A myeloma microscopic cell image cancer cell counting method

Abstract

The application discloses a myeloma microscopic cell image cancer cell counting method and belongs to the technical field of medical image processing. In order to solve the problem of realizing automatic and accurate counting of myeloma cells, the application collects myeloma multi-channel microscopic image data for adaptive preprocessing, fuses noise suppression and convolution enhancement, obtains pixel gray value of the preprocessed myeloma multi-channel microscopic image, then carries out rough segmentation of cancer cells based on edge constraint to obtain a rough segmentation mask of the image, then carries out deep feature extraction to obtain the deep feature of the pixel of the preprocessed myeloma multi-channel microscopic image, then calculates the average deep feature of the cells, then combines Bayesian prior correction to calculate the probability of the cancer cells through MLP deep modeling, and finally, the application verifies the cancer cell counting result through correlation coefficient calibration combined with clinical prior constraint to obtain the final myeloma cancer cell counting result. The application provides technical support for standardization and normalization of myeloma diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a method for counting cancer cells in microscopic images of myeloma cells. Background Technology

[0002] Myeloma, a malignant plasma cell disease, relies heavily on the accurate counting of cancer cells in pathological sections for diagnosis and prognosis assessment. This indicator directly determines treatment selection, efficacy monitoring, and patient survival risk assessment. Currently, in clinical practice, pathologists manually count cancer cells using high-powered microscopes. This is not only time-consuming and laborious (averaging over 20 minutes per section), but also susceptible to subjective experience and visual fatigue, leading to counting errors as high as 15%-20%. The misjudgment rate is further increased, especially in scenarios where myeloma cells and normal plasma cells have similar morphologies and dense overlap, severely impacting diagnostic accuracy. With the increasing incidence of myeloma year by year, the clinical need for efficient and accurate automated counting methods is becoming increasingly urgent.

[0003] Existing automated cancer cell counting methods suffer from significant technical bottlenecks, making it difficult to meet the stringent requirements of clinical diagnosis. Traditional computer vision methods (such as threshold segmentation and morphological filtering) rely on manually designed features, which cannot adapt to the uneven staining and blurred edges of myeloma cells, and have weak decoupling ability for overlapping cells, often resulting in overcounting or undercounting. Early machine learning methods (such as SVM and random forest) require manual extraction of texture and shape features, have poor generalization ability, and struggle to distinguish morphologically similar cancer cells from normal plasma cells. While some deep learning methods have improved segmentation accuracy, their models are complex and computationally intensive, and they are not optimized for the clinical morphological characteristics of myeloma cells, resulting in insufficient adaptability to images acquired in special environments such as high-altitude areas, limiting their practical application. These limitations make it difficult for existing methods to be applied on a large scale to routine clinical testing.

[0004] In recent years, the rapid development of AI technology and the upgrading of microscopic imaging equipment have provided new opportunities to solve the above problems. The combination of high-resolution CMOS cameras and multi-channel staining technology can stably output multi-dimensional microscopic images containing cell morphology, texture, and spectral information, providing a data foundation for accurate segmentation and classification. The emergence of lightweight deep learning models (such as lightweight U-Net and VisionTransformer) has achieved a balance between "high precision and high efficiency," adapting to the real-time processing needs of clinical equipment. Against this backdrop, there is an urgent need to develop a cancer cell counting method that integrates AI technology and clinical pathology priors to specifically address the core pain points of "difficulty in distinguishing morphologically similar cells, difficulty in decoupling overlapping cells, and low efficiency of manual counting," achieving automated and accurate counting of myeloma cells and providing reliable technical support for clinical diagnosis. Summary of the Invention

[0005] The problem this invention aims to solve is to achieve automated and precise counting of myeloma cells, and proposes a method for counting cancer cells in myeloma microscopic cell images.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for counting cancer cells in myeloma microscopic cell images, comprising the following steps:

[0008] S1. Collect basic data, including multi-channel microscopic image data of myeloma, multi-channel microscopic image data of myeloma annotated by 0.1% of doctors, and prior data on cell morphology;

[0009] S2. Adaptive preprocessing is performed on the multi-channel microscopic image data of myeloma obtained in step S1, and noise suppression and convolution enhancement are fused to obtain the pixel grayscale values ​​of the preprocessed multi-channel microscopic image of myeloma.

[0010] S3. Perform edge-constraint-based coarse segmentation of cancer cells on the pixel gray values ​​of the preprocessed multi-channel microscopic image of myeloma obtained in step S2 to obtain the coarse segmentation mask of the image.

[0011] S4. Extract depth features from the coarse segmentation mask of the image obtained in step S3 to obtain the depth features of pixels in the preprocessed multi-channel microscopic image of myeloma.

[0012] S5. Calculate the cell mean depth features of the pixels in the preprocessed multi-channel microscopic image of myeloma obtained in step S4, and then calculate the probability of cancer cells by combining the obtained cell mean depth features with MLP depth modeling and Bayesian prior correction.

[0013] S6. The probability of cancer cells obtained in step S5 is verified by correlation coefficient calibration combined with clinical prior constraints to obtain the final myeloma cancer cell count result.

[0014] Furthermore, the specific implementation method of step S1 includes the following steps:

[0015] S1.1. Acquire multi-channel microscopic image data of myeloma, generated from pathological sections by capturing images with a high-powered microscope coupled to a CMOS camera. The resolution of a single image is 1024×1024 pixels. Obtain the pixel grayscale values ​​of the myeloma microscopic image in the m-th row, n-th column, and c-th channel. c=1 corresponds to the H channel, c=2 corresponds to the E channel, and c=3 corresponds to the DAPI channel;

[0016] S1.2. Collect 0.1% of the myeloma multichannel microscopic image data labeled by doctors. Three pathologists labeled 0.1% of the total pixels using annotation software.

[0017] S1.3. Collect prior data on cell morphology The shape parameters of myeloma cancer cells were obtained through statistical analysis of 1000 labeled samples, specifically: mean area 120 μm², roundness 0.75, and texture entropy 3.2.

[0018] Furthermore, the calculation formula for adaptive preprocessing of the myeloma multichannel microscopic image data from step S1 in step S2 is as follows:

[0019]

[0020] in, The pixel grayscale values ​​of the preprocessed myeloma microscopic image in row m, column n, and channel c are: is the staining correction coefficient for channel c; This is the formula for calculating the error unit, used for adaptive suppression of outlier noise; The median grayscale value of all pixels in the c-th channel; Indicates the interquartile range; This is a 2D convolution operation. The kernel is a convolution kernel, specifically a 3×3 adaptive kernel.

[0021] Furthermore, the expression for image segmentation of the pixel grayscale values ​​of the preprocessed multichannel microscopic image of myeloma obtained in step S2 in step S3 is as follows:

[0022]

[0023] in, This is the coarse segmentation mask for the image pixels in the m-th row and n-th column; It is a mapping function. It is a lightweight U-Net network, based on input. Output pixel-level prediction scores; The edge constraint coefficient; For edge detection operators; These are the corresponding pixel values ​​in the attention weight matrix.

[0024] Furthermore, the expression for extracting deep features from the coarse segmentation mask of the obtained image in step S4 is as follows:

[0025]

[0026] in, It is the k-th dimension depth feature of the image pixel in the m-th row and n-th column; It is the basic block of VisionTransformer; For layer normalization; This is the morphological constraint coefficient; A Gaussian mixture model, used based on prior data of cell morphology. Calculate the probability feature that a pixel belongs to a cancer cell.

[0027] Furthermore, the specific implementation method of step S5 includes the following steps:

[0028]

[0029] in, Let be the average depth feature of the t-th cell instance. Let be the total number of pixels contained in the t-th cell instance. The classification result represents the t-th cell instance;

[0030] S5.2. The average cell depth features obtained in step S5.1 are used for MLP depth modeling combined with Bayesian prior correction to calculate the probability of cancer cells. The expression is as follows:

[0031]

[0032] in, Let be the class probability of the t-th cell instance; The function is a probability transformation tool used in the cell classification stage, which maps 3D scores to probability distributions; It is a 3-layer fully connected network, whose input is the average depth feature of the t-th cell instance, and whose output is a 3-dimensional score; These are the Bayesian correction coefficients; Calculate the Bayesian posterior probability;

[0033] A value of 0 indicates the background. A value of 1 indicates a normal cell. A value of 2 indicates cancer cells.

[0034] Furthermore, the specific implementation method of step S6 includes the following steps:

[0035] S6.1. First, perform initial statistics, counting all... The sum of class probabilities of 2 ;

[0036] S6.2. Correct the initial statistics, combine them with clinicopathological priors, and output the final myeloma cancer cell count result, expressed as:

[0037]

[0038] in, For the final myeloma cancer cell count; To round to the nearest integer; For calibration coefficients; Mark the classification results for doctors within the specified range. Results with doctor's annotation The calculated correlation coefficient is used to assess reliability.

[0039] The beneficial effects of this invention are:

[0040] This invention discloses a method for counting cancer cells in myeloma microscopic cell images, addressing a core challenge in clinical cell counting and improving diagnostic reliability. This method employs a lightweight network combined with edge constraints and attention-weighted coarse segmentation design to enhance the identification of cell edges and high-response regions. Further, it utilizes depth feature-guided and distance-attenuation-constrained decoupling calculations to effectively separate densely clustered cells. Simultaneously, it incorporates a Gaussian mixture model and depth feature extraction based on prior cancer cell morphology to accurately distinguish morphologically similar myeloma cells from normal plasma cells, significantly reducing counting bias compared to traditional methods. The results directly match the accuracy requirements of clinical diagnosis.

[0041] This invention discloses a method for counting cancer cells in myeloma microscopic cell images, achieving a balance between high accuracy and high efficiency, and is suitable for routine clinical testing scenarios. This method employs a combined architecture of a lightweight network (optimized with depthwise separable convolution) and a lightweight vision module, significantly simplifying the computational process. Simultaneously, layer normalization stabilizes model training, improving the processing efficiency of a single image by hundreds of times compared to manual counting. Furthermore, the model requires no additional hardware support and can be directly embedded into existing pathology analysis systems, lowering the technical and cost barriers to clinical implementation.

[0042] This invention discloses a method for counting cancer cells in microscopic images of myeloma, enhancing the method's generalization and adaptability to cover complex clinical data scenarios. The method employs multi-channel adaptive denoising and staining normalization calculations to effectively handle uneven staining and noise interference from slides from different hospitals and batches. Step five combines Bayesian posterior probability calculation with clinical morphological priors, enabling the method to maintain stable recognition capabilities for images with different cell densities and imaging conditions, avoiding the scenario limitations of traditional methods.

[0043] This invention discloses a method for counting cancer cells in myeloma microscopic cell images, enhancing its practical value in clinical applications and assisting in the standardization of pathological diagnosis. This method achieves accurate classification and counting of cell instances through a pixel-to-cell instance mapping design; the final counting result is an intuitive integer, requiring no additional interpretation by pathologists and directly compatible with clinical reporting processes. This reduces diagnostic discrepancies caused by physicians' subjective experience and significantly lowers their workload, providing technical support for the standardization and normalization of myeloma diagnosis. Attached Figure Description

[0044] Figure 1This is a flowchart of a method for counting cancer cells in a microscopic image of myeloma as described in this invention;

[0045] Figure 2 Example image of multichannel microscopic images of myeloma collected;

[0046] Figure 3 for Figure 2 Enlarged image. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.

[0048] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.

[0049] To further understand the invention's content, features, and effects, the following specific embodiments are provided, along with accompanying drawings. Figure 1 -Appendix Figure 3 Detailed explanation is as follows:

[0050] Example 1:

[0051] A method for counting cancer cells in myeloma microscopic cell images, comprising the following steps:

[0052] S1. Collect basic data, including multi-channel microscopic image data of myeloma, multi-channel microscopic image data of myeloma annotated by 0.1% of doctors, and prior data on cell morphology;

[0053] Furthermore, the specific implementation method of step S1 includes the following steps:

[0054] S1.1. Acquire multi-channel microscopic image data of myeloma, generated from pathological sections by capturing images with a high-powered microscope coupled to a CMOS camera. The resolution of a single image is 1024×1024 pixels. Obtain the pixel grayscale values ​​of the myeloma microscopic image in the m-th row, n-th column, and c-th channel. c=1 corresponds to the H channel, c=2 corresponds to the E channel, and c=3 corresponds to the DAPI channel;

[0055] S1.2. Collect 0.1% of the myeloma multichannel microscopic image data labeled by doctors. Three pathologists labeled 0.1% of the total pixels using annotation software.

[0056] S1.3. Collect prior data on cell morphology The shape parameters of myeloma cancer cells were obtained through statistical analysis of 1000 labeled samples, specifically: mean area 120 μm², roundness 0.75, and texture entropy 3.2.

[0057] S2. Adaptive preprocessing is performed on the multi-channel microscopic image data of myeloma obtained in step S1, and noise suppression and convolution enhancement are fused to obtain the pixel grayscale values ​​of the preprocessed multi-channel microscopic image of myeloma.

[0058] Furthermore, the calculation formula for adaptive preprocessing of the myeloma multichannel microscopic image data from step S1 in step S2 is as follows:

[0059]

[0060] in, The pixel grayscale values ​​of the preprocessed myeloma microscopic image in row m, column n, and channel c are: is the staining correction coefficient for channel c; This is the formula for calculating the error unit, used for adaptive suppression of outlier noise; The median grayscale value of all pixels in the c-th channel; Indicates the interquartile range; This is a 2D convolution operation. The kernel is a convolution kernel, specifically a 3×3 adaptive kernel.

[0061] The calculation of GELU is expressed as follows:

[0062] ;

[0063] Furthermore, traditional methods use only a single Gaussian filter for noise reduction, ignoring issues such as uneven staining and blurred cell edges. This step eliminates microscope noise and interference from uneven staining, generating a standardized image that provides clean input for subsequent segmentation. The corresponding reference values ​​are: H=1.05, E=0.98, DAPI=1.12;

[0064] Error cell calculation, adaptive suppression of outlier noise; x is used to explain the calculation process of Gaussian error cell and has no meaning. This is a dispersion index designed to resist extreme values. This step combines noise suppression and convolution enhancement, improving the image signal-to-noise ratio and cell edge preservation rate, effectively addressing subsequent segmentation errors caused by inconsistent staining depths in myeloma cells.

[0065] S3. Perform edge-constraint-based coarse segmentation of cancer cells on the pixel gray values ​​of the preprocessed multi-channel microscopic image of myeloma obtained in step S2 to obtain the coarse segmentation mask of the image.

[0066] Furthermore, the expression for image segmentation of the pixel grayscale values ​​of the preprocessed multichannel microscopic image of myeloma obtained in step S2 in step S3 is as follows:

[0067]

[0068] in, This is the coarse segmentation mask for the image pixels in the m-th row and n-th column; It is a mapping function. It is a lightweight U-Net network, based on input. Output pixel-level prediction scores; The edge constraint coefficient; For edge detection operators; These are the corresponding pixel values ​​in the attention weight matrix.

[0069] Furthermore, traditional threshold segmentation (such as Otsu) cannot handle cell overlap and edge adhesion, with a segmentation accuracy of only about 60%. This step captures deep features through U-Net, combining edge constraints and attention weighting to improve coarse segmentation accuracy and reduce false negative rate, especially suitable for scenarios with dense distribution of myeloma cells. Initial separation of cancer cells, normal cells, and background is performed, generating pixel-level candidate region masks. The value ranges from 0 to 1, with higher values ​​indicating a greater likelihood of cancer cells. The reference value is 0.35; Extract the cell edges of the E channel (c=2) (values ​​0-1).

[0070] Mapping function The calculation method is as follows: The predicted score is mapped to the 0-1 interval. Here, x is used to explain the calculation process of the mapping and has no meaning.

[0071] Lightweight U-Net Network First, the input image is preprocessed by standardization, size unification, and channel expansion. Then, features are extracted and saved by a 4-level encoder with depthwise separable convolution, batch normalization, ReLU activation, and average pooling. After fusing deep semantic features through two depthwise separable convolutions at the bottleneck layer, the decoder upsamples and enlarges the feature map through 4-level lightweight transposed convolutions. The feature map is then cropped, aligned, and stitched with the corresponding level feature map of the encoder. After further fusion by depthwise separable convolution, a pixel-level segmentation mask with the same size as the input is generated by a 1×1 convolution and Sigmoid activation. The entire process is designed with lightweight features to reduce computation and parameters while retaining core segmentation capabilities.

[0072] S4. Extract depth features from the coarse segmentation mask of the image obtained in step S3 to obtain the depth features of pixels in the preprocessed multi-channel microscopic image of myeloma.

[0073] Furthermore, the expression for extracting deep features from the coarse segmentation mask of the obtained image in step S4 is as follows:

[0074]

[0075] in, It is the k-th dimension depth feature of the image pixel in the m-th row and n-th column; It is the basic block of VisionTransformer; For layer normalization; This is the morphological constraint coefficient; A Gaussian mixture model, used based on prior data of cell morphology. Calculate the probability feature that a pixel belongs to a cancer cell.

[0076] Furthermore, traditional handcrafted features (such as HOG and SIFT) can only capture superficial morphology and are easily affected by cell deformation. By extracting global semantic features and fusing morphological constraints, the discriminative power of features is improved, effectively distinguishing between morphologically similar normal plasma cells and myeloma cancer cells. Extracting higher-order features enhances the differences between cancer cells and normal cells, providing a basis for classification. k=64; Output 64-dimensional features; Used to stabilize feature distribution; The reference value is 0.28. Further, the VisionTransformer base block... First, the input normalized image is concatenated with a coarse segmentation mask to form a fused feature map. Then, the fused map is divided into fixed-size image blocks, and a learnable positional encoding is added to each image block to preserve spatial information. The image blocks with positional encoding are then converted into sequence vectors, which are then input into a multi-head self-attention layer. Each vector learns the association weights with other vectors and updates its features. The output of the attention layer is then fed into a feedforward neural network for nonlinear transformation. At the same time, layer normalization and residual connections are added before and after the attention layer and the feedforward layer, respectively, to stabilize the training and preserve the original features. Finally, a 64-dimensional deep feature map is output.

[0077] The layer normalization calculation process is as follows: For the input features (such as the 64-dimensional features output by ViT_Block), calculate the mean and variance of all feature elements for each sample separately. To avoid division by zero, add a very small constant (such as 0.0001) to the variance. Then, subtract the mean from each feature element and divide it by the square root of (variance plus the very small constant) to complete the standardization. Finally, multiply the standardized result by a learnable scaling parameter and add a learnable offset parameter to obtain the normalized features. This stabilizes the mean and variance of the features and improves the stability of model training.

[0078] Gaussian mixture model Using the pixel and cancer cell morphology priors of the normalized image as input, first based on Several Gaussian components that adapt to the morphological distribution of cancer cells are preset and the prior probability, mean and variance of each component are initialized. Then, by iterating these parameters, the posterior probability of each pixel belonging to each Gaussian component is calculated. Finally, the posterior probability of each component is weighted and fused with the corresponding morphological matching features to output the probability feature corresponding to the pixel (for subsequent combination with the output features of ViT_Block).

[0079] S5. Calculate the cell mean depth features of the pixels in the preprocessed multi-channel microscopic image of myeloma obtained in step S4, and then calculate the probability of cancer cells by combining the obtained cell mean depth features with MLP depth modeling and Bayesian prior correction.

[0080] Furthermore, the specific implementation method of step S5 includes the following steps:

[0081] S5.1. Calculate the average cell depth feature from the pixel depth features of the preprocessed multichannel microscopic image of myeloma obtained in step S4. The expression is:

[0082]

[0083] in, Let be the average depth feature of the t-th cell instance. Let be the total number of pixels contained in the t-th cell instance. The classification result represents the t-th cell instance;

[0084] S5.2. The average cell depth features obtained in step S5.1 are used for MLP depth modeling combined with Bayesian prior correction to calculate the probability of cancer cells. The expression is as follows:

[0085]

[0086] in, Let be the class probability of the t-th cell instance; The function is a probability transformation tool used in the cell classification stage, which maps 3D scores to probability distributions; It is a 3-layer fully connected network, whose input is the average depth feature of the t-th cell instance, and whose output is a 3-dimensional score; These are the Bayesian correction coefficients; Calculate the Bayesian posterior probability; The reference value is 0.35;

[0087] A value of 0 indicates the background. A value of 1 indicates a normal cell. A value of 2 indicates cancer cells.

[0088] Furthermore, traditional classifiers such as SVM and Random Forest have weak capabilities in modeling high-dimensional features, achieving an accuracy of approximately 80% in classifying cancer cells. This step improves classification accuracy by using deep modeling with MLP and Bayesian prior correction, reducing the probability of misclassifying normal plasma cells as cancer cells. Initial counting logic: statistics The initial count is obtained by counting the number of instances with the highest probability category, "cancer cells" (category 2), and a probability > 0.85. The reference value is 0.31; the reference value for the morphological prior parameter is 0.35.

[0089] Bayesian posterior probability calculation The calculation process is as follows: taking the prior morphology of cancer cells and the features of the t-th cell instance as input, firstly, the prior probability of "this instance belongs to cancer cells" is determined from the morphological prior. Then, combined with the distribution rule corresponding to the morphological prior, the probability that the features of this instance conform to the distribution of cancer cell features is calculated. Next, through the total probability method, the probability of "the instance is a cancer cell" and "the instance is not a cancer cell" are fused with the prior probability to obtain the overall probability of the occurrence of the features of this instance. Finally, the probability of "the probability that the features conform to the distribution of cancer cells" is multiplied by the prior probability of "the instance is a cancer cell" and then divided by the calculated overall probability of occurrence to finally obtain the probability that the instance belongs to cancer cells.

[0090] S6. The probability of cancer cells obtained in step S5 is verified by correlation coefficient calibration combined with clinical prior constraints to obtain the final myeloma cancer cell count result.

[0091] Furthermore, the specific implementation method of step S6 includes the following steps:

[0092] S6.1. First, perform initial statistics, counting all... The sum of class probabilities of 2 ;

[0093] S6.2. Correct the initial statistics, combine them with clinicopathological priors, and output the final myeloma cancer cell count result, expressed as:

[0094]

[0095] in, For the final myeloma cancer cell count; To round to the nearest integer; For calibration coefficients; Mark the classification results for doctors within the specified range. Results with doctor's annotation The calculated correlation coefficient is used to assess reliability;

[0096] Furthermore, this step can be omitted when the quantity is small.

[0097] This embodiment will be illustrated by an example:

[0098] The acquired multi-channel microscopic images of myeloma were set to 5×5 pixel images (rows m=1-5, columns n=1-5). Figure 2 and Figure 3 As shown, Figure 3 The numbers in the array represent the row and column numbers, for example, (3,3) represents the third row and third column;

[0099] The pixel values ​​of the image were read and used to replace the original image, as shown in Table 1:

[0100] Table 1

[0101]

[0102] Image instance annotations (corresponding grid diagrams) are as follows:

[0103] Example 1 (suspected myeloma cancer cells): Row 1, Column 1; Row 1, Column 2; Row 2, Column 1; Row 2, Column 2 (4 pixels, H channel grayscale 120-135, higher than other areas);

[0104] Example 2 (suspected myeloma cancer cells): row 3 column 3, row 3 column 4, row 4 column 3, row 4 column 4 (4 pixels, H channel grayscale 110-125);

[0105] Example 3 (normal plasma cells): Row 1, Column 3, Row 1, Column 4, Row 2, Column 3, Row 2, Column 4 (4 pixels, H channel grayscale 90-98, lower than examples 1 and 2);

[0106] Background: Row 1, Column 5; Row 2, Column 5; Row 3, Column 1; Row 3, Column 2; Row 4, Column 1; Row 4, Column 2; Row 5, Columns 1-5 (13 pixels remaining, H channel grayscale 40-78, lowest);

[0107] The cancer cell count was calculated from a microscopic image of myeloma as follows:

[0108] Step 1: Multi-channel image adaptive preprocessing (taking the "row 1, column 1" pixel of instance 1 as an example), noise removal, color unevenness correction, and output of standardized pixel value calculation process:

[0109] Take the original grayscale value of the H channel of this pixel, 120 (from row 1, column 1 of the grid), multiply it by the H channel color correction factor of 1.05 (from the source table), and get: 120 × 1.05 = 126;

[0110] Calculate the median and interquartile range (IQR) of the H-channel pixels in the entire image:

[0111] H channel values ​​for the entire image: 120, 125, 90, 95, 50, 130, 135, 92, 98, 52, 70, 75, 110, 115, 55, 72, 78, 120, 125, 58, 45, 48, 50, 52, 40 (25 in total, from all H channels of the grid image);

[0112] After sorting: 40,45,48,50,50,52,52,55,58,70,72,75,78,90,92,95,98,110,115,120,120,125,125,130,135;

[0113] Median (13th value): 78; IQR (19th value - 7th value): 115 - 52 = 63;

[0114] Suppressing noise using error units: 1 - (126 - 78) / (63 × 2.5) = 1 - 48 / 157.5 ≈ 1 - 0.305 = 0.695;

[0115] 3×3 adaptive convolution enhancement (convolution kernel weights preset to [0.1,0.2,0.1;0.2,0.1,0.2;0.1,0.2,0.1]): Take the H channel values ​​of the 3×3 region surrounding the pixel (rows 1-3, columns 1-3: 120,125,90,130,135,92,70,75,110), and sum them by weight to get: 120×0.1+125×0.2+90×0.1+130×0.2+135×0.1+92×0.2+70×0.1+75×0.2+110×0.1=105.9;

[0116] The normalized value of the H channel of this pixel is: 126×0.695+105.9×0.305≈87.57+32.29≈119.86 (120 is retained).

[0117] Repeat the above operation to process the E channel (80×0.98≈78.4) and DAPI channel (150×1.12≈168) of this pixel.

[0118] Step 2: Coarse segmentation of cancer cells based on edge constraints, outlining pixels suspected of being cancer cells, and outputting 0-1 mask values. Calculation process:

[0119] Input the normalized image from step 1 into the lightweight network:

[0120] Network downsampling: 5×5 image is scaled to 256×256 (adapting network), and features are extracted through 4 levels of depthwise separable convolution;

[0121] Network sampling: Restored to 5×5 size, output the prediction score for each pixel (pixel score of instance 1: 1.8-2.2, instance 2: 1.5-1.8, instance 3: 0.5-0.8, background: 0.1-0.3).

[0122] Canny edge detection: Extract the edges of the E channel (the edge value between row 1, column 1 and row 1, column 2 of instance 1 is 0.9, and the edge value between row 1, column 2 and row 2, column 2 is 0.85).

[0123] Edge constraint and attention weighting: Score of each pixel + edge value × 0.35 (edge ​​constraint coefficient) × attention weight (weight of instance 1 is 0.9, and weight of background is 0.1);

[0124] Example: Instance 1, row 1, column 1, pixel: 2.0 + 0.9 × 0.35 × 0.9 ≈ 2.0 + 0.2835 = 2.2835;

[0125] Background (5 rows, 1 column, 1 pixel): 0.2 + 0 × 0.35 × 0.1 = 0.2;

[0126] Mapping function transformation (0-1 range): Pixel value = 1 / (1+e^(-fraction));

[0127] Instance 1, row 1, column 1: 1 / (1+e^(-2.2835))≈0.905 (high likelihood, mask value 0.9);

[0128] Instance 3, row 1, column 3: 1 / (1+e^(-0.8))≈0.69 (low likelihood, mask value 0.7);

[0129] Background row 5 column 1: 1 / (1+e^(-0.2))≈0.55 (not suspected, mask value 0.5);

[0130] Final candidate masks: Instance 1 and 2 have pixel mask values ​​> 0.8 (marked as "suspected cancer cells"), instance 3 has a value < 0.8, and the background has a value < 0.6.

[0131] Step 3: Deep Feature Extraction and Cancer Cell Significance Enhancement. 64-dimensional features are extracted for each instance to enhance the calculation process of differences between cancer cells and normal cells.

[0132] Image block segmentation and position encoding: Divide the 5×5 image into 2×2 image blocks (the block containing instance 1 is "row 1-2 column 1-2"), and add position encoding to the block (marked as "located in the upper left region of the image, belonging to a dense pixel group").

[0133] The association weights of instance block 1 and its surrounding blocks (instance block 3 in rows 1-2 and columns 3-4) are calculated. The pixel weights within instance block 1 are 0.8-0.9 (strong association) and the pixel weights with instance block 3 are 0.2-0.3 (weak association).

[0134] Feedforward neural network: Converts the attention output into 64-dimensional features (such as "H channel grayscale mean 127, roundness 0.73, texture entropy 3.1", which closely approximates the prior morphology of cancer cells in the document);

[0135] Calculate the mean (assume it to be 0.5) and variance (assume it to be 0.1) of the 64-dimensional features of instance 1, and add a minimal constant 1e-6 to avoid division by zero;

[0136] Standardization: (eigenvalue - 0.5) / (0.1 + 1e-6) 1 / 2 ≈(eigenvalue - 0.5) / 0.316;

[0137] Add a learnable parameter: standardized value × 0.9 (scaling parameter) + 0.1 (offset parameter) to obtain stable features;

[0138] Based on the prior morphology of cancer cells (area 120 μm², roundness 0.75), the probability that the features of instance 1 match the distribution of cancer cells is calculated to be 0.85.

[0139] Step 4: Cancer cell probability classification and initial counting, determining whether each instance is a cancer cell, statistical calculation of the initial number, and instance feature aggregation:

[0140] The 64-dimensional features of the 4 pixels of instance 1 are averaged to obtain the "average features of instance 1" (e.g., "H channel mean 127, roundness 0.73").

[0141] Similarly, the average features of instance 2 are: "H channel mean 118, roundness 0.72"; and the average features of instance 3 are: "H channel mean 94, roundness 0.65".

[0142] Fully connected network prediction: Input the average features of instance 1 into a 3-layer MLP and output 3-dimensional scores (background 0.08, normal cells 0.12, cancer cells 0.80).

[0143] Instance 2 output scores: background 0.10, normal cells 0.15, cancer cells 0.75; Instance 3 output scores: background 0.12, normal cells 0.78, cancer cells 0.10.

[0144] Bayesian posterior probability correction: Example 1: Prior probability 0.35 (source table) × likelihood probability 0.85 (feature matches cancer cell distribution) ÷ marginal probability 0.4 (total probability calculation: 0.35×0.85+0.65×0.15≈0.4)≈0.74;

[0145] Corrected score: MLP output cancer cell score 0.80 + Bayesian probability 0.74 × 0.31 (Bayesian correction coefficient) ≈ 0.80 + 0.229 ≈ 1.029;

[0146] Mapped to probability (Softmax): Probability of cancer cell in instance 1 ≈ 0.88 (> 0.85 threshold, source table);

[0147] Instance classification results:

[0148] Instance 1: Cancer cell (probability 0.88); Instance 2: Cancer cell (probability 0.86); Instance 3: Normal cell (probability 0.10).

[0149] Initial count: 2 (number 1 + number 2).

[0150] Step 5: Count Calibration and Accurate Output:

[0151] Objective: To obtain the final count calculation process by incorporating doctor annotations and error correction.

[0152] Correlation coefficient calculation: Compare the classification results from step 4 (1 = cancer, 2 = cancer, 3 = normal) with the doctor's annotation (gold standard: 1 = cancer, 2 = cancer, 3 = normal). The Pearson correlation coefficient is 0.95 (>0.9, triggering calibration).

[0153] Calibration calculation:

[0154] Calibration factor: 1 + 0.05 (calibration coefficient) × (0.95 - 0.9) = 1 + 0.05 × 0.05 = 1.0025;

[0155] Count after calibration: 2 × 1.0025 ≈ 2.005;

[0156] Rounding to the nearest integer: Final myeloma cancer cell count = 2.

[0157] By combining the calculations of the 5×5 pixel grid, two myeloma cancer cells (examples 1 and 2), one normal plasma cell (example 3), and no background pixels were accurately identified.

[0158] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0159] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for counting cancer cells in a microscopic image of myeloma, characterized in that, Includes the following steps: S1. Collect basic data, including multi-channel microscopic image data of myeloma, multi-channel microscopic image data of myeloma annotated by 0.1% of doctors, and prior data on cell morphology; S2. Adaptive preprocessing is performed on the multi-channel microscopic image data of myeloma obtained in step S1, and noise suppression and convolution enhancement are fused to obtain the pixel grayscale values ​​of the preprocessed multi-channel microscopic image of myeloma. S3. Perform edge-constraint-based coarse segmentation of cancer cells on the pixel gray values ​​of the preprocessed multi-channel microscopic image of myeloma obtained in step S2 to obtain the coarse segmentation mask of the image. S4. Extract depth features from the coarse segmentation mask of the image obtained in step S3 to obtain the depth features of pixels in the preprocessed multi-channel microscopic image of myeloma. The expression for extracting deep features from the coarse segmentation mask of the obtained image in step S4 is as follows: ； in, It is the k-th dimension depth feature of the image pixel in the m-th row and n-th column; It is the basic block of VisionTransformer; For layer normalization; This is the morphological constraint coefficient; A Gaussian mixture model, used based on prior data of cell morphology. Calculate the probability feature that a pixel belongs to a cancer cell; These are the pixel grayscale values ​​of the preprocessed myeloma microscopic image in the m-th row, n-th column, and c-th channel. It is the coarse segmentation mask of the image pixels in the m-th row and n-th column; S5. Calculate the cell mean depth features of the pixels in the preprocessed multi-channel microscopic image of myeloma obtained in step S4, and then calculate the probability of cancer cells by combining the obtained cell mean depth features with MLP depth modeling and Bayesian prior correction. The specific implementation method of step S5 includes the following steps: S5.

1. Calculate the average cell depth feature from the pixel depth features of the preprocessed multichannel microscopic image of myeloma obtained in step S4, and obtain the average depth feature of the t-th cell instance. ; S5.

2. The average cell depth features obtained in step S5.1 are used for MLP depth modeling combined with Bayesian prior correction to calculate the probability of cancer cells. The expression is as follows: ； in, Let be the class probability of the t-th cell instance; The function is a probability transformation tool used in the cell classification stage, which maps 3D scores to probability distributions; It is a 3-layer fully connected network, whose input is the average depth feature of the t-th cell instance, and whose output is a 3-dimensional score; These are the Bayesian correction coefficients; Calculate the Bayesian posterior probability; A value of 0 indicates the background. A value of 1 indicates a normal cell. A value of 2 indicates cancer cells; S6. The probability of cancer cells obtained in step S5 is verified by correlation coefficient calibration combined with clinical prior constraints to obtain the final myeloma cancer cell count result.

2. The method for counting cancer cells in a microscopic image of myeloma according to claim 1, characterized in that, The specific implementation method of step S1 includes the following steps: S1.

1. Acquire multi-channel microscopic image data of myeloma, generated from pathological sections by capturing images with a high-powered microscope coupled to a CMOS camera. The resolution of a single image is 1024×1024 pixels. Obtain the pixel grayscale values ​​of the myeloma microscopic image in the m-th row, n-th column, and c-th channel. c=1 corresponds to the H channel, c=2 corresponds to the E channel, and c=3 corresponds to the DAPI channel; S1.

2. Collect 0.1% of the myeloma multichannel microscopic image data labeled by doctors. Three pathologists labeled 0.1% of the total pixels using annotation software. S1.

3. Collect prior data on cell morphology The shape parameters of myeloma cancer cells were obtained through statistical analysis of 1000 labeled samples, specifically: mean area 120 μm², roundness 0.75, and texture entropy 3.

2.

3. The method for counting cancer cells in a microscopic image of myeloma according to claim 2, characterized in that, The calculation formula for adaptive preprocessing of the myeloma multichannel microscopic image data from step S1 in step S2 is as follows: ； in, The pixel grayscale values ​​of the preprocessed myeloma microscopic image in row m, column n, and channel c are: is the staining correction coefficient for channel c; This is the formula for calculating the error unit, used for adaptive suppression of outlier noise; The median grayscale value of all pixels in the c-th channel; Indicates the interquartile range; This is a 2D convolution operation. The kernel is a convolution kernel, specifically a 3×3 adaptive kernel.

4. The method for counting cancer cells in a microscopic image of myeloma according to claim 3, characterized in that, The expression for image segmentation of the pixel grayscale values ​​of the preprocessed multichannel microscopic image of myeloma obtained in step S2, step S3, is as follows: ； in, This is the coarse segmentation mask for the image pixels in the m-th row and n-th column; It is a mapping function. It is a lightweight U-Net network, based on input. Output pixel-level prediction scores; The edge constraint coefficient; For edge detection operators; These are the corresponding pixel values ​​in the attention weight matrix.

5. The method for counting cancer cells in a microscopic image of myeloma according to claim 4, characterized in that, In step S5, the average cell depth feature is calculated. The expression is: ； in, Let be the average depth feature of the t-th cell instance. Let be the total number of pixels contained in the t-th cell instance. This represents the classification result for the t-th cell instance.

6. The method for counting cancer cells in a microscopic image of myeloma according to claim 5, characterized in that, The specific implementation method of step S6 includes the following steps: S6.

1. First, perform initial statistics, counting all... The sum of class probabilities of 2 ; S6.

2. Correct the initial statistics, combine them with clinicopathological priors, and output the final myeloma cancer cell count result, expressed as: ； in, For the final myeloma cancer cell count; To round to the nearest integer; For calibration coefficients; Mark the classification results for doctors within the specified range. Results with doctor's annotation The calculated correlation coefficient is used to assess reliability.

Images