Tumor protein quantitative evaluation method and system based on image recognition

Abstract

The invention relates to the technical field of medical image processing and analysis, and discloses an oncoprotein quantitative evaluation method and system based on image recognition, and the method comprises the steps: obtaining a panoramic digital image of an IHC slice, extracting a region of interest in the panoramic digital image, segmenting the region of interest to obtain high-resolution image blocks, and carrying out the segmentation of the high-resolution image blocks; inputting the high-resolution image blocks into a pre-trained FCN model, outputting a preliminary positive region, performing SAPD algorithm optimization according to the preliminary positive region to obtain a positive region, performing parameter extraction on the positive region to obtain the pixel number of an MAFG positive region, the pixel number of an HMOX1 positive region and the area of a single pixel, and based on the pixel number of the MAFG positive region, the pixel number of the HMOX1 positive region and the single pixel area, carrying out calculation to obtain tumor protein quantitative analysis parameters. The method can realize quantitative evaluation of oncoprotein.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing and analysis, and in particular to a method and system for quantitatively evaluating tumor proteins based on image recognition. Background Art

[0002] Immunohistochemistry (IHC) technology is a key pathological diagnostic method and is widely used in cancer detection and tumor prognosis assessment. By staining pathological sections and observing positive reaction areas, IHC technology can reflect the expression level and distribution of target proteins (such as MAFG and HMOX1). However, traditional staining and analysis methods, such as hematoxylin and eosin (HE) staining and DAB staining, still rely on manual observation and qualitative analysis by pathologists. This manual method is highly subjective and has problems such as low efficiency, easy errors and poor reproducibility of results, which makes it difficult to meet the needs of modern medicine for efficient and accurate diagnosis.

[0003] Currently, quantitative assessment of tumor proteins relies primarily on traditional manual methods. These methods involve microscopic observation of sections, manual identification and marking of positive regions, and the use of simple statistical tools to estimate the area of ​​positive regions and protein expression levels. These traditional methods have multiple limitations: low manual efficiency, reliance on operator judgment, difficulty in high-throughput processing of large numbers of samples, and difficulty ensuring measurement consistency across sample batches.

[0004] The limitations of traditional manual methods are primarily low efficiency, subjective results, and poor consistency. This not only increases experimental timelines but also calls into question the reliability of the results. This is particularly true for quantitative assessment of protein-positive areas in tumors, where traditional methods struggle to provide highly accurate and consistent results. Summary of the Invention

[0005] The present invention provides a method and system for quantitative evaluation of tumor proteins based on image recognition to solve the problem that traditional manual methods cannot quantitatively evaluate tumor protein positive areas.

[0006] In a first aspect, in order to solve the above technical problems, the present invention provides a method for quantitatively evaluating tumor proteins based on image recognition, comprising:

[0007] Acquiring a panoramic digital image of the IHC slice, and extracting a region of interest in the panoramic digital image;

[0008] Segmenting the region of interest to obtain a high-resolution image block;

[0009] Input the high-resolution image block into a pre-trained FCN model to output a preliminary positive area;

[0010] According to the preliminary positive area, SAPD algorithm optimization is performed to obtain a positive area;

[0011] Parameter extraction was performed on the positive area to obtain the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel;

[0012] Based on the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel, a calculation is performed to obtain tumor protein quantitative analysis parameters, wherein the tumor protein quantitative analysis parameters include: the area of ​​the MAFG-positive area, the area of ​​the HMOX1-positive area, the intra-batch coefficient of variation of MAFG, the inter-batch coefficient of variation of MAFG, the intra-batch coefficient of variation of HMOX1, and the inter-batch coefficient of variation of HMOX1. Preferably, the region of interest is segmented to obtain a high-resolution image block, including:

[0013] Analyzing the region of interest based on an edge detection method to obtain preliminary boundary information of the region of interest;

[0014] The preliminary boundary information of the region of interest is combined with image features for analysis to obtain a high-resolution image block.

[0015] Preferably, inputting the high-resolution image block into a pre-trained FCN model and outputting the positive area includes:

[0016] The FCN model is a fully convolutional neural network model;

[0017] The FCN model training process is as follows:

[0018] Optimize the ATSS algorithm on the pre-prepared training set to obtain the optimized training set;

[0019] Based on the FNC model, convolution calculation is performed layer by layer on the optimized training set to obtain multi-layer features of the image;

[0020] Restoring the multi-layer features of the image to the original image size through a sampling operation to obtain a corresponding pixel-level feature map;

[0021] Based on the pixel-level feature map, the category probability of each pixel is calculated by a classification activation function to obtain a pixel-level classification result;

[0022] Obtaining a region label through region aggregation according to the pixel-level classification result;

[0023] Compare the region labels with the true labels and calculate the classification loss value;

[0024] Based on the classification loss value, a back propagation algorithm is executed to update the FCN model parameters and optimize the FCN model;

[0025] Repeat the above process and continuously optimize the FCN model until the classification accuracy of the FCN model on the training set reaches the preset classification accuracy requirement, and then stop training.

[0026] Preferably, the SAPD algorithm is optimized based on the preliminary positive area to obtain the positive area including:

[0027] For the preliminary positive regions, extracting the position and size information of the bounding box of each preliminary positive region to obtain a preliminary label of the preliminary positive region;

[0028] Based on the preliminary marking of the preliminary positive areas, the soft label classification prediction is performed in combination with the SAPD algorithm to obtain the distribution information of the preliminary positive areas;

[0029] Based on the distribution information of the preliminary positive area, regional optimization is performed to obtain the positive area.

[0030] Preferably, the parameter extraction of the positive area to obtain the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area and the area of ​​a single pixel comprises:

[0031] Perform pixel counting on the positive area to obtain the number of pixels in the MAFG-positive area and the number of pixels in the HMOX1-positive area;

[0032] Based on the resolution of the image and the predetermined value of the pixel unit area, calculation is performed to obtain the area parameter of a single pixel.

[0033] Preferably, the tumor protein quantitative analysis parameters obtained by calculation based on the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel include:

[0034] The area of ​​MAFG-positive region was calculated by the following formula:

[0035] A MAFG =P MAFG ×S

[0036] Among them, A MAFG is the area of ​​MAFG positive region, P MAFG is the number of pixels in the MAFG-positive area, and S is the area of ​​a single pixel;

[0037] The HMOX1-positive area was calculated using the following formula:

[0038] A HMOX1 =P HMOX1 ×S

[0039] Among them, A HMOX1 is the area of ​​HMOX1 positive region, P HMOX1 is the number of pixels in the HMOX1-positive area.

[0040] Preferably, the calculation based on the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel to obtain the tumor protein quantitative analysis parameter includes:

[0041] The intra-assay coefficient of variation of MAFG was calculated using the following formula:

[0042]

[0043] Among them, CV intra_MAFG is the intra-batch coefficient of variation of MAFG, A MAFG,i is the area of ​​MAFG positive region of sample i, μ MAFG is the mean of the MAFG-positive area within the batch, and n is the total number of samples in a single batch;

[0044] The intra-assay coefficient of variation of HMOX1 was calculated using the following formula:

[0045]

[0046] Among them, CV intra_HMOX1 is the intra-assay coefficient of variation of HMOX1, A HMOX1,i is the HMOX1 positive area of ​​the i-th sample, μ HMOX1 is the mean of the HMOX1-positive area within the batch;

[0047] The MAFG inter-assay coefficient of variation was calculated using the following formula:

[0048]

[0049] Among them, CV inter_MAFG is the MAFG inter-batch variation coefficient, μ MAFG,j is the mean area of ​​MAFG positive regions of the jth batch of samples, μ MAFG is the overall mean of the MAFG positive reaction area of ​​all batches, and m is the total number of different batches;

[0050] The inter-assay coefficient of variation of HMOX1 was calculated using the following formula:

[0051]

[0052] Among them, CV inter_HMOX1 is the inter-batch variation coefficient of HMOX1, μ HMOX1,jis the mean area of ​​HMOX1 positive regions in the jth batch of samples, μ HMOX1 It is the overall mean of the HMOX1 positive reaction area of ​​all batches.

[0053] In a second aspect, the present invention provides a tumor protein quantitative assessment system based on image recognition, comprising:

[0054] a data acquisition module, configured to acquire a panoramic digital image of the IHC slice and extract a region of interest from the panoramic digital image;

[0055] High-resolution image blocks, used to segment the region of interest to obtain high-resolution image blocks;

[0056] An FCN module is used to input the high-resolution image block into a pre-trained FCN model and output a preliminary positive area;

[0057] A positive region module is used to optimize the SAPD algorithm based on the preliminary positive region to obtain a positive region;

[0058] A parameter extraction module performs parameter extraction on the positive area to obtain the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel;

[0059] A quantitative calculation module is configured to calculate, based on the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​each pixel, to obtain a quantitative analysis parameter for tumor protein. In a third aspect, the present invention further provides an electronic device comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements any of the aforementioned methods for quantitative tumor protein assessment based on image recognition.

[0060] In a fourth aspect, the present invention also provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned image recognition-based tumor protein quantitative assessment methods.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] (1) The present invention obtains a panoramic digital image of an IHC slice, further processes it to obtain a high-resolution image block, and then outputs a preliminary positive area, thereby obtaining a positive area, and then extracting parameters to obtain the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel, and finally calculates the quantitative analysis parameters of tumor protein. This overcomes the problems of high subjectivity, low efficiency, easy error, and poor reproducibility of results in the manual method of the prior art, and greatly meets the needs of modern medicine for efficient and accurate diagnosis.

[0063] (2) The present invention utilizes a fully convolutional neural network (FCN) model to output preliminary positive regions. This process is the core of the entire method and enables accurate identification and quantification of tumor protein MAFG and HMOX1 positive regions, providing a foundation for subsequent SAPD algorithm optimization and positive region quantitative analysis. This marks the realization of a complete closed loop from image input to preliminary classification, laying the core foundation for the successful application of the overall technical solution.

[0064] (3) The present invention utilizes the SAPD algorithm for optimization to accurately determine the boundaries and size information of the positive region. The output positive region data is not only closer to the true value in spatial distribution, but also has a higher classification confidence. This provides accurate data input for subsequent quantitative analysis of positive regions, significantly improving the reliability and repeatability of the calculation results, and meeting the high standard requirements for quantitative analysis of tumor protein positive regions.

[0065] (4) The present invention combines the number of pixels in the positive area with the area of ​​a single pixel to meet the high requirements of quantitative analysis of tumor protein-positive areas. The calculated positive area of ​​MAFG and HMOX1 is an important basic data for subsequent analysis. These data are not only used to describe the size of the positive area, but also provide key parameters for the calculation of intra- and inter-batch coefficients of variation. Through this detailed calculation process, accurate quantification of the positive area can be achieved, laying a solid foundation for further quantitative analysis.

[0066] (5) The present invention calculates the intra-batch coefficient of variation and the inter-batch coefficient of variation of MAFG and HMOX1, which can evaluate the stability between samples of the same batch and the measurement consistency between different batches, thereby providing a scientific basis for subsequent quantitative analysis of tumor proteins. It can not only quantify the measurement consistency between samples of the same batch, but also verify the cross-batch repeatability of the experiment, providing a scientific basis for subsequent data analysis and decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a schematic flow chart of the method for quantitatively evaluating tumor protein provided by the first embodiment of the present invention;

[0068] Figure 2FIG. 1 is a schematic diagram of a tumor protein quantitative evaluation system provided in a second embodiment of the present invention. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0070] Reference Figure 1 The first embodiment of the present invention provides a method for quantitatively evaluating tumor proteins based on image recognition, comprising the following steps:

[0071] S11, acquiring a panoramic digital image of the IHC slice, and extracting a region of interest in the panoramic digital image;

[0072] S12, segmenting the region of interest to obtain high-resolution image blocks;

[0073] S13, inputting the high-resolution image block into a pre-trained FCN model to output a preliminary positive area;

[0074] S14, performing SAPD algorithm optimization based on the preliminary positive area to obtain a positive area;

[0075] S15, performing parameter extraction on the positive area to obtain the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel;

[0076] S16, performing calculation based on the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel to obtain a tumor protein quantitative analysis parameter.

[0077] In step S11, a panoramic digital image of the IHC slice is acquired, and a region of interest in the panoramic digital image is extracted;

[0078] It is worth noting that in step S11, obtaining a panoramic digital image of the IHC slice and extracting the region of interest in the panoramic digital image are crucial basic steps in the quantitative evaluation of tumor proteins and play a decisive role in the accuracy and reliability of subsequent analysis.

[0079] First, IHC sections are digitally captured using a high-precision digital scanning device (a dedicated panoramic scanner), converting the physical sections into high-resolution digital images. The sections must be rigorously prepared to ensure that the stained area is clear and undamaged, and the section surface is clean and free of bubbles. The sections are fixed on the scanner stage, and the scanner parameters are adjusted, with the resolution set to 800x optical magnification to ensure sufficient clarity of image detail. Subsequently, a line-by-line scan is performed to obtain a high-resolution digital image of the entire section. To ensure color accuracy, the scanner's light source and sensor must also be calibrated to ensure consistent color across different scan batches.

[0080] During scanning, appropriate lighting intensity and focus mode must be set. Light intensity should be adjusted to highlight DAB-stained areas while avoiding overexposure or darkening the image. Focus mode should be selected using autofocus based on the slice thickness and staining characteristics, ensuring that every detail in the image is in optimal focus. Digital images should be saved with a high bit depth, such as 16-bit, and in a lossless format, such as TIFF, to ensure that detailed information is not lost during subsequent data processing.

[0081] After acquiring panoramic digital images, image preprocessing is required to optimize image quality and make them suitable for subsequent algorithmic analysis. First, the digital images are denoised to remove random noise introduced during the scanning process, such as Gaussian noise or artifacts. Next, color normalization is performed to correct for color deviations caused by differences in staining batches or light sources, maintaining a consistent color distribution across all images. Next, brightness and contrast are adjusted to enhance the contrast between the DAB-stained area and the background, making the target area more prominent.

[0082] After preprocessing, an image segmentation algorithm is used to extract regions of interest from the panoramic digital image. This extraction process is based on DAB staining features, identifying target regions through a color distribution threshold segmentation method. Specifically, the algorithm analyzes the color value of each pixel in the image. When a pixel's color value falls within a dark brown threshold range, it is marked as part of the region of interest. Subsequently, morphological operations (such as erosion and dilation) are combined to optimize the boundaries of the marked region, remove isolated noise points, and fill in broken areas, making the extracted region of interest more complete and accurate.

[0083] Finally, the extracted region of interest is stored as a mask and paired with the original panoramic digital image. The mask file records the boundary location, size, and pixel distribution characteristics of the region of interest. This data provides accurate input for subsequent image segmentation and tumor protein quantification, ensuring the scientific and reliable analysis results. This step lays a solid foundation for the entire quantitative evaluation process and provides reliable data support for the precise analysis and quantification of tumor proteins.

[0084] In step S12, the region of interest is segmented to obtain high-resolution image blocks, including:

[0085] Analyzing the region of interest based on an edge detection method to obtain preliminary boundary information of the region of interest;

[0086] The preliminary boundary information of the region of interest is combined with image features for analysis to obtain a high-resolution image block.

[0087] It is worth noting that segmenting the region of interest and obtaining high-resolution image blocks in step S12 is a critical step in the entire tumor protein quantitative assessment process, directly determining the accuracy and reliability of subsequent analysis and quantitative results. To achieve this goal, the following specific measures and methods are required.

[0088] First, the region of interest (ROI) is analyzed based on edge detection methods to obtain preliminary boundary information. Specifically, the Canny edge detection algorithm, an efficient and accurate edge detection algorithm, is used. By calculating pixel gradients on the input panoramic digital image, edge points with significant grayscale changes within the image are identified. During this processing, appropriate high and low thresholds are set to ensure that the extracted edges accurately reflect the actual boundaries of the region of interest. Furthermore, to avoid noise interference and false detection of non-target edges, the image is Gaussian filtered before edge detection to smooth the image and reduce the impact of subtle noise.

[0089] Next, based on the edge detection results, preliminary boundary information for the region of interest is obtained. In this stage, clustering is performed on the edge points, connecting them into closed boundary curves to ensure that the extracted preliminary boundary fully covers the region of interest. Simultaneously, morphological methods, such as erosion and dilation, are used to optimize the boundary. Erosion removes isolated noisy edge points, while dilation fills in any gaps in the boundary, making the final preliminary boundary information more accurate and complete.

[0090] Subsequently, the preliminary boundary information of the region of interest is combined with image features for further analysis to obtain a high-resolution image patch. Image features used in this process primarily include pixel intensity distribution, color features, and texture features. First, by analyzing the pixel intensity distribution within the boundary, the salient features of the DAB-stained area are identified, thereby distinguishing the target area from the background. Secondly, through color feature analysis, it is further verified that the area within the preliminary boundary matches the color characteristics of the target area, such as dark brown DAB staining. Finally, through texture feature analysis, it is ensured that the texture distribution within the target area is consistent with the expected characteristics of the region of interest.

[0091] After completing the image feature analysis, the original panoramic image is cropped based on the preliminary boundary information to obtain a high-resolution image block. During the cropping process, the background outside the boundary area must be removed, retaining only the complete content of the target area. At the same time, the resolution and pixel density of the high-resolution image block must meet the requirements of the subsequent FCN model input. In this embodiment of the present invention, the resolution is set to 800 times and the pixel density is set to 1200 pixels per millimeter. The size of the high-resolution image block should fully cover the boundary of the area of ​​interest and provide clear image input for subsequent model recognition and classification.

[0092] The segmentation steps described above not only accurately locate the region of interest (ROI) but also minimize interference from non-target regions on subsequent analysis, providing reliable image data input for tumor protein quantitative analysis. The accuracy of this step directly determines the scientific and reliable nature of subsequent positive region extraction and quantitative analysis.

[0093] In step S13, the high-resolution image block is input into a pre-trained FCN model, and the positive area is output, including:

[0094] The FCN model is a fully convolutional neural network model;

[0095] The FCN model training process is as follows:

[0096] Optimize the ATSS algorithm on the pre-prepared training set to obtain the optimized training set;

[0097] Based on the FNC model, convolution calculation is performed layer by layer on the optimized training set to obtain multi-layer features of the image;

[0098] Restoring the multi-layer features of the image to the original image size through a sampling operation to obtain a corresponding pixel-level feature map;

[0099] Based on the pixel-level feature map, the category probability of each pixel is calculated by a classification activation function to obtain a pixel-level classification result;

[0100] Obtaining a region label through region aggregation according to the pixel-level classification result;

[0101] Compare the region labels with the true labels and calculate the classification loss value;

[0102] Based on the classification loss value, a back propagation algorithm is executed to update the FCN model parameters and optimize the FCN model;

[0103] Repeat the above process and continuously optimize the FCN model until the classification accuracy of the FCN model on the training set reaches the preset classification accuracy requirement, and then stop training.

[0104] It is worth noting that in step S13, the high-resolution image blocks are input into a pre-trained fully convolutional neural network model (FCN) to ultimately output preliminary positive areas. This process is the core of the entire method, aiming to achieve accurate identification and quantification of positive areas for tumor proteins MAFG and HMOX1. The essence of the FCN model is a pixel-level classification network, which inputs high-resolution pathology slice image blocks and outputs classified preliminary positive areas. In order to ensure the classification accuracy and generalization ability of the model, it is necessary to rely on a high-quality training set and ensure the robustness and accuracy of the classification model through a dynamic optimization strategy during the training process. The quality of the training set, the representativeness of the samples, and the setting of the classification accuracy during the training process all play a decisive role in the final performance of the model.

[0105] The training set, the core foundation for FCN model training, is obtained by scanning IHC slides using a high-precision microscope. The scanning resolution is uniformly set to 0.25 microns per pixel to ensure that subtle features of pathological tissue can be captured. This resolution was chosen based on the clear boundaries of stained regions required by current pathology research, while also taking into account the processability of the scanned data. The scanned slide images are manually annotated, primarily to identify areas positive for expression and areas negative for the tumor proteins MAFG and HMOX1. The annotated boundaries are accurate to the pixel level to ensure that the training set covers a wide range of sample characteristics, including differences in staining depth and blurred boundaries between positive regions.

[0106] In order to make up for the limitation of insufficient learning of complex sample features in the original training set samples, the training data was further optimized by the ATSS (Adaptive Training Sample Selection) algorithm. The ATSS algorithm adjusts the sample weights according to the difficulty of sample classification, so that samples with greater classification difficulty occupy a higher proportion in the training set. For example, for image blocks with uneven staining, blurred boundaries, or too small positive expression areas, the ATSS algorithm will dynamically increase their weights to ensure that the model can focus on learning the features of these complex samples. The optimized training set has been significantly improved in the balance and representativeness of classification samples, ensuring that the model can cope with a variety of complex pathological images in practical applications. In addition, in order to ensure that the model can take into account both classification accuracy and computational efficiency, the optimized training set still retains a certain proportion of simple samples to maintain the stability of the training process.

[0107] Based on the optimized training set, the FCN model extracts multiple layers of image features through layer-by-layer convolution calculations, including low-level features (edges, textures), mid-level features (regional contours, shapes), and high-level features (staining density, positive distribution, etc.). The model uses a standard convolutional neural network architecture, and each layer of convolution operation extracts features from the input data through a trainable convolution kernel. Feature extraction follows a bottom-up approach, accumulating and optimizing layer by layer. Low-level features provide preliminary structural information, while mid-level and high-level features focus on the classification and boundary location of positive areas. Through this process, the FCN model can achieve preliminary labeling of positive areas at the pixel level, laying the foundation for subsequent classification optimization.

[0108] During model training, to ensure that the final classification results meet clinical needs, a classification accuracy of 95% was set. This accuracy was based on a comprehensive analysis of the tolerance for clinical pathological diagnostic errors and the technical feasibility of the FCN model. The training process optimizes the convolution kernel weights using a backpropagation algorithm, gradually reducing the classification error with each iteration. Model training is terminated when the classification accuracy reaches a preset value or the training loss function reaches a plateau. A classification accuracy of 95% demonstrates sufficient reliability in identifying positive regions in clinical pathological diagnoses, meeting the practical requirements for classifying tumor protein MAFG and HMOX1-positive regions.

[0109] In the trained FCN model, high-resolution image patches are input into the model for prediction, and the model generates preliminary positive regions through pixel-level classification. Specifically, the model first assigns a classification label to each pixel in the input image and generates a pixel-level distribution map of preliminary positive regions based on the classification label. Subsequently, the preliminary positive regions are further processed using a regional aggregation algorithm to integrate continuous positive pixels into complete regions while removing isolated noise pixels to ensure the integrity and accuracy of the output results. These preliminary positive regions provide the basis for subsequent SAPD algorithm optimization and quantitative analysis of positive regions, marking the realization of a complete closed-loop from image input to preliminary classification, and laying the core foundation for the successful application of the overall technical solution.

[0110] In step S14, the SAPD algorithm is optimized based on the preliminary positive area, and the positive area obtained includes:

[0111] For the preliminary positive regions, extracting the position and size information of the bounding box of each preliminary positive region to obtain a preliminary label of the preliminary positive region;

[0112] Based on the preliminary marking of the preliminary positive areas, the soft label classification prediction is performed in combination with the SAPD algorithm to obtain the distribution information of the preliminary positive areas;

[0113] Based on the distribution information of the preliminary positive area, regional optimization is performed to obtain the positive area.

[0114] It is worth noting that, in step S14, based on preliminary positive area, by SAPD algorithm, it is optimized, with the accurate border and the size information of finally obtaining positive area.This process first need carry out bounding box extraction to preliminary positive area, clear and definite boundary position and the size information of each preliminary positive area.These bounding box information are the intuitive reflection of preliminary positive area, have comprised spatial position and the preliminary area scope of positive area in image, are the basis of subsequent optimization.After bounding box extraction is finished, for further improving the accuracy of classification and boundary detection, need carry out preliminary marking to preliminary positive area.Preliminary marking comprises the classification label assignment to each preliminary positive area and the classification of corresponding area attribute, so that for follow-up SAPD algorithm provides preliminary regional description.

[0115] Next, the preliminary positive areas are optimized in combination with the SAPD algorithm. The SAPD algorithm (Soft Anchor Point Detection) is an optimization method that combines a soft label mechanism and bounding box regression. Its core lies in accurately identifying positive areas by adjusting the classification probability and bounding box of the candidate area. Specifically, the SAPD algorithm takes the preliminary marked positive areas as input, generates a classification prediction value for each candidate area based on the soft label mechanism, and determines its category by calculating the classification probability. At the same time, the SAPD algorithm will fine-tune the boundary position and size of the area based on the regression value of the preliminary bounding box to ensure that the bounding box can more accurately cover the actual positive area.

[0116] During the classification prediction phase, the SAPD algorithm uses a trained classification network combined with forward propagation to obtain the classification probability distribution and corresponding category label for each preliminary positive region. During the bounding box optimization phase, the SAPD algorithm uses a loss function in the form of region regression to adjust the position and size of the preliminary bounding box. Through multiple rounds of iterations, convergence is achieved, ultimately outputting a more accurate bounding box for the positive region. It is worth noting that the SAPD algorithm is executed based on the trained model parameters, and its optimization process is achieved by gradually reducing the classification and boundary prediction errors.

[0117] Ultimately, after optimization using the SAPD algorithm, the boundaries and size of the positive regions were precisely determined. The outputted positive region data not only more closely approximated the true spatial distribution but also achieved higher classification confidence. This provided precise data input for subsequent quantitative analysis of positive regions, significantly improving the reliability and reproducibility of the calculation results and meeting the high standards required for quantitative analysis of tumor protein-positive regions.

[0118] In step S15, parameters are extracted for the positive area to obtain the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel, including:

[0119] Perform pixel counting on the positive area to obtain the number of pixels in the MAFG-positive area and the number of pixels in the HMOX1-positive area;

[0120] Based on the resolution of the image and the predetermined value of the pixel unit area, calculation is performed to obtain the area parameter of a single pixel.

[0121] It is worth noting that in step S15, based on the positive area obtained, a detailed parameter extraction operation is performed on it, the purpose of which is to obtain the number of pixels and the single pixel area of ​​MAFG and HMOX1 positive areas, so that basic data is provided for subsequent quantitative calculation. First, it is necessary to perform pixel-by-pixel statistics on the positive areas to clearly define the number of pixels in each positive area. This operation is performed by scanning each pixel of the positive area point by point, and cumulative calculation is performed according to its specific classification label, thereby accurately obtaining the number of pixels in the MAFG and HMOX1 positive areas. The statistics of these pixel numbers can intuitively reflect the preliminary scope of the positive areas, and provide the necessary basis for the analysis of area calculation and batch coefficient of variation.

[0122] Next, in order to further calculate the area parameters of the positive area, it is necessary to clarify the area of ​​a single pixel. The calculation of the area of ​​a single pixel depends on the resolution of the image and the predetermined pixel unit area value. The resolution of the image is obtained by the setting of the image acquisition device and is expressed as the number of pixels per unit length. For example, at a 400-fold microscopic magnification, the actual physical unit length corresponding to each pixel is 0.25 microns. Based on this resolution and the unit area value calibrated by the equipment, the area formula can be used for accurate calculation to obtain the area value of a single pixel. After the calculation is completed, the area of ​​a single pixel becomes one of the important parameters for the subsequent total area calculation.

[0123] After counting the number of pixels in the positive region and calculating the area of ​​each pixel, the two are combined to obtain the area parameter of the positive region. The area parameter is derived by multiplying the number of pixels within the positive region by the area of ​​each pixel. This calculation not only provides high accuracy but also allows for consistency and reliability in repeated experiments, thus meeting the stringent requirements of quantitative analysis of tumor protein-positive areas. Ultimately, these extracted key parameters lay a solid foundation for the subsequent calculation of the positive area and batch coefficient of variation for MAFG and HMOX1.

[0124] In step S16, calculations are performed based on the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel to obtain tumor protein quantitative analysis parameters including:

[0125] The area of ​​MAFG-positive region was calculated by the following formula:

[0126] A MAFG =P MAFG ×S

[0127] Among them, A MAFG is the area of ​​MAFG positive region, P MAFG is the number of pixels in the MAFG-positive area, and S is the area of ​​a single pixel;

[0128] The HMOX1-positive area was calculated using the following formula:

[0129] A HMOX1 =P HMOX1 ×S

[0130] Among them, A HMOX1 is the area of ​​HMOX1 positive region, P HMOX1 is the number of pixels in the HMOX1-positive area.

[0131] It is worth noting that in step S16, based on the positive region pixel number and the single pixel area extracted, the positive region area of ​​MAFG and HMOX1 is calculated respectively.First, the pixel number of the positive region is obtained by carrying out pixel-by-pixel classification statistics to the positive region in the image. This step is in the positive region after classification is completed, and the positive pixel number belonging to MAFG and HMOX1 is counted one by one based on the pixel classification label. The pixel point of each positive region is all labeled as MAFG or HMOX1 positive region according to the output of classification model, and by adding up one by one the pixel points of these marks, the positive pixel number in each region can be obtained.

[0132] Calculating the area of ​​a single pixel requires the resolution and calibration data of the image acquisition device. The device's resolution is provided when the image is acquired. For example, in images acquired by a microscope, the acquisition magnification and resolution are known fixed values. This fixed value corresponds to the actual physical length represented by each pixel. For example, at a microscope magnification of 400x, the actual physical length of each pixel is 0.25 microns. Based on this physical length, the actual area of ​​each pixel can be accurately calculated. The value of the single pixel area is provided by the acquisition device after calibration, ensuring consistency in each calculation and avoiding the introduction of human errors.

[0133] Next, by combining the number of pixels and the area per pixel in the positive region, the areas of MAFG and HMOX1 positive regions can be calculated separately. Specifically, the number of pixels in the MAFG-positive region represents the number of pixels in that region, while the area per pixel provides the actual physical area corresponding to each pixel. Therefore, by multiplying the number of pixels in the MAFG-positive region by the area per pixel, the actual area of ​​the MAFG-positive region can be calculated. Similarly, by performing the same process on the number of pixels in the HMOX1-positive region, the area of ​​the HMOX1-positive region can be calculated.

[0134] Throughout the entire process, the accuracy of the classification model and the precision of the resolution calibration directly determine the accuracy of the positive area calculation. Because the classification model has undergone rigorous training and testing, it ensures high-precision classification and labeling of positive areas in the image. The precision of the resolution calibration ensures accurate calculation of the area of ​​each pixel, providing a scientific basis for the final area calculation.

[0135] Ultimately, the calculated positive area for MAFG and HMOX1 serves as crucial foundational data for subsequent analysis. These data not only describe the size of the positive area but also provide key parameters for calculating the intra- and inter-assay coefficients of variation. This detailed calculation process enables precise quantification of the positive area, laying a solid foundation for further quantitative analysis.

[0136] In step S16, calculations are performed based on the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel to obtain tumor protein quantitative analysis parameters, including:

[0137] The intra-assay coefficient of variation of MAFG was calculated using the following formula:

[0138]

[0139] Among them, CV intra_MAFG is the intra-batch coefficient of variation of MAFG, A MAFG,i is the area of ​​MAFG positive region of sample i, μ MAFG is the mean of the MAFG-positive area within the batch, and n is the total number of samples in a single batch;

[0140] The intra-assay coefficient of variation of HMOX1 was calculated using the following formula:

[0141]

[0142] Among them, CV intra_HMOX1 is the intra-assay coefficient of variation of HMOX1, A HMOX1,iis the HMOX1 positive area of ​​the i-th sample, μ HMOX1 is the mean of the HMOX1-positive area within the batch;

[0143] The MAFG inter-assay coefficient of variation was calculated using the following formula:

[0144]

[0145] Among them, CV inter_MAFG is the MAFG inter-batch variation coefficient, μ MAFG,j is the mean area of ​​MAFG positive regions of the jth batch of samples, μ MAFG is the overall mean of the MAFG positive reaction area of ​​all batches, and m is the total number of different batches;

[0146] The inter-assay coefficient of variation of HMOX1 was calculated using the following formula:

[0147]

[0148] Among them, CV inter_HMOX1 is the inter-batch variation coefficient of HMOX1, μ HMOX1,j is the mean area of ​​HMOX1 positive regions in the jth batch of samples, μ HMOX1 It is the overall mean of the HMOX1 positive reaction area of ​​all batches.

[0149] It is worth noting that in this step, tumor protein quantification analysis parameters are calculated based on the number of pixels and individual pixel areas extracted from MAFG and HMOX1-positive regions, as well as the intra- and inter-batch sample distribution. These parameters primarily include the calculation of the intra- and inter-batch coefficients of variation for MAFG and HMOX1. These parameters are calculated to assess stability within samples within the same batch and measurement consistency between different batches, thereby providing a scientific basis for subsequent tumor protein quantification analysis.

[0150] First, the intra-batch coefficient of variation (CV) is calculated to measure the measurement variability between samples within the same batch. The positive area for MAFG and HMOX1 is an important indicator for quantitative analysis, and its distribution within a single batch reflects the consistency of measurements within that batch. By calculating the positive area for each sample and comparing it with the batch mean, the deviation of each sample from the mean can be determined. The squared deviations of all samples are summed and averaged, and the results are then normalized to effectively quantify the degree of variability between samples within the batch. A smaller intra-batch coefficient of variation indicates more consistent measurement results within the batch and a higher reliability of the experimental results.

[0151] Secondly, the inter-batch coefficient of variation is calculated to assess the stability of measurement results between different batches of samples. Since the experimental process involves the processing of multiple batches of samples, the inter-batch coefficient of variation can reflect whether the experimental conditions of different batches are consistent. Specifically, by calculating the mean of the positive area of ​​each batch of samples and comparing it with the overall mean of all batches, the mean deviation between different batches can be assessed. The inter-batch coefficient of variation is finally obtained by summing the squares of the deviations from the batch mean and taking the average value, and then normalizing it. A lower inter-batch coefficient of variation indicates that the experimental results of different batches are highly consistent, thereby improving the credibility of the experiment and the reproducibility of the results.

[0152] The calculation of these coefficients of variation relies on the core parameter of the positive area, and the accuracy of the positive area directly depends on the number of pixels previously extracted and the area of ​​a single pixel. By incorporating the number of samples within a batch (n) or the total number of batches (m) into the calculation of the coefficient of variation, the overall characteristics of the sample distribution can be effectively reflected, and errors caused by a single or too few samples can be avoided. The setting of the above parameters and the calculation method have been verified and can provide comprehensive and reliable data support for quantitative analysis. This process can not only quantify the measurement consistency between samples in the same batch, but also verify the cross-batch repeatability of the experiment, providing a scientific basis for subsequent data analysis and decision-making.

[0153] Reference Figure 2 The second embodiment of the present invention provides a tumor protein quantitative evaluation system based on image recognition, comprising:

[0154] a data acquisition module, configured to acquire a panoramic digital image of the IHC slice and extract a region of interest from the panoramic digital image;

[0155] High-resolution image blocks, used to segment the region of interest to obtain high-resolution image blocks;

[0156] An FCN module is used to input the high-resolution image block into a pre-trained FCN model and output a preliminary positive area;

[0157] A positive region module is used to optimize the SAPD algorithm based on the preliminary positive region to obtain a positive region;

[0158] A parameter extraction module performs parameter extraction on the positive area to obtain the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel;

[0159] Quantitative calculation module: Based on the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area and the area of ​​a single pixel, a calculation is performed to obtain a quantitative analysis parameter of tumor protein.

[0160] It should be noted that the image recognition-based tumor protein quantitative assessment system provided in an embodiment of the present invention is used to execute all the process steps of the image recognition-based tumor protein quantitative assessment method in the above embodiment. The working principles and beneficial effects of the two correspond one to one, and therefore will not be repeated here.

[0161] An embodiment of the present invention further provides an electronic device. The electronic device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a tumor protein quantitative assessment program. When the processor executes the computer program, the steps of each of the above-mentioned embodiments of the tumor protein quantitative assessment method based on image recognition are implemented, such as Figure 1 Alternatively, when the processor executes the computer program, the functions of the modules / units in the above-mentioned device embodiments are realized, such as the quantitative calculation module.

[0162] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of implementing specific functions, and the instruction segments are used to describe the execution process of the computer program in the electronic device.

[0163] The electronic device may be a computing device such as a desktop computer, notebook, PDA, or smart tablet. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will appreciate that the aforementioned components are merely examples of electronic devices and do not constitute a limitation of the electronic device. The electronic device may include more or fewer components than those described above, or a combination of certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, and the like.

[0164] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the electronic device, connecting various parts of the entire electronic device using various interfaces and lines.

[0165] The memory can be used to store the computer programs and / or modules, and the processor realizes various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created based on the use of the mobile phone (such as audio data, a phone book, etc.). In addition, the memory can include a high-speed random access memory and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0166] Wherein, if the module / unit integrated in the electronic device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by a computer program to instruct the relevant hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, it can implement the steps of each of the above-mentioned method embodiments. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0167] It should be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art can understand and implement the present invention without inventive effort.

[0168] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for quantitative evaluation of tumor proteins based on image recognition, characterized in that: include: Acquiring a panoramic digital image of the IHC slice, and extracting a region of interest in the panoramic digital image; Segmenting the region of interest to obtain a high-resolution image block; Input the high-resolution image block into a pre-trained FCN model to output a preliminary positive area; According to the preliminary positive area, SAPD algorithm optimization is performed to obtain a positive area; Parameter extraction was performed on the positive area to obtain the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel; Based on the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel, calculations are performed to obtain tumor protein quantitative analysis parameters, wherein the tumor protein quantitative analysis parameters include: the area of ​​the MAFG-positive area, the area of ​​the HMOX1-positive area, the intra-batch coefficient of variation of MAFG, the inter-batch coefficient of variation of MAFG, the intra-batch coefficient of variation of HMOX1, and the inter-batch coefficient of variation of HMOX1.

2. The method for quantitative evaluation of tumor proteins based on image recognition according to claim 1, characterized in that: Segmenting the region of interest to obtain high-resolution image blocks includes: Analyzing the region of interest based on an edge detection method to obtain preliminary boundary information of the region of interest; The preliminary boundary information of the region of interest is combined with image features for analysis to obtain a high-resolution image block.

3. The method for quantitative evaluation of tumor proteins based on image recognition according to claim 1, characterized in that: The high-resolution image block is input into a pre-trained FCN model to output the positive area, including: The FCN model is a fully convolutional neural network model; The FCN model training process is as follows: Optimize the ATSS algorithm on the pre-prepared training set to obtain the optimized training set; Based on the FNC model, convolution calculation is performed layer by layer on the optimized training set to obtain multi-layer features of the image; Restoring the multi-layer features of the image to the original image size through a sampling operation to obtain a corresponding pixel-level feature map; Based on the pixel-level feature map, the category probability of each pixel is calculated by a classification activation function to obtain a pixel-level classification result; Obtaining a region label through region aggregation according to the pixel-level classification result; Compare the region labels with the true labels and calculate the classification loss value; Based on the classification loss value, a back propagation algorithm is executed to update the FCN model parameters and optimize the FCN model; Repeat the above process and continuously optimize the FCN model until the classification accuracy of the FCN model on the training set reaches the preset classification accuracy requirement, and then stop training.

4. The method for quantitative evaluation of tumor proteins based on image recognition according to claim 1, characterized in that: The SAPD algorithm is optimized based on the preliminary positive area, and the positive area obtained includes: For the preliminary positive regions, extracting the position and size information of the bounding box of each preliminary positive region to obtain a preliminary label of the preliminary positive region; Based on the preliminary marking of the preliminary positive areas, the soft label classification prediction is performed in combination with the SAPD algorithm to obtain the distribution information of the preliminary positive areas; Based on the distribution information of the preliminary positive area, regional optimization is performed to obtain the positive area.

5. The method for quantitative evaluation of tumor proteins based on image recognition according to claim 1, characterized in that: The parameter extraction of the positive area to obtain the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area and the area of ​​a single pixel includes: Perform pixel counting on the positive area to obtain the number of pixels in the MAFG-positive area and the number of pixels in the HMOX1-positive area; Based on the resolution of the image and the predetermined value of the pixel unit area, calculation is performed to obtain the area parameter of a single pixel.

6. The method for quantitative evaluation of tumor proteins based on image recognition according to claim 1, characterized in that: The tumor protein quantitative analysis parameters obtained by calculation based on the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel include: The area of ​​MAFG-positive region was calculated by the following formula: A MAFG =P MAFG ×S Among them, A MAFG is the area of ​​MAFG positive region, P MAFG is the number of pixels in the MAFG-positive area, and S is the area of ​​a single pixel; The HMOX1-positive area was calculated using the following formula: A HMOX1 =P HMOX1 ×S Among them, A HMOX1 is the area of ​​HMOX1 positive region, P HMOX1 is the number of pixels in the HMOX1-positive area.

7. The method for quantitative evaluation of tumor proteins based on image recognition according to claim 1, characterized in that: The calculation based on the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel to obtain tumor protein quantitative analysis parameters includes: The intra-assay coefficient of variation of MAFG was calculated using the following formula: Among them, CV intra_MAFG is the intra-batch coefficient of variation of MAFG, A MAFG,i is the area of ​​MAFG positive region of sample i, μ MAFG is the mean of the MAFG-positive area within the batch, and n is the total number of samples in a single batch; The intra-assay coefficient of variation of HMOX1 was calculated using the following formula: Among them, CV intra_HMOX1 is the intra-assay coefficient of variation of HMOX1, A HMOX1,i is the HMOX1 positive area of ​​the i-th sample, μ HMOX1 is the mean of the HMOX1-positive area within the batch; The MAFG inter-assay coefficient of variation was calculated using the following formula: Among them, CV inter_MAFG is the MAFG inter-batch variation coefficient, μ MAFG,j is the mean area of ​​MAFG positive regions of the jth batch of samples, μ MAFG is the overall mean of the MAFG positive reaction area of ​​all batches, and m is the total number of different batches; The inter-assay coefficient of variation of HMOX1 was calculated using the following formula: Among them, CV inter_HMOX1 is the inter-batch variation coefficient of HMOX1, μ HMOX1,j is the mean area of ​​HMOX1 positive regions in the jth batch of samples, μ HMOX1 It is the overall mean of the HMOX1 positive reaction area of ​​all batches.

8. A tumor protein quantitative assessment system based on image recognition, characterized in that: include: a data acquisition module, configured to acquire a panoramic digital image of the IHC slice and extract a region of interest from the panoramic digital image; High-resolution image blocks, used to segment the region of interest to obtain high-resolution image blocks; An FCN module is used to input the high-resolution image block into a pre-trained FCN model and output a preliminary positive area; A positive region module is used to optimize the SAPD algorithm based on the preliminary positive region to obtain a positive region; A parameter extraction module performs parameter extraction on the positive area to obtain the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area, and the area of ​​a single pixel; Quantitative calculation module: Based on the number of pixels in the MAFG-positive area, the number of pixels in the HMOX1-positive area and the area of ​​a single pixel, a calculation is performed to obtain a quantitative analysis parameter of tumor protein.

9. An electronic device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the method for quantitatively evaluating tumor protein based on image recognition according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the method for quantitative evaluation of tumor proteins based on image recognition according to any one of claims 1 to 7.

Images