Artificial Intelligence-Based Pathological Image Data Analysis Method and System

Abstract

The present application discloses a method for analyzing pathological image data based on artificial intelligence, and the present application belongs to the field of artificial intelligence. The method includes: performing cancer type annotation and multi-label feature annotation on each pathological image in the first pathological image dataset, and updating the first pathological image dataset; creating a multi-task learning model; creating a joint loss function; training the multi-task learning model according to the joint loss function and the updated first pathological image dataset, and if the multi-task learning model reaches a preset model determination criterion, the training is completed; identifying the type information, pathological features, and cancer markers of the pathological image according to the multi-task learning model, determining a diagnosis result based on the type information, pathological features, and cancer markers, and sending it to the intelligent terminal device of the doctor. This solution enables doctors to obtain more comprehensive diagnostic information from one model, reduces the time for multiple diagnoses, and improves the response speed and diagnostic efficiency of medical services.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular, to a method and system for analyzing pathological image data based on artificial intelligence. Background Art

[0002] With the rapid development of medical technology, pathological diagnosis, as an important standard for disease diagnosis, its accuracy and efficiency are crucial for the treatment of patients.

[0003] In the prior art, doctors need to observe pathological images with the naked eye, paying attention to features such as the morphology, arrangement of cells, and the size of nucleoli. Using professional knowledge, identifying abnormal features in the lesion area, and combining the patient's medical history, clinical manifestations, and laboratory test results, doctors conduct a comprehensive analysis of pathological images to judge the nature and type of the lesion.

[0004] However, relying solely on doctors to analyze pathological images may be affected by doctors' personal experience and visual fatigue, resulting in subjective differences in diagnostic results. And it requires a large amount of time and energy from doctors. Especially when dealing with a large number of pathological images, it will affect the diagnostic efficiency and speed, increasing medical costs and patients' waiting time. Summary of the Invention

[0005] The embodiments of this application provide a method for analyzing pathological image data based on artificial intelligence, which solves the problems in the prior art that when doctors analyze pathological images, it may be affected by doctors' personal experience and visual fatigue, resulting in subjective differences in diagnostic results. And it requires a large amount of time and energy from doctors. Especially when dealing with a large number of pathological images, it will affect the diagnostic efficiency and speed, increasing medical costs and patients' waiting time.

[0006] In a first aspect, the embodiments of this application provide a method for analyzing pathological image data based on artificial intelligence, and the method includes:

[0007] Obtain a first pathological image dataset, perform cancer type annotation and multi-label feature annotation on each pathological image in the first pathological image dataset, and update the first pathological image dataset according to the annotated pathological images; wherein, the multi-label feature annotation includes pathological feature annotation and cancer biomarker annotation;

[0008] Create a type recognition output head and a multi-task output head, and create a multi-task learning model according to the type recognition output head and the multi-task output head;

[0009] Create a combined loss function according to the cross-entropy loss function and a preset multi-task output head loss function;

[0010] Train a multi-task learning model according to the joint loss function and the updated first pathological image dataset. If the multi-task learning model meets the preset model determination criteria, it is determined that the training of the multi-task learning model is completed;

[0011] Identify the type information, pathological features, and cancer markers of the pathological image according to the multi-task learning model, determine the diagnosis result according to the type information, pathological features, and cancer markers, and send the diagnosis result to the doctor's intelligent terminal device.

[0012] Further, the preset multi-task output head loss function is:

[0013]

[0014] Among them, is the multi-task output head loss function, which is used to measure the predicted value and the actual label y i The difference between them; i is the sample index, indicating the i-th sample in the dataset; C is the number of samples, indicating the total number of samples in the dataset; w i is the weight of the i-th sample, which is used to adjust the loss contribution degree of each sample; y i is the true label, indicating the actual category of the i-th sample; is the predicted probability; (1 - y i ) is the complement of the true label y i , that is, if y i = 1, then (1 - y i ) = 0.

[0015] Further, the joint loss function is:

[0016]

[0017] Among them, Total Loss is the total loss, which represents the optimization target of the model during training and is composed of type loss and multi-task loss; λ 1 is the weight of the type loss, which is used to adjust the contribution degree of type recognition in the total loss; TypeLoss is the type loss, and in this solution, it is the cross-entropy loss function; λ 2 is the weight of the multi-task loss, which is used to adjust the contribution degree of multi-task learning in the total loss; is the preset multi-task output head loss function.

[0018] Further, after training the multi-task learning model according to the joint loss function and the updated first pathological image dataset, the method further includes:

[0019] If the multi-task learning model does not meet the preset model determination criteria, obtain the first training loss value of the type recognition output head and the second training loss value of the multi-task output head;

[0020] If the first training loss value exceeds the preset loss threshold, continuously adjust the weight of the type loss, and use the adjusted weight of the type loss, the joint loss function, and the updated first pathological image dataset to continuously retrain the multi-task learning model until the multi-task learning model meets the preset model determination criteria.

[0021] Further, after obtaining the first training loss value of the type recognition output head and the second training loss value of the multi-task output head, the method further includes:

[0022] If the second training loss value exceeds the preset loss threshold, continuously adjust the weight of the multi-task loss, and use the adjusted weight of the multi-task loss, the joint loss function, and the updated first pathological image dataset to continuously retrain the multi-task learning model;

[0023] If the multi-task learning model still does not meet the preset model determination criteria, continuously adjust the weights of the samples in the preset multi-task output head loss function, and use the adjusted weights of the samples in the preset multi-task output head loss function to continuously update the preset multi-task output head loss function. Use the updated preset multi-task output head loss function, the adjusted weight of the multi-task loss, the joint loss function, and the updated first pathological image dataset to continuously retrain the multi-task learning model until the multi-task learning model meets the preset model determination criteria.

[0024] Further, after sending the diagnosis result to the doctor's intelligent terminal device, the method further includes:

[0025] Obtain the pathological image feature map output by the last convolutional layer of the multi-task learning model, and calculate the gradient information of the type information, pathological features, and cancer markers with respect to the pathological image feature map;

[0026] Weight the pathological image feature map according to the gradient information to obtain a weighted feature map, and normalize the weighted feature map to obtain a heat map;

[0027] Map the heat map to the pathological image to obtain a pathological image mapping map, obtain the pixel values of each pixel in the pathological image mapping map, and determine the area composed of pixels with pixel values higher than the preset pixel threshold as the key attention area. Send the key attention area and the pathological image mapping map to the doctor's intelligent terminal device.

[0028] Further, weighting the pathological image feature map according to the gradient information to obtain a weighted feature map, including:

[0029] Determine the weight information of each channel in the feature map according to the gradient information and a preset feature map weighting formula, and weight the pathological image feature map according to the weight information of each channel to obtain a weighted feature map.

[0030] Further, the preset feature map weighting formula is:

[0031]

[0032] where w i is the weight of channel i; H is the height of the feature map; W is the width of the feature map; represents the gradient of the i-th channel in the feature map at the position (x, y).

[0033] Further, after determining that the multi-task learning model is trained, the method further includes:

[0034] Real-time identify whether the preset model update duration is reached. After reaching the preset model update duration, re-obtain a second pathological image data set including multiple cancer types and markers, perform cancer type annotation and multi-label feature annotation on each pathological image in the second pathological image data set, and update the second pathological image data set according to the annotated pathological images; wherein, the multi-label feature annotation includes pathological feature annotation and cancer marker annotation;

[0035] Retrain the multi-task learning model according to the joint loss function and the updated second pathological image data set. If the multi-task learning model meets the preset model determination criteria, determine that the multi-task learning model is trained.

[0036] According to a second aspect of the present application, there is provided a pathological image data analysis system based on artificial intelligence, the system includes:

[0037] A data set creation module, configured to obtain a first pathological image data set, perform cancer type annotation and multi-label feature annotation on each pathological image in the first pathological image data set, and update the first pathological image data set according to the annotated pathological images; wherein, the multi-label feature annotation includes pathological feature annotation and cancer marker annotation;

[0038] A multi-task learning model creation module, configured to create a type recognition output head and a multi-task output head, and create a multi-task learning model according to the type recognition output head and the multi-task output head;

[0039] A combined loss function creation module for creating a combined loss function based on a cross-entropy loss function and a preset multi-task output head loss function;

[0040] A model training module for training a multi-task learning model according to the combined loss function and an updated first pathological image dataset. If the multi-task learning model meets a preset model determination criterion, it is determined that the training of the multi-task learning model is completed;

[0041] A diagnosis result determination module for identifying the type information, pathological features, and cancer markers of a pathological image according to the multi-task learning model, determining a diagnosis result according to the type information, pathological features, and cancer markers, and sending the diagnosis result to the intelligent terminal device of a doctor.

[0042] In an embodiment of the present application, a first pathological image dataset is obtained, each pathological image in the first pathological image dataset is labeled with a cancer type and multi-label features, and the first pathological image dataset is updated according to the labeled pathological images; wherein, the multi-label feature labeling includes pathological feature labeling and cancer marker labeling; a type recognition output head and a multi-task output head are created, and a multi-task learning model is created according to the type recognition output head and the multi-task output head; a combined loss function is created according to a cross-entropy loss function and a preset multi-task output head loss function; the multi-task learning model is trained according to the combined loss function and the updated first pathological image dataset. If the multi-task learning model meets a preset model determination criterion, it is determined that the training of the multi-task learning model is completed; the type information, pathological features, and cancer markers of a pathological image are identified according to the multi-task learning model, a diagnosis result is determined according to the type information, pathological features, and cancer markers, and the diagnosis result is sent to the intelligent terminal device of a doctor. Through the above artificial intelligence-based pathological image data analysis method, the multi-task learning model can simultaneously output diagnosis results of cancer types, pathological features, and cancer markers in a single model, enabling a doctor to obtain more comprehensive diagnosis information from one model. This can reduce the time for a doctor to perform multiple diagnoses and improve the response speed and diagnosis efficiency of medical services. Description of the Drawings

[0043] Figure 1 is a flowchart of an artificial intelligence-based pathological image data analysis method provided in Embodiment 1 of the present application;

[0044] Figure 2 is a flowchart of an artificial intelligence-based pathological image data analysis method provided in Embodiment 2 of the present application;

[0045] Figure 3It is a schematic flowchart of the method for analyzing pathological image data based on artificial intelligence provided in Embodiment 3 of the present application;

[0046] Figure 4 It is a schematic structural diagram of the system for analyzing pathological image data based on artificial intelligence provided in Embodiment 4 of the present application; Detailed implementation manners

[0047] In order to make the objectives, technical solutions, and advantages of the present application clearer, the following further describes the specific embodiments of the present application in detail with reference to the accompanying drawings. It can be understood that the specific embodiments described herein are only used to explain the present application, rather than limiting the present application. Additionally, it should be noted that for the sake of convenience of description, only the parts related to the present application are shown in the drawings rather than all the content. Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations can be implemented in parallel, concurrently, or simultaneously. In addition, the order of the operations can be rearranged. When the operations are completed, the process can be terminated, but there can also be additional steps not included in the drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0048] The following will clearly describe the technical solutions in the embodiments of the present application with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art belong to the scope of protection of the present application.

[0049] The terms "first", "second", etc. in the description and claims of the present application are used to distinguish similar objects, rather than to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are usually of the same type, and the number of objects is not limited. For example, the first object can be one or multiple. In addition, "and / or" in the description and claims means at least one of the connected objects, and the character " / " generally means an "or" relationship between the associated objects before and after.

[0050] The following will describe in detail the method for analyzing pathological image data based on artificial intelligence provided in the embodiments of the present application with reference to the accompanying drawings, through specific embodiments and their application scenarios.

[0051] Embodiment 1

[0052] Figure 1It is a schematic flowchart of a method for analyzing pathological image data based on artificial intelligence provided in the first embodiment of the present application. As Figure 1 shown, the specific steps are as follows:

[0053] S101. Obtain a first pathological image dataset, perform cancer type annotation and multi-label feature annotation on each pathological image in the first pathological image dataset, and update the first pathological image dataset according to the annotated pathological images; wherein, the multi-label feature annotation includes pathological feature annotation and cancer biomarker annotation.

[0054] First, the usage scenario of this solution can be a scenario where a local pathological image dataset is obtained, cancer type annotation and multi-label feature annotation are performed on each pathological image, a multi-task learning model is trained according to the annotated pathological images, and after the multi-task learning model meets the training standard, the multi-task learning model is used to identify the type information, pathological features, and cancer biomarkers of the pathological image, and the combined diagnosis result is sent to the intelligent terminal device of a doctor.

[0055] Based on the above usage scenario, it can be understood that the execution subject of the present application can be a pathological image data analysis system integrating the function of training the model and the function of determining the diagnosis result by applying the model, and no excessive limitation is made here.

[0056] The first pathological image dataset can be a set of initially collected pathological images containing multiple cancer types. These images can be from different cases. Specifically, they can include multiple cancer types and different pathological features. These image data can be from hospitals, medical research institutions, or public medical image databases.

[0057] Cancer type annotation can be to annotate each pathological image with its corresponding cancer type. This is a classification task, and the annotation result can be one or more predefined cancer type labels, such as breast cancer, lung cancer, and prostate cancer, etc.

[0058] Multi-label feature annotation can be to annotate each pathological image with multiple related features. These features can exist simultaneously and are not mutually exclusive. Multi-label feature annotation includes pathological feature annotation and cancer biomarker annotation.

[0059] Pathological feature annotation can be to annotate each pathological image with features related to the lesion. For example, cell density, morphological changes, and degree of inflammation, etc. These features can help describe the lesion situation of the pathological image in more detail.

[0060] Cancer biomarker annotation can be to annotate specific cancer biomarkers for each pathological image. These biomarkers can be certain specific proteins, gene expression levels, or other biomolecular markers, such as HER2, ER, and PR, etc. These biomarkers are of great significance for cancer diagnosis and treatment.

[0061] The database of a hospital or medical research institution can be accessed to collect pathological images containing different cancer types, ensuring that the image data has high resolution and sufficient diversity, covering different cancer types and pathological features. Then each pathological image is reviewed. According to the lesion characteristics in the image, one or more cancer type labels are assigned to each image. Using annotation tools or software, these labels are recorded in the metadata of the image. And a list of features to be annotated is determined, including pathological features and cancer biomarkers. Then, according to the predefined feature list, each image can be annotated in detail. Using annotation tools or software, these feature annotations are recorded in the metadata of the image. Specifically, pathological features can be annotated such as cell density, morphological changes, and degree of inflammation, etc., and these features are annotated one by one according to the image content. Specifically, cancer biomarkers can be annotated such as cancer biomarkers like HER2, ER, and PR, etc., and these biomarkers are annotated according to the results of specific detection methods. When the annotation is completed, all the annotated pathological images can be used to update the first set of pathological images, so that the pathological images in the first set of pathological images are all annotated pathological images.

[0062] S102, create a type recognition output head and a multi-task output head, and create a multi-task learning model according to the type recognition output head and the multi-task output head.

[0063] The type recognition output head can refer to the part of the model used to identify the cancer type in the pathological image. Specifically, this output head can be the last few fully connected layers, and its output is the probability distribution of different cancer types. The design of the type recognition output head is usually for a single task, that is, to determine which cancer type the pathological image belongs to.

[0064] The multi-task output head can refer to the output part of the model used to process multiple related tasks. Specifically, these tasks can include the prediction of multiple pathological features, such as cell density, morphological changes, degree of inflammation, and the detection of specific cancer biomarkers. The design of the multi-task output head allows the model to process multiple tasks simultaneously in one training process, improving the comprehensive ability and generalization performance of the model.

[0065] A multi-task learning model can refer to a neural network model that can handle multiple related tasks simultaneously. Compared with traditional single-task models, a multi-task learning model can share most of the network structure, transfer information and features between multiple tasks, and improve the overall performance. Such a model can not only identify cancer types but also predict multiple pathological features and cancer markers simultaneously.

[0066] When constructing a neural network model, the basic convolutional layer and pooling layer can be designed first to extract low-level features of pathological images. Before the final fully connected layer, an output head for type recognition can be added. Specifically, it can be one or more fully connected layers, and finally, a Softmax layer is output, and the number of nodes is equal to the number of cancer types. The output of this output head is the probability distribution of different cancer types. Then, based on the same network, multiple output heads for different tasks are added. The output head for each task can contain one or more fully connected layers, and finally, a Sigmoid layer or other suitable activation functions are output. These output heads are respectively used to predict different pathological features and cancer markers. The number of nodes in each output head is determined according to the requirements of the task, such as the number of pathological features or markers. Then, the above type recognition output head and multi-task output heads are connected to the same basic network to form a complete multi-task learning model.

[0067] S103, create a combined loss function according to the cross-entropy loss function and the preset loss function for the multi-task output heads.

[0068] The cross-entropy loss function can be a commonly used loss function in classification tasks, mainly used to measure the difference between the probability distribution predicted by the model and the actual label. Specifically, it can be:

[0069]

[0070] where, y i is the actual label, is the probability predicted by the model, and N is the number of classes.

[0071] The loss function for the multi-task output heads can be a separate loss function defined for each task in multi-task learning.

[0072] The combined loss function can be to combine the loss functions of multiple tasks and is used to optimize the loss of the multi-task model simultaneously.

[0073] For the type recognition task, the cross-entropy loss function can be used. For each multi-task output head, the weighted binary cross-entropy loss function can be used. Then, the cross-entropy loss function and the loss function for the multi-task output heads are combined to form a combined loss function.

[0074] Based on the above technical solution, optionally, the preset multi-task output head loss function is:

[0075]

[0076] where is the multi-task output head loss function, used to measure the difference between the predicted value and the actual label y i ; i is the sample index, representing the i-th sample in the dataset; C is the number of samples, representing the total number of samples in the dataset; w i is the weight of the i-th sample, used to adjust the loss contribution degree of each sample; y i is the true label, representing the actual category of the i-th sample; is the predicted probability; (1 - y i ) is the complement of the true label y i , that is, if y i = 1, then (1 - y i ) = 0.

[0077] In this solution, for each sample i, the binary cross-entropy loss term can be calculated. If y i = 1 (positive example), the loss term is If y i = 0, the loss term is In this solution, since each output head can be processed using this formula, the positive example can be the presence of a specific pathological feature or lesion in the pathological image. For example, for tumor pathological images, the positive examples can include the morphological features, karyotype features, and cell density of tumor cells. It can also be the presence of clear cancer-related markers or known cancer markers in the pathological image. For example, in breast cancer pathological images, positive examples can be marked to indicate HER2-positive cell membrane expression or other histological features. Normal tissues (category 0) may be more common than cancer tissues (category 1), so higher weights may be assigned to cancer tissues to ensure that the model can better learn and distinguish this rare situation. Certain pathological features may have higher discriminability and importance in diagnosis, such as specific cell structures, nuclear morphologies, or abnormal tissue structures. These features contribute significantly to disease diagnosis and analysis and may require higher weights to ensure that the model can accurately capture and utilize these key features.

[0078] Based on the above technical solution, optionally, the joint loss function is:

[0079]

[0080] Among them, Total Loss is the total loss, representing the objective to be optimized during the training of the model, which consists of the type loss and the multi-task loss; λ 1 is the weight of the type loss, used to adjust the contribution degree of type recognition in the total loss; TypeLoss is the type loss, and in this solution, it is the cross-entropy loss function; λ 2 is the weight of the multi-task loss, used to adjust the contribution degree of multi-task learning in the total loss; is the preset loss function for the multi-task output head.

[0081] In this solution, for each sample in the training set, its cross-entropy loss is calculated, based on the model prediction value and the actual label. According to the specific multi-task loss function, the loss of each sample on all tasks is calculated. The cross-entropy loss and the multi-task loss of each sample are weighted and combined according to the set weight coefficients λ1 and λ2 to obtain the total loss of each sample. During the training process, the average value of the total losses of all samples can be calculated as the optimization objective, and the backpropagation algorithm is used to minimize this average total loss. The model parameters are updated through the backpropagation algorithm to minimize the total loss function. This process will continue for multiple epochs until the model converges or reaches the preset stopping condition.

[0082] Based on the above technical solution, optionally, after training the multi-task learning model according to the joint loss function and the updated first pathological image dataset, the method further includes:

[0083] If the multi-task learning model does not meet the preset model determination criteria, obtain the first training loss value of the type recognition output head and the second training loss value of the multi-task output head;

[0084] If the first training loss value exceeds the preset loss threshold, continuously adjust the weight of the type loss, and use the adjusted weight of the type loss, the joint loss function, and the updated first pathological image dataset to continuously retrain the multi-task learning model until the multi-task learning model meets the preset model determination criteria.

[0085] In this solution, the first training loss value can be the training loss value of the type recognition output head obtained after the first training of the multi-task learning model. This loss value can be the result calculated by the cross-entropy loss function and is used to measure the error rate of the model in the classification task.

[0086] The multi-task learning model can contain multiple output heads, and each output head corresponds to a specific task or prediction target. The second training loss value can refer to the loss values calculated during the training of other output heads except the type recognition output head in multi-task learning, that is, pathological feature prediction and cancer biomarker prediction.

[0087] The preset loss threshold can be a predefined upper limit of the loss value. If the first training loss value exceeds this threshold, it indicates that the type recognition performance of the model is not ideal enough and further optimization is required.

[0088] The weight of the type loss can refer to the coefficient used to weight the loss of the type recognition output head in the joint loss function. By adjusting the weight of the type loss, the balance between the classification task and other tasks of the model can be adjusted to expect better model performance.

[0089] After the first training, the loss value of the type recognition output head and the loss values of other multi-task output heads can be calculated. Check whether the loss value of the type recognition output head exceeds the preset loss threshold. If it exceeds, proceed to the next step. Increase or decrease the weight of the type loss, and according to experiments or experience, reset the weight coefficient of the type loss in the joint loss function. Use the adjusted type loss weight, the updated first pathological image dataset, and the redefined joint loss function to retrain the multi-task learning model. After each training iteration, evaluate the performance of the model on the validation set or test set to determine whether the preset model determination criteria are met. If the model does not meet the preset model determination criteria, repeat the above steps, continuously adjusting the weight of the type loss and other parameters until the expected performance requirements are met or the maximum number of training iterations is reached.

[0090] In this solution, by monitoring and adjusting the loss value during the training process, the weaknesses of the model in the classification task can be discovered in a timely manner, and targeted adjustments and optimizations can be made to improve the classification accuracy and the overall performance of the model.

[0091] Based on the above technical solution, optionally, after obtaining the first training loss value of the type recognition output head and the second training loss value of the multi-task output head, the method further includes:

[0092] If the second training loss value exceeds the preset loss threshold, continuously adjust the weight of the multi-task loss, and use the adjusted weight of the multi-task loss, the joint loss function, and the updated first pathological image dataset to continuously retrain the multi-task learning model;

[0093] If the multi-task learning model still does not meet the preset model determination criteria, continuously adjust the weights of each sample in the preset multi-task output head loss function, and use the adjusted weights of each sample in the preset multi-task output head loss function to continuously update the preset multi-task output head loss function. Use the updated preset multi-task output head loss function, the adjusted weight of the multi-task loss, the joint loss function, and the updated first pathological image dataset to continuously retrain the multi-task learning model until the multi-task learning model meets the preset model determination criteria.

[0094] In this solution, the weights of the multi-task losses can be the weight parameters of each loss function in the multi-task learning model. For example, if there are two tasks, namely cancer biomarker prediction and pathological feature prediction, a weight can be assigned to each task to control its relative importance in the loss function of the multi-task output head.

[0095] The weights of each sample can be used to adjust the contribution of each sample to the total loss in the multi-task output head in the loss function, and can be used to handle unbalanced datasets or focus on the importance of specific samples. For example, for rare classes or specific key samples, their weights can be increased so that the model pays more attention to the learning effects of these samples.

[0096] The performance of each task can be evaluated according to the situation of the second training loss value. If the loss value of a certain task is high, its weight can be increased or decreased as needed to adjust the balance between tasks of the model. The updated multi-task loss weights will affect the calculation of the joint loss function, and thus affect the direction and effect of model training. If the multi-task learning model cannot meet the preset model determination criteria no matter how it is updated, then according to the performance of the model on the samples, the importance of each sample or its contribution to model training is evaluated. The weights of each sample in the preset multi-task output head loss function are adjusted so that the model pays more attention to those samples that are helpful for the overall model learning or have special importance. The updated loss function reflects these adjusted weights, and the joint loss function is recalculated for subsequent model training. The multi-task learning model is continuously retrained using the adjusted loss function and weights, and the updated first pathological image dataset. Monitor the performance of the model on the training set and validation set in each training iteration until the model meets the preset model determination criteria. This process may require multiple iterations and adjustments until the multi-task learning model meets the preset model determination criteria.

[0097] In this solution, by reasonably allocating task weights and sample weights, unnecessary calculations in the training process of the model can be reduced, and the training efficiency and convergence speed can be improved. The setting of sample weights can help handle the problem of class imbalance in the dataset, ensure that the model has better learning effects on minority classes or important samples, and prevent the model from being overly biased towards majority classes.

[0098] S104. Train the multi-task learning model according to the joint loss function and the updated first pathological image dataset. If the multi-task learning model meets the preset model determination criteria, it is determined that the training of the multi-task learning model is completed.

[0099] The preset model determination criteria can refer to the criteria and metrics used to evaluate the performance of the model during the training process. Specifically, it can include accuracy, which is used to measure the proportion of correct predictions by the classification model. Precision, which is used to measure the proportion of samples predicted as positive classes that are actually positive classes. Recall, which is used to measure the proportion of samples that are actually positive classes and are correctly predicted as positive classes by the model. F1 score, which is the harmonic mean of precision and recall. Loss value, where the loss value on the validation set reaches a preset threshold. Training time or number of iterations, where the model reaches a stable state within the preset time or number of iterations.

[0100] Before training, the model determination criteria can be defined. The updated first pathological image dataset is input into the model for forward propagation, the loss function is calculated, and then the gradient is backpropagated to update the model parameters. After each training epoch, the performance of the model is evaluated using the validation set to ensure the generalization ability of the model on unseen data. If the performance of the model on the validation set meets the preset model determination criteria, the training is stopped or the current model parameters are saved.

[0101] S105, Identify the type information, pathological features, and cancer markers of the pathological image according to the multi-task learning model, determine the diagnosis result according to the type information, pathological features, and cancer markers, and send the diagnosis result to the intelligent terminal device of the doctor.

[0102] Pathological images can be digital images of tissue sections obtained through pathological examinations. These images can be used for diagnosis and research and can show the microscopic structure of cells and tissues.

[0103] The type information can be the type of cancer in the pathological image, that is, to identify which specific type of cancer the lesion shown in the image belongs to, such as lung cancer, breast cancer, and colorectal cancer, etc.

[0104] Pathological features can refer to the morphological, structural, or histological features extracted from pathological images, such as cell morphology, mitotic index, and cell arrangement morphology, etc.

[0105] Cancer markers can refer to biological markers that can indicate the presence of cancer or pathological status, such as the expression of specific proteins and gene mutations, etc.

[0106] The diagnosis result can be the conclusion drawn from analyzing the pathological image by the model. Specifically, it can include cancer type information, description of pathological features, analysis of cancer marker status, and final disease diagnosis and prognosis assessment.

[0107] The intelligent terminal device can be a mobile device or a dedicated display used by doctors, such as a smartphone, a tablet computer, or a medical dedicated display. These devices can receive and display pathological images and their diagnosis results, facilitating doctors to view remotely or make real-time diagnoses.

[0108] Each new pathological image can be preprocessed and then input into the trained model to obtain the predicted results of the cancer type, pathological feature description, and cancer markers output by the model. A complete diagnostic report is generated by combining the information output by the model, including the cancer type, pathological feature description, status of the analyzed cancer markers, and the final disease diagnosis and prognosis assessment. The generated diagnostic results are sent to the doctor's intelligent terminal device through network transmission or storage services to ensure that the doctor can obtain and view the diagnostic information in a timely manner.

[0109] For the technical solution provided in this embodiment, a first pathological image dataset is obtained, and each pathological image in the first pathological image dataset is labeled with a cancer type and multi-label features, and the first pathological image dataset is updated according to the labeled pathological images; wherein, the multi-label feature labeling includes pathological feature labeling and cancer marker labeling; a type recognition output head and a multi-task output head are created, and a multi-task learning model is created according to the type recognition output head and the multi-task output head; a joint loss function is created according to the cross-entropy loss function and a preset multi-task output head loss function; the multi-task learning model is trained according to the joint loss function and the updated first pathological image dataset, and if the multi-task learning model meets the preset model determination criteria, it is determined that the training of the multi-task learning model is completed; the type information, pathological features, and cancer markers of the pathological image are identified according to the multi-task learning model, and a diagnostic result is determined according to the type information, pathological features, and cancer markers, and the diagnostic result is sent to the doctor's intelligent terminal device. Through the above artificial intelligence-based pathological image data analysis method, the multi-task learning model can simultaneously output the diagnostic results of cancer type, pathological features, and cancer markers in a single model, enabling doctors to obtain more comprehensive diagnostic information from one model. This can reduce the time for doctors to perform multiple diagnoses and improve the response speed and diagnostic efficiency of medical services.

[0110] Embodiment 2

[0111] Figure 2 is a schematic flowchart of the artificial intelligence-based pathological image data analysis method provided in Embodiment 2 of this application, as Figure 2 shown, and the specific method includes the following steps:

[0112] S201, obtain a first pathological image dataset, label each pathological image in the first pathological image dataset with a cancer type and multi-label features, and update the first pathological image dataset according to the labeled pathological images; wherein, the multi-label feature labeling includes pathological feature labeling and cancer marker labeling.

[0113] S202, create a type recognition output header and a multi-task output header, and create a multi-task learning model according to the type recognition output header and the multi-task output header.

[0114] S203, create a combined loss function according to the cross-entropy loss function and a preset multi-task output header loss function.

[0115] S204, train the multi-task learning model according to the combined loss function and the updated first pathological image dataset. If the multi-task learning model meets the preset model determination criteria, it is determined that the training of the multi-task learning model is completed.

[0116] S205, identify the type information, pathological features, and cancer markers of the pathological image according to the multi-task learning model, determine the diagnosis result according to the type information, pathological features, and cancer markers, and send the diagnosis result to the doctor's intelligent terminal device.

[0117] S206, obtain the pathological image feature map output by the last convolutional layer of the multi-task learning model, and calculate the gradient information of the type information, pathological features, and cancer markers with respect to the pathological image feature map.

[0118] In a neural network, a convolutional layer can be composed of a set of convolutional kernels. Each convolutional kernel performs a convolution operation with the input data to generate a corresponding output feature map. The parameters of the convolutional layer are learned through backpropagation so that the network can extract more effective feature representations.

[0119] In pathological image analysis, a pathological image feature map can be a feature map obtained after processing by a convolutional neural network. These feature maps show the abstract features extracted in different convolutional layers, which help the network perform classification or other tasks.

[0120] Gradient information can be the gradient of the loss function with respect to the model parameters. In pathological image analysis, if it is necessary to explain the model's decision or analyze the model's attention to specific tasks, such as type information, pathological features, and cancer markers, the gradient of the feature map can be calculated. These gradients show how the model uses the feature map for classification or prediction, and can be used to visualize the process of the model's decision or generate an explanatory heat map.

[0121] The name or index of the last convolutional layer in the multi-task learning model can be determined. Specifically, it can be viewed from the model architecture or identified through the model definition. A trained multi-task learning model can be used to input a pathological image into the model and obtain the output of the last convolutional layer. Specifically, these outputs can be feature maps containing rich pathological image features. For the feature maps, calculate the gradient information of the type information, pathological features, and cancer markers with respect to the feature maps. Specifically, backpropagation can be used to calculate the gradients of the type information, pathological features, and cancer markers on the feature maps. In the multi-task learning model, the backpropagation path of the gradient can be defined according to the loss function of each task. For each task, namely the type information, pathological features, and cancer markers, calculate the gradient of its loss function with respect to the feature maps. These gradients represent the influence degree of each task on the feature maps.

[0122] S207, weight the pathological image feature map according to the gradient information to obtain a weighted feature map, and normalize the weighted feature map to obtain a heat map.

[0123] The weighted feature map can refer to the image obtained by weighting the feature map according to the gradient information of the model. In deep learning, the feature map output by each convolutional layer can be regarded as a representation of different abstraction levels of the input image. After these feature maps are weighted, the features of specific regions can be strengthened or weakened, so as to better understand the focus points and decision-making basis of the model for the image.

[0124] The heat map is the result obtained by superimposing the weighted feature map on the original input image and normalizing it. It generates a new image by superimposing the values of the weighted feature map at the corresponding positions of the original image, and is used to intuitively display the important regions that the model focuses on in the input image. The main role of the heat map is to provide a visualization method to help explain which regions play a key role in the final prediction during the model's decision-making process.

[0125] For each channel in the feature map, calculate the global average pooling of its gradient in the spatial dimension to obtain the weight of each channel. These weights reflect the influence degree of each channel in the feature map on the model output. Use the calculated weights to weight the original feature map, that is, multiply the feature map of each channel by the corresponding weight, so as to obtain the weighted feature map. Normalize the weighted feature map so that the pixel value range of the image is within [0, 1]. Specifically, it can include subtracting the minimum value from the pixel value and dividing by the maximum value minus the minimum value. Superimposing the normalized weighted feature map on the original pathological image can intuitively display the important regions that the model focuses on. This superimposing process can be completed by performing element-level weighted superimposition of the normalized feature map and the original pathological image.

[0126] Based on the above technical solution, optionally, weighting the pathological image feature map according to the gradient information to obtain a weighted feature map, including:

[0127] Determine the weight information of each channel in the feature map according to the gradient information and a preset feature map weighting formula, and weight the pathological image feature map according to the weight information of each channel to obtain a weighted feature map.

[0128] In this solution, the preset feature map weighting formula can be based on the gradient information of each channel and is used to calculate the weight of each channel. Specifically, global average pooling can be used to calculate the average value of the gradients of each channel in the spatial dimension as the weight.

[0129] A multi-task learning model can be used to obtain the output of the last convolutional layer of the pathological image feature map, calculate the gradient of each channel using the preset feature map weighting formula, perform global average pooling on the gradients of each channel to obtain the weights of each channel, and multiply the feature map of each channel by its corresponding weight to obtain a weighted feature map.

[0130] In this solution, the weighted feature map can highlight the regions of interest in the pathological image for the model, enabling doctors to more intuitively understand the decision-making basis of the model. By weighting the feature map, the importance of different pathological features and cancer markers for the model can be strengthened, helping doctors identify important lesion regions.

[0131] Based on the above technical solution, optionally, the preset feature map weighting formula is:

[0132]

[0133] where w i is the weight of channel i; H is the height of the feature map; W is the width of the feature map; represents the gradient of the i-th channel in the feature map at position (x, y).

[0134] In this solution, for each channel i in the feature map, calculate its gradient values in the spatial dimensions x and y This represents the gradient of the i-th channel in the feature map at each pixel position (x, y). For each position (x, y), take the gradient absolute value Sum the absolute values of the gradients at all positions (x, y) and divide by the total number of pixels H×W of the feature map. This is equivalent to performing global average pooling on the gradients in the spatial dimension. The resulting value is the weight w i of channel i, which represents the contribution degree of this channel to the entire feature map in terms of gradient. Generally, the weight w iThe larger it is, the greater the influence of the information in this channel in the feature map on the output of the model.

[0135] S208, map the heat map to the pathological image to obtain a pathological image mapping map, obtain the pixel values of each pixel in the pathological image mapping map, determine the area composed of pixels with pixel values higher than a preset pixel threshold as the key area of concern, and send the key area of concern and the pathological image mapping map to the intelligent terminal device of the doctor.

[0136] The pathological image mapping map may refer to the resulting image obtained by superimposing the heat map on the original pathological image. It reflects the important areas that the model focuses on in the pathological image and can be a grayscale image, where the intensity of the pixels represents the degree of attention of the model to the pixel area.

[0137] The value of each pixel in the pathological image mapping map represents the importance or confidence of the pixel area in the model's decision-making. Higher pixel values usually correspond to areas that the model considers important.

[0138] The preset pixel threshold can be a predefined value used to determine which pixels are considered part of the key area of concern. Pixels with pixel values higher than the preset pixel threshold are regarded as part of the key area of concern.

[0139] The key area of concern may refer to the area composed of pixels with pixel values higher than the preset pixel threshold in the pathological image mapping map. These areas are considered areas that the model pays special attention to and are of great significance for diagnosis.

[0140] The feature map can be obtained from the last convolutional layer of the multi-task learning model, and the heat map is weighted according to the gradient information. The normalized heat map is superimposed on the original pathological image at the pixel level to form a pathological image mapping map. Obtain the numerical values of each pixel in the pathological image mapping map, which reflect the degree of attention of the model to each pixel area. According to the preset pixel threshold, determine the area composed of pixels with pixel values higher than the threshold in the pathological image mapping map as the key area of concern. Send the information of the pathological image mapping map and the key area of concern to the intelligent terminal device of the doctor for the doctor to further analyze and diagnose.

[0141] In this embodiment, the heat map can intuitively display the important areas that the model focuses on in the pathological image, enabling the doctor to understand the decision-making process and basis of the deep learning model, and enhancing the interpretability of the model. By determining the key area of concern, it can help the doctor quickly locate the possible lesions or abnormal areas in the pathological image, and improve the accuracy and efficiency of diagnosis.

[0142] Embodiment Three

[0143] Figure 3 is a schematic flowchart of a method for analyzing pathological image data based on artificial intelligence provided in Embodiment 3 of the present application. As Figure 3 shown, the specific method includes the following steps:

[0144] S301, obtain a first pathological image dataset, perform cancer type annotation and multi-label feature annotation on each pathological image in the first pathological image dataset, and update the first pathological image dataset according to the annotated pathological images; wherein, the multi-label feature annotation includes pathological feature annotation and cancer marker annotation.

[0145] S302, create a type recognition output head and a multi-task output head, and create a multi-task learning model according to the type recognition output head and the multi-task output head.

[0146] S303, create a combined loss function according to the cross-entropy loss function and a preset multi-task output head loss function.

[0147] S304, train the multi-task learning model according to the combined loss function and the updated first pathological image dataset. If the multi-task learning model meets the preset model determination criteria, it is determined that the training of the multi-task learning model is completed.

[0148] S305, real-time identify whether the preset model update duration is reached. After reaching the preset model update duration, re-obtain a second pathological image dataset including multiple cancer types and markers, perform cancer type annotation and multi-label feature annotation on each pathological image in the second pathological image dataset, and update the second pathological image dataset according to the annotated pathological images; wherein, the multi-label feature annotation includes pathological feature annotation and cancer marker annotation.

[0149] The preset model update duration may refer to the time interval set in real-time applications, that is, the time period during which the model needs to be updated or retrained regularly.

[0150] The second pathological image dataset may refer to the pathological image dataset re-obtained and annotated after the preset model update duration is reached. These datasets may be composed of the first pathological image dataset and the new pathological images generated from the application of the model until the preset model update duration is reached.

[0151] When the preset model update duration arrives, a new second pathological image dataset can be re-collected. Perform detailed annotation on each pathological image in the second pathological image dataset. Specifically, it may include annotating the cancer type and related multi-label features of each image, such as pathological features and cancer markers. Use the annotated second pathological image dataset to update the existing second pathological image dataset.

[0152] S306. Retrain the multi-task learning model according to the combined loss function and the updated second pathological image dataset. If the multi-task learning model meets the preset model determination criteria, it is determined that the training of the multi-task learning model is completed.

[0153] The multi-task learning model can be recompiled based on the updated second pathological image dataset and the defined combined loss function. Specifically, it can include selecting a suitable optimizer, learning rate, and network structure configuration. Use the updated second pathological image dataset and the defined combined loss function to train the multi-task learning model. During the training process, the model will optimize its parameters according to the loss function to gradually improve its performance on pathological image classification and feature prediction tasks. Monitor the performance metrics of the model during the training process, such as training loss, validation loss, and accuracy. These metrics can help evaluate the learning effect and generalization ability of the model on the updated dataset. According to the preset model determination criteria, monitor whether the model has reached the expected performance level. Specifically, it can include meeting the standards of accuracy, recall rate, F1 score, or other relevant metrics. When the multi-task learning model meets or exceeds the preset model determination criteria, confirm that the model training is completed. After the model training is completed, it can be deployed to actual applications for the diagnosis and analysis of pathological images.

[0154] S307. Identify the type information, pathological features, and cancer markers of the pathological image according to the multi-task learning model, determine the diagnosis result according to the type information, pathological features, and cancer markers, and send the diagnosis result to the doctor's intelligent terminal device.

[0155] In this embodiment, regularly retraining the model with the updated dataset can enable the model to continuously learn and adapt to new data features and pattern changes, thereby improving its performance and accuracy in pathological image analysis tasks.

[0156] Embodiment 4

[0157] Figure 4 is a schematic structural diagram of an artificial intelligence-based pathological image data analysis system provided by Embodiment 4 of the present application. As Figure 4 shown, this system is used to implement the method of an artificial intelligence-based pathological image data analysis system provided in Embodiments 1, 2, and 3. Specifically, this system includes the following:

[0158] A dataset creation module 401, configured to obtain a first pathological image dataset, perform cancer type annotation and multi-label feature annotation on each pathological image in the first pathological image dataset, and update the first pathological image dataset according to the annotated pathological images; wherein, the multi-label feature annotation includes pathological feature annotation and cancer marker annotation;

[0159] The multi-task learning model creation module 402 is used to create a type recognition output head and a multi-task output head, and create a multi-task learning model according to the type recognition output head and the multi-task output head;

[0160] The joint loss function creation module 403 is used to create a joint loss function according to the cross-entropy loss function and a preset multi-task output head loss function;

[0161] The model training module 404 is used to train the multi-task learning model according to the joint loss function and the updated first pathological image dataset. If the multi-task learning model meets the preset model determination criteria, it is determined that the multi-task learning model training is completed;

[0162] The diagnosis result determination module 405 is used to identify the type information, pathological features and cancer markers of the pathological image according to the multi-task learning model, determine the diagnosis result according to the type information, pathological features and cancer markers, and send the diagnosis result to the doctor's intelligent terminal device.

[0163] In the embodiment of the present application, the dataset creation module is used to obtain the first pathological image dataset, perform cancer type annotation and multi-label feature annotation on each pathological image in the first pathological image dataset, and update the first pathological image dataset according to the annotated pathological images; wherein, the multi-label feature annotation includes pathological feature annotation and cancer marker annotation; the multi-task learning model creation module is used to create a type recognition output head and a multi-task output head, and create a multi-task learning model according to the type recognition output head and the multi-task output head; the joint loss function creation module is used to create a joint loss function according to the cross-entropy loss function and a preset multi-task output head loss function; the model training module is used to train the multi-task learning model according to the joint loss function and the updated first pathological image dataset. If the multi-task learning model meets the preset model determination criteria, it is determined that the multi-task learning model training is completed; the diagnosis result determination module is used to identify the type information, pathological features and cancer markers of the pathological image according to the multi-task learning model, determine the diagnosis result according to the type information, pathological features and cancer markers, and send the diagnosis result to the doctor's intelligent terminal device. Through the above artificial intelligence-based pathological image data analysis system, the multi-task learning model can simultaneously output the diagnosis results of cancer type, pathological features and cancer markers in a single model, so that doctors can obtain more comprehensive diagnosis information from one model. This can reduce the time for doctors to conduct multiple diagnoses, improve the response speed of medical services and the efficiency of diagnosis.

[0164] The above are only the preferred embodiments of the present application and the technical principles applied. The present application is not limited to the specific embodiments described herein. Various obvious changes, re-adjustments, and substitutions that can be made by those skilled in the art will not depart from the protection scope of the present application. Therefore, although the present application has been described in more detail through the above embodiments, the present application is not limited to the above embodiments. Without departing from the concept of the present application, it may also include more other equivalent embodiments, and the scope of the present application is determined by the scope of the claims.

Claims

1. A pathological image data analysis method based on artificial intelligence, characterized in that: The method comprises: Acquire a first pathological image dataset, annotate each pathological image in the first pathological image dataset with a cancer type and a multi-label feature, and update the first pathological image dataset according to the annotated pathological images; wherein the multi-label feature annotation includes pathological feature annotation and cancer marker annotation; Creating a type recognition output head and a multi-task output head, and creating a multi-task learning model based on the type recognition output head and the multi-task output head; wherein the type recognition output head refers to the part of the multi-task learning model used to identify the cancer type in the pathological image, which is the fully connected layer of the last few layers, and the output is the probability distribution of different cancer types; the multi-task output head refers to the output part of the multi-task learning model used to process multiple related tasks, and these related tasks include the prediction of multiple pathological features; A joint loss function is created according to the cross entropy loss function and the preset multi-task output head loss function; wherein the preset multi-task output head loss function is: in, is the multi-task output head loss function, used to measure the prediction value and the actual label y i The difference between them; i is the sample index, which indicates the i-th sample in the data set; C is the number of samples, which indicates the total number of samples in the data set; w i is the weight of the i-th sample, which is used to adjust the loss contribution of each sample; y i is the true label, indicating the actual category of the i-th sample; is the predicted probability; (1-y i ) is the true label y i The complement of i =1, then (1-yi)=0; The multi-task learning model is trained according to the joint loss function and the updated first pathological image data set. If the multi-task learning model meets the preset model judgment standard, it is determined that the training of the multi-task learning model is completed; wherein the joint loss function is: Among them, TotalLoss is the total loss, which represents the goal optimized by the model during training, and is composed of type loss and multi-task loss; λ1 is the weight of type loss, which is used to adjust the contribution of type recognition to the total loss; Type Loss is the type loss, which is the cross entropy loss function in this scheme; λ2 is the weight of multi-task loss, which is used to adjust the contribution of multi-task learning to the total loss; is the preset multi-task output head loss function; The type information, pathological features and cancer markers of the pathological image are identified according to the multi-task learning model, a diagnosis result is determined according to the type information, pathological features and cancer markers, and the diagnosis result is sent to the doctor's smart terminal device.

2. The pathological image data analysis method based on artificial intelligence according to claim 1, characterized in that: After training the multi-task learning model according to the joint loss function and the updated first pathological image dataset, the method further includes: If the multi-task learning model does not meet the preset model judgment standard, obtaining a first training loss value of the type recognition output head and a second training loss value of the multi-task output head; If the first training loss value exceeds a preset loss threshold, the weight of the type loss is continuously adjusted, and the multi-task learning model is continuously retrained using the adjusted type loss weight, the joint loss function, and the updated first pathological image data set until the multi-task learning model meets the preset model judgment criteria.

3. The pathological image data analysis method based on artificial intelligence according to claim 2 is characterized in that: After obtaining the first training loss value of the type recognition output head and the second training loss value of the multi-task output head, the method further includes: If the second training loss value exceeds a preset loss threshold, continuously adjusting the weight of the multi-task loss, and continuously retraining the multi-task learning model using the adjusted multi-task loss weight, the joint loss function, and the updated first pathological image data set; If the multi-task learning model still fails to meet the preset model judgment standard, the weight of each sample in the preset multi-task output head loss function is continuously adjusted, and the preset multi-task output head loss function is continuously updated using the adjusted weight of each sample in the preset multi-task output head loss function, and the multi-task learning model is continuously retrained using the updated preset multi-task output head loss function, the adjusted multi-task loss weight, the joint loss function and the updated first pathological image data set until the multi-task learning model meets the preset model judgment standard.

4. The method for analyzing pathological image data based on artificial intelligence according to claim 1, characterized in that: After sending the diagnosis result to the doctor's smart terminal device, the method further includes: Obtain the pathological image feature map output by the last convolutional layer of the multi-task learning model, and calculate the type information, pathological features, and gradient information of cancer markers relative to the pathological image feature map; Weighting the pathological image feature map according to the gradient information to obtain a weighted feature map, and normalizing the weighted feature map to obtain a thermal map; The thermal map is mapped into the pathological image to obtain a pathological image mapping map, the pixel value of each pixel in the pathological image mapping map is obtained, and the area composed of pixels whose pixel values ​​are higher than a preset pixel threshold is determined as a key focus area, and the key focus area and the pathological image mapping map are sent to the doctor's smart terminal device.

5. The method for analyzing pathological image data based on artificial intelligence according to claim 4, characterized in that: The pathological image feature map is weighted according to the gradient information to obtain a weighted feature map, including: The weight information of each channel in the feature map is determined according to the gradient information and a preset feature map weighting formula, and the pathological image feature map is weighted according to the weight information of each channel to obtain a weighted feature map.

6. The method for analyzing pathological image data based on artificial intelligence according to claim 5, characterized in that: The preset feature map weighting formula is: Among them, w i is the weight of channel i; H is the height of the feature map; W is the width of the feature map; Represents the gradient of the i-th channel in the feature map at position (x, y).

7. The method for analyzing pathological image data based on artificial intelligence according to claim 1, characterized in that: After determining that the multi-task learning model training is completed, the method further includes: identifying in real time whether a preset model update time has been reached, reacquiring a second pathology image dataset containing multiple cancer types and markers after the preset model update time has been reached, annotating each pathology image in the second pathology image dataset with a cancer type and a multi-label feature, and updating the second pathology image dataset according to the annotated pathology images; wherein the multi-label feature annotation includes pathology feature annotation and cancer marker annotation; The multi-task learning model is retrained according to the joint loss function and the updated second pathological image data set. If the multi-task learning model meets the preset model judgment standard, it is determined that the training of the multi-task learning model is completed.

8. A pathological image data analysis system based on artificial intelligence, characterized in that: The system comprises: A data set creation module, used to obtain a first pathological image data set, annotate each pathological image in the first pathological image data set with a cancer type and a multi-label feature, and update the first pathological image data set according to the annotated pathological images; wherein the multi-label feature annotation includes pathological feature annotation and cancer marker annotation; A multi-task learning model creation module, used to create a type recognition output head and a multi-task output head, and create a multi-task learning model based on the type recognition output head and the multi-task output head; wherein the type recognition output head refers to the part of the multi-task learning model used to identify the cancer type in the pathological image, which is the fully connected layer of the last few layers, and the output is the probability distribution of different cancer types; the multi-task output head refers to the output part of the multi-task learning model used to process multiple related tasks, and these related tasks include the prediction of multiple pathological features; A joint loss function creation module is used to create a joint loss function based on a cross entropy loss function and a preset multi-task output head loss function; wherein the preset multi-task output head loss function is: in, is the multi-task output head loss function, used to measure the prediction value and the actual label y i The difference between them; i is the sample index, which indicates the i-th sample in the data set; C is the number of samples, which indicates the total number of samples in the data set; w i is the weight of the i-th sample, which is used to adjust the loss contribution of each sample; y i is the true label, indicating the actual category of the i-th sample; is the predicted probability; (1-y i ) is the true label y i The complement of i =1, then (1-yi)=0; A model training module is used to train a multi-task learning model according to the joint loss function and the updated first pathological image data set, and if the multi-task learning model meets the preset model judgment standard, it is determined that the training of the multi-task learning model is completed; wherein the joint loss function is: Among them, TotalLoss is the total loss, which represents the goal optimized by the model during training, and is composed of type loss and multi-task loss; λ1 is the weight of type loss, which is used to adjust the contribution of type recognition to the total loss; Type Loss is the type loss, which is the cross entropy loss function in this scheme; λ2 is the weight of multi-task loss, which is used to adjust the contribution of multi-task learning to the total loss; is the preset multi-task output head loss function; The diagnosis result determination module is used to identify the type information, pathological characteristics and cancer markers of the pathological image according to the multi-task learning model, determine the diagnosis result according to the type information, pathological characteristics and cancer markers, and send the diagnosis result to the doctor's smart terminal device.

Images