Automatic processing method of thyroid tissue digital pathological image
By employing superpixel segmentation and color distance fusion techniques, the segmentation challenge of follicular and papillary carcinoma cells in digital pathological images of thyroid cancer was solved, achieving more accurate display and segmentation of pathological features.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI TONGREN HOSPITAL
- Filing Date
- 2026-01-08
- Publication Date
- 2026-05-05
AI Technical Summary
Existing digital pathology image segmentation technology for thyroid cancer has difficulty accurately identifying the air bubbles and tissue fluid portions of follicular cancer cells, and the nuclei and cytoplasm portions of papillary cancer cells, resulting in poor segmentation results.
A superpixel segmentation method was used to segment digital pathological images of thyroid cancer into multiple superpixel blocks. Bubble, cell nucleus and cytoplasm regions were determined by gradient feature values and position parameters. Adjacent superpixel blocks were fused by color distance to improve segmentation accuracy.
It improves the segmentation accuracy and reliability of digital pathology images of thyroid cancer, better displays pathological features, and enhances segmentation results.
Smart Images
Figure CN121982304A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image segmentation technology, specifically to an automatic processing method for digital pathological images of thyroid tissue. Background Technology
[0002] The incidence of thyroid cancer is gradually increasing worldwide, and its ranking among malignant tumors is also rising. Its impact is becoming increasingly significant. Thyroid cancer can be diagnosed based on the pathology of thyroid cancer cells. Digital pathological images in pathological diagnosis are usually formed by scanning and fusing pathological tissues after dehydration, embedding, and sectioning. Unlike traditional microscopic observation of lesions, digital pathological images do not rely on optical microscopes and can be directly displayed and observed on any device. They can also be magnified at high magnification in specific areas, helping doctors to observe tissue cells more clearly.
[0003] Digital pathology images typically have high resolution, are large overall, and contain numerous pixels. Since a single pixel lacks any medical characteristics, superpixel segmentation can be used to divide digital pathology images into multiple superpixel blocks, displaying medical features based on these blocks. However, in digital pathology images, the colors of air bubbles and tissue fluid in follicular thyroid cancer cells differ. Superpixel segmentation cannot group these air bubbles and tissue fluid into the same superpixel block, thus failing to display pathological features. Therefore, it is difficult to directly identify cancerous areas through superpixel segmentation. Furthermore, papillary thyroid cancer cells exhibit color differences between their nuclei and cytoplasm and have irregular shapes, making direct identification as pathological areas through superpixel blocks insufficient in accuracy and reliability. This results in poor segmentation performance for digital pathology images of thyroid cancer. Summary of the Invention
[0004] This invention provides an automated processing method for digital pathological images of thyroid tissue to solve existing problems.
[0005] The present invention provides an automated processing method for digital pathological images of thyroid tissue, employing the following technical solution: Obtain a digital pathological image of thyroid cancer and a LAB image of the pathological image. Use superpixel segmentation to divide the LAB image into multiple superpixel blocks, each superpixel block corresponding to a region in the LAB image. The gradient feature value of each pixel is determined based on the distribution difference of the channel value of pixels in the same channel within different superpixel blocks and the gradient value of the channel value of each pixel in each channel. The location parameters of each region are determined based on the gradient direction of the gradient eigenvalues of the edge points within each region. Based on the location parameters of each region, the enclosed bubble region and cell nucleus region, as well as the corresponding cytoplasm region and tissue fluid region, are determined. The bubble region and cell nucleus region in the image are then fused with the corresponding tissue fluid region and cytoplasm region to obtain multiple initial target regions. The initial target superpixel blocks are determined based on the gradient feature values of the superpixel blocks corresponding to each initial target region; The color distance between the initial target superpixel block and its neighboring superpixel blocks is obtained based on the channel value and gradient feature value of the center point of each initial target superpixel block and its neighboring superpixel blocks. The initial target superpixel block and its neighboring superpixel blocks are then fused together based on the color distance to obtain the target superpixel block. The region corresponding to the target superpixel block in the digital pathological image of thyroid cancer is taken as the target region.
[0006] Preferably, the step of segmenting the LAB image into multiple superpixel blocks using superpixel segmentation includes: The digital pathological images of thyroid cancer are converted to grayscale to obtain grayscale images. The number of superpixel blocks is determined based on the clustering results of the grayscale values of the pixels in the grayscale image. Based on the number of superpixel blocks, the LAB image is divided into multiple superpixel blocks using a superpixel segmentation algorithm.
[0007] Preferably, the method for determining the number of superpixel blocks is as follows: The gray values of all pixels in the grayscale image are used as input to the AP clustering algorithm. The distance metric during clustering is the difference between the gray values of two pixels. The AP clustering algorithm is used to divide all pixels in the grayscale image into different clusters, and the number of clusters is counted as the number of superpixel blocks.
[0008] Preferably, the method for determining the gradient feature value of each pixel is as follows: The weighting of each channel is determined based on the similarity of the distribution of pixel channel values among different superpixel blocks; For any channel in the LAB image, sort the channel values of all pixels in each channel according to the position of the pixels to obtain the channel value matrix of each channel, and use the Sobel operator to obtain the gradient value of the channel value of each pixel. The gradient values of each pixel in each channel are linearly weighted based on the weighted values of each channel. The result of the linear weighting is the gradient feature value of each pixel.
[0009] Preferably, the method for determining the weighting of each channel is as follows: Calculate the distribution histogram of channel values for all pixels in each superpixel block in each channel, and calculate the Bach distance between the distribution histograms of channel values for any two different superpixel blocks in the same channel. The normalized result of each Bach distance is used as the distribution distance between the two corresponding superpixel blocks in the same channel; The average of the ratio of the distribution distance of two superpixel blocks in each channel to the sum of the distribution distances of the two superpixel blocks in all three channels is used as the weighting weight for each channel.
[0010] Preferably, the method for determining the location parameters of each region is as follows: Set the sign function value of the gradient feature value in each region to 1, pointing the gradient direction of the gradient feature value to the edge pixel outside each region. Set the sign function value of the gradient feature value in each region to -1, pointing the gradient direction of the gradient feature value to the edge pixel inside each region. The mean of the sign function values of all edge pixels within each region is used as the location parameter for each region.
[0011] Preferably, the step of determining the surrounded bubble region and cell nucleus region, as well as the corresponding cytoplasmic region and tissue fluid region, includes: If the location parameter of the region is 1, then the region is regarded as the enclosed bubble region or cell nucleus region. The area adjacent to the bubble region is the tissue fluid region; The region adjacent to the nucleus is the cytoplasm.
[0012] Preferably, the step of obtaining multiple initial target regions includes: The bubble region is surrounded by multiple tissue fluid regions; any one of the multiple tissue fluid regions is selected and merged with the corresponding bubble region; The nucleus is surrounded by multiple cytoplasmic regions; any one of these cytoplasmic regions is selected and fused with its corresponding nucleus region.
[0013] Preferably, the method for obtaining the color distance between the initial target superpixel block and its adjacent superpixel blocks is as follows: Calculate the squared value of the channel value difference between the center point of the initial target superpixel block and the center point of the adjacent superpixel block in each channel. The sum of the squared values of the differences in the three channels and the sum of the squared values of the differences between the center point of the initial target superpixel block and the center point of the adjacent superpixel block is taken as the color distance between the initial target superpixel block and the adjacent superpixel block.
[0014] Preferably, the method for fusing the initial target superpixel block with adjacent superpixel blocks based on color distance to obtain the target superpixel block is as follows: The color distances between all initial target superpixel blocks and their neighboring superpixel blocks are calculated, and the Otsu thresholding algorithm is used to obtain the segmentation thresholds for all the color distances as the color distance thresholds. When the color distance is less than the color distance threshold, the initial target superpixel block is fused with the adjacent superpixel blocks, and the fused superpixel block is recorded as the target superpixel block.
[0015] The beneficial effects of this invention are as follows: The automatic processing method for digital pathological images of thyroid tissue of this invention determines the number of superpixel blocks by determining the number of regions obtained through edge detection, avoiding the problem of inaccurate segmentation due to improper selection of the number of superpixel blocks, and improving the adaptive effect of superpixel segmentation; by calculating positional parameters, the enclosed bubble and cell nucleus regions are determined, and the enclosed regions are fused with the corresponding enclosing regions, enabling the bubble and cell nucleus regions with large color differences, which should have been classified into the corresponding cancerous parts, to be re-fused and segmented into the corresponding cancerous parts, making the subsequent segmentation results more accurate; by using the color difference between thyroid cancer cells and normal cells in the pathological image, the initial target superpixel blocks corresponding to thyroid cancer cells can be screened, reducing the interference of superpixel blocks corresponding to normal cells during subsequent fusion; simultaneously, the same cancerous part that was segmented into multiple superpixel blocks during the superpixel segmentation process is re-fused into a single superpixel block using color distance and positional relationship, making the final superpixel block more accurate and able to better display the medical features in the pathological image, improving the accuracy and reliability of the original image segmentation of thyroid cancer digital pathology, and enhancing the segmentation effect. Attached Figure Description
[0016] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a flowchart illustrating the implementation of an automated processing method for digital pathological images of thyroid tissue according to the present invention. Detailed Implementation
[0018] 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.
[0019] An embodiment of the automatic processing method for digital pathological images of thyroid tissue according to the present invention is as follows: Figure 1 As shown, the method includes: S1. Obtain the original digital pathology image and LAB image of thyroid cancer. Use superpixel segmentation to divide the LAB image into multiple superpixel blocks, each superpixel block corresponding to a region in the LAB image.
[0020] For superpixel segmentation, the choice of the number of superpixel blocks has a significant impact on the segmentation results. Too many or too few blocks cannot accurately reflect the characteristics of the superpixel blocks and the segmentation results. Traditional superpixel segmentation often involves multiple block number selections for segmentation comparison. This method is computationally intensive and cannot adapt well to different types of images. By leveraging the color differentiation of different tissue regions in digital pathology images, we use clustering to cluster the pixels in digital pathology images of thyroid cancer to achieve the initial segmentation effect. The number of clusters is counted and used as the basis for selecting the number of blocks for superpixel segmentation. This reduces the computational load while being able to adapt to digital pathology images of thyroid cancer with different lesion degrees.
[0021] Specifically, a digital pathological image of thyroid cancer is acquired, and the image is converted to grayscale. The grayscale values of all pixels in the grayscale image are used as input to the Affinity Propagation (AP) clustering algorithm. The distance metric during clustering is the difference between the grayscale values of two pixels. The AP clustering algorithm is used to divide all pixels in the grayscale image into different clusters, and the number of clusters, m, is counted. AP clustering is a well-known technique, and its specific process will not be elaborated further.
[0022] It should be noted that AP clustering can determine high-quality cluster centers from the data without specifying the number of clusters in advance. It can adapt to the gray value distribution of different tissue regions in the digital pathology original image of thyroid cancer after graying, and use the number m as the basis for setting the number of superpixel blocks in subsequent superpixel segmentation.
[0023] Furthermore, the digital pathological image of thyroid cancer is converted to the LAB color space to obtain a LAB image. The Simple Linear Iterative Clustering (SLIC) algorithm is then used to perform superpixel segmentation on the LAB image, with the number of superpixel blocks set to m, and each superpixel block corresponding to a region in the LAB image. The SLIC algorithm is a well-known technique in the field of image processing, and its specific process will not be elaborated further.
[0024] S2. Obtain the gradient feature value of each pixel based on the channel value of each pixel in different channels, and obtain the position parameter of the corresponding region based on the gradient direction of the gradient feature value of the edge point in each region.
[0025] In follicular carcinoma cells, the air bubbles are completely surrounded by the tissue fluid, while in papillary carcinoma cells, the nuclei are completely surrounded by the cytoplasm. During traditional superpixel segmentation, because the air bubbles in follicles and the nuclei of papillary carcinoma cells differ significantly from their respective cancerous regions in color and spatial distance, they cannot be classified into the superpixel blocks of their respective cancerous regions. Therefore, the positional distribution of each superpixel block can be determined based on the edge color gradient changes and directions, allowing for superpixel block fusion.
[0026] Specifically, for each region corresponding to each superpixel block, the edge points of each region are obtained, and the gradient change direction is defined as positive from within the region to outside the region, otherwise it is negative.
[0027] For a LAB image, for any channel, taking the L channel value of a pixel as an example, the channel values of all pixels in the L channel are sorted according to their positions to obtain the channel value matrix of the L channel. Then, the Sobel operator is used to calculate the gradient value at each pixel. The Sobel operator for calculating gradient magnitude is a commonly used technique in image processing; the specific process will not be elaborated further.
[0028] Furthermore, following the same processing method as the channel values in the L channel, the gradient values of each pixel in the a and b channels are calculated respectively.
[0029] Specifically, the similarity of pixel channel values among superpixel blocks is statistically analyzed to determine the ability of each channel value to distinguish different superpixel blocks, and the weighting weight of each channel is determined based on this.
[0030] Secondly, for each superpixel block, a histogram of the distribution of channel values of all pixels within each superpixel block on the L channel is plotted. The Barthel distance between the histograms of channel values of any two different superpixel blocks on the L channel is calculated to characterize the difference in the distribution of channel values on the L channel; the larger the Barthel distance, the greater the distribution difference. The Barthel distances between the histograms of channel values of all the superpixel blocks on the L channel are calculated and normalized. The normalized result of each Barthel distance is taken as the distribution distance between the corresponding two superpixel blocks on the L channel.
[0031] Furthermore, according to the above processing flow, the distribution distance of the two superpixel blocks on channels a and b is obtained respectively.
[0032] It should be noted that normalization is a well-known technique in the field of data processing, and the specific process will not be elaborated here. Commonly used normalization methods include, but are not limited to, Max-Min normalization and Z-score normalization. Preferably, Max-Min normalization is used to normalize the Bach distance and the metric distance respectively.
[0033] Here, the weighted weight of the L channel is expressed as: :
[0034] In the formula, m represents the number of superpixel blocks. , , These represent the distribution distances between the k-th superpixel block and the i-th superpixel block in the L channel, a channel, and b channel, respectively.
[0035] Furthermore, the weighting of each channel is determined based on the similarity of pixel channel values among superpixel blocks, thereby weighting the three gradient magnitudes to obtain the gradient feature value of each pixel. The gradient feature value of the y-th pixel is represented as... :
[0036] In the formula, , as well as Let L, a, and b represent the gradient values of the y-th pixel in the L, a, and b channels, respectively. , as well as These represent the weighted weights of the L channel, a channel, and b channel, respectively.
[0037] Furthermore, the gradient feature value of each pixel is calculated separately and arranged according to the position of the pixel. The matrix composed of the gradient feature values is used as the eigenvalue matrix. The Sobel operator is used again to process each element in the eigenvalue matrix to obtain the gradient magnitude and gradient direction of the gradient feature value of the corresponding pixel for each element.
[0038] Then, the location parameters of each region are obtained using the gradient magnitude and gradient direction of the gradient feature values of the edge points within each region. The location parameters of the p-th region are represented as follows: :
[0039] In the formula, N represents the number of edge points in the p-th region. The function represents a sign function with a value of 1 or -1. The function value is 1 when the gradient direction of the gradient eigenvalue of the q-th edge point in the p-th region points out of the p-th region; and -1 when the gradient direction of the gradient eigenvalue of the q-th edge point in the p-th region points in the p-th region.
[0040] Where the p-th region is a fully enclosed region, the gradient direction of the gradient feature values at the edge points within this region should be consistent, and the calculated result of the position parameter is -1 or 1. If the p-th region is not a fully enclosed region, the gradient direction of the gradient feature values at the edge pixels will be significantly different, and the position parameter... .
[0041] S3. Determine the enclosed bubble region and cell nucleus region, as well as the corresponding cytoplasm region and tissue fluid region, based on the position parameters of each region. Merge the bubble region and cell nucleus region in the image with the corresponding tissue fluid region and cytoplasm region to obtain multiple initial target regions.
[0042] Based on the pathological images, it can be seen that the air bubbles of follicular carcinoma cells are surrounded by tissue fluid, and the nuclei of papillary carcinoma cells are surrounded by cytoplasm. Therefore, the completely surrounded areas are the air bubbles of follicular carcinoma cells and the nuclei of papillary carcinoma cells. The area adjacent to the air bubble area is the tissue fluid area, and the area adjacent to the nucleus area is the cytoplasm area. The completely surrounded area and the surrounding area are merged. The air bubble area and the nucleus area are surrounded by multiple tissue fluid areas, and the nucleus area is surrounded by multiple cytoplasm areas. Any one of the multiple tissue fluid areas or cytoplasm areas is selected and merged with the corresponding air bubble area or nucleus area. After fusion, n initial target areas are obtained. Selecting any one, instead of merging all of them, is because when performing color distance fusion later, the color distance of the superpixel blocks of the same cancerous part will not differ too much, so the same cancerous part is divided into different superpixel blocks.
[0043] S4. Determine the initial target superpixel block based on the gradient feature value of the superpixel block corresponding to each initial target region.
[0044] In thyroid cancer pathological images, follicular cancer cells appear lighter in color, with those containing air bubbles appearing even lighter. Papillary cancer cells have cytoplasm that is nearly white, and while the nuclei are darker, their area is relatively small. Therefore, after region fusion, the superpixel blocks corresponding to papillary cancer cells are lighter in color compared to the normal region. Thus, a color channel threshold is set to obtain the superpixel blocks corresponding to each initial target region. When the gradient feature value of a superpixel block is greater than the color channel threshold, it is considered that the superpixel block is sufficiently light in color and may be a follicular or papillary cancer cell. Superpixel blocks with gradient feature values greater than the color channel threshold are then identified as the initial target superpixel blocks. The color channel threshold is obtained by calculating the first quartile of the gradient feature values of all pixels as the color channel threshold.
[0045] S5. Obtain the color distance between the initial target superpixel block and the adjacent superpixel block based on the color flux of each initial target superpixel block and the adjacent superpixel block, and fuse the initial target superpixel block and the adjacent superpixel block according to the color distance to obtain the target superpixel block.
[0046] Superpixel segmentation is essentially an application of clustering algorithms based on color and spatial distances. Therefore, after segmentation, the original pixels within each superpixel block are actually distributed within a joint centroid cluster representing both color and spatial distance. For any superpixel block, its positional relationships and color representation can be obtained based on the joint centroid. The color flux of the joint centroid of a superpixel block is denoted as . The color flux is a sequence of three-channel values at the center point of a superpixel block in the Lab color space.
[0047] Specifically, the formula for calculating the color distance between the initial target superpixel block and its neighboring superpixel blocks (including the darker superpixel blocks filtered out in step S3) is as follows:
[0048] In the formula, This represents the color distance between the i-th initial target superpixel block and the j-th adjacent superpixel block. , as well as These represent the three channel values of the center point of the i-th initial target superpixel block in the Lab color space; , as well as These represent the three channel values of the center point of the j-th adjacent superpixel block in the Lab color space, respectively. This represents the difference between the gradient feature values of the center point of the i-th initial target superpixel block and the center point of the j-th adjacent superpixel block.
[0049] Furthermore, the color distances between all initial target superpixel blocks and their adjacent superpixel blocks are statistically analyzed. The Otsu thresholding algorithm is used to obtain a segmentation threshold for all these color distances. When the color distance is less than the threshold, it indicates that the initial target superpixel block and its adjacent superpixel blocks have similar colors and, in reality, belong to the same cancerous site. These two superpixel blocks should be classified into the same category and fused. Conversely, when the color distance exceeds the threshold, it indicates that the two superpixel blocks differ significantly and do not belong to the same cancerous site. When judging adjacent superpixel blocks, superpixel blocks that have already been judged are not fused again; the fused superpixel block is recorded as the target superpixel block. The Otsu thresholding method is a well-known technique, and its specific process will not be elaborated further.
[0050] Furthermore, the target superpixel blocks in the LAB image are mapped to the original image to obtain the target region in the original image, thus obtaining the automatic segmentation result for the digital pathology image of thyroid cancer.
[0051] In summary, this invention provides an automatic processing method for digital pathological images of thyroid tissue. It determines the number of superpixel blocks by identifying the number of regions obtained through edge detection, avoiding inaccurate segmentation due to improper selection of the number of superpixel blocks and improving the adaptive effect of superpixel segmentation. By calculating positional parameters, it identifies the enclosed bubble and cell nucleus regions, and then merges the enclosed regions with their corresponding enclosing regions. This allows bubble and cell nucleus regions with significant color differences, which should have been classified as cancerous areas, to be re-merged and segmented into their corresponding cancerous areas, resulting in more accurate subsequent segmentation results. Furthermore, by analyzing the color differences between thyroid cancer cells and normal cells in the pathological image, it can screen out the initial target superpixel blocks corresponding to thyroid cancer cells, reducing interference from superpixel blocks corresponding to normal cells during subsequent fusion. Simultaneously, it re-merges the same cancerous region, which was segmented into multiple superpixel blocks during the superpixel segmentation process, into a single superpixel block using color distance and positional relationships, resulting in more accurate superpixel blocks that effectively display the medical features in the pathological image.
[0052] It should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
[0053] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
Claims
1. An automated processing method for digital pathological images of thyroid tissue, characterized in that: Obtain a digital pathological image of thyroid cancer and a LAB image of the pathological image. Use superpixel segmentation to divide the LAB image into multiple superpixel blocks, each superpixel block corresponding to a region in the LAB image. The gradient feature value of each pixel is determined based on the distribution difference of the channel value of pixels in the same channel within different superpixel blocks and the gradient value of the channel value of each pixel in each channel. The location parameters of each region are determined based on the gradient direction of the gradient eigenvalues of the edge points within each region. Based on the location parameters of each region, the enclosed bubble region and cell nucleus region, as well as the corresponding cytoplasm region and tissue fluid region, are determined. The bubble region and cell nucleus region in the image are then fused with the corresponding tissue fluid region and cytoplasm region to obtain multiple initial target regions. The initial target superpixel blocks are determined based on the gradient feature values of the superpixel blocks corresponding to each initial target region; The color distance between the initial target superpixel block and its neighboring superpixel blocks is obtained based on the channel value and gradient feature value of the center point of each initial target superpixel block and its neighboring superpixel blocks. The initial target superpixel block and its neighboring superpixel blocks are then fused together based on the color distance to obtain the target superpixel block. The region corresponding to the target superpixel block in the digital pathological image of thyroid cancer is taken as the target region.
2. The automatic processing method for digital pathological images of thyroid tissue according to claim 1, characterized in that, The step of segmenting the LAB image into multiple superpixel blocks using superpixel segmentation includes: The digital pathological images of thyroid cancer are converted to grayscale to obtain grayscale images. The number of superpixel blocks is determined based on the clustering results of the grayscale values of the pixels in the grayscale image. Based on the number of superpixel blocks, the LAB image is divided into multiple superpixel blocks using a superpixel segmentation algorithm.
3. The automatic processing method for digital pathological images of thyroid tissue according to claim 2, characterized in that, The method for determining the number of superpixel blocks is as follows: The gray values of all pixels in the grayscale image are used as input to the AP clustering algorithm. The distance metric during clustering is the difference between the gray values of two pixels. The AP clustering algorithm is used to divide all pixels in the grayscale image into different clusters, and the number of clusters is counted as the number of superpixel blocks.
4. The automatic processing method for digital pathological images of thyroid tissue according to claim 1, characterized in that, The method for determining the gradient feature value of each pixel is as follows: The weighting of each channel is determined based on the similarity of the distribution of pixel channel values among different superpixel blocks; For any channel in the LAB image, sort the channel values of all pixels in each channel according to the position of the pixels to obtain the channel value matrix of each channel, and use the Sobel operator to obtain the gradient value of the channel value of each pixel. The gradient values of each pixel in each channel are linearly weighted based on the weighted values of each channel. The result of the linear weighting is the gradient feature value of each pixel.
5. The automatic processing method for digital pathological images of thyroid tissue according to claim 4, characterized in that, The method for determining the weighting of each channel is as follows: Calculate the distribution histogram of channel values for all pixels in each superpixel block in each channel, and calculate the Bach distance between the distribution histograms of channel values for any two different superpixel blocks in the same channel. The normalized result of each Bach distance is used as the distribution distance between the two corresponding superpixel blocks in the same channel; The average of the ratio of the distribution distance of two superpixel blocks in each channel to the sum of the distribution distances of the two superpixel blocks in all three channels is used as the weighting weight for each channel.
6. The automatic processing method for digital pathological images of thyroid tissue according to claim 1, characterized in that, The method for determining the location parameters of each region is as follows: Set the sign function value of the gradient feature value in each region to 1, pointing the gradient direction of the gradient feature value to the edge pixel outside each region. Set the sign function value of the gradient feature value in each region to -1, pointing the gradient direction of the gradient feature value to the edge pixel inside each region. The mean of the sign function values of all edge pixels within each region is used as the location parameter for each region.
7. The automatic processing method for digital pathological images of thyroid tissue according to claim 1, characterized in that, The step of determining the surrounded bubble region and cell nucleus region, as well as the corresponding cytoplasmic region and tissue fluid region, includes: If the location parameter of the region is 1, then the region is regarded as the enclosed bubble region or cell nucleus region. The area adjacent to the bubble region is the tissue fluid region; The region adjacent to the nucleus is the cytoplasm.
8. The automatic processing method for digital pathological images of thyroid tissue according to claim 1, characterized in that, The step of obtaining multiple initial target regions includes: The bubble region is surrounded by multiple tissue fluid regions; any one of the multiple tissue fluid regions is selected and merged with the corresponding bubble region; The nucleus is surrounded by multiple cytoplasmic regions; any one of these cytoplasmic regions is selected and fused with its corresponding nucleus region.
9. The automatic processing method for digital pathological images of thyroid tissue according to claim 1, characterized in that, The method for obtaining the color distance between the initial target superpixel block and its adjacent superpixel blocks is as follows: Calculate the squared value of the channel value difference between the center point of the initial target superpixel block and the center point of the adjacent superpixel block in each channel. The sum of the squared values of the differences in the three channels and the sum of the squared values of the differences between the center point of the initial target superpixel block and the center point of the adjacent superpixel block is taken as the color distance between the initial target superpixel block and the adjacent superpixel block.
10. The automatic processing method for digital pathological images of thyroid tissue according to claim 1, characterized in that, The method for fusing the initial target superpixel block with adjacent superpixel blocks based on color distance to obtain the target superpixel block is as follows: The color distances between all initial target superpixel blocks and their neighboring superpixel blocks are calculated, and the Otsu thresholding algorithm is used to obtain the segmentation thresholds for all the color distances as the color distance thresholds. When the color distance is less than the color distance threshold, the initial target superpixel block is fused with the adjacent superpixel blocks, and the fused superpixel block is recorded as the target superpixel block.