Counting method of Ki-67 proliferation index based on improved UNet network

Abstract

The invention relates to a ki-67 proliferation index counting method based on an improved UNet network. The method comprises the following steps: acquiring a pathological section image of ki-67 positive tumor cells; calibrating the pathological section image according to a preset standard optical parameter set to obtain a calibrated section image; extracting a region mean value and a region variance of the positive staining signal of the calibration slice image, and the cell density of the calibration slice image; identifying a hot spot region of the calibration slice image based on a region mean value, a region variance and a cell density; based on the hot spot region, a pre-trained improved UNet network is adopted to carry out ki-67 positive tumor cell recognition, and a ki-67 positive tumor cell recognition result is obtained; and calculating a Ki-67 proliferation index based on an identification result. By adopting the method, the accuracy and efficiency of ki-67 proliferation index counting can be improved, manual intervention is reduced through automatic image processing, and objective evaluation of tumor proliferation activity is facilitated.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence-assisted medical image processing, and in particular relates to a method for counting the ki-67 proliferation index based on an improved UNet network. Background Technology

[0002] With the development of medical image processing technology and artificial intelligence-assisted diagnostic technology, computer vision-based automated cell counting technology has emerged, which is characterized by high efficiency, objectivity, and reproducibility. In traditional techniques, Ki-67 proliferation index counting mainly relies on pathologists manually observing pathological slide images under a microscope, identifying and counting Ki-67-positive tumor cells, and then calculating their ratio to the total number of tumor cells. The process involves visually screening hotspot areas, judging staining intensity cell by cell, and relying on subjective evaluation based on experience; the entire process is time-consuming and requires highly specialized training. Current methods based on manual counting or traditional image processing have many problems: manual counting is easily affected by subjective factors such as fatigue and experience differences, leading to poor consistency and low reproducibility; traditional image processing methods often rely on fixed threshold segmentation, making it difficult to adapt to optical differences and staining variations in different slides, easily resulting in false positives or false negatives; the entire process is inefficient, unable to meet the needs of large-scale clinical screening, and lacks standardized objective indicators, affecting the reliability of tumor proliferation activity assessment. Summary of the Invention

[0003] Therefore, it is necessary to provide a counting method for the ki-67 proliferation index based on the improved UNet network that can solve the above problems.

[0004] In a first aspect, this application provides a method for counting the ki-67 proliferation index based on an improved UNet network, including:

[0005] Obtain pathological images of Ki-67 positive tumor cells;

[0006] The pathological slide images are calibrated according to a preset set of standard optical parameters to obtain calibrated slide images.

[0007] Extract the regional mean and regional variance of positive staining signals from calibration slide images, as well as the cell density of calibration slide images;

[0008] Based on the region mean, region variance, and cell density, hotspot regions in the calibration slice image are identified.

[0009] Based on hotspot regions, a pre-trained improved UNet network was used to identify ki-67 positive tumor cells, and the identification results of ki-67 positive tumor cells were obtained.

[0010] The Ki-67 proliferation index was calculated based on the Ki-67 positive tumor cell identification results.

[0011] In one embodiment, extracting the region mean and region variance of the positive staining signal from the calibration slide image, as well as the cell density of the calibration slide image, includes:

[0012] The calibration slice image is subjected to RGB channel separation processing to obtain images of each channel;

[0013] A color deconvolution algorithm is used to decompose each channel image into multiple color components;

[0014] The positive staining component is obtained by extracting the staining component corresponding to the positive staining signal from multiple staining components.

[0015] Based on the positive staining components, a staining density map is generated;

[0016] Based on the staining density map, the calibration slice image is divided into regions according to a preset density threshold to obtain strong signal regions;

[0017] Based on the strong signal region, the average value of the positive staining component is calculated as the region mean;

[0018] Based on the strong signal region, the variance of the positive staining component is calculated as the region variance;

[0019] Cell density is generated based on the regional mean and regional variance.

[0020] In one embodiment, hotspot regions of the calibration slice image are identified based on region mean, region variance, and cell density, including:

[0021] Based on the region mean, region variance, and cell density, the dynamic intensity threshold and dynamic density threshold are calculated using the following formulas:

[0022]

[0023]

[0024] Where Q is the dynamic intensity threshold, To obtain the 75th percentile of the region mean, The 25th percentile of the regional variance. M represents the maximum value of the regional variance, and M is the dynamic density threshold. The 80th percentile of cell density;

[0025] Select calibration slice images where the region mean J ≥ dynamic intensity threshold Q and cell density is [not specified]. ≥Dynamic density threshold Strong signal areas are considered potential hotspots.

[0026] Calculate the neighborhood cell density difference of potential hotspot areas, and based on the neighborhood cell density difference, select qualified potential areas according to a preset cell density difference threshold;

[0027] Morphological opening operations are performed on potential qualified regions to obtain hotspot regions.

[0028] In one embodiment, the pre-trained improved UNet network includes an encoder and a decoder. The encoder consists of four downsampling blocks, each of which uses a depthwise separable convolutional layer. The number of convolutional kernels in the depthwise separable convolutional layers are 64, 128, 256, and 512, respectively. A skip connection is set between the decoder and the encoder, and the skip connection uses the CBAM attention mechanism.

[0029] Based on hotspot regions, a pre-trained improved UNet network was used to identify Ki-67 positive tumor cells, and the Ki-67 positive tumor cell identification results were obtained, including:

[0030] The hotspot region is fed into a pre-trained improved UNet network, and the encoder extracts the encoded features.

[0031] A decoder is used, combined with the CBAM attention mechanism in the skip connection, to process the encoded features and obtain the decoded feature map;

[0032] Scan the decoded feature map according to the preset connectivity judgment criteria to identify all connected regions;

[0033] Based on a preset area threshold, connected regions larger than the preset area threshold are selected as the results of Ki-67 positive tumor cell identification.

[0034] In one embodiment, the preset standard optical parameter set includes a set of reference optical parameters, multiple sets of calibration optical parameters, and multiple deviation levels, wherein each deviation level corresponds to a set of calibration optical parameters.

[0035] The calibration slice image is calibrated according to a preset set of standard optical parameters to obtain the calibration slice image, including:

[0036] The pathological slide image is divided into multiple image sub-regions according to a preset grid.

[0037] Extract the actual optical parameters of each image sub-region;

[0038] Calculate the deviation between the actual optical parameters and the reference optical parameters;

[0039] Based on the deviation value, determine the deviation level of each image sub-region;

[0040] For each image sub-region, calibration optical parameters with corresponding deviation levels are used for calibration to obtain the calibrated sub-region;

[0041] Edge pixel fusion processing is performed on each calibrated sub-region to obtain the calibration slice image.

[0042] In one embodiment, after obtaining the Ki-67 positive tumor cell identification results and before calculating the Ki-67 proliferation index, the method further includes:

[0043] Morphological characteristics of each identified positive tumor cell were extracted from the Ki-67 positive tumor cell identification results;

[0044] Based on the preset morphological feature threshold, target cells whose morphological features meet the preset morphological feature threshold are selected;

[0045] Overlap detection is performed between target cells to obtain overlapping regions;

[0046] The non-maximum suppression algorithm was used to process the overlapping regions to obtain optimized Ki-67 positive tumor cell identification results, which were then used to calculate the Ki-67 proliferation index.

[0047] In one embodiment, the Ki-67 proliferation index is calculated based on the Ki-67 positive tumor cell identification results using the following formula:

[0048]

[0049] in, Ki-67 proliferation index, The number of positive tumor cells in the optimized Ki-67 positive tumor cell identification results. Weights are matched for morphological features. Here, is the overlap correction factor, and n is the total number of connected regions. Let be the actual pixel area of ​​the i-th connected region that meets the preset area threshold. For the preset area threshold, Let be the cell density within the i-th connected region that meets the preset area threshold.

[0050] Secondly, this application also provides a counting device for the ki-67 proliferation index based on an improved UNet network, comprising:

[0051] The slide image acquisition module is used to acquire pathological slide images of Ki-67 positive tumor cells;

[0052] The slide image calibration module is used to calibrate pathological slide images according to a preset set of standard optical parameters to obtain calibrated slide images;

[0053] The signal and density extraction module is used to extract the regional mean and regional variance of the positive staining signal in the calibration slide image, as well as the cell density in the calibration slide image.

[0054] The hotspot region identification module is used to identify hotspot regions in the calibration slice image based on the region mean, region variance, and cell density.

[0055] The tumor cell identification module is used to identify ki-67 positive tumor cells based on hotspot regions using a pre-trained improved UNet network, and obtain ki-67 positive tumor cell identification results.

[0056] The proliferation index calculation module is used to calculate the Ki-67 proliferation index based on the Ki-67 positive tumor cell identification results.

[0057] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described counting method for the ki-67 proliferation index based on the improved UNet network.

[0058] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described counting method for the ki-67 proliferation index based on the improved UNet network.

[0059] The aforementioned Ki-67 proliferation index counting method based on an improved UNet network acquires pathological slide images of Ki-67 positive tumor cells and calibrates them using a pre-set set of standard optical parameters. This standardizes the image input, eliminates inconsistencies caused by differences in slide preparation and optical conditions, and solves the problem of false positives or false negatives caused by the inability of traditional methods to adapt to staining variations due to fixed thresholds. The method extracts the regional mean, regional variance, and cell density of positive staining signals from the calibrated images, replacing subjective experience and avoiding subjective bias in manual counting. Based on these features, it intelligently identifies hotspot regions, prioritizing areas with dense cells and strong staining, thus improving the targeting of the count. A pre-trained improved UNet network is used for automated cell identification. Its encoder-decoder structure, combined with an attention mechanism, segments Ki-67 positive tumor cells, significantly reducing manual intervention time and subjective errors, and improving processing efficiency. The proliferation index is calculated based on the identification results, and quantitative analysis ensures the objectivity and repeatability of the assessment. This achieves the beneficial effects of improving counting accuracy, efficiency, and standardization. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart of the ki-67 proliferation index counting method based on the improved UNet network of the present invention;

[0062] Figure 2 This is a structural diagram of the ki-67 proliferation index counting device based on the improved UNet network of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0064] In one embodiment, such as Figure 1 As shown, a method for counting the Ki-67 proliferation index based on an improved UNet network is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. The implementation environment of this invention includes a pathological microscope or digital slide scanner with image acquisition capabilities, a computer terminal, and a server. Application scenarios include: when a hospital pathology department needs to efficiently and objectively assess tumor proliferative activity, the acquisition device acquires Ki-67 stained pathological slide images and transmits them to the terminal. After the terminal completes image calibration and feature extraction, it uploads the hotspot region data to a cloud server. The server uses a pre-trained improved UNet network to identify positive cells and sends the identification results and proliferation index back to the local terminal for pathologists to refer to for diagnosis. In this embodiment, the method includes the following steps:

[0065] S01, Obtain pathological section images of Ki-67 positive tumor cells.

[0066] Among them, pathological slide images are digital images of tumor tissue slides after immunohistochemical staining, which can be captured by image acquisition equipment such as pathological microscopes or digital slide scanners.

[0067] S02, calibrate the pathological slide image according to the preset standard optical parameter set to obtain the calibrated slide image.

[0068] The preset standard optical parameter set is a collection of reference optical parameters, calibration optical parameters, and deviation levels, used to standardize the optical conditions of pathological slide images. The reference optical parameters represent the ideal imaging standard, while the calibration optical parameters correspond to adjustable parameter sets with different deviation levels to adapt to various imaging variations. The calibration slide image is a standardized image obtained through calibration processing, designed to eliminate inconsistencies caused by slide preparation, staining differences, or optical equipment. During calibration, the pathological slide image is divided into multiple sub-regions, the actual optical parameters of each sub-region are extracted, their deviation values ​​from the reference parameters are calculated, the corresponding deviation level is determined based on the deviation range, and the corresponding level of calibration optical parameters is applied for local correction. Pixel fusion processing integrates the sub-regions to obtain the calibrated slide image, ensuring the reliability of subsequent feature extraction and providing a consistent input basis for automated recognition.

[0069] S03, extract the regional mean and regional variance of the positive staining signal from the calibration slide image, as well as the cell density of the calibration slide image.

[0070] The region mean is the statistical average of the positive staining signal within the selected region, reflecting the staining intensity; the region variance represents the signal dispersion in the corresponding region, used to assess staining uniformity; cell density is generated based on the region mean and variance, characterizing the cell density per unit area. In implementation, positive staining signals can be extracted using a color deconvolution algorithm to generate a staining density map, and strong signal regions can be divided based on a preset threshold to calculate the mean and variance; cell density can be derived using a statistical model (such as combining region feature weights). Automated image processing replaces manual judgment, providing a data foundation for hotspot region identification.

[0071] S04 identifies hotspot regions in the calibration slice image based on the region mean, region variance, and cell density.

[0072] Hotspot regions are localized areas in the image with dense cell distribution and strong positive staining signals, used for priority focusing and counting analysis. Their identification can be achieved through dynamic threshold calculation and region filtering. For example, dynamic intensity thresholds and dynamic density thresholds are calculated based on region mean, region variance, and cell density to filter potential regions that meet the threshold conditions. Neighborhood density difference analysis and morphological processing are combined to optimize boundaries, automatically locating hotspot regions and improving the targeting and reliability of the counting. This method reduces human subjectivity through adaptive adjustment of statistical features.

[0073] S05, based on hotspot regions, uses a pre-trained improved UNet network to identify ki-67 positive tumor cells and obtain ki-67 positive tumor cell identification results.

[0074] The improved UNet network is a semantic segmentation network that incorporates an encoder-decoder structure and an attention mechanism to accurately segment target cells from images. The Ki-67 positive tumor cell identification result refers to the binary map or set of regions output by the network, containing all identified positive cell regions. This is achieved by inputting hotspot regions into a pre-trained network. The encoder extracts multi-level features through multi-layer downsampling (e.g., using depthwise separable convolution), while the decoder reconstructs spatial details through upsampling and an attention mechanism in skip connections (e.g., CBAM). By performing connected component analysis and area thresholding on the output feature map, positive cell regions meeting morphological criteria are automatically identified. This method adaptively learns cell features through a deep learning model, improving the accuracy and robustness of identification.

[0075] S06, Calculate the Ki-67 proliferation index based on the Ki-67 positive tumor cell identification results.

[0076] The Ki-67 proliferation index is a quantitative ratio representing the proliferation activity of Ki-67 positive tumor cells to the total proliferation activity of tumor cells. Its calculation can be achieved by integrating the identification results of Ki-67 positive tumor cells. For example, it can be calculated based on the number of positive tumor cells, combined with morphological feature matching weights, overlap correction coefficients, and parameters such as regional area and cell density, and output the proliferation index as a percentage to achieve an objective and quantitative assessment of tumor proliferation activity.

[0077] The aforementioned Ki-67 proliferation index counting method based on an improved UNet network overcomes image inconsistencies caused by differences in slide preparation and optical conditions by acquiring pathological slide images and calibrating them using a preset set of standard optical parameters. It eliminates subjective bias by extracting the regional mean, regional variance, and cell density of positive staining signals to replace traditional manual judgment. Based on these features, it intelligently identifies hotspot regions, prioritizing areas with dense cells and strong staining, thus improving the targeting of the count. A pre-trained improved UNet network is used for automated cell identification; its encoder-decoder structure combined with an attention mechanism improves the accuracy and robustness of positive tumor cell segmentation. Quantitative calculation of the proliferation index ensures the objectivity and reproducibility of the evaluation results. This method, through fully automated processing, improves the accuracy, efficiency, and standardization of Ki-67 proliferation index counting, solving problems such as poor reproducibility, weak adaptability, and low efficiency in traditional methods.

[0078] In one embodiment, extracting the region mean and region variance of the positive staining signal from the calibration slide image, as well as the cell density of the calibration slide image, includes:

[0079] S11, Perform RGB channel separation processing on the calibration slice image to obtain images of each channel;

[0080] S12, the color deconvolution algorithm is used to decompose the image of each channel into multiple color components;

[0081] S13, extract the staining component corresponding to the positive staining signal from multiple staining components to obtain the positive staining component;

[0082] S14, Generate a staining density map based on the positive staining components;

[0083] S15, Based on the staining density map, the calibration slice image is divided into regions according to a preset density threshold to obtain strong signal regions;

[0084] S16, Based on the strong signal region, calculate the average value of the positive staining component as the region mean;

[0085] S17, Based on the strong signal region, calculate the variance of the positive staining component as the region variance;

[0086] S18 generates cell density based on the regional mean and regional variance.

[0087] For example, when extracting the regional mean, regional variance, and cell density of the positive staining signal from the calibration slide image, image processing algorithms such as the split function in OpenCV can be used to separate the red, green, and blue (RGB) channels of the pathological slide image, resulting in three independent channel images. Color deconvolution algorithms based on principal component analysis (PCA) or independent component analysis (ICA) can be used to further decompose each channel image into multiple non-interfering staining components, each corresponding to the distribution of different staining substances in the image. Based on the spectral characteristics or grayscale range corresponding to Ki-67 positive staining determined experimentally beforehand, staining components matching these characteristics can be selected from the multiple decomposed staining components; these are the positive staining components. Based on the grayscale value of each pixel in the positive staining component, its... The staining density map is mapped to the corresponding staining concentration value, forming a staining density map in pixels (the higher the gray value, the greater the staining density). Using a preset density threshold (determined through statistical analysis of a large number of samples), areas in the staining density map with pixel density values ​​greater than the threshold are defined as strong signal regions, thus achieving regional division of the pathological slide image. For the defined strong signal regions, the gray values ​​of the positive staining components corresponding to all pixels in the region are statistically analyzed and their arithmetic mean is calculated as the region mean. At the same time, the average of the sum of squares of the deviations between the gray values ​​of the positive staining components of all pixels in the region and the region mean is calculated as the region variance. The obtained region mean and region variance can be substituted into the mapping model (trained through a large number of labeled samples, containing the correspondence between region mean, region variance and cell density) to output the corresponding cell density value through calculation.

[0088] In one embodiment, hotspot regions of the calibration slice image are identified based on region mean, region variance, and cell density, including:

[0089] S21, based on the region mean, region variance, and cell density, calculate the dynamic intensity threshold and dynamic density threshold using the following formulas:

[0090]

[0091]

[0092] Where Q is the dynamic intensity threshold, To obtain the 75th percentile of the region mean, The 25th percentile of the regional variance. M represents the maximum value of the regional variance, and M is the dynamic density threshold. The 80th percentile of cell density;

[0093] S22, Filter out regions in the calibration slice image whose mean J ≥ dynamic intensity threshold Q and cell density is... ≥Dynamic density threshold Strong signal areas are considered potential hotspots.

[0094] S23, calculate the neighborhood cell density difference of potential hotspot areas, and based on the neighborhood cell density difference, select qualified potential areas according to the preset cell density difference threshold;

[0095] S24, perform morphological opening operation on the qualified potential region to obtain hotspot region.

[0096] Specifically, for all the identified strong signal areas, statistical tools can be used to sort the regional means of each strong signal area in ascending order, and the value at the 75th percentile of the sorted results can be taken as the 75th percentile of the regional means. After sorting the regional variances of each strong signal region in ascending order, the value at the 25th percentile is taken as the 25th percentile of the regional variance. ), and extract the maximum value of the regional variance among all strong signal regions as the maximum regional variance ( ),Will , , Substitute into the formula " "The dynamic intensity threshold (Q) is calculated to measure whether the staining intensity of a region meets the intensity requirements of a hotspot region; at the same time, the cell density of each strong signal region is sorted in ascending order, and the value at the 80th percentile of the sorting results is taken as the 80th percentile of the cell density." The strong signal regions are directly used as the dynamic density threshold (M, used to measure whether the cell density of a region meets the density requirements of a hotspot region). All strong signal regions in the scanned calibration slice image are checked, and it is determined whether the mean (J) of each strong signal region is greater than or equal to the dynamic intensity threshold (Q), and whether the cell density (X) of that region is greater than or equal to the dynamic density threshold (M). Strong signal regions that simultaneously meet both conditions are marked as potential hotspot regions. Based on the neighborhood range (e.g., a rectangular area extending outwards by 3×3 or 5×5 pixels from the boundary of the potential hotspot region; the specific range is determined through previous sample validation), the difference between the cell density of each potential hotspot region and the average cell density of all non-potential hotspot regions within that neighborhood (i.e., neighborhood cell density difference) is calculated. A preset cell density difference threshold (e.g., difference ≥ 25%, this threshold can be fine-tuned according to the slice type) is set, and potential hotspot regions with a neighborhood cell density difference greater than or equal to this preset threshold are selected and marked as qualified potential regions. Morphological opening operations are performed on the target potential regions: For example, a 3×3 pixel rectangular structuring element (the size of the structuring element can be adjusted according to the average diameter of cells in the slice) is used. First, an erosion operation is performed on the target potential regions to remove small noise blocks at the edge of the regions (such as isolated regions with an area of ​​less than 5 pixels). Then, a dilation operation is performed on the eroded regions to restore the main outline of the target potential regions, resulting in hotspot regions with complete boundaries and no noise interference.

[0097] In one embodiment, the pre-trained improved UNet network includes an encoder and a decoder. The encoder consists of four downsampling blocks, each of which uses a depthwise separable convolutional layer. The number of convolutional kernels in the depthwise separable convolutional layers are 64, 128, 256, and 512, respectively. A skip connection is set between the decoder and the encoder, and the skip connection uses the CBAM attention mechanism.

[0098] Based on hotspot regions, a pre-trained improved UNet network was used to identify Ki-67 positive tumor cells, and the Ki-67 positive tumor cell identification results were obtained, including:

[0099] S31, input the hotspot region into the pre-trained improved UNet network, and extract the encoded features through the encoder;

[0100] S32 uses a decoder and combines the CBAM attention mechanism in the skip connection to process the encoded features and obtain the decoded feature map;

[0101] S33, scan and decode the feature map according to the preset connectivity judgment criteria to identify all connected regions;

[0102] S34. Based on a preset area threshold, connected regions larger than the preset area threshold are selected as the ki-67 positive tumor cell identification results.

[0103] For example, the structure of the improved UNet network includes an encoder and a decoder. The encoder consists of four downsampling blocks, each of which uses a depthwise separable convolutional layer with a 3×3 kernel, a stride of 2, and same padding. The number of kernels in each depthwise separable convolutional layer is 64, 128, 256, and 512, respectively. Each depthwise separable convolutional layer is followed by a batch normalization (BN) layer and a ReLU activation function to enhance feature extraction capabilities. The decoder and encoder have a symmetrical structure, with each decoding block using a 3×3 kernel. Upsampling is achieved using transposed convolutional layers with a kernel size of 2, a stride of 2, and same padding. Skip connections are set between the corresponding downsampling blocks of the decoder and encoder. These skip connections integrate the CBAM attention mechanism (first, channel features are obtained through global average pooling and global max pooling, then channel attention weighting is achieved by outputting channel weights through an MLP with one hidden layer, then spatial features are obtained through channel-dimensional average pooling and max pooling, and spatial attention weighting is achieved by outputting spatial weights through a 3×3 convolution) to enhance the transmission and utilization of key features. The recognition process is as follows: The hotspot region image is preprocessed to the same size as the input image of the pre-trained network (e.g., 256×256 pixels, pixel values ​​normalized to the range [0,1]). After being input into the pre-trained improved UNet network, it is processed sequentially by four downsampling blocks of the encoder: The first downsampling block (64 convolutional kernels) performs depthwise separable convolution on the input image, outputting a 128×128×64 feature map, which is then activated by BN and ReLU and passed to the second downsampling block; the second downsampling block (128 convolutional kernels) outputs a 64×64×128 feature map, which is similarly activated by BN and ReLU and passed to the third downsampling block; the third downsampling block (256 convolutional kernels) outputs a 32×32×256 feature map, which is activated by BN and ReLU and passed to the fourth downsampling block; the fourth downsampling block (512 convolutional kernels) outputs a 16×16×512 feature map. The above multi-scale feature maps together constitute the encoded features.The decoder starts processing the 16×16×512 encoded features output by the encoder: the first decoding block upsamples its size to 32×32×256 through transposed convolution, and simultaneously calls skip connections to weight the 32×32×256 feature map output by the encoder's third downsampled block through CBAM attention mechanism, and then concatenates it with the upsampled feature map in the channel dimension (to obtain a 32×32×512 feature map), and then adjusts the number of channels to 256 through 1×1 convolution, and after BN and ReLU activation, it is passed to the next decoding block; the subsequent second decoding block (processed to 64×64×128), third decoding block (processed to 128×128×64), and fourth decoding block (processed to 256×256×1) all repeat the above process of upsampling, CBAM attention weighting, channel concatenation, and convolution activation, and output a single-channel decoded feature map with the same size as the input hotspot region (the pixel value represents the probability of Ki-67 positive tumor cells at that location). The preset connectivity criterion is 8-neighbor connectivity (i.e., adjacent pixels in the top, bottom, left, right, and four diagonal directions of a pixel are considered connected). The decoded feature map is binarized (with a probability threshold of 0.5; pixels with a value ≥ 0.5 are set to 1, otherwise to 0). Then, a flood fill algorithm is used to scan the binarized feature map, marking all sets of pixels with a value of 1 and that are interconnected as independent connected regions, thus identifying all connected regions. A preset area threshold (determined by statistically analyzing the average pixel area of ​​positive cells in a large number of labeled Ki-67 positive tumor cell slices; for example, if the average area is 80 pixels, the preset threshold is set to 50 pixels to exclude noise regions) is used to scan all identified connected regions, calculate the number of pixels in each region (i.e., the region area), and filter out connected regions with an area greater than the preset area threshold. These regions can be used as the Ki-67 positive tumor cell identification results.

[0104] In one embodiment, the preset standard optical parameter set includes a set of reference optical parameters, multiple sets of calibration optical parameters, and multiple deviation levels, wherein each deviation level corresponds to a set of calibration optical parameters.

[0105] Based on a preset set of standard optical parameters, the pathological slide images are calibrated to obtain calibrated slide images, including:

[0106] S41, divide the pathological slide image into multiple image sub-regions according to a preset grid;

[0107] S42, extract the actual optical parameters of each image sub-region;

[0108] S43, calculate the deviation between the actual optical parameters and the reference optical parameters;

[0109] S44, Determine the deviation level of each image sub-region based on the deviation value;

[0110] S45, for each image sub-region, calibration is performed using calibration optical parameters corresponding to the deviation level to obtain the calibrated sub-region;

[0111] S46. Edge pixel fusion processing is performed on each calibrated sub-region to obtain the calibration slice image.

[0112] Specifically, the baseline optical parameters can be obtained by statistically analyzing the average grayscale values ​​(e.g., R channel 128, G channel 110, B channel 95) and average color saturation (35%) of 100 standard Ki-67 pathological slides acquired under standard imaging conditions (light source intensity 5000 lux, color temperature 5500 K, objective lens magnification 20x). Multiple sets of calibration optical parameters are pre-calibrated parameter adjustment values ​​for different imaging deviation scenarios. For example, when the light source intensity is insufficient, the corresponding compensation is +20 for R channel grayscale value, +15 for G channel, +10 for B channel, and +8% for color saturation. Multiple deviation levels are levels divided according to the deviation range between the actual optical parameters and the baseline parameters (e.g., deviation value 0-10 is level 1, 11-20 is level 2, and 21-30 is level 3), and each deviation level uniquely corresponds to a set of calibration optical parameters. When calibrating pathological slide images based on this preset standard optical parameter set, OpenCV's image segmentation function can be used to uniformly divide the pathological slide image into multiple non-overlapping image sub-regions according to a preset grid size of 200×200 pixels (this size has been determined through testing to balance calibration accuracy and computational efficiency). If the remaining portion of the image edge is less than 200×200 pixels, the edge region is merged with the adjacent sub-region to ensure coverage of the entire image. For each image sub-region, the RGB values ​​of all pixels within the sub-region can be read using Python's PIL library, and the average grayscale value of the RGB three channels of the sub-region (the arithmetic mean of the grayscale values ​​of all pixels in each channel) can be calculated. The average color saturation of the sub-region is extracted using an RGB to HSV color space conversion algorithm. These two parameters are used as the actual optical parameters of the sub-region. Then, S43 calculates the deviation value using the Euclidean distance formula.

[0113]

[0114] The overall deviation between the actual optical parameters and the reference optical parameters is comprehensively quantified. The calculated deviation value is matched with a preset deviation level range, such as a deviation value of 8 corresponding to level 1, a deviation value of 18 corresponding to level 2, and a deviation value of 25 corresponding to level 3, to determine the deviation level of each sub-region. Based on the deviation level of the sub-region, the corresponding calibration optical parameters from the preset standard optical parameter set are called to adjust the RGB three-channel grayscale values ​​and color saturation of the sub-region. For example, if the calibration parameters corresponding to deviation level 2 are R+18, G+14, B+9, and saturation +7%, then the actual R of the sub-region will be adjusted accordingly. The mean value is increased by 18, the mean value of G is increased by 14, the mean value of B is increased by 9, and the mean value of saturation is increased by 7%. The RGB values ​​of all pixels in the sub-region are corrected according to the adjusted parameters to obtain the calibrated sub-region. A Gaussian blur fusion algorithm can be used to calculate the weighted average RGB value of the pixels in the overlapping 10-pixel wide edge of the adjacent calibrated sub-regions (this width has been determined by testing and can eliminate splicing marks). The weight of the pixels in the overlapping area is calculated according to the rule that the pixel closer to the center of the sub-region has a weight of 0.8 and the pixel closer to the edge has a weight of 0.2. This process is performed on the edges of all adjacent sub-regions, and the images are spliced ​​to form a calibrated slice image with unified optical parameters.

[0115] In one embodiment, after obtaining the Ki-67 positive tumor cell identification results and before calculating the Ki-67 proliferation index, the method further includes:

[0116] S51, extract the morphological characteristics of each identified positive tumor cell from the Ki-67 positive tumor cell identification results;

[0117] S52, based on the preset morphological feature threshold, select target cells whose morphological features meet the preset morphological feature threshold;

[0118] S53, performs overlap detection between target cells to obtain overlapping areas;

[0119] S54. The non-maximum suppression algorithm is used to process the overlapping region to obtain the optimized Ki-67 positive tumor cell identification results. The optimized Ki-67 positive tumor cell identification results are used for subsequent calculation of Ki-67 proliferation index.

[0120] For example, before calculating the Ki-67 proliferation index, the recognition results can be optimized as follows: The cv2.findContours function in OpenCV can be used to perform contour detection on the connected regions of each identified positive tumor cell to obtain the set of pixel coordinates of the cell contour. Based on this set of coordinates, the morphological features of each positive tumor cell are calculated. The morphological features include the cell region area (the total number of pixels enclosed by the contour), perimeter (the pixel length of the contour), roundness (calculated by the formula 4π×area / perimeter², the closer the value is to 1, the closer the cell is to round), and the aspect ratio of the cell nucleus (obtained by the cv2.minAreaRect function, which obtains the minimum bounding rectangle of the cell and calculates the ratio of the long side length to the short side length of the rectangle). At the same time, the probability value when the improved UNet network outputs the connected region is retained as the cell recognition confidence. The preset morphological feature thresholds can be determined by statistically analyzing the morphological feature distribution of a large number of Ki-67 positive tumor cell samples labeled by pathologists. The range is as follows: cell area 20-150 pixels (excluding excessively small noise blocks and excessively large cell aggregation areas), roundness 0.4-1.0 (excluding irregularly shaped non-cellular areas), nucleus aspect ratio 1.0-3.0 (conforming to the morphological proportions of normal tumor cell nuclei), and cell recognition confidence ≥0.7 (ensuring reliable recognition). All positive tumor cells are traversed, and cells whose morphological features fully meet the above preset threshold range are selected and marked as target cells. The intersection-over-union (IoU) algorithm can be used to calculate the degree of overlap between any two target cells. The formula for calculating the IoU is:

[0121]

[0122] An overlap threshold of 0.3 was set (tested and confirmed to effectively distinguish between normal adjacent cells and excessively overlapping duplicate recognition areas). If the IoU between two target cells is ≥0.3, they are considered to overlap, and the pixel coordinates of the intersection of the two cell regions are recorded as the overlapping area. All target cells are sorted from high to low cell recognition confidence, and the target cell with the highest confidence is selected as the baseline cell. The remaining target cells are then iterated. If the IoU between a target cell and the baseline cell is ≥0.3 (i.e., there is an overlapping area), the target cell is suppressed (to exclude duplicate recognition). If the IoU is <0.3, the cell is retained. From the remaining untreated and unsuppressed target cells, the cell with the highest confidence is selected again as the new baseline cell, and the above iterative suppression operation is repeated until all target cells are treated. The set of retained target cells is the optimized Ki-67 positive tumor cell recognition result, which is used for subsequent Ki-67 proliferation index calculation.

[0123] In one embodiment, S61, based on the Ki-67 positive tumor cell identification results, the Ki-67 proliferation index is calculated using the following formula:

[0124]

[0125] in, Ki-67 proliferation index, The number of positive tumor cells in the optimized Ki-67 positive tumor cell identification results. Weights are matched for morphological features. Here, is the overlap correction factor, and n is the total number of connected regions. Let be the actual pixel area of ​​the i-th connected region that meets the preset area threshold. For the preset area threshold, Let be the cell density within the i-th connected region that meets the preset area threshold.

[0126] Specifically, Ki-67 proliferation index, output as a percentage, is used to quantify tumor cell proliferation activity; The number of positive tumor cells in the optimized Ki-67 positive tumor cell identification results can be determined by counting the total number of target cells in the optimized target cell set (e.g., if 80 target cells are retained after optimization, then...). =80); The morphological feature matching weight is set based on the degree of matching between the target cell morphological features and the standard Ki-67 positive tumor cell morphological features. The standard morphological features are determined statistically through a large number of positive cell samples labeled by pathologists (e.g., cell roundness 0.6-0.9, aspect ratio 1.2-2.5). If all morphological features of the target cell (roundness, aspect ratio, etc.) fully meet the standard range, Assign a value of 1.0; if there is one feature that is close to the standard range (e.g., roundness 0.58, close to the lower limit of 0.6). Assign a value of 0.9; if two or more features are close to the standard range, The value is assigned to 0.8, and the average of the morphological feature matching weights of all target cells is taken as the final value. (If 60 out of 80 target cells completely match and 20 have similar characteristics in one aspect, then...) =(60×1.0+20×0.9) / 80=0.975). This is the overlap correction factor, used to correct for the slight overlap effect that may remain after non-maximum suppression. It is calculated as follows:

[0127] )

[0128] The residual overlap region is detected by comparing the contour coordinates of the optimized target cells, identifying incompletely suppressed intersection pixel regions and counting the total number of pixels (e.g., if the total number of residual overlap pixels is 50 and the total number of pixels in all target cells is 4000, then Coverlap = 1 - 50 / 4000 = 0.9875); n is the total number of connected regions that meet the preset area threshold (i.e., before morphological screening and overlap optimization, all regions with an area greater than a certain threshold are considered to have a certain number of overlapping pixels). The number of connected components, e.g., if there are 120, then n=120). The actual pixel area of ​​the i-th connected region that meets the preset area threshold is obtained by counting the pixels of the contour of each connected region using OpenCV's contour area calculation function (e.g., if the contour of the first connected region contains 60 pixels, then...). ); The preset area threshold is a threshold determined by statistically analyzing the average pixel area of ​​standard positive cells (e.g., 50 pixels). ); Let be the cell density within the i-th connected region that meets the preset area threshold (e.g., if the cell density of the first connected region is 0.8 cells / pixel, then...). According to the formula Perform calculations: Calculate the numerator. (e.g., 80 × 0.975 × 0.9875 ≈ 77.025); Calculate the denominator, for each i from 1 to n, calculate sequentially. (For example, the first connected region is 60 / 50×0.8=0.96, the second connected region is 70 / 50×0.7=0.98, and so on), and sum all the results (e.g., the sum is 96.28); divide the numerator by the denominator and multiply by 100% to get the Ki-67 proliferation index (e.g., (77.025 / 96.28)×100%≈80.0%).

[0129] The aforementioned Ki-67 proliferation index counting method based on the improved UNet network acquires Ki-67 positive tumor cell pathological slide images. Using a preset set of standard optical parameters including baseline optical parameters, multiple sets of calibrated optical parameters, and corresponding deviation levels, the image is divided into sub-regions using a grid. Actual optical parameters are extracted, and deviations from the baseline parameters are calculated. After matching the deviation levels, each sub-region is calibrated using the corresponding calibration parameters, and edge pixels are fused to obtain a standardized calibrated slide image, eliminating image inconsistencies caused by differences in slide preparation and optical conditions. Positive staining components are extracted using RGB channel separation and color deconvolution algorithms to generate a staining density map. Strong signal regions are divided, and the region mean, region variance, and cell density are calculated, replacing subjective human judgment to avoid bias. Based on these features, hotspot regions are identified through dynamic intensity and density thresholding, neighborhood cell density difference analysis, and morphological opening operations, focusing on high-value counting regions to improve targeting. An improved UNet network with four depthwise separable convolutional downsampling blocks (64, 128, 256, and 512 kernels respectively) and a CBAM attention mechanism skip connection between the decoder and encoder is used to extract encoded features from hotspot regions and process them using the attention mechanism to obtain decoded feature maps. Positive tumor cell identification results are obtained through connected component recognition and area thresholding. The identification results are then optimized through morphological feature filtering, overlap detection, and non-maximum suppression. The proliferation index is calculated based on a formula that integrates parameters such as the number of positive cells, morphological matching weights, and overlap correction coefficients. This fully automated process solves the problems of large subjective errors and low repeatability in manual counting, and the poor adaptability, easy false negatives, and low efficiency of traditional image processing with fixed thresholds. It improves the accuracy, efficiency, and standardization of Ki-67 proliferation index counting, providing an objective and reliable basis for assessing tumor proliferative activity.

[0130] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0131] Based on the same inventive concept, this application also provides a counting device for the ki-67 proliferation index based on an improved UNet network for implementing the above-described counting method for the ki-67 proliferation index based on the improved UNet network. The solution provided by this device is similar to the implementation described in the above-described method. Therefore, the specific limitations of one or more embodiments of the counting device for the ki-67 proliferation index based on an improved UNet network provided below can be found in the limitations of the counting method for the ki-67 proliferation index based on the improved UNet network described above, and will not be repeated here.

[0132] In one exemplary embodiment, such as Figure 2 As shown, a counting device for the ki-67 multiplication index based on an improved UNet network is provided, comprising:

[0133] The slide image acquisition module 101 is used to acquire pathological slide images of Ki-67 positive tumor cells;

[0134] The slide image calibration module 102 is used to calibrate the pathological slide image according to a preset standard optical parameter set to obtain a calibrated slide image;

[0135] The signal and density extraction module 103 is used to extract the regional mean and regional variance of the positive staining signal of the calibration slide image, as well as the cell density of the calibration slide image;

[0136] Hotspot region identification module 104 is used to identify hotspot regions in the calibration slice image based on the region mean, region variance and cell density;

[0137] The tumor cell identification module 105 is used to identify ki-67 positive tumor cells based on hotspot regions using a pre-trained improved UNet network, and obtain ki-67 positive tumor cell identification results.

[0138] The proliferation index calculation module 106 is used to calculate the Ki-67 proliferation index based on the Ki-67 positive tumor cell identification results.

[0139] In one embodiment, the signal and density extraction module 103 is further configured to:

[0140] The calibration slice image is subjected to RGB channel separation processing to obtain images of each channel;

[0141] A color deconvolution algorithm is used to decompose each channel image into multiple color components;

[0142] The positive staining component is obtained by extracting the staining component corresponding to the positive staining signal from multiple staining components.

[0143] Based on the positive staining components, a staining density map is generated;

[0144] Based on the staining density map, the calibration slice image is divided into regions according to a preset density threshold to obtain strong signal regions;

[0145] Based on the strong signal region, the average value of the positive staining component is calculated as the region mean;

[0146] Based on the strong signal region, the variance of the positive staining component is calculated as the region variance;

[0147] Cell density is generated based on the regional mean and regional variance.

[0148] In one embodiment, the hotspot area identification module 104 is further configured to:

[0149] Based on the region mean, region variance, and cell density, the dynamic intensity threshold and dynamic density threshold are calculated using the following formulas:

[0150]

[0151]

[0152] Where Q is the dynamic intensity threshold, To obtain the 75th percentile of the region mean, The 25th percentile of the regional variance. M represents the maximum value of the regional variance, and M is the dynamic density threshold. The 80th percentile of cell density;

[0153] Select calibration slice images where the region mean J ≥ dynamic intensity threshold Q and cell density is [not specified]. ≥Dynamic density threshold Strong signal areas are considered potential hotspots.

[0154] Calculate the neighborhood cell density difference of potential hotspot areas, and based on the neighborhood cell density difference, select qualified potential areas according to a preset cell density difference threshold;

[0155] Morphological opening operations are performed on potential qualified regions to obtain hotspot regions.

[0156] In one embodiment, the pre-trained improved UNet network of the tumor cell recognition module 105 includes an encoder and a decoder. The encoder consists of four downsampling blocks, each of which uses a depthwise separable convolutional layer. The number of convolutional kernels in the depthwise separable convolutional layer is 64, 128, 256, and 512, respectively. A skip connection is set between the decoder and the encoder, and the skip connection uses the CBAM attention mechanism.

[0157] Based on hotspot regions, a pre-trained improved UNet network was used to identify Ki-67 positive tumor cells, and the Ki-67 positive tumor cell identification results were obtained, including:

[0158] The hotspot region is fed into a pre-trained improved UNet network, and the encoder extracts the encoded features.

[0159] A decoder is used, combined with the CBAM attention mechanism in the skip connection, to process the encoded features and obtain the decoded feature map;

[0160] Scan the decoded feature map according to the preset connectivity judgment criteria to identify all connected regions;

[0161] Based on a preset area threshold, connected regions larger than the preset area threshold are selected as the results of Ki-67 positive tumor cell identification.

[0162] In one embodiment, the slice image calibration module 102 includes a set of reference optical parameters, multiple sets of calibration optical parameters, and multiple deviation levels, wherein each deviation level corresponds to a set of calibration optical parameters.

[0163] Based on a preset set of standard optical parameters, the pathological slide images are calibrated to obtain calibrated slide images, including:

[0164] The pathological slide image is divided into multiple image sub-regions according to a preset grid.

[0165] Extract the actual optical parameters of each image sub-region;

[0166] Calculate the deviation between the actual optical parameters and the reference optical parameters;

[0167] Based on the deviation value, determine the deviation level of each image sub-region;

[0168] For each image sub-region, calibration optical parameters with corresponding deviation levels are used for calibration to obtain the calibrated sub-region;

[0169] Edge pixel fusion processing is performed on each calibrated sub-region to obtain the calibration slice image.

[0170] In one embodiment, the proliferation index calculation module 106 is further configured to:

[0171] Morphological characteristics of each identified positive tumor cell were extracted from the Ki-67 positive tumor cell identification results;

[0172] Based on the preset morphological feature threshold, target cells whose morphological features meet the preset morphological feature threshold are selected;

[0173] Overlap detection is performed between target cells to obtain overlapping regions;

[0174] The non-maximum suppression algorithm was used to process the overlapping regions to obtain optimized Ki-67 positive tumor cell identification results, which were then used to calculate the Ki-67 proliferation index.

[0175] In one embodiment, the proliferation index calculation module 106 calculates the Ki-67 proliferation index based on the Ki-67 positive tumor cell identification results using the following formula:

[0176]

[0177] in, Ki-67 proliferation index, The number of positive tumor cells in the optimized Ki-67 positive tumor cell identification results. Weights are matched for morphological features. Here, is the overlap correction factor, and n is the total number of connected regions. Let be the actual pixel area of ​​the i-th connected region that meets the preset area threshold. For the preset area threshold, Let be the cell density within the i-th connected region that meets the preset area threshold.

[0178] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the ki-67 proliferation index counting method based on the improved UNet network as described above.

[0179] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0180] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0181] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A counting method of ki-67 proliferation index based on improved UNet network, characterized in that, The method comprises: acquiring a pathological section image of ki-67 positive tumor cells; calibrating the pathological section image according to a preset standard optical parameter set to obtain a calibrated section image; extracting the regional mean and regional variance of the positive staining signal of the calibrated section image, and the cell density of the calibrated section image; based on the regional mean, regional variance and cell density, identifying a hotspot region of the calibrated section image; based on the hotspot region, using a pre-trained improved UNet network to identify ki-67 positive tumor cells to obtain a ki-67 positive tumor cell identification result; based on the ki-67 positive tumor cell identification result, calculating a ki-67 proliferation index.

2. The method of claim 1, wherein, The extraction of the regional mean and regional variance of the positive staining signal of the calibrated section image, and the cell density of the calibrated section image comprises: performing RGB channel separation processing on the calibrated section image to obtain each channel image; using a color deconvolution algorithm to decompose each channel image into multiple staining components; extracting a staining component corresponding to the positive staining signal from the multiple staining components to obtain a positive staining component; generating a staining density map according to the positive staining component; based on the staining density map, dividing the calibrated section image into regions according to a preset density threshold to obtain a strong signal region; based on the strong signal region, calculating the average value of the positive staining component as the regional mean; based on the strong signal region, calculating the variance of the positive staining component as the regional variance; generating the cell density according to the regional mean and the regional variance.

3. The method of claim 2, wherein, The identification of the hotspot region of the calibrated section image based on the regional mean, regional variance and cell density comprises: based on the regional mean, regional variance and cell density, calculating a dynamic intensity threshold and a dynamic density threshold through the following formula: where Q is a dynamic intensity threshold, is the 75th percentile of the regional mean, is the 25th percentile of the regional variance, is the maximum of the regional variance, M is a dynamic density threshold, is the 80th percentile of the cell density; strong signal regions in the calibration slice image where the region mean J is greater than the dynamic intensity threshold Q and the cell density is greater than the dynamic density threshold are selected as potential hot spot regions; calculating the neighborhood cell density difference of the potential hotspot region, and based on the neighborhood cell density difference, filtering out a qualified potential region according to a preset cell density difference threshold; performing a morphological opening operation on the qualified potential region to obtain the hotspot region.

4. The method of claim 3, wherein, The pre-trained improved UNet network comprises an encoder and a decoder, wherein the encoder is composed of four downsampling blocks, each downsampling block uses a depth separable convolution layer, the number of convolution kernels of the depth separable convolution layer is 64, 128, 256 and 512 in turn, a skip connection is arranged between the decoder and the encoder, and the skip connection uses a CBAM attention mechanism; The ki-67 positive tumor cell identification based on the hotspot region using the pre-trained improved UNet network to obtain a ki-67 positive tumor cell identification result comprises: inputting the hotspot region into the pre-trained improved UNet network to extract encoding features through the encoder; using the decoder to process the encoding features in combination with the CBAM attention mechanism in the skip connection to obtain a decoding feature map; scan the decoding feature map according to the preset connectivity determination criterion, and identify all connected regions; According to the preset area threshold, the connected region larger than the preset area threshold is screened out as the ki-67 positive tumor cell recognition result.

5. The method of claim 1, wherein, The preset standard optical parameter set includes a group of reference optical parameters, a plurality of sets of calibration optical parameters, and a plurality of deviation levels, wherein each deviation level corresponds to a set of calibration optical parameters; The calibration slice image is obtained by calibrating the pathological slice image according to the preset standard optical parameter set, including: The pathological slice image is divided into a plurality of image sub-regions according to a preset grid; The actual optical parameters of each image sub-region are extracted; The deviation value of the actual optical parameter and the reference optical parameter is calculated; According to the deviation value, the deviation level of each image sub-region is determined; For each image sub-region, the calibration optical parameter corresponding to the deviation level is used for calibration to obtain a calibrated sub-region; The edge pixel fusion processing is performed on each calibrated sub-region to obtain the calibrated slice image.

6. The method of claim 4, wherein, After obtaining the ki-67 positive tumor cell recognition result, before calculating the ki-67 proliferation index, the method further includes: Extract the morphological features of each recognized positive tumor cell in the ki-67 positive tumor cell recognition result; According to the preset morphological feature threshold, the target cells whose morphological features meet the preset morphological feature threshold are screened out; The overlapping detection is performed between the target cells to obtain an overlapping region; The non-maximum suppression algorithm is used to process the overlapping region to obtain an optimized ki-67 positive tumor cell recognition result, which is used for subsequent calculation of the ki-67 proliferation index.

7. The method of claim 6, wherein, The ki-67 proliferation index is calculated based on the ki-67 positive tumor cell recognition result, which is realized by the following formula: wherein, is a ki-67 proliferation index, is a number of positive tumor cells in the optimized ki-67 positive tumor cell recognition result, is a morphological feature matching weight, is an overlap correction coefficient, n is a total number of connected regions, is an actual pixel area of the ith connected region meeting the preset area threshold, is a preset area threshold, is a cell density in the ith connected region meeting the preset area threshold. 8.A counting device of ki-67 proliferation index based on an improved UNet network, characterized in that, The device includes: A slice image acquisition module for acquiring a pathological slice image of ki-67 positive tumor cells; A slice image calibration module for calibrating the pathological slice image according to a preset standard optical parameter set to obtain a calibrated slice image; A signal and density extraction module for extracting the regional mean and regional variance of the positive staining signal of the calibrated slice image, and the cell density of the calibrated slice image; A hotspot region identification module for identifying the hotspot region of the calibrated slice image based on the regional mean, regional variance and cell density; A tumor cell identification module for identifying ki-67 positive tumor cells based on the hotspot region using a pre-trained improved UNet network to obtain a ki-67 positive tumor cell recognition result; A proliferation index calculation module for calculating the ki-67 proliferation index based on the ki-67 positive tumor cell recognition result. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8. The processor executes the computer program to realize the steps of the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method of any one of claims 1 to 7.

Images