Intelligent typing method and system for M protein in immunofixation electrophoresis image

Abstract

The invention discloses an intelligent typing method and system for M protein in an immunofixation electrophoresis image, and belongs to the technical field of immunofixation electrophoresis image analysis. According to the intelligent typing method and system for M protein in the immunofixation electrophoresis image, an expert knowledge-driven labeling specification of a double-blind labeling process and a labeling verification algorithm embedded into electrophoresis mobility physical constraint is established; according to the method, the image is subjected to standardization processing through the Vahadane dyeing normalization and strip ROI dynamic extraction technology, intelligent noise suppression processing is carried out on image data, the data quality is improved, standardized interpretation can be carried out on the electrophoresis image through a hybrid neural network deep learning algorithm and an AI model, the influence of human factors is reduced, and the method is suitable for being applied to the field of electrophoresis. And the model can reproduce the specificity and accuracy of the expert level. The problems that traditional naked eye interpretation is high in subjectivity, low in accuracy rate, low in efficiency of searching for expert consultation for complex and difficult cases and the like are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of immunofixation electrophoresis image analysis, and in particular relates to an intelligent typing method and system for M protein in immunofixation electrophoresis images. Background Art

[0002] Monoclonal immunoglobulins, also known as M-proteins, are immunoglobulin molecules and their fragments (such as light and heavy chains) with identical structure and electrophoretic mobility, produced by the abnormal proliferation of monoclonal B lymphocytes or plasma cells. M-proteins are important markers for diseases such as multiple myeloma, macroglobulinemia, and malignant lymphoma. Currently, clinical diagnosis of monoclonal immunoglobulin typing primarily relies on immunofixation electrophoresis, and the results provide important clinical guidance for the diagnosis and treatment of these diseases, particularly multiple myeloma.

[0003] The traditional immunofixation electrophoresis typing diagnosis has the following main problems:

[0004] Low diagnostic efficiency: The diagnostic efficacy of traditional immunofixation electrophoresis testing is subject to two core issues. The first is subjective dependence: the diagnostic results rely on the naked eye interpretation of the electrophoresis film by clinical laboratory technicians, and the efficiency and quality of the interpretation are positively correlated with the technician's level of experience. The second is objective resource limitations: for complex and difficult cases (such as special band typing or weakly positive samples), it is usually necessary to seek consultation and discussion with senior technicians, clinical physicians or specialists. However, at present, the country lacks a systematic electrophoresis pattern interpretation training system, and there is a limited number of senior technicians and experts who are proficient in clinical electrophoresis pattern diagnosis and uneven regional distribution. Some difficult cases even require cross-hospital, cross-provincial and cross-city assistance to complete the diagnosis. The above multiple factors lead to a prolonged diagnosis cycle and low work efficiency.

[0005] Significant interference from human factors: The interpretation of immunofixation electrophoresis patterns is highly dependent on the subjective experience of clinical laboratory technicians. Different technicians may make different interpretations of the same pattern due to differences in work experience, professional knowledge level, and film reading habits. This inconsistency in interpretation results not only affects the reliability of the diagnosis, but may also delay the patient's diagnostic typing and subsequent treatment.

[0006] Insufficient data utilization: With the increasing number of clinical immunofixation electrophoresis tests, massive amounts of electrophoresis data are stored in traditional image format, lacking effective data management and in-depth analysis methods. This untapped data potential limits its application value in disease diagnosis and treatment optimization and medical research, hindering further improvement in clinical diagnostic capabilities. Summary of the Invention

[0007] In response to the above-mentioned deficiencies in the prior art, the present application provides an intelligent typing method and system for M protein in immunofixation electrophoresis images.

[0008] In a first aspect, the present application proposes an intelligent typing method for M protein in immunofixation electrophoresis images, comprising the following steps:

[0009] Collect the immunofixation electrophoresis film scan images, perform double-blind annotation and standardization to obtain the scan images to be analyzed;

[0010] Construct an initial neural network model, wherein the architecture of the initial neural network model is a CBAM-ResNet50 convolutional backbone network;

[0011] The initial neural network model is subjected to classification training using the scanned image to be analyzed, wherein the classification training includes pre-training and fine-tuning training. The pre-training is performed through puzzle restoration and rotation prediction self-supervised tasks. During the fine-tuning training phase, the labeled data is used to perform multi-task weight dynamic balance and clinical rule constraint reinforcement training to obtain an intelligent classification model to be tested;

[0012] Obtaining a standardized immunofixation electrophoresis film scan image as a test image, inputting the test image into the intelligent typing model to be tested, calculating a performance index of the intelligent typing model to be tested, and selecting a model whose performance index exceeds a preset value as the final intelligent typing model;

[0013] The final intelligent typing model is deployed in the electrophoresis pattern intelligent expert analysis system. The electrophoresis scanning image generated by the electrophoresis instrument supporting software system stored in the laboratory terminal is captured through file system monitoring. The electrophoresis scanning image is subjected to standardized preprocessing. The processed image is analyzed by the final intelligent typing model to generate the intelligent typing result of the M protein.

[0014] In some embodiments, the normalization process includes staining normalization, strip ROI extraction, and grayscale normalization;

[0015] The staining normalization is as follows: converting the immunofixation electrophoresis film scanning image obtained by double-blind annotation into an optical density space for quantitative analysis of staining intensity;

[0016] The Vahadane algorithm is used to separate the optical density vectors of multiple dyes through sparse non-negative matrices and extract the dye vector and concentration matrix , maintain the concentration matrix No change, the dye The resulting color vector Replace with dye The resulting color vector , then reconstruct the normalized image and inversely convert it back to RGB space to ensure that images under different staining conditions have consistent staining effects;

[0017] The strip ROI extraction is as follows: performing Gaussian filtering to denoise the image after staining normalization to retain the main structural information of the image, using contrast-limited adaptive histogram equalization to enhance the contrast of the strip, locating the boundary of the strip by edge detection, and performing binarization processing on the strip to convert the image into a black and white binary image;

[0018] The grayscale normalization is as follows: matching the black and white binary image to the reference template, adjusting the mean, standard deviation and nonlinear brightness of the image, and ensuring that the grayscale values ​​of different images have the same range and distribution through grayscale normalization processing to obtain a standardized processing result image.

[0019] In some embodiments, the initial neural network model includes an input layer, an initial convolutional layer, four residual stages, a multi-scale feature fusion module, and a multi-label classification head module, and the initial convolutional layer, four residual stages, multi-scale feature fusion module, and multi-label classification head module together constitute the CBAM-ResNet50 convolutional backbone network;

[0020] The initial convolutional layer includes a convolutional layer, a batch normalization layer and a pooling layer;

[0021] The four residual stages include residual stage 1, residual stage 2, residual stage 3 and residual stage 4. Each stage consists of multiple residual blocks, and the number of residual blocks is [3, 4, 6, 3]. Each residual block is inserted into the CBAM module after the three basic convolutional layers. The CBAM module includes a channel attention mechanism and a spatial attention mechanism, a compression ratio of 16, an input convolution kernel of 7×7, a stride of 2, and a spatial attention convolution kernel of 7×7.

[0022] The multi-scale feature fusion module includes upsampling, downsampling and feature fusion;

[0023] The multi-label classification head module includes a global average pooling layer, a fully connected layer and multiple independent binary classifiers.

[0024] In some embodiments, the pre-training uses a self-supervised learning task to enable the model to learn the basic features and structure of the image without labeled data. The pre-training puzzle restoration task includes:

[0025] The image received by the input layer is divided into Grid, get 9 's sub-block;

[0026] Randomly shuffle the order of the sub-blocks to generate a shuffled image;

[0027] Each sub-block is processed by the initial convolutional layer and the output is The feature maps of all sub-blocks are concatenated in the channel dimension to form The fusion feature map is output through the global average pooling layer and the fully connected layer to predict the probability distribution of the order of the image sub-blocks ;

[0028] Calculated using the first cross entropy loss function:

[0029]

[0030] in is the first true label, To obtain the first predicted probability, the parameters of the initial convolutional layer and the classification head are optimized by backpropagation.

[0031] In some embodiments, the pre-trained rotation prediction task includes:

[0032] Apply random rotation to the image received by the input layer to generate a rotated image;

[0033] After the initial convolutional layer and four residual stages, the output The deep features of the deep features are output through the global average pooling layer and the fully connected layer to output a 4-dimensional probability distribution ;

[0034] Calculated using the second cross entropy loss function:

[0035]

[0036] in is the second true label, For the second predicted probability, the parameters of the initial convolutional layer and the classification head are optimized by back-propagation.

[0037] In some embodiments, the fine-tuning training is performed using labeled data to adapt the model to a specific classification task. The specific steps are as follows:

[0038] The input layer receives the normalized processing result image and converts it into Feature map of

[0039] The feature map passes through four residual stages in sequence. After each residual block in each residual stage outputs the feature map, it is immediately passed through the CBAM module for attention weighting;

[0040] After all residual stages are processed, the feature maps of different scales are adjusted through the multi-scale feature fusion module and spliced ​​in the channel dimension to form The fusion feature map of

[0041] The fused feature map passes through the global average pooling layer, the fully connected layer and the independent binary classifier to output the predicted probability of each label ;

[0042] Calculated using the third cross entropy loss function:

[0043]

[0044] in, is the final true label, To finally predict the probability, adjust the model parameters through backpropagation, including weights, biases, scaling factors, offset factors, and attention mechanism parameters, so that the prediction results are close to the annotation labels;

[0045] In some embodiments, the multi-task weight dynamic balance training of the fine-tuning training includes:

[0046] Calculate the loss value independently for each puzzle restoration task or rotation prediction task , calculate the gradient of each task loss with respect to the shared parameters , adjust the task weight according to the gradient amplitude ratio , aligning the gradient magnitudes of each task;

[0047] Calculate the weighted total loss function: , and update the parameters in the back-propagation phase.

[0048] In some embodiments, the clinical rule-constrained intensive training of the fine-tuning training includes:

[0049] Convert the diagnostic rule into a differentiable mathematical expression and use it directly as a regularization term Added to the model's loss function, , the model parameters are updated by gradient descent in the back-propagation phase.

[0050] In a second aspect, the present application proposes an intelligent typing system for M protein in immunofixation electrophoresis images, comprising an image acquisition module, an image preprocessing module, a neural network model module, a diagnostic report generation module, a user interface module, and a continuous learning and online update module;

[0051] The image acquisition module is used to capture the immunofixation electrophoresis film scan image generated from the electrophoresis instrument supporting software system. This module uses the file system monitoring mechanism to detect new image files in the specified directory in real time and transmit them to the image preprocessing module;

[0052] The image preprocessing module is used to standardize the collected images, including staining normalization, strip ROI extraction, Gaussian filtering denoising, contrast enhancement, edge detection, binarization and grayscale normalization. This module ensures that the images input to the neural network model are consistent with the images in the training phase in terms of feature extraction and classification tasks;

[0053] The neural network model module is used to deploy the trained neural network model into the system and is responsible for analyzing and classifying the preprocessed images. The module includes an input layer, an initial convolutional layer, four residual stages, a multi-scale feature fusion module, and a multi-label classification head module, which outputs the predicted probability of each label.

[0054] The diagnostic report generation module is used to generate a detailed diagnostic report based on the prediction results of the neural network model. The module integrates the predicted immunoglobulin type and its probability value into the report and provides a visual electrophoresis pattern and band analysis results for reference by clinical technicians and experts;

[0055] The user interface module is used to provide a user-friendly interface to facilitate technicians and experts to view diagnostic reports, mark error cases and provide feedback. This module supports multi-user concurrent operation to ensure the efficiency and convenience of the system;

[0056] The continuous learning and online update module is responsible for collecting error cases and newly labeled data from user feedback, and optimizing the neural network model through continuous learning and online update mechanisms. This module dynamically updates the model weights and clinical rule base to improve the classification accuracy and clinical applicability of the model.

[0057] In a third aspect, the present application proposes an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0058] In a fourth aspect, the present application proposes a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0059] Beneficial effects of the present invention:

[0060] An expert-driven annotation standard has been established, featuring a double-blind annotation process and an annotation verification algorithm embedded with physical constraints on electrophoretic mobility. Vahadane stain normalization and dynamic band ROI extraction technologies standardize images and intelligently suppress noise, improving data quality and providing a reliable data foundation for subsequent AI model training and diagnosis. Using a hybrid neural network deep learning algorithm, the AI ​​model can standardize the interpretation of electrophoresis images, reducing the influence of human factors and enabling the model to replicate expert-level specificity and accuracy. For the analysis and judgment of difficult cases, the intelligent diagnostic system can also provide diagnostic evidence and treatment recommendations, helping to improve the diagnostic skills and professional capabilities of technicians in primary healthcare institutions. This addresses the high subjectivity and low accuracy of traditional visual interpretation, such as band overlap and weak positive results, as well as the inefficiency of seeking expert consultation for complex and difficult cases. The application of an AI intelligent diagnostic system based on deep learning algorithms has enabled immunofixation electrophoresis testing to break through the traditional "experience-intensive" technical barriers and provide a universal solution for early screening and diagnostic typing of diseases such as multiple myeloma. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is the overall flow chart of the present invention.

[0062] Figure 2 This is a system principle block diagram of the present invention. DETAILED DESCRIPTION

[0063] The following will describe exemplary embodiments of the present invention in more detail with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein; rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0064] In a first aspect, the present application proposes an intelligent typing method for M protein in immunofixation electrophoresis images, comprising the following steps:

[0065] S100: Collect the immunofixation electrophoresis film scan images, perform double-blind annotation and standardization to obtain the scan images to be analyzed;

[0066] Double-blind labeling process

[0067] Data Assignment: 2000 electrophoresis film scans were randomly divided into two groups, 1000 in each group, and independently annotated by two clinical electrophoresis experts (Expert A and Expert B). The annotated labels included nine types, including monoclonal IgGκ, IgGλ, and IgAκ, as well as biclonal combinations.

[0068] Annotation tools: Using a customized annotation system, experts generate structured labels by selecting the location and type of bands on the spectrum, and the system automatically records the annotation results.

[0069] Consistency check:

[0070] Calculation of the Kappa coefficient for double-blind labeling: The system automatically compared the two sets of labeling results and calculated the Kappa coefficient to be 0.93 (>0.90 threshold), judging the overall labeling consistency as qualified.

[0071] Disagreement sample processing: For 150 samples with a Kappa coefficient < 0.90 (e.g., blurred bands, weak positive results), the arbitration mechanism is activated:

[0072] First arbitration: The third expert C independently interprets the case. 120 of the cases are consistent with the conclusions of experts A or B (e.g., expert C agrees with A in 90 cases and with B in 30 cases). The majority result is directly adopted.

[0073] Second arbitration: Of the remaining 30 samples, expert C proposed 20 new conclusions (such as revising "monoclonal IgGκ" to "oligoclonal"), triggering additional expert discussions: a five-person team consisting of experts A, B, C and two chief technicians discussed the controversial samples one by one according to the "Guidelines for Clinical Electrophoresis Diagnosis" and finally reached a consensus through voting (such as a 4:1 judgment of "biclonal IgGκ+IgAλ").

[0074] Annotation data processing:

[0075] Label structuring: Convert expert annotation results into multi-label vectors (such as the vector [1,0,0,…,0] corresponding to monoclonal IgGκ).

[0076] Data cleaning: 5 samples that could not reach a consensus after arbitration (e.g., severe band distortion) were removed, and 1995 valid data sets were retained.

[0077] In some embodiments, the normalization process includes staining normalization, strip ROI extraction, and grayscale normalization;

[0078] The staining normalization is as follows: converting the immunofixation electrophoresis film scanning image obtained by double-blind annotation into an optical density space for quantitative analysis of staining intensity;

[0079] The Vahadane algorithm is used to separate the optical density vectors of multiple dyes through sparse non-negative matrices and extract the dye vector and concentration matrix , maintain the concentration matrix No change, the dye The resulting color vector Replace with dye The resulting color vector , then reconstruct the normalized image and inversely convert it back to RGB space to ensure that images under different staining conditions have consistent staining effects;

[0080] The strip ROI extraction is as follows: performing Gaussian filtering to denoise the image after staining normalization to retain the main structural information of the image, using contrast-limited adaptive histogram equalization to enhance the contrast of the strip, locating the boundary of the strip by edge detection, and performing binarization processing on the strip to convert the image into a black and white binary image;

[0081] The grayscale normalization is as follows: matching the black and white binary image to the reference template, adjusting the mean, standard deviation and nonlinear brightness of the image, and ensuring that the grayscale values ​​of different images have the same range and distribution through grayscale normalization processing to obtain a standardized processing result image.

[0082] The 1995 electrophoresis scan images were standardized and preprocessed to eliminate staining differences and noise interference. The RGB images were converted to optical density space and the staining vectors were extracted. and concentration matrix , and reconstruct a normalized image. Strip ROI extraction is then performed on the normalized image. Gaussian filtering for denoising, contrast enhancement with CLAHE, edge detection, and binarization are used to accurately extract the strip regions. Finally, the ROI histogram is matched to a reference template, and the mean, standard deviation, and nonlinear brightness are adjusted to achieve grayscale normalization. These steps effectively standardize the image, providing a high-quality data foundation for subsequent neural network model training and image analysis.

[0083] S200: Constructing an initial neural network model, wherein the architecture of the initial neural network model is a CBAM-ResNet50 convolutional backbone network;

[0084] In some embodiments, the initial neural network model includes an input layer, an initial convolutional layer, four residual stages, a multi-scale feature fusion module, and a multi-label classification head module, and the initial convolutional layer, four residual stages, multi-scale feature fusion module, and multi-label classification head module together constitute the CBAM-ResNet50 convolutional backbone network;

[0085] The initial convolutional layer includes a convolutional layer, a batch normalization layer and a pooling layer;

[0086] The four residual stages include residual stage 1, residual stage 2, residual stage 3 and residual stage 4. Each stage consists of multiple residual blocks, and the number of residual blocks is [3, 4, 6, 3]. Each residual block is inserted into the CBAM module after the three basic convolutional layers. The CBAM module includes a channel attention mechanism and a spatial attention mechanism, a compression ratio of 16, an input convolution kernel of 7×7, a stride of 2, and a spatial attention convolution kernel of 7×7.

[0087] The multi-scale feature fusion module includes upsampling, downsampling and feature fusion;

[0088] The multi-label classification head module includes a global average pooling layer, a fully connected layer and multiple independent binary classifiers.

[0089] 1. Input layer:

[0090] The received data are immunofixation electrophoresis film images, which are grayscale images of size after the normalization process in step S100, or RGB images without normalization. The function of the input layer is to pass these image data to the subsequent convolutional layers for processing.

[0091] 2. Initial convolutional layer:

[0092] Convolutional layer Conv1: The input image is passed through a The convolution kernel is used for convolution operation, and the step size is ,filling , outputting a feature map of 64 channels.

[0093] Batch normalization layer: average the feature maps of 64 channels and variance Normalization of the dataset, learning scaling and offset parameters.

[0094] Max pooling layer: Use The pooling kernel, stride ,filling , downsample the feature map and output feature map.

[0095] 3. Four residual stages:

[0096] Residual stage 1: contains 3 residual blocks, each of which is inserted into the CBAM module after the 3 basic convolutional layers. The compression ratio of the CBAM module is 16, and the input convolution kernel is , step length , the spatial attention convolution kernel is .

[0097] Residual stage stage2: contains 4 residual blocks, each of which is inserted into the CBAM module after the 3 basic convolutional layers.

[0098] Residual stage stage3: contains 6 residual blocks, each of which is inserted into the CBAM module after 3 basic convolutional layers.

[0099] Residual stage stage4: contains 3 residual blocks, each of which is inserted into the CBAM module after the 3 basic convolutional layers.

[0100] Residual stage: It includes four residual stages (stage1, stage2, stage3, stage4). Each stage consists of multiple residual blocks, and the number of residual blocks is [3, 4, 6, 3]. Each residual block is inserted into the CBAM module after the three basic convolutional layers. The CBAM module includes a channel attention mechanism and a spatial attention mechanism, a compression ratio of 16, an input convolution kernel of 7×7, and a stride of , the spatial attention convolution kernel is 7 × 7. The CBAM module weights the feature map by calculating the channel attention weight and spatial attention weight to enhance the expression ability of important features.

[0101] 4. Multi-scale feature fusion module:

[0102] Upsampling: Use bilinear interpolation or transposed convolution to upsample the low-resolution feature map to the same size as the high-resolution feature map.

[0103] Downsampling: Use pooling operations to downsample the high-resolution feature map to make it consistent with the size of the low-resolution feature map.

[0104] Feature fusion: Feature maps of different scales are concatenated in the channel dimension to form a 28×28×256 fused feature map.

[0105] Applied after all residual stages, this approach resizes feature maps of different scales to a consistent size using upsampling (bilinear interpolation or transposed convolution) or downsampling (pooling). Specifically, stage 1 outputs high-resolution detail features (56×56), stage 2 outputs medium-resolution structural features (28×28), stage 3 outputs low-resolution semantic features (14×14), and stage 4 outputs deep abstract features (7×7). The output feature maps of each stage are uniformly resized to 28×28, with 1×1 convolutions used to unify the number of channels to 256. After concatenation along the channel dimension, the convolutions are fused to produce a 28×28×256 feature map.

[0106] 5. Multi-label classification head module:

[0107] Global average pooling layer GAP: performs global average pooling on the fused feature map and outputs a 256-dimensional feature vector.

[0108] Fully connected layer FC: Pass the 256-dimensional feature vector through a fully connected layer and output the predicted probability of each label.

[0109] Independent binary classifier: Each label corresponds to an independent binary classifier, outputting a single clone 、 、 、 、 、 、 、 and oligoclonal probability .

[0110] As the final decision module of the model, it includes a global average pooling layer (GAP), a shared fully connected layer (FC) and multiple independent binary classifiers (one for each label). 、 、 、 、 、 、 、 Oligoclonals are classified using single labels, while biclonals are classified using multi-label classification. Each clone type is treated as an independent binary classification task. Specifically, a global average pooling layer performs a global average on the feature map, outputting a 256-dimensional feature vector. A fully connected layer maps the 256-dimensional feature vector to the label space. Multiple independent binary classifiers perform binary classification on each label, outputting the probability of each label.

[0111] A specific implementation case is used to illustrate the construction process of the neural network model. Suppose there is a standardized immunofixation electrophoresis film image with a size of 224*224*1. First, the image passes through the initial convolution layer, a 7×7 convolution kernel, a batch normalization layer, and a maximum pooling layer to output a 56×56×64 feature map. Then, the feature map passes through four residual stages in sequence. Each residual stage contains multiple residual blocks. Each residual block is inserted into the CBAM module after the basic convolution layer for attention weighting. After all residual stages are processed, the feature map enters the multi-scale feature fusion module, adjusts the size of feature maps of different scales by upsampling or downsampling, and splices them in the channel dimension to form a 28×28×256 fused feature map. Finally, the fused feature map passes through the global average pooling layer, the fully connected layer, and the independent binary classifier to output the predicted probability of each label. Through these steps, the neural network model is able to accurately extract image features and perform classification.

[0112] S300: Performing classification training on the initial neural network model using the scanned image to be analyzed, wherein the classification training includes pre-training and fine-tuning training. The pre-training is performed through puzzle restoration and rotation prediction self-supervised tasks. During the fine-tuning training phase, the labeled data is used to perform multi-task weight dynamic balance and clinical rule constraint reinforcement training to obtain an intelligent classification model to be tested.

[0113] In some embodiments, the pre-training uses a self-supervised learning task to enable the model to learn the basic features and structure of the image without labeled data. The pre-training puzzle restoration task includes:

[0114] The image received by the input layer is divided into Grid, get 9 's sub-block;

[0115] Randomly shuffle the order of the sub-blocks to generate a shuffled image;

[0116] Each sub-block is processed by the initial convolutional layer and the output is The feature maps of all sub-blocks are concatenated in the channel dimension to form The fusion feature map is output through the global average pooling layer and the fully connected layer to predict the probability distribution of the order of the image sub-blocks ;

[0117] Calculated using the first cross entropy loss function:

[0118]

[0119] in is the first true label, To obtain the first predicted probability, the parameters of the initial convolutional layer and the classification head are optimized by backpropagation.

[0120] In some embodiments, the pre-trained rotation prediction task includes:

[0121] Apply random rotation to the image received by the input layer to generate a rotated image;

[0122] After the initial convolutional layer and four residual stages, the output The deep features of the deep features are output through the global average pooling layer and the fully connected layer to output a 4-dimensional probability distribution ;

[0123] Calculated using the second cross entropy loss function:

[0124]

[0125] in is the second true label, For the second predicted probability, the parameters of the initial convolutional layer and the classification head are optimized by back-propagation.

[0126] In some embodiments, the fine-tuning training is performed using labeled data to adapt the model to a specific classification task. The specific steps are as follows:

[0127] The input layer receives the normalized processing result image and converts it into Feature map of

[0128] The feature map passes through four residual stages in sequence. After each residual block in each residual stage outputs the feature map, it is immediately passed through the CBAM module for attention weighting;

[0129] After all residual stages are processed, the feature maps of different scales are adjusted through the multi-scale feature fusion module and spliced ​​in the channel dimension to form The fusion feature map of

[0130] The fused feature map passes through the global average pooling layer, the fully connected layer and the independent binary classifier to output the predicted probability of each label ;

[0131] Calculated using the third cross entropy loss function:

[0132]

[0133] in, is the final true label, To finally predict the probability, adjust the model parameters through backpropagation, including weights, biases, scaling factors, offset factors, and attention mechanism parameters, so that the prediction results are close to the annotation labels;

[0134] In some embodiments, the multi-task weight dynamic balance training of the fine-tuning training includes:

[0135] Calculate the loss value independently for each puzzle restoration task or rotation prediction task , calculate the gradient of each task loss with respect to the shared parameters , adjust the task weight according to the gradient amplitude ratio , aligning the gradient magnitudes of each task;

[0136] Calculate the weighted total loss function: , and update the parameters in the back-propagation phase.

[0137] In some embodiments, the clinical rule-constrained intensive training of the fine-tuning training includes:

[0138] Convert the diagnostic rule into a differentiable mathematical expression and use it directly as a regularization term Added to the model's loss function, , the model parameters are updated by gradient descent in the back-propagation phase.

[0139] Suppose there is an unlabeled immunofixed electrophoresis film image with a size of 224*224*1. First, in the pre-training phase, the model learns the basic features and structure of the image through the puzzle restoration task and the rotation prediction task. Then, in the fine-tuning phase, the model is trained using labeled data and outputs the predicted probability of each label through the initial convolutional layer, four residual stages, multi-scale feature fusion module and multi-label classification head module. Through dynamic balancing of multi-task weights and strengthening of clinical rule constraints, the model can accurately extract image features and perform classification, obtaining an intelligent classification model to be tested.

[0140] S400: Acquire a normalized immunofixation electrophoresis film scan image as a test image, input the test image into the intelligent typing model to be tested, calculate the performance index of the intelligent typing model to be tested, and select the model whose performance index exceeds a preset value as the final intelligent typing model;

[0141] During the model testing phase, the double-blind annotated electrophoresis images were processed using the same standardized preprocessing method as in the training phase. These images were not used in the training phase. The testing process includes the following steps:

[0142] Image preprocessing: The test images are subjected to the same standardization processing as in the training phase, including staining normalization, strip ROI extraction, Gaussian filtering denoising, contrast enhancement, edge detection, binarization, and grayscale normalization, to ensure consistency between the test images and the training images in feature extraction and classification tasks.

[0143] Model inference: The preprocessed image is input into the trained neural network model. The model performs inference through the initial convolutional layer, four residual stages, multi-scale feature fusion module and multi-label classification head module, and outputs the predicted probability of each label. .

[0144] Performance indicator calculation: Based on the model's prediction results and the double-blind labeled true labels, the following performance indicators are calculated:

[0145] Accuracy:

[0146]

[0147] in, For a real example, For a true counterexample, For a false positive example, is a false negative example. The accuracy reflects the proportion of correctly classified samples among all samples.

[0148] Precision:

[0149]

[0150] Precision reflects the proportion of samples that the model predicts to be positive that are actually positive.

[0151] Recall:

[0152]

[0153] The recall rate reflects the proportion of samples that the model predicts as positive examples among the samples that are actually positive examples.

[0154] F1-Score:

[0155]

[0156] The qualified model achieved a recall rate ≥ 90%, a precision rate ≥ 88%, an F1-Score ≥ 0.85, and an accuracy rate ≥ 85%. The excellent model achieved a recall rate and precision rate ≥ 95%, an F1-Score ≥ 0.92, an accuracy rate ≥ 92%, and a clinical rule violation rate of < 1%.

[0157] In order to improve the performance of the model and adapt to dynamically changing clinical needs, the model is optimized using a continuous learning and online update mechanism. Specific optimization methods include:

[0158] Continuous learning: By introducing new annotated data, the model is allowed to perform incremental learning based on existing knowledge. The specific steps are as follows:

[0159] New data collection: New immunofixation electrophoresis film images were collected and subjected to the same normalization and double-blind annotation as in the training phase.

[0160] Model fine-tuning: Input new data into the existing model and update the model parameters through fine-tuning training so that the model can adapt to the new data distribution and clinical needs.

[0161] Online update: Through the online learning mechanism, the model is allowed to dynamically update weights in actual applications. The specific steps are as follows:

[0162] Error case marking: Technicians mark error cases during system use and feed these cases back to the intelligent analysis system.

[0163] Online fine-tuning: The system fine-tunes the model online based on the feedback of error cases and adjusts the model parameters to improve classification accuracy.

[0164] Clinical rule base maintenance: Dynamically maintain the clinical rule base and domain knowledge to ensure that the model follows the latest clinical rules and diagnostic standards during the reasoning process. The specific steps are as follows:

[0165] Rule update: Update the rules and knowledge in the clinical rule base based on the latest clinical research and diagnostic guidelines.

[0166] Rule constraints: The updated clinical rules are converted into differentiable mathematical expressions, added as regularization terms to the model's loss function, and the model parameters are updated through gradient descent.

[0167] A specific implementation case illustrates the model testing and optimization process: the preprocessed image is input into a trained neural network model. The model performs inference through an initial convolutional layer, four residual stages, a multi-scale feature fusion module, and a multi-label classification head module, outputting the predicted probability for each label. Based on the model's predictions and the double-blind labeled ground truth labels, performance metrics such as accuracy, precision, recall, and F1-Score are calculated to evaluate the model's performance. If the model performance does not meet the expected standards, the model is optimized through continuous learning and online update mechanisms. New labeled data and error case feedback are introduced, and the model weights and clinical rule base are dynamically updated to improve the model's classification accuracy and clinical applicability. The final intelligent classification model is obtained, with performance metrics exceeding the preset values.

[0168] S500: Deploy the final intelligent typing model into the electrophoresis pattern intelligent expert analysis system, capture the electrophoresis scanning image generated by the electrophoresis instrument supporting software system stored in the laboratory terminal through file system monitoring, perform standardized preprocessing on the electrophoresis scanning image, analyze the processed image using the final intelligent typing model and generate the intelligent typing result of the M protein.

[0169] The specific implementation steps of system integration are as follows:

[0170] Environment Configuration: Configure the system operating environment on a server or high-performance computing platform, including the operating system, deep learning framework (such as TensorFlow or PyTorch), database management system, etc. Ensure that the system has sufficient computing resources and storage space to support large-scale image processing and high-concurrency user access.

[0171] Model deployment: Deploy the trained neural network model to the system. The specific steps include:

[0172] Model export: Export the trained model to a deployable format (such as SavedModel or ONNX).

[0173] Model loading: Load the exported model into the system and perform necessary initialization operations.

[0174] Interface encapsulation: Provide a unified API interface for the model to facilitate calls from other modules.

[0175] Module integration: Integrate the image acquisition module, image preprocessing module, neural network model module, diagnostic report generation module, user interface module, and continuous learning and online update module into the system. Ensure that data transmission and call logic between modules are correct.

[0176] System testing: Conduct comprehensive testing on the integrated system, including functional testing, performance testing, compatibility testing, etc., to ensure that the system can operate stably in different scenarios and meet the expected performance and accuracy requirements.

[0177] User training: Provide system usage training to clinical technicians and experts to ensure that they can operate the system proficiently, correctly interpret diagnostic reports, and effectively feedback error cases and opinions.

[0178] System launch: The system is officially deployed in a live work environment to support clinical diagnosis. Regular system maintenance and updates are performed to ensure long-term stable operation and continuous optimization of the system.

[0179] In a second aspect, the present application proposes an intelligent typing system for M protein in immunofixation electrophoresis images, comprising an image acquisition module, an image preprocessing module, a neural network model module, a diagnostic report generation module, a user interface module, and a continuous learning and online update module;

[0180] The image acquisition module is used to capture the immunofixation electrophoresis film scan image generated from the electrophoresis instrument supporting software system. This module uses the file system monitoring mechanism to detect new image files in the specified directory in real time and transmit them to the image preprocessing module;

[0181] The image preprocessing module is used to standardize the collected images, including staining normalization, strip ROI extraction, Gaussian filtering denoising, contrast enhancement, edge detection, binarization and grayscale normalization. This module ensures that the images input to the neural network model are consistent with the images in the training phase in terms of feature extraction and classification tasks;

[0182] Immunofixation electrophoresis film scans, including normal samples and samples for monoclonal immunoglobulin typing in diseases such as multiple myeloma, were collected and double-blindedly annotated by two or more clinical technicians and experts. Stain normalization was implemented to account for differences in stains used by different electrophoresis instruments (Helena electrophoresis instruments: acid violet stain; Sebia electrophoresis instruments: crystal violet stain), film specifications, and scanning devices (Helena electrophoresis instruments: external EPSON scanner; Sebia electrophoresis instruments: built-in scanner). Electrophoresis image normalization employed Vahadane stain normalization to account for differences in stains and devices, band ROI extraction, and grayscale normalization. This enabled identification of film scans from different electrophoresis instrument brands, improving data quality and usability, and facilitating the widespread adoption of the subsequent intelligent diagnostic analysis system across clinical laboratories.

[0183] The neural network model module is used to deploy the trained neural network model into the system and is responsible for analyzing and classifying the preprocessed images. The module includes an input layer, an initial convolutional layer, four residual stages, a multi-scale feature fusion module, and a multi-label classification head module, which outputs the predicted probability of each label.

[0184] A neural network model is constructed based on a deep learning algorithm. The convolutional ResNet50 backbone network is embedded with a convolutional attention module (CBAM), a multi-scale feature fusion module, and a multi-label classification head. The convolutional ResNet50 backbone network is adapted for electrophoresis band features. The CBAM module uses a dual attention mechanism (channel + spatial) to focus the network on the electrophoresis band region. The multi-scale feature fusion module aggregates the multi-scale features processed by CBAM at each stage to generate a fused global feature, enabling the model to focus on details while understanding the overall structure. In electrophoresis images, shallow high-resolution features are used to detect faint bands, addressing the weak signal of low-concentration monoclonal protein bands. Deep features help identify broad polyclonal immunoglobulin bands. The fusion module integrates features at all scales to determine the type. The multi-label classification head performs multi-label prediction based on the multi-scale feature fusion feature graph. By explicitly modeling label coexistence relationships, it significantly improves the detection capabilities of complex cases such as biclonal and oligoclonal cases.

[0185] The model automatically learns the features and patterns in immunofixation electrophoresis data, thereby achieving automatic interpretation and diagnosis of monoclonal immunoglobulin typing.

[0186] The diagnostic report generation module is used to generate a detailed diagnostic report based on the prediction results of the neural network model. The module integrates the predicted immunoglobulin type and its probability value into the report and provides a visual electrophoresis pattern and band analysis results for reference by clinical technicians and experts;

[0187] The user interface module is used to provide a user-friendly interface to facilitate technicians and experts to view diagnostic reports, mark error cases and provide feedback. This module supports multi-user concurrent operation to ensure the efficiency and convenience of the system;

[0188] The continuous learning and online update module is responsible for collecting error cases and newly labeled data from user feedback, and optimizing the neural network model through continuous learning and online update mechanisms. This module dynamically updates the model weights and clinical rule base to improve the classification accuracy and clinical applicability of the model.

[0189] A two-stage training strategy, from pre-training to fine-tuning, is employed, with an adaptive model training framework embedded in the training process. The pre-training phase uses standardized images mixed with 10%-15% original images, forcing the model to learn device-independent features and enhancing robustness. The electrophoresis images learned in the pre-training phase do not require annotation, and universal features of electrophoresis images are learned through self-supervised learning (SimCLR). The fine-tuning phase trains on the annotated standardized images, leveraging the annotated data to provide clear clinical diagnostic targets. Combined with transfer learning and data augmentation, the model ensures the accuracy and rationality of clinical diagnostic output. Through self-supervised pre-training and supervised fine-tuning, the model achieves an optimal training path for immunofixation electrophoresis pattern analysis, enabling accurate recognition of immunofixation electrophoresis images and monoclonal immunoglobulin typing. The adaptive model training framework dynamically senses the training status during training and adjusts the training strategy in real time based on feedback signals. The adaptive fine-tuning training strategy utilizes dynamic balancing of multi-task weights, reinforcement of clinical rule constraints, and active learning for difficult samples. The dynamic balancing of multi-task weights automatically adjusts the loss weights according to the convergence speed of multiple tasks such as immunoglobulin typing and classification, band positioning, image segmentation, and quality assessment. The clinical rule constraint reinforcement strategy is based on the identification rules of the monoclonal immunoglobulin laboratory diagnostic guidelines to ensure that the diagnostic results output by the model comply with the identification rules. Active learning is performed on difficult samples with low model prediction confidence (<0.6).

[0190] In a third aspect, the present application proposes an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0191] In a fourth aspect, the present application proposes a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0192] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0193] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0194] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0195] In the embodiments provided in the present disclosure, it should be understood that the disclosed apparatus / computer equipment and methods can be implemented in other ways. For example, the apparatus / computer equipment embodiments described above are merely schematic. For example, the division of modules or units is merely a logical function division. In actual implementation, there may be other division methods. Multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection of the apparatus or unit, which may be electrical, mechanical or other forms.

[0196] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0197] In addition, the functional units in the various embodiments of the present disclosure may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0198] If the integrated module / unit is implemented as a software functional unit and sold or used as a standalone product, it can be stored in a computer-readable storage medium. Based on this understanding, the present disclosure can implement all or part of the process steps in the above-mentioned method embodiments by using a computer program to instruct the relevant hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program can include computer program code, which can be in source code form, object code form, executable file, or some intermediate form. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, removable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunications signals, and software distribution media. It should be noted that the content included in computer-readable media can be appropriately increased or decreased based on the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, based on legislation and patent practice, computer-readable media does not include electrical carrier signals and telecommunications signals.

[0199] The above are only preferred embodiments of the present invention. It should be pointed out that various modifications and improvements made by those skilled in the art without departing from the present technical solution should also be deemed to fall within the scope of protection required by this solution.

Claims

1. An intelligent typing method for M protein in immunofixation electrophoresis images, characterized by: The following steps are involved: Collect the immunofixation electrophoresis film scan images, perform double-blind annotation and standardization to obtain the scan images to be analyzed; Construct an initial neural network model, wherein the architecture of the initial neural network model is a CBAM-ResNet50 convolutional backbone network; The initial neural network model is subjected to classification training using the scanned image to be analyzed, wherein the classification training includes pre-training and fine-tuning training. The pre-training is performed through puzzle restoration and rotation prediction self-supervised tasks. During the fine-tuning training phase, the labeled data is used to perform multi-task weight dynamic balance and clinical rule constraint reinforcement training to obtain an intelligent classification model to be tested; Obtaining a standardized immunofixation electrophoresis film scan image as a test image, inputting the test image into the intelligent typing model to be tested, calculating a performance index of the intelligent typing model to be tested, and selecting a model whose performance index exceeds a preset value as the final intelligent typing model; The final intelligent typing model is deployed in the electrophoresis pattern intelligent expert analysis system. The electrophoresis scanning image generated by the electrophoresis instrument supporting software system stored in the laboratory terminal is captured through file system monitoring. The electrophoresis scanning image is subjected to standardized preprocessing. The processed image is analyzed by the final intelligent typing model to generate the intelligent typing result of the M protein.

2. The method according to claim 1, wherein: The standardization process includes staining normalization, strip ROI extraction and grayscale normalization; The staining normalization is as follows: converting the immunofixation electrophoresis film scanning image obtained by double-blind annotation into an optical density space for quantitative analysis of staining intensity; The Vahadane algorithm is used to separate the optical density vectors of multiple dyes through sparse non-negative matrices and extract the dye vector and concentration matrix , maintain the concentration matrix No change, the dye The resulting color vector Replace with dye The resulting color vector , then reconstruct the normalized image and inversely convert it back to RGB space to ensure that images under different staining conditions have consistent staining effects; The strip ROI extraction is as follows: performing Gaussian filtering to denoise the image after staining normalization to retain the main structural information of the image, using contrast-limited adaptive histogram equalization to enhance the contrast of the strip, locating the boundary of the strip by edge detection, and performing binarization processing on the strip to convert the image into a black and white binary image; The grayscale normalization is as follows: matching the black and white binary image to the reference template, adjusting the mean, standard deviation and nonlinear brightness of the image, and ensuring that the grayscale values ​​of different images have the same range and distribution through grayscale normalization processing to obtain a standardized processing result image.

3. The method according to claim 2, wherein: The initial neural network model includes an input layer, an initial convolutional layer, four residual stages, a multi-scale feature fusion module and a multi-label classification head module, which together constitute the CBAM-ResNet50 convolutional backbone network; The initial convolutional layer includes a convolutional layer, a batch normalization layer and a pooling layer; The four residual stages include residual stage 1, residual stage 2, residual stage 3 and residual stage 4. Each stage consists of multiple residual blocks, and the number of residual blocks is [3, 4, 6, 3]. Each residual block is inserted into the CBAM module after the three basic convolutional layers. The CBAM module includes a channel attention mechanism and a spatial attention mechanism, a compression ratio of 16, an input convolution kernel of 7×7, a stride of 2, and a spatial attention convolution kernel of 7×7. The multi-scale feature fusion module includes upsampling, downsampling and feature fusion; The multi-label classification head module includes a global average pooling layer, a fully connected layer and multiple independent binary classifiers.

4. The method according to claim 3, wherein: The pre-training uses a self-supervised learning task to enable the model to learn the basic features and structure of the image without labeled data. The pre-training puzzle restoration task includes: The image received by the input layer is divided into Grid, get 9 's sub-block; Randomly shuffle the order of the sub-blocks to generate a shuffled image; Each sub-block is processed by the initial convolutional layer and the output is The feature maps of all sub-blocks are concatenated in the channel dimension to form The fusion feature map is output through the global average pooling layer and the fully connected layer to predict the probability distribution of the order of the image sub-blocks ; Calculated using the first cross entropy loss function: in is the first true label, To obtain the first predicted probability, the parameters of the initial convolutional layer and the classification head are optimized by backpropagation.

5. The method according to claim 4, characterized in that: The pre-trained rotation prediction task includes: Apply random rotation to the image received by the input layer to generate a rotated image; After the initial convolutional layer and four residual stages, the output The deep features of the deep features are output through the global average pooling layer and the fully connected layer to output a 4-dimensional probability distribution ; Calculated using the second cross entropy loss function: in is the second true label, For the second predicted probability, the parameters of the initial convolutional layer and the classification head are optimized by back-propagation.

6. The method according to claim 5, characterized in that: The fine-tuning training is performed using labeled data to adapt the model to specific classification tasks. The specific steps are as follows: The input layer receives the normalized processing result image and converts it into Feature map of The feature map passes through four residual stages in sequence. After each residual block in each residual stage outputs the feature map, it is immediately passed through the CBAM module for attention weighting; After all residual stages are processed, the feature maps of different scales are adjusted through the multi-scale feature fusion module and spliced ​​in the channel dimension to form The fusion feature map of The fused feature map passes through the global average pooling layer, the fully connected layer and the independent binary classifier to output the predicted probability of each label ; Calculated using the third cross entropy loss function: in, is the final true label, To finally predict the probability, the model parameters, including weights, biases, scaling factors, offset factors, and attention mechanism parameters, are adjusted through backpropagation so that the prediction results are approximately the same as the annotation labels.

7. The method according to claim 6, characterized in that: The multi-task weight dynamic balance training of the fine-tuning training includes: Calculate the loss value independently for each puzzle restoration task or rotation prediction task , calculate the gradient of each task loss with respect to the shared parameters , adjust the task weight according to the gradient amplitude ratio , aligning the gradient magnitudes of each task; Calculate the weighted total loss function: , and update the parameters in the back-propagation phase.

8. The method according to claim 7, wherein: The clinical rule-constrained intensive training of the fine-tuning training includes: Convert the diagnostic rule into a differentiable mathematical expression and use it directly as a regularization term Added to the model's loss function, , the model parameters are updated by gradient descent in the back-propagation phase.

9. An intelligent typing system for M protein in immunofixation electrophoresis images, characterized in that: It includes image acquisition module, image preprocessing module, neural network model module, diagnosis report generation module, user interface module and continuous learning and online update module; The image acquisition module is used to capture the immunofixation electrophoresis film scan image generated from the electrophoresis instrument supporting software system. This module uses the file system monitoring mechanism to detect new image files in the specified directory in real time and transmit them to the image preprocessing module; The image preprocessing module is used to standardize the collected images, including staining normalization, strip ROI extraction, Gaussian filtering denoising, contrast enhancement, edge detection, binarization and grayscale normalization. This module ensures that the images input to the neural network model are consistent with the images in the training phase in terms of feature extraction and classification tasks; The neural network model module is used to deploy the trained neural network model into the system and is responsible for analyzing and classifying the preprocessed images. The module includes an input layer, an initial convolutional layer, four residual stages, a multi-scale feature fusion module, and a multi-label classification head module, which outputs the predicted probability of each label. The diagnostic report generation module is used to generate a detailed diagnostic report based on the prediction results of the neural network model. The module integrates the predicted immunoglobulin type and its probability value into the report and provides a visual electrophoresis pattern and band analysis results for reference by clinical technicians and experts; The user interface module is used to provide a user-friendly interface to facilitate technicians and experts to view diagnostic reports, mark error cases and provide feedback. This module supports multi-user concurrent operation to ensure the efficiency and convenience of the system; The continuous learning and online update module is responsible for collecting error cases and newly labeled data from user feedback, and optimizing the neural network model through continuous learning and online update mechanisms. This module dynamically updates the model weights and clinical rule base to improve the classification accuracy and clinical applicability of the model.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images