coordinates back to the corresponding record in the read-only index table and retain the revision evidence and timestamp.

[0110] Step 73: Invoke the fixed decision rules for severity total sequence, evidence priority, and category coding frozen in Step 51 to select candidate labels. The candidate labels are the union of direct and aggregated labels of the object at the cell, cell cluster, lesion region, and slide levels. The unique label is determined according to the following order: severity total sequence first, evidence priority second, and category coding last. When the object has a cross-modal anchor record label associated with the histological lesion region label, the label evidence priority is the priority corresponding to the histological evidence and overrides the cytological evidence. Generate a category revision record and write the object key, old unique label, unique label, triggering rule, reviser, and Coordinated Universal Time timestamp.

[0111] Step 74: Write the slice unique identifier, object key, baseline coordinates, unique label, category code, severity, evidence priority, cross-modal anchor record identifier, reviser and Coordinated Universal Time timestamp into a read-only index table in a fixed field order, and concatenate the record into a byte sequence in a fixed field order to calculate an integrity digest. The integrity digest is a fixed-length hash value used to verify that the content entered into the database has not been tampered with; establish a foreign key association between the integrity digest and the slice unique identifier. Version numbers are updated according to a three-stage rule: major version, minor version, and patch version. The major version is used to identify changes in field structure or reference system caliber, the minor version is used to identify changes in thresholds or decision rules, and the patch version is used to identify single-object revisions. A data dictionary is output, which is a structured list of field definitions, units, coordinate calibers, rounding rules, and version information. A consistency report is output, which is a structured result of the key-level intersection-union index and key-level kappa consistency value, along with their thresholds, sample sizes, and version numbers. A cross-modal anchor snapshot is output, which is a finalized copy of the cross-modal anchor record identifier, cell-level unique identifier, histological lesion region identifier, and its baseline layer paired coordinates. The corresponding review pending items are marked as completed, and geometric and categorical revision evidence is backfilled.

[0112] This embodiment achieves dual-modal fixed-point verification by jointly locating the to-do list and the cross-modal anchor point list; it revises coordinates and unique labels by unifying mask aperture and rule-based decision-making; and finally, it solidifies the output data dictionary, consistency report and cross-modal anchor point snapshot with a read-only index table, integrity summary and three-segment version number to ensure that the database can be traced and updated.

[0113] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0114] The embodiments of the present invention have been described above, but the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of the present embodiments, all of which are within the protection scope of the present embodiments.

Claims

1. A method for constructing a standardized cervical cancer pathological image database, characterized in that, Includes the following steps: Step 1: Hash the original specimen number corresponding to the cervical cancer pathology image to obtain a pseudo-anonymous specimen number, and combine it with the slide acquisition identification data to generate a unique slide identifier. Combine the unique slide identifier with the cell frame coordinates to generate a cell-level unique identifier. Step 2: Establish a unified reference system, determine the pixel spacing of the reference layer, the fixed downsampling factor, and the coordinate mapping relationship of scaling, rotation, and translation from any layer to the reference layer; Step 3: Convert the sliced ​​image from color space to optical density space and normalize it according to the target staining base to obtain a standardized image; Step 4: Manually compare the cytological images and histological whole-section images of the same subject, and establish cross-modal anchor points by pairing the cell-level unique identifier with the histological lesion area identifier and its coordinates under the reference layer, forming a list of cross-modal anchor points; Step 5: Define the total severity order of lesion categories, and combine multiple annotations of the same object at the cell, cell cluster, lesion area and slice level into a unique label to generate a hierarchical label set; Step 6: Perform consistency measurement on the hierarchical label set according to the combined granularity of image level, lesion category, and equipment configuration; calculate the cross-union ratio (CUI) and kappa consistency value; and generate a review pending list based on the cross-modal anchor point list to locate objects, including: Step 61: Combine image level, lesion category, and device configuration to form a key, and retrieve the object set from the hierarchical label set and multi-expert annotation based on the key; wherein, the image level is fixed as cell level, cell cluster, lesion area, and slice; Step 62: Based on the coordinate mapping relationship, convert the geometric annotations in the object set to the reference layer, and generate a uniform resolution binary mask according to the pixel spacing of the reference layer. Step 63: For each object, take the intersection pixel count of all expert binary masks as the numerator and the union pixel count as the denominator to obtain the object-level intersection-union ratio index; calculate the arithmetic mean of the object-level intersection-union ratio indices of all objects to obtain the combinatorial bond-level intersection-union ratio index; for each object in the object set, count the expert counts of each category to obtain the object-level observation consistency rate, and calculate the arithmetic mean of the object-level observation consistency rates of all objects as the combinatorial bond-level observation consistency rate; summarize the category counts of all objects and all experts into a marginal distribution, and sum the squares of the marginal proportions of each category as the combinatorial bond-level coupling consistency rate; subtract the combinatorial bond-level coupling consistency rate from the combinatorial bond-level observation consistency rate to obtain the numerator, and use the difference between the numerator and the combinatorial bond-level coupling consistency rate as the denominator to obtain the combinatorial bond-level kappa consistency value; Step 64: When the cross-union ratio index at the composite key level is lower than the preset cross-union ratio threshold, select a preset number of objects from the object set according to the object-level cross-union ratio index from low to high to generate the first review set; when the cross-union ratio index at the composite key level is lower than the preset consistency threshold, generate the second review set according to the same rule; merge the two sets to remove duplicates, and generate the review to-do list based on the cross-modal anchor list. Step 7: Perform point-to-point verification of the objects in the review to-do list between cytological images and histological whole-section images, make rule-based revisions, and output a data dictionary, consistency report, and cross-modal anchor point snapshot.

2. The method for constructing a standardized cervical cancer pathological image database according to claim 1, characterized in that, The original specimen numbers corresponding to cervical cancer pathology images are hashed to obtain pseudo-anonymized specimen numbers, and unique slice identifiers and cell-level unique identifiers are generated, including: Step 11: Calculate the pseudo-anonymous specimen number using a keyed message authentication hash algorithm on the original specimen number; Step 12: Concatenate the pseudo-anonymous specimen number and the slice acquisition identifier data into a byte sequence in a fixed order and calculate using a message authentication hash algorithm to generate a unique slice identifier; wherein, the slice acquisition identifier data includes: site number, slice number, scan timestamp, and magnification; Step 13: Represent the cell frame coordinates in pixel coordinates, with the origin at the top left corner. The cell frame coordinates include the horizontal start point, vertical start point, width, and height. Concatenate the slice unique identifier with the horizontal start point, vertical start point, width, and height in a fixed order to form a byte sequence and calculate it using a message authentication hash algorithm to obtain the cell-level unique identifier.

3. The method for constructing a standardized cervical cancer pathological image database according to claim 1, characterized in that, Determine the pixel spacing of the reference layer, fix the downsampling factor, and the coordinate mapping relationship from any layer to the reference layer for scaling, rotation, and translation, including: Step 21: Select the highest resolution layer as the reference layer using the unique identifier of the slice as the index, determine the pixel spacing of the reference layer, set a fixed downsampling factor, and calculate the pixel spacing of each layer by multiplying the pixel spacing of the reference layer by the corresponding power of the fixed downsampling factor. Step 22: Obtain the image orientation based on the scan metadata and discretize it into a preset set of angles; determine the rotation parameters; determine the scaling parameters of each layer according to a fixed downsampling factor and layer number; determine the translation parameters with the upper left corner of the full-frame image of the reference layer as the origin and in pixels; and freeze the scaling parameters, rotation parameters, and translation parameters by associating them with the unique identifier of the slice using a read-only index table. Step 23: Generate the coordinate mapping relationship from any layer to the reference layer in the order of scaling, rotation and translation. Convert the pixel coordinates and cell box coordinates to the reference layer according to the coordinate mapping relationship. The cell box coordinates are converted and reconstructed into the minimum bounding rectangle according to the four vertices. The coordinates are rounded down. The conversion from pixel length to physical length is completed according to the pixel spacing of the reference layer.

4. The method for constructing a standardized cervical cancer pathological image database according to claim 1, characterized in that, The sliced ​​image is converted from color space to optical density space and normalized according to the target staining base to obtain a normalized image, including: Step 31: Locate the slice image using the slice unique identifier, extract the high quantile of the three-channel intensity of the blank background pixels as the reference white field, fix the numerical stability constant, and obtain the reference white field and the numerical stability constant. Step 32: Convert the sliced ​​image from color space to optical density space, divide the three-channel intensity of each pixel by the three-channel intensity of the reference white field element by element, and take the negative of the natural logarithm to obtain the optical density matrix. Step 33: Within the tissue region, singular value decomposition is used to obtain the source chromatin matrix composed of two principal directions, and non-negative least squares is used to calculate the chromatin coefficient matrix. Step 34: Normalize the amplitude of the staining coefficient matrix using the target staining base, calculate the quantile ratio of each channel and scale the staining coefficient matrix channel by channel, and reconstruct the normalized optical density matrix under the target staining base. Step 35: The normalized optical density matrix is ​​inversely transformed by the channel exponent and multiplied element-wise with the reference white field, cropped to the effective numerical domain, and the standardized image is obtained and stored in association with the unique identifier of the slice.

5. The method for constructing a standardized cervical cancer pathological image database according to claim 4, characterized in that, Step 34, the determination of the target chromatin, includes: The device model, magnification, scanning timestamp, white balance number and staining reagent batch number are encoded in a fixed order and the device configuration fingerprint is calculated using a message authentication hash algorithm with a key. The device configuration fingerprint is used as the key to aggregate the slice set. For each slice in the slice set, calculate the channel quantile vector using a fixed quantile for the coloring coefficient matrix, and then calculate the median quantile vector for each channel within the set. The sum of the absolute differences between the channel quantile vector and the median quantile vector of each slice is used as the distance metric. The slice with the smallest distance metric is selected as the representative slice, and the source chromatogram matrix of the representative slice is used as the target chromatogram corresponding to the device configuration fingerprint.

6. The method for constructing a standardized cervical cancer pathological image database according to claim 1, characterized in that, Establish cross-modal anchor points and create a list of cross-modal anchor points, including: Step 41: Use the pseudo-anonymous specimen number to retrieve the unique identifier of the slices on both sides; when the pseudo-anonymous specimen numbers are the same, the dissection sources are the same, and the time difference does not exceed the upper limit of the preset time window, generate a candidate pairing set; Step 42: Based on the coordinate mapping relationship, convert the pixel coordinates and cell box coordinates from any layer to their respective reference layers; convert the cell box coordinates according to the four vertices and reconstruct them into the minimum bounding rectangle in the reference layer, round the coordinates down and label the coordinate system. Step 43: Manually verify the cell-level unique identifier and histological lesion region identifier within the candidate pairing set. Concatenate the pseudo-anonymous specimen number, bilateral slice unique identifier, paired coordinates, and source rules into a byte sequence according to a fixed field order, and calculate the cross-modal anchor record identifier using a message authentication hash algorithm with a key. Store the record after deduplication according to the identifier, and establish a foreign key association with the bilateral slice unique identifier to establish a cross-modal anchor, forming a cross-modal anchor list, and generating a version number and timestamp.

7. The method for constructing a standardized cervical cancer pathological image database according to claim 1, characterized in that, Define the total severity order of lesion categories, and synthesize multiple annotations of the same object at the cell, cell cluster, lesion region, and slide level into a unique label, generating a hierarchical label set, including: Step 51: Freeze the lesion category set and severity sequence, set category codes and evidence priorities, with histological evidence taking precedence over cytological evidence; Step 52: Define a cell cluster object using a polygon represented by the reference layer coordinates. When all four vertices of the cell frame coordinates are within the polygon, add the corresponding cell-level unique identifier to the member set of the cell cluster object. Merge the candidate labels of the member set with the direct labels of the cell cluster object to form a candidate label set for the cell cluster object. Step 53: Perform a total severity check of the parent-child hierarchy for each histological lesion region identifier and each slide unique identifier. When the total severity of the parent-level unique label is lower than that of any child-level unique label, generate a hierarchy conflict item and record the histological lesion region identifier, slide unique identifier, parent-level unique label, child-level unique label and timestamp.

8. The method for constructing a standardized cervical cancer pathological image database according to claim 1, characterized in that, After establishing cross-modal anchor points, quality grading is performed, including: The cell frame coordinates are converted from any layer to the cytological reference layer according to the coordinate mapping relationship. The corresponding point pairs associated with the cross-modal anchor point record are identified and the homography mapping is solved in the local neighborhood. The cell frame coordinates are then converted to the histological reference layer according to the homography mapping. The cross-union index is calculated according to the cross-union index in step 63. The coverage cross-union index is calculated for the two masks and used as the coverage. The distance ratio is calculated as the ratio of the geometric centroid distance between the cell frame and the histological lesion region in the coordinates of the histological reference layer to the square root of the area of ​​the histological lesion region, and the time ratio is calculated as the ratio of the scan time difference to the upper limit of the preset time window. The anchor quality score is obtained by weighted summation of coverage, distance ratio and time ratio. When the anchor quality score is lower than the preset score threshold, the object corresponding to the cross-modal anchor is added to the priority queue of the review task list and the anchor quality score is recorded.

9. The method for constructing a standardized cervical cancer pathological image database according to claim 1, characterized in that, The system performs point-to-point verification of objects on the review task list between cytological images and histological whole-section images, revises coordinates and categories according to rules, and outputs a data dictionary, consistency report, and cross-modal anchor point snapshots, including: Step 71: Read the object key, slide unique identifier, baseline layer coordinates and cross-modal anchor point record identifier from the review to-do list; obtain the cell-level unique identifier and histological lesion area identifier from the cross-modal anchor point list; call the coordinate mapping relationship to convert the coordinates of any layer to the baseline layer, and complete the synchronous positioning of the cytological image and the histological whole slide image. Step 72: Verify the geometry of cell-level unique identifiers and histological lesion region identifiers in the baseline layer; when the geometry needs to be revised, record the new polygon and reconstruct it into the smallest bounding rectangle, round the coordinates down, generate a binary mask of the old and new geometry, calculate the area difference and the intersection-union ratio of the old and new geometry, and write the revised coordinates back to the baseline layer. Step 73: Based on the fixed decision-making rules of total severity, evidence priority, and category coding, candidate labels are selected to determine a unique label; when there is a cross-modal anchor point record that is associated with the histological lesion area, the evidence priority is marked and a category revision record is generated. Step 74: Write the slice unique identifier, object key, baseline coordinates, unique label, category code, severity, evidence priority, cross-modal anchor record identifier, reviser and Coordinated Universal Time timestamp into a read-only index table in a fixed field order, generate an integrity summary and associate it with the slice unique identifier foreign key; update the version number according to the three-stage rule of major version, minor version and patch version, and output the data dictionary, consistency report and cross-modal anchor snapshot.

Images