Internet feedback mining method and system applied to animal remote diagnosis platform

Abstract

The embodiment of the application provides an internet feedback mining method and system applied to an animal remote diagnosis platform, realizes effective utilization of internet remote case feedback data uploaded to the animal remote diagnosis platform, and can generate accurate animal case diagnosis decision results by loading the feedback data into a specially trained animal case diagnosis decision network. The animal case diagnosis decision network combines diversified training case feedback data and corresponding case diagnosis heat maps for joint model learning, significantly improving the accuracy and reliability of diagnosis. In particular, by distinguishing and updating the training case feedback data and the corresponding heat maps, the learning effect of the animal case diagnosis decision network is further optimized. Ultimately, based on the generated diagnosis decision results, relevant information can be timely pushed to the animal remote diagnosis platform, greatly improving the efficiency and practicality of remote diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, in particular to an Internet feedback mining method and system applied to an animal remote diagnosis platform. BACKGROUND

[0002] With the popularity of the Internet and the development of telemedicine technology, animal remote diagnosis platforms have gradually become an important tool for breeders and veterinarians to diagnose animal diseases. These platforms collect and organize animal case data from various places, providing veterinarians with rich case resources, which helps them more accurately diagnose diseases and develop treatment plans. However, in actual application, due to the complexity and diversity of animal cases, relying solely on traditional case data analysis and diagnosis methods often fails to meet actual needs.

[0003] In order to improve the accuracy and efficiency of animal remote diagnosis, people have begun to apply machine learning and deep learning technologies to animal case diagnosis. These technologies can learn and analyze a large amount of case data to mine potential disease characteristics and patterns, thereby providing veterinarians with more accurate diagnostic support. However, existing animal case diagnosis methods often have some problems, such as insufficient model training data, low diagnostic accuracy, and insufficient ability to handle complex cases. SUMMARY

[0004] Therefore, the purpose of the present application is to provide an Internet feedback mining method and system applied to an animal remote diagnosis platform.

[0005] According to the first aspect of the present application, an Internet feedback mining method applied to an animal remote diagnosis platform is provided, the method comprising:

[0006] obtaining Internet remote case feedback data uploaded by the animal remote diagnosis platform;

[0007] loading the Internet remote case feedback data into an animal case diagnosis decision network to generate an animal case diagnosis decision result of the Internet remote case feedback data, wherein the animal case diagnosis decision network is generated based on joint model learning of first training case feedback data, a first sample case diagnosis heat map corresponding to the first training case feedback data, second training case feedback data, and a second sample case diagnosis heat map corresponding to the second training case feedback data, the first training case feedback data and the second training case feedback data are distinguished based on an initial network learning result, the second sample case diagnosis heat map is updated based on the initial network learning result, and the initial network learning result is generated by joint model learning of basic training case feedback data of the animal case diagnosis decision network.

[0008] push information to the animal remote diagnosis platform based on the animal case diagnosis decision result of the internet remote case feedback data.

[0009] In a possible implementation of the first aspect, before the internet remote case feedback data uploaded by the animal remote diagnosis platform is acquired, the method further includes:

[0010] determining a basic training case feedback data sequence of the animal case diagnosis decision network, wherein the basic training case feedback data sequence includes a plurality of basic training case feedback data, and each basic training case feedback data has a corresponding sample case diagnosis heat map;

[0011] loading the plurality of basic training case feedback data into the animal case diagnosis decision network based on the sample case diagnosis heat map, to generate an initial network learning result;

[0012] determining the initial network learning result corresponding to each basic training case feedback data, and determining an average learning result corresponding to the initial network learning result;

[0013] distinguishing a first training case feedback data and a second training case feedback data from the plurality of basic training case feedback data according to the average learning result;

[0014] taking the sample case diagnosis heat map corresponding to the first training case feedback data as a first sample case diagnosis heat map, and updating the sample case diagnosis heat map corresponding to the second training case feedback data as a second sample case diagnosis heat map;

[0015] performing network parameter learning on the animal case diagnosis decision network according to the first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data, and the second sample case diagnosis heat map, to generate an animal case diagnosis decision network with completed network parameter learning.

[0016] In a possible implementation of the first aspect, the initial network learning result is a network learning error parameter, and the average learning result is a network learning average error parameter.

[0017] Correspondingly, the distinguishing a first training case feedback data and a second training case feedback data from the plurality of basic training case feedback data according to the average learning result includes:

[0018] determining the network learning error parameter corresponding to each basic training case feedback data;

[0019] the basic training case feedback data with the network learning error parameter being not less than the network learning average error parameter is determined as first training case feedback data;

[0020] the basic training case feedback data with the network learning error parameter being less than the network learning average error parameter is determined as second training case feedback data.

[0021] In a possible implementation of the first aspect, the updating of the sample case diagnosis heat map corresponding to the second training case feedback data to a second sample case diagnosis heat map comprises:

[0022] determining the sample case diagnosis heat map corresponding to the second training case feedback data;

[0023] updating the sample case diagnosis heat map corresponding to the second training case feedback data according to a preset heat value updating strategy to generate a second sample case diagnosis heat map, wherein the heat value of the sample case diagnosis heat map corresponding to the second training case feedback data is greater than the second sample case diagnosis heat map.

[0024] In a possible implementation of the first aspect, the updating of the sample case diagnosis heat map corresponding to the second training case feedback data according to a preset heat value updating strategy to generate a second sample case diagnosis heat map comprises:

[0025] updating the sample case diagnosis heat map corresponding to the second training case feedback data to a preset first case diagnosis heat map, and taking the preset first case diagnosis heat map as the second sample case diagnosis heat map; or

[0026] calculating an updated sample case diagnosis heat map according to the sample case diagnosis heat map corresponding to the second training case feedback data and the initial network learning result corresponding to the second training case feedback data, and updating the sample case diagnosis heat map corresponding to the second training case feedback data by using the updated sample case diagnosis heat map to generate a second sample case diagnosis heat map.

[0027] In a possible implementation of the first aspect, the network parameter learning of the animal case diagnosis decision network according to the first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data, and the second sample case diagnosis heat map to generate an animal case diagnosis decision network with completed network parameter learning comprises:

[0028] select first target training case feedback data from the first training case feedback data according to the first sample case diagnosis heat map, and select second target training case feedback data from the second training case feedback data according to the second sample case diagnosis heat map;

[0029] perform network parameter learning on the animal case diagnosis decision network according to the first target training case feedback data and the second target training case feedback data, and generate a network parameter learning result;

[0030] continue to train the animal case diagnosis decision network by iteratively performing the steps of determining the first training case feedback data and the second training case feedback data from the plurality of basic training case feedback data, and performing the steps of updating the first sample case diagnosis heat map corresponding to the first training case feedback data and updating the second sample case diagnosis heat map corresponding to the second training case feedback data, until the animal case diagnosis decision network converges, and generate a completed network parameter learning animal case diagnosis decision network.

[0031] In a possible implementation of the first aspect, the method further includes:

[0032] determine network learning weights corresponding to the second training case feedback data, and update the network learning weights based on the second sample case diagnosis heat map to generate updated network learning weights;

[0033] The network parameter learning on the animal case diagnosis decision network according to the first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data, and the second sample case diagnosis heat map to generate a completed network parameter learning animal case diagnosis decision network includes:

[0034] perform network parameter learning on the animal case diagnosis decision network according to the first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data, the second sample case diagnosis heat map, and the updated network learning weights to generate a completed network parameter learning animal case diagnosis decision network.

[0035] In a possible implementation of the first aspect, the network parameter learning on the animal case diagnosis decision network according to the first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data, and the second sample case diagnosis heat map to generate a completed network parameter learning animal case diagnosis decision network includes:

[0036] determine a current network learning progress of the animal case diagnosis decision network;

[0037] update the first sample case diagnosis heat map to a preset second case diagnosis heat map to obtain an updated first sample case diagnosis heat map when the current network learning progress reaches a set progress; and

[0038] update the second sample case diagnosis heat map to the preset second case diagnosis heat map to obtain an updated second sample case diagnosis heat map;

[0039] perform network parameter learning on the animal case diagnosis decision network according to the first training case feedback data, the updated first sample case diagnosis heat map, the second training case feedback data, and the updated second sample case diagnosis heat map, to generate an animal case diagnosis decision network with completed network parameter learning.

[0040] In a possible implementation of the first aspect, the loading of the internet remote case feedback data into the animal case diagnosis decision network and the generation of the animal case diagnosis decision result of the internet remote case feedback data include:

[0041] determining an initialization neural network corresponding to the animal case diagnosis decision network;

[0042] loading the internet remote case feedback data into the initialization neural network to generate an initialization diagnosis result of the internet remote case feedback data;

[0043] loading the initialization diagnosis result into the animal case diagnosis decision network to generate the animal case diagnosis decision result of the internet remote case feedback data.

[0044] According to a second aspect of the present application, an internet feedback mining system applied to an animal remote diagnosis platform is provided, which comprises a processor and a readable storage medium, and the readable storage medium stores a program which, when executed by the processor, implements the aforementioned internet feedback mining method applied to the animal remote diagnosis platform.

[0045] According to a third aspect of the present application, a computer readable storage medium is provided, and the computer readable storage medium stores computer executable instructions which, when executed, implement the aforementioned internet feedback mining method applied to the animal remote diagnosis platform.

[0046] Based on any of the above aspects, embodiments of this application achieve effective utilization of remote case feedback data uploaded to an animal remote diagnosis platform. By loading the feedback data into a specially trained animal case diagnosis decision network, accurate animal case diagnosis decision results can be generated. The animal case diagnosis decision network combines diverse training case feedback data and their corresponding case diagnosis heatmaps for joint model learning, significantly improving the accuracy and reliability of diagnosis. In particular, by distinguishing and updating the training case feedback data and their corresponding heatmaps, the learning effect of the animal case diagnosis decision network is further optimized. Finally, based on the generated diagnostic decision results, relevant information can be promptly pushed to the animal remote diagnosis platform, greatly improving the efficiency and practicality of remote diagnosis. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A flowchart illustrating the Internet feedback mining method for animal remote diagnostic platforms provided in this application embodiment is shown.

[0049] Figure 2 This illustration shows a schematic diagram of the component structure of an Internet feedback mining system for an animal remote diagnostic platform, which is used to implement the above-described Internet feedback mining method for an animal remote diagnostic platform, according to an embodiment of this application. Detailed Implementation

[0050] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.

[0051] It is to be understood that the terms "first", "second", and the like used in the specification and claims of the application and the above description of the drawings do not necessarily have an ordinal or chronological significance. It is to be understood that where the application is indicated to include a process, method, system, product or apparatus involving steps or units, the steps or units, as appropriate, can be implemented in hardware, software, or a combination of both. It is to be understood that the terms "including", "comprising", and variations thereof, are intended to cover all possible combinations of the listed steps or units, and that the use of "consisting of" or "consisting essentially of" is intended to cover only those combinations of the listed steps or units that do not involve additional steps or units.

[0052] Figure 1 The flowchart of the internet feedback mining method applied to the animal remote diagnosis platform provided by the embodiments of the application is shown. It should be understood that in other embodiments, the order of some steps of the internet feedback mining method applied to the animal remote diagnosis platform can be exchanged according to actual needs, or some steps can be omitted or deleted. The detailed steps of the internet feedback mining method applied to the animal remote diagnosis platform are introduced as follows.

[0053] In step S110, internet remote case feedback data uploaded by the animal remote diagnosis platform is acquired.

[0054] In this embodiment, the animal remote diagnosis platform can provide remote diagnosis services for many animal medical institutions, breeders, and the like. In this scenario, the animal remote diagnosis platform will receive animal case related information from all over the country every day. These information includes basic information of animals (such as species, age, gender, etc.), symptom performance (for example, whether the animal has fever, cough, diarrhea, and other symptoms, the duration and severity of the symptoms, etc.), past medical history (whether it has suffered from certain diseases before, what kind of treatment has been received, etc.), and possible environmental factors (whether the living environment has changed, whether it has been in contact with other sick animals, etc.).

[0055] When various clients (such as computer terminals of animal hospitals, mobile devices of breeders, etc.) input these case information into the animal remote diagnosis platform, the platform will preliminarily arrange and format the data, and convert it into a format that can be recognized and processed by the server. For example, the basic information of the animal is classified and stored according to specific fields, and the symptom performance is represented in a specific coding manner to represent different symptom types and severity.

[0056] Then, the animal remote diagnosis platform uploads the sorted internet remote case feedback data to the server. The server end sets a special receiving module, which always listens to the connection request from the animal remote diagnosis platform. Once a new data upload request is detected, the receiving module will receive data according to the pre-set communication protocol. For example, if the HTTP protocol is used, the receiving module will parse the information in the request header according to the HTTP protocol specification, obtain the length, type and other metadata of the data, then gradually receive the data content, and store the received data in the temporary buffer area of the server. After receiving, the data is checked for integrity and accuracy to ensure that the data is not damaged or lost during transmission. If the check passes, the server will store these internet remote case feedback data in a special database for subsequent processing.

[0057] Step S120, loading the internet remote case feedback data to the animal case diagnosis decision network to generate the animal case diagnosis decision result of the internet remote case feedback data, wherein the animal case diagnosis decision network is generated based on joint model learning of the first training case feedback data, the first sample case diagnosis heat map corresponding to the first training case feedback data, the second training case feedback data, and the second sample case diagnosis heat map corresponding to the second training case feedback data, the first training case feedback data and the second training case feedback data are distinguished based on the initial network learning result, the second sample case diagnosis heat map is updated based on the initial network learning result, and the initial network learning result is generated by joint model learning of the basic training case feedback data of the animal case diagnosis decision network.

[0058] In this embodiment, the server needs to determine the basic training case feedback data sequence before constructing the animal case diagnosis decision network. For example, the server collects a large amount of animal case data from animal medical research institutions, past animal case databases and other data sources. It is assumed that 1000 basic training case feedback data of different animal cases are collected. For each case feedback data, there is a corresponding sample case diagnosis heat map. This sample case diagnosis heat map is generated by professional animal medical experts according to various symptoms, examination results and other information in the case to label the possible disease diagnosis results. For example, for an animal case with pneumonia, the symptom area related to pneumonia (such as the feature area corresponding to fever, cough, and rapid breathing) in the sample case diagnosis heat map will be marked as a higher heat value, indicating that these features have a higher correlation with pneumonia diagnosis.

[0059] The server loads these multiple basic training case feedback data into an animal case diagnosis decision network (assuming this is a deep learning-based neural network model, such as a convolutional neural network CNN). During the loading process, each feature of each case feedback data (such as the basic information of the animal, the symptom performance, etc.) is mapped to the input layer node of the neural network. Then, the neural network performs forward propagation calculation according to the pre-set network structure (including multiple hidden layers, each hidden layer contains a number of neurons, and the neurons are connected through different weights). During the calculation process, each neuron will perform linear transformation (such as calculating the weighted sum) on the input data, and then perform non-linear transformation through an activation function (such as the ReLU function) to obtain the output result. This output result is compared with the diagnosis result in the corresponding sample case diagnosis heat map to calculate the network learning error parameter (initial network learning result). For example, for a certain case, the disease result predicted by the neural network is different from the result in the sample case diagnosis heat map labeled by the expert, and this difference is quantified as the network learning error parameter.

[0060] After the server determines the network learning error parameter corresponding to each basic training case feedback data, it calculates the average value of these error parameters to obtain the network learning average error parameter.

[0061] For the network learning error parameter corresponding to each basic training case feedback data, the server compares it with the network learning average error parameter. For example, for 500 basic training case feedback data, if their network learning error parameters are not less than the network learning average error parameter, the server determines these data as the first training case feedback data. These first training case feedback data may contain some relatively complex or special case situations, which makes it difficult for the neural network to accurately predict during the initial learning process. For another 500 basic training case feedback data whose network learning error parameters are less than the network learning average error parameter, the server determines them as the second training case feedback data. These second training case feedback data are relatively easy for the neural network to handle during the initial learning.

[0062] For each case determined as the second training case feedback data, the server first determines its corresponding sample case diagnosis heat map. Then, it is updated according to a preset heat value update strategy. For example, one update strategy is to update the sample case diagnosis heat map corresponding to the second training case feedback data into a preset first case diagnosis heat map. Suppose that the preset first case diagnosis heat map is an optimized standard heat map, which has more explicit annotations for the key features of disease diagnosis. In this case, for a case originally about animal digestive system diseases, the original sample case diagnosis heat map may have a lower annotation heat value for some symptoms (such as mild diarrhea), and the preset first case diagnosis heat map will increase the heat value corresponding to the symptom of diarrhea according to more medical experience and data statistics, to more accurately reflect the importance of the symptom in the diagnosis of digestive system diseases.

[0063] Another update strategy is to obtain an updated sample case diagnosis heat map by calculating the sample case diagnosis heat map corresponding to the second training case feedback data and its corresponding initial network learning result. For example, the server analyzes which features the neural network is prone to misjudge according to the error information in the initial network learning result. If it is found that for a certain second training case feedback data, the neural network overemphasizes the weight change of the animal and ignores the fecal egg situation when judging whether the animal is infected with parasites. Then when updating the sample case diagnosis heat map, the heat value corresponding to the weight change feature will be appropriately reduced, and the heat value corresponding to the fecal egg situation will be appropriately increased, and then the original sample case diagnosis heat map corresponding to the second training case feedback data is updated using the updated sample case diagnosis heat map to generate a second sample case diagnosis heat map.

[0064] The server selects first target training case feedback data from the first training case feedback data according to the first sample case diagnosis heat map, and selects second target training case feedback data from the second training case feedback data according to the second sample case diagnosis heat map. For example, in the first training case feedback data, according to the annotations of disease features in the first sample case diagnosis heat map, those cases with typical features (such as for a certain infectious disease, selecting cases with typical symptom combinations) are selected as the first target training case feedback data. Similarly, in the second training case feedback data, according to the updated second sample case diagnosis heat map, the cases that best reflect the optimized diagnosis features are selected as the second target training case feedback data.

[0065] The server performs network parameter learning on the animal case diagnosis decision network according to the first target training case feedback data and the second target training case feedback data. In this process, a back propagation algorithm is adopted. First, the first target training case feedback data and the second target training case feedback data are input into the animal case diagnosis decision network for forward propagation calculation to obtain a prediction result. Then, according to the error between the prediction result and the actual diagnosis result (from the first sample case diagnosis heat map and the second sample case diagnosis heat map), the gradient of each neuron connection weight is calculated. For example, if the output result of a certain neuron leads to a large difference between the probability of predicting the disease and the probability in the actual diagnosis result, then the gradient of the weight of the neuron connected to the next layer of neurons will be larger. Next, according to the calculated gradient, the weight parameters in the network are updated according to a certain learning rate (such as 0.01), and this process is iterated until a certain stopping condition is reached (such as the number of iterations reaching a set value or the error converging to a certain range), and the network parameter learning result is generated.

[0066] Using this network parameter learning result, the server iteratively performs the steps of determining the first training case feedback data and the second training case feedback data from the plurality of basic training case feedback data, and performing the steps of updating the first sample case diagnosis heat map corresponding to the first training case feedback data and updating the second sample case diagnosis heat map corresponding to the second training case feedback data, to continue training the animal case diagnosis decision network. For example, in each iteration process, the network learning error parameters corresponding to each basic training case feedback data are recalculated, the first training case feedback data and the second training case feedback data are reclassified, and then the first sample case diagnosis heat map and the second sample case diagnosis heat map are updated according to the new classification result, and the next round of network parameter learning is performed. This process is repeated until the animal case diagnosis decision network converges, and the completed network parameter learning animal case diagnosis decision network is generated.

[0067] In the process of network learning, the server also determines the current network learning progress of the animal case diagnosis decision network. When the current network learning progress reaches a set progress, for example, when the learning progress reaches 80%, the first sample case diagnosis heat map is updated to a preset second case diagnosis heat map to obtain an updated first sample case diagnosis heat map. At the same time, the second sample case diagnosis heat map is also updated to the preset second case diagnosis heat map to obtain an updated second sample case diagnosis heat map. This preset second case diagnosis heat map can be a more comprehensive and accurate heat map that integrates more medical knowledge and the latest research results. Then, the animal case diagnosis decision network is subjected to network parameter learning based on the first training case feedback data, the updated first sample case diagnosis heat map, the second training case feedback data, and the updated second sample case diagnosis heat map, to generate an animal case diagnosis decision network that has completed network parameter learning.

[0068] When the server receives the internet remote case feedback data, it first determines the initialization neural network corresponding to the animal case diagnosis decision network. This initialization neural network is a neural network with certain initial weights and structures obtained after completing network parameter learning. Then, the internet remote case feedback data is loaded into this initialization neural network. For example, the animal basic information, symptom performance, and other data in the case are mapped according to the input format of the neural network. In the initialization neural network, the data will be calculated through multiple hidden layers, and each neuron will process the data according to its own weights and activation functions, and finally obtain an initialization diagnosis result. This initialization diagnosis result can be a preliminary prediction result about the diseases that the animal may have, for example, for a certain animal case, the initialization diagnosis result can be a probability distribution of the animal having a certain disease, such as a probability of 30% of having pneumonia and a probability of 20% of having other respiratory diseases.

[0069] Then, the initialization diagnosis result is loaded into the animal case diagnosis decision network. The animal case diagnosis decision network will further analyze and adjust the initialization diagnosis result based on the knowledge and model parameters learned before. For example, considering some special symptoms or environmental factors in the animal case, the previous probability distribution is corrected, and finally the animal case diagnosis decision result of the internet remote case feedback data is generated. This result can be a clear disease diagnosis conclusion, such as the animal having pneumonia, and can also include some related suggestions, such as suggesting further examination (such as X-ray examination, blood examination, etc.) or treatment plan (such as using specific drugs for treatment, etc.).

[0070] In step S130, the animal case diagnosis decision result based on the internet remote case feedback data is pushed to the animal remote diagnosis platform.

[0071] In this embodiment, after obtaining the animal case diagnosis decision result of the Internet remote case feedback data, the server needs to push the relevant information to the animal remote diagnosis platform. For example, if the diagnosis result shows that the animal has a certain disease, the server will arrange the information such as the disease name, the basis for diagnosis (such as which symptoms and examination results support this diagnosis), and the recommended treatment plan. Then, according to the communication protocol between the animal remote diagnosis platform, a connection is established and the information is sent to the animal remote diagnosis platform.

[0072] After receiving the information pushed by the server, the animal remote diagnosis platform can display these information to the client who submits the case (such as the doctor of the animal hospital or the breeder). For example, on the terminal interface of the animal hospital, the doctor can see the detailed diagnosis result and treatment suggestion, so as to provide more accurate medical service for the animal. For the breeder, they can take corresponding measures according to these information, such as isolating and treating the sick animals, adjusting the breeding environment, etc.

[0073] At the same time, the server may also push information in different ways according to the urgency or special circumstances of the animal case. For example, for some urgent diseases (such as acute infectious diseases), the server will set special identification in the pushed information, so that the animal remote diagnosis platform can timely remind the client to pay attention to the emergency treatment. In addition, if it is a relatively rare disease or a newly discovered disease, the server may attach some related research materials or reference cases in the pushed information, to help the client better understand and handle such cases.

[0074] Based on the above steps, the embodiment of the present application realizes the effective utilization of the Internet remote case feedback data uploaded to the animal remote diagnosis platform. By loading the feedback data into the animal case diagnosis decision network specially trained, accurate animal case diagnosis decision results can be generated. The animal case diagnosis decision network combines diversified training case feedback data and its corresponding case diagnosis heat map for joint model learning, which significantly improves the accuracy and reliability of diagnosis. In particular, by distinguishing and updating the training case feedback data and the corresponding heat map, the learning effect of the animal case diagnosis decision network is further optimized. Finally, based on the generated diagnosis decision result, relevant information can be timely pushed to the animal remote diagnosis platform, greatly improving the efficiency and practicality of remote diagnosis.

[0075] In a possible implementation, before step S110, the method further includes:

[0076] Step S101, determine a basic training case feedback data sequence of the animal case diagnosis decision network, wherein the basic training case feedback data sequence contains a plurality of basic training case feedback data, and each basic training case feedback data has a corresponding sample case diagnosis heat map.

[0077] Step S102, load the plurality of basic training case feedback data into the animal case diagnosis decision network based on the sample case diagnosis heat map, and generate an initial network learning result.

[0078] Step S103, determine the initial network learning result corresponding to each basic training case feedback data, and determine an average learning result corresponding to the initial network learning result.

[0079] Step S104, according to the average learning result, distinguish first training case feedback data and second training case feedback data from the plurality of basic training case feedback data.

[0080] Step S105, take the sample case diagnosis heat map corresponding to the first training case feedback data as a first sample case diagnosis heat map, and update the sample case diagnosis heat map corresponding to the second training case feedback data as a second sample case diagnosis heat map.

[0081] Step S106, according to the first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data, and the second sample case diagnosis heat map, perform network parameter learning on the animal case diagnosis decision network, and generate an animal case diagnosis decision network with completed network parameter learning.

[0082] In this embodiment, first, the basic training case feedback data sequence of the animal case diagnosis decision network is determined. In the storage system of the server, there are a large number of animal case related data collected from different channels, which together constitute the basic training case feedback data sequence. For example, case records in the past years are obtained from the case archives of animal hospitals in various places, which record the disease information of different types of animals (such as cats, dogs, cows, sheep, etc.) in detail. At the same time, some special cases or valuable case data are also obtained from animal medical research institutions. Each of these data is a basic training case feedback data, and each basic training case feedback data has a corresponding sample case diagnosis heat map. This sample case diagnosis heat map is carefully drawn by a team of professional animal medical experts according to the detailed information in the case. Taking a dog's pneumonia case as an example, the sample case diagnosis heat map will mark the areas corresponding to typical symptoms such as cough, fever, and rapid breathing as higher heat values, which reflect the degree of association between these symptoms and the diagnosis of pneumonia. Other symptoms that may exist but have weak association with pneumonia are marked as lower heat values.

[0083] Then, based on the sample case diagnosis heat map, the multiple basic training case feedback data are loaded into the animal case diagnosis decision network. This animal case diagnosis decision network is a complex neural network structure, which may include an input layer, multiple hidden layers, and an output layer. When loading the basic training case feedback data, various feature information in the case (such as animal species, age, symptom performance, etc.) is mapped to the input layer nodes of the network. For example, the age information of the dog is input to the corresponding input node according to a certain numerical encoding method, and the symptom performance may be converted into a combination of numerical values of multiple input nodes through a specific encoding method. Then, the animal case diagnosis decision network performs a series of mathematical operations in the hidden layer according to the pre-set connection weights and activation functions, and finally obtains a result in the output layer. This result is compared with the diagnosis result in the sample case diagnosis heat map, and the difference between the two is calculated, which is the initial network learning result. For example, in a case of a cat's urinary system disease, the disease possibility distribution output by the animal case diagnosis decision network deviates from the standard diagnosis result in the sample case diagnosis heat map, and the quantitative value of this deviation is the initial network learning result for this case.

[0084] Afterwards, the server needs to determine the initial network learning result corresponding to each basic training case feedback data, and determine the average learning result corresponding to the initial network learning result. For each basic training case feedback data, a specific initial network learning result can be obtained through the above calculation process. Then, all these initial network learning results are summed and divided by the total number of basic training case feedback data to obtain the average learning result. For example, if there are 100 basic training case feedback data, each data has its own initial network learning result, and the average learning result is obtained by adding the 100 results and dividing by 100.

[0085] According to this average learning result, the first training case feedback data and the second training case feedback data are distinguished from the plurality of basic training case feedback data. For the initial network learning result corresponding to each basic training case feedback data, it is compared with the average learning result. The basic training case feedback data whose initial network learning result is not less than the average learning result is determined as the first training case feedback data. These first training case feedback data are often cases that the network is difficult to accurately process at the initial learning. For example, some cases with complex symptom combinations or rare disease manifestations may have larger initial network learning results, so they are classified as first training case feedback data. And the basic training case feedback data whose initial network learning result is less than the average learning result is determined as the second training case feedback data, and these cases are relatively better handled by the network at the initial learning.

[0086] Then, the sample case diagnosis heat map corresponding to the first training case feedback data is taken as the first sample case diagnosis heat map, and the sample case diagnosis heat map corresponding to the second training case feedback data is updated to obtain the second sample case diagnosis heat map. Taking a complex disease case of a cow in the first training case feedback data as an example, its corresponding sample case diagnosis heat map is directly set as the first sample case diagnosis heat map, which will be used as an important reference for network training in the future. For a sheep's ordinary disease case in the second training case feedback data, assuming that the original sample case diagnosis heat map is not accurate or optimized enough in marking the heat value of some symptoms, it is updated according to the preset rules. For example, the heat value of the sheep's mild diarrhea symptom in the original heat map is low in the diagnosis of this disease. According to more medical experience and data analysis, it is found that this symptom actually has more important significance in the diagnosis of this disease, so the heat value corresponding to this symptom is increased to obtain the updated second sample case diagnosis heat map.

[0087] Finally, the network parameter learning of the animal case diagnosis decision network is performed according to the first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data and the second sample case diagnosis heat map, and the animal case diagnosis decision network with completed network parameter learning is generated. From the first training case feedback data, according to the annotation of the first sample case diagnosis heat map, those cases with representative features are selected as the first target training case feedback data for network parameter learning. Similarly, from the second training case feedback data, appropriate cases are selected as the second target training case feedback data according to the second sample case diagnosis heat map. Then, the first target training case feedback data and the second target training case feedback data are input into the animal case diagnosis decision network. The network performs forward propagation calculation according to these input data to obtain the prediction result. Then, the prediction result is compared with the actual diagnosis result in the first sample case diagnosis heat map and the second sample case diagnosis heat map, and the error is calculated. According to the error, the gradient of each connection weight in the network is calculated using the back propagation algorithm, and the weight parameter is updated according to a certain learning rate. This process is repeated constantly, and the performance of the network is re-evaluated after each update of the weight parameter. In the network parameter learning process, some other factors are also considered, such as the regularization method to prevent overfitting. With the increase of the number of iterations, the error between the prediction result of the animal case diagnosis decision network and the actual diagnosis result gradually decreases, and when the error reaches an acceptable range or the performance of the network no longer improves significantly, it is considered that the network converges, and the animal case diagnosis decision network with completed network parameter learning is generated at this time. This network will be able to accurately diagnose and decide on subsequent Internet remote case feedback data.

[0088] In a possible implementation, the initial network learning result is a network learning error parameter, and the average learning result is a network learning average error parameter.

[0089] Step S104 includes:

[0090] Step S1041, determining the network learning error parameter corresponding to each basic training case feedback data.

[0091] Step S1042, determining the basic training case feedback data with the network learning error parameter not less than the network learning average error parameter as the first training case feedback data.

[0092] Step S1043, determining the basic training case feedback data with the network learning error parameter less than the network learning average error parameter as the second training case feedback data.

[0093] In a possible implementation, step S105 includes:

[0094] Step S1051, determine the sample case diagnosis heat map corresponding to the second training case feedback data.

[0095] Step S1052, update the sample case diagnosis heat map corresponding to the second training case feedback data according to the preset heat value update strategy, generate a second sample case diagnosis heat map, wherein the heat value of the sample case diagnosis heat map corresponding to the second training case feedback data is greater than the second sample case diagnosis heat map.

[0096] In this embodiment, in the step of determining the network learning error parameter corresponding to each basic training case feedback data, the server will process each basic training case feedback data separately. For each basic training case feedback data, when it is loaded into the animal case diagnosis decision network and calculated, a result compared with the sample case diagnosis heat map will be obtained, which is the network learning error parameter. Taking the cat's gastrointestinal disease in the animal case as an example, the server inputs the basic training case feedback data containing the cat's age, symptoms (such as vomiting frequency, stool state, etc.), medical history and other information into the animal case diagnosis decision network. The network calculates according to its own structure and parameters to obtain a prediction result about disease diagnosis. This prediction result is compared with the standard diagnosis result represented by the sample case diagnosis heat map drawn by animal medical experts. Assuming that the sample case diagnosis heat map shows that the probability of the cat suffering from a certain gastrointestinal disease is 80%, while the network prediction probability is 60%, the difference value between the two is calculated by a specific error calculation method (such as mean square error calculation method), which is the network learning error parameter corresponding to this basic training case feedback data.

[0097] Then, the first training case feedback data and the second training case feedback data are distinguished according to the network learning average error parameter. After obtaining the network learning error parameters corresponding to all the basic training case feedback data, the server calculates the network learning average error parameter. For example, assuming that there are 1000 basic training case feedback data, each data has its own network learning error parameter, and the network learning average error parameter is obtained by adding the 1000 error parameters and dividing by 1000. Then, for the network learning error parameter corresponding to each basic training case feedback data, it is compared with the network learning average error parameter. The basic training case feedback data whose network learning error parameter is not less than the network learning average error parameter is determined as the first training case feedback data. For example, in some complex animal disease cases, such as cases of dogs with multiple complications, due to the complex relationship between symptoms, it is difficult for the animal case diagnosis decision network to accurately grasp in the learning process, and its network learning error parameter is often large, not less than the network learning average error parameter, so the basic training case feedback data of such cases is determined as the first training case feedback data. The basic training case feedback data whose network learning error parameter is less than the network learning average error parameter is determined as the second training case feedback data. For example, some animal disease cases with relatively single and clear symptoms, such as simple rabbit skin infection cases, the animal case diagnosis decision network can better learn and predict, and its network learning error parameter is small, so it is determined as the second training case feedback data.

[0098] Then, for the operation of updating the sample case diagnosis heat map corresponding to the second training case feedback data to the second sample case diagnosis heat map, the sample case diagnosis heat map corresponding to the second training case feedback data is first determined. Taking a second training case feedback data of a chicken respiratory disease as an example, the sample case diagnosis heat map corresponding thereto is drawn by an animal medical expert according to the symptoms of the chicken (such as cough, sneezing, respiratory sound, etc.), the growth environment of the chicken, whether it has been in contact with the source of the disease, etc. In this sample case diagnosis heat map, different symptoms and factors are marked with different heat values according to their importance to disease diagnosis. For example, the cough symptom may be marked with a higher heat value, and the chicken growth environment factor is marked with a relatively lower heat value.

[0099] Then, the sample case diagnosis heat map is updated according to a preset heat value updating strategy, to generate a second sample case diagnosis heat map. The preset heat value updating strategy is formulated based on a large amount of animal case data statistical analysis and the latest animal medical research results. For example, according to more chicken respiratory disease case studies, it is found that the ventilation condition in the growth environment of the chicken has a more important significance for the diagnosis of the disease than previously thought. In the original sample case diagnosis heat map, the heat value corresponding to the ventilation condition factor is low. According to the preset heat value updating strategy, the heat value corresponding to the ventilation condition factor will be appropriately increased, and the heat values of other factors may be fine-tuned according to the relationship between the factors. After such updating operation, the new sample case diagnosis heat map obtained is the second sample case diagnosis heat map, and the heat value sum of the updated second sample case diagnosis heat map is less than the heat value sum of the sample case diagnosis heat map corresponding to the second training case feedback data. This is because the updating strategy aims to more accurately allocate heat values, highlight factors that are really important to disease diagnosis, and remove some possible redundant or inaccurate heat value labels, so that the sample case diagnosis heat map can provide more accurate guidance information in the subsequent training of the animal case diagnosis decision network.

[0100] In a possible implementation, the step S1051 comprises:

[0101] updating the sample case diagnosis heat map corresponding to the second training case feedback data to a preset first case diagnosis heat map, and taking the preset first case diagnosis heat map as a second sample case diagnosis heat map. Or

[0102] According to the sample case diagnosis heat map corresponding to the second training case feedback data and the initial network learning result corresponding to the second training case feedback data, an updated sample case diagnosis heat map is calculated and obtained, and the sample case diagnosis heat map corresponding to the second training case feedback data is updated using the updated sample case diagnosis heat map, to generate a second sample case diagnosis heat map.

[0103] In the process of updating the sample case diagnosis heat map corresponding to the second training case feedback data according to the preset heat value updating strategy to generate the second sample case diagnosis heat map, there are two ways.

[0104] The first way is to update the sample case diagnosis heat map corresponding to the second training case feedback data to a preset first case diagnosis heat map, and then take the preset first case diagnosis heat map as the second sample case diagnosis heat map. Taking the second training case feedback data of a certain infectious disease of pigs as an example, the sample case diagnosis heat map corresponding thereto contains heat values corresponding to factors such as body temperature, skin symptoms, and diet of pigs. The preset first case diagnosis heat map is a standardized heat map obtained through a large amount of case data verification and expert experience summary. This preset first case diagnosis heat map has a more optimized factor weight distribution in the diagnosis of the infectious disease of pigs. For example, in the original sample case diagnosis heat map, the skin symptoms of pigs may be assigned a relatively low heat value due to early recognition limitations, but in the preset first case diagnosis heat map, based on more research, it is found that skin symptoms play a more critical role in the early diagnosis of this infectious disease, so they are assigned a higher heat value. When the sample case diagnosis heat map corresponding to the second training case feedback data of pigs is updated to the preset first case diagnosis heat map, it is equivalent to adopting a more standardized and optimized diagnosis reference mode. This new preset first case diagnosis heat map becomes the second sample case diagnosis heat map, which is used for subsequent training of the animal case diagnosis decision network.

[0105] The second way is to calculate an updated sample case diagnosis heat map according to the sample case diagnosis heat map corresponding to the second training case feedback data and the initial network learning result corresponding to the second training case feedback data, and update the sample case diagnosis heat map corresponding to the second training case feedback data using the updated sample case diagnosis heat map, thereby generating the second sample case diagnosis heat map. Still taking the infectious disease case of pigs as an example, suppose that in the original sample case diagnosis heat map, a higher heat value is assigned to the diet of pigs, but according to the initial network learning result, the animal case diagnosis decision network relies on the diet factor when learning this case, resulting in certain errors. This may be because in the actual case, although the diet has changed, it is not the core basis for diagnosing this infectious disease. Therefore, according to the error analysis in the initial network learning result, the weights of the diet and other factors in diagnosis are recalculated. For example, through a specific algorithm (such as an algorithm based on the error backpropagation principle to adjust the weights), the weight of the diet is reduced, and it may be found that the importance of the feces state of pigs in diagnosis is underestimated, and accordingly the weight of the feces state is increased, thereby obtaining the updated sample case diagnosis heat map. Then, the original sample case diagnosis heat map is updated using the updated sample case diagnosis heat map, and the second sample case diagnosis heat map is finally generated.

[0106] In a possible implementation, step S106 includes:

[0107] Step S1061, selecting first target training case feedback data from the first training case feedback data according to the first sample case diagnosis heat map, and selecting second target training case feedback data from the second training case feedback data according to the second sample case diagnosis heat map.

[0108] Step S1062, performing network parameter learning on the animal case diagnosis decision network according to the first target training case feedback data and the second target training case feedback data, and generating a network parameter learning result.

[0109] Step S1063, using the network parameter learning result, iteratively performing the steps of determining first training case feedback data and second training case feedback data from the plurality of basic training case feedback data, and performing the steps of updating the first sample case diagnosis heat map corresponding to the first training case feedback data and updating the second sample case diagnosis heat map corresponding to the second training case feedback data, and continuing training the animal case diagnosis decision network until the animal case diagnosis decision network converges, and generating a completed network parameter learning animal case diagnosis decision network.

[0110] In this embodiment, first, first target training case feedback data is selected from first training case feedback data according to a first sample case diagnosis heat map, and second target training case feedback data is selected from second training case feedback data according to a second sample case diagnosis heat map. Taking a complex case of a plurality of diseases of a dog as an example of the first training case feedback data, the first sample case diagnosis heat map details the importance of different symptoms and factors in diagnosing different diseases. According to this first sample case diagnosis heat map, those cases with typical symptom combinations or important features in disease diagnosis are selected as the first target training case feedback data. For example, in a mixed disease of heart disease and arthritis of a dog, those cases with typical heart function index abnormalities and joint swelling symptoms are selected. For the case of common diseases of sheep as the second training case feedback data, according to the second sample case diagnosis heat map, this heat map more accurately reflects the key factors of disease diagnosis after the previous update operation. Those cases that can reflect the updated diagnosis key factors are selected as the second target training case feedback data, such as those cases with typical digestive system disease symptoms of sheep and in line with the updated diagnosis factor weight.

[0111] Then, the animal case diagnosis decision network is trained according to the first target training case feedback data and the second target training case feedback data to generate a network parameter learning result. The first target training case feedback data and the second target training case feedback data are input into the animal case diagnosis decision network, which includes an input layer, multiple hidden layers, and an output layer. In the input layer, various information (such as animal species, age, symptoms, etc.) in the case is converted into numerical values recognizable by the network according to a predetermined encoding method. Then, in the hidden layer, neurons perform a series of linear and nonlinear transformations on the input data according to the pre-set weights and activation functions. For example, in a hidden layer, neurons perform weighted summation on the input data according to the connection weights with the previous layer, then perform nonlinear transformation through the ReLU activation function, and then pass the result to the next layer. Finally, the prediction result is obtained in the output layer. The prediction result is compared with the standard diagnosis result in the first example case diagnosis heat map and the second example case diagnosis heat map, and the error is calculated by a specific error calculation method (such as cross-entropy error calculation method). According to the error, the gradient of each connection weight in the network is calculated using the backpropagation algorithm, and then the weight parameters are updated according to a certain learning rate (such as 0.01). This process is repeated continuously, and the error is recalculated after each weight update until the error reaches an acceptable range or a predetermined number of iterations is reached, at which point the network parameter learning result is obtained.

[0112] Finally, based on the network parameter learning result, the steps of determining the first training case feedback data and the second training case feedback data from the plurality of basic training case feedback data and updating the first sample case diagnosis heat map corresponding to the first training case feedback data and the second sample case diagnosis heat map corresponding to the second training case feedback data are iteratively performed, and the animal case diagnosis decision network is continuously trained until the animal case diagnosis decision network converges, and the animal case diagnosis decision network with completed network parameter learning is generated. For example, after the first network parameter learning result is obtained, the initial network learning result (here, the network learning error parameter) corresponding to each basic training case feedback data is recalculated according to the new network parameters, and then the network learning average error parameter is recalculated. The first training case feedback data and the second training case feedback data are distinguished again according to the new network learning average error parameter. For the newly determined first training case feedback data, the first sample case diagnosis heat map corresponding thereto may need to be updated according to the new case situation and the network learning situation. For the newly determined second training case feedback data, the second sample case diagnosis heat map corresponding thereto is also updated according to the method mentioned above. Then, the network parameter learning is performed again according to the updated first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data and the second sample case diagnosis heat map. This iterative process is repeated continuously, and with the increase of the iteration number, the error between the prediction result of the animal case diagnosis decision network and the standard diagnosis result becomes smaller and smaller. When the error no longer significantly decreases or reaches a preset minimum value, it is considered that the animal case diagnosis decision network converges, and the network generated at this time is the animal case diagnosis decision network with completed network parameter learning, which can accurately diagnose and decide the animal case.

[0113] In a possible implementation, the method further includes determining the network learning weight corresponding to the second training case feedback data, and updating the network learning weight based on the second sample case diagnosis heat map to generate an updated network learning weight.

[0114] In a possible implementation, the step S106 can further include:

[0115] According to the first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data, the second sample case diagnosis heat map and the updated network learning weight, the network parameter learning of the animal case diagnosis decision network is performed to generate the animal case diagnosis decision network with completed network parameter learning.

[0116] In a possible implementation, the step S106 can further include:

[0117] determine a current network learning progress of the animal case diagnosis decision network.

[0118] update the first sample case diagnosis heat map to a preset second case diagnosis heat map to obtain an updated first sample case diagnosis heat map when the current network learning progress reaches a set progress.

[0119] update the second sample case diagnosis heat map to the preset second case diagnosis heat map to obtain an updated second sample case diagnosis heat map.

[0120] perform network parameter learning on the animal case diagnosis decision network according to the first training case feedback data, the updated first sample case diagnosis heat map, the second training case feedback data, and the updated second sample case diagnosis heat map, to generate an animal case diagnosis decision network with completed network parameter learning.

[0121] First, determine the network learning weight corresponding to the second training case feedback data, and update the network learning weight based on the second sample case diagnosis heat map to generate an updated network learning weight. Taking the second training case feedback data of a certain disease of sheep as an example, in the animal case diagnosis decision network, each training case feedback data has a corresponding network learning weight in the network learning process. These network learning weights are initially allocated according to the initial settings of the network, and they reflect the relative importance of each case data in network learning. For this case of sheep, its corresponding network learning weight may be determined based on the proportion of the disease in the overall cases, the complexity of the symptoms, and other factors.

[0122] When updating the network learning weight based on the second sample case diagnosis heat map, the information in the second sample case diagnosis heat map will be considered. For example, in the second sample case diagnosis heat map, if it is found that the importance of a certain symptom (such as a specific behavior) of sheep in disease diagnosis is re-evaluated (possibly due to new research results or statistical analysis of more cases), and this symptom was not fully valued in the previous network learning weight determination, the network learning weight of this case will be adjusted accordingly. Assuming that the original network learning weight was allocated according to the traditional understanding of symptoms, without considering the critical significance of this specific behavior to diagnosis, and now according to the information of the second sample case diagnosis heat map, this behavior is considered a very important diagnostic basis, then the weight of this case in network learning will be increased, thereby generating an updated network learning weight.

[0123] Then, the animal case diagnosis decision network is trained according to the first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data, the second sample case diagnosis heat map and the updated network learning weight value, so as to generate the animal case diagnosis decision network after network parameter learning.

[0124] In this process, the complex disease of the dog is taken as the first training case feedback data, and the disease of the sheep is taken as the second training case feedback data. These data are used for network parameter learning together with the corresponding first sample case diagnosis heat map, the second sample case diagnosis heat map and the updated network learning weight value. First, the first training case feedback data and the second training case feedback data are processed according to the input format of the network, and various information in the case (such as the basic information of the animal and the symptom performance) is converted into a numerical form that can be recognized by the network. Then, these data are input into the animal case diagnosis decision network. This network is composed of an input layer, multiple hidden layers and an output layer.

[0125] In the input layer, the data are distributed to each node. Then, the data are propagated and calculated in the hidden layer. The neurons of each hidden layer process the input data according to the pre-set connection weight and activation function. For example, in a hidden layer, the neurons will multiply the input data from the previous layer by the corresponding connection weight, then add these products, and then perform a nonlinear transformation through an activation function (such as a Sigmoid function or a ReLU function) to obtain the output result, and pass this result to the next layer. In this process, the updated network learning weight value will affect the processing method of the data. For those second training case feedback data with higher network learning weight value, their influence in the network learning process will be relatively greater, and the animal case diagnosis decision network will pay more attention to the feature information in these data.

[0126] When the data propagate in the network to the output layer, a prediction result is obtained. This prediction result is compared with the standard diagnosis result in the first sample case diagnosis heat map and the second sample case diagnosis heat map to calculate the error. After obtaining the error value through a specific error calculation method (such as the mean square error method), the back propagation algorithm is used to update the connection weight and other parameters in the network. The back propagation algorithm will calculate the gradient of each connection weight according to the error, and then adjust the weight according to a certain learning rate (such as 0.001). This process will be repeated continuously, and the error is recalculated after adjusting the weight each time, until the error reaches an acceptable range or the network converges, so as to generate the network parameter learning result.

[0127] In the network parameter learning process, it is also necessary to determine the current network learning progress of the animal case diagnosis decision network. For example, the current network learning progress can be represented by the ratio of the number of iterations that have been completed to the preset total number of iterations. When the current network learning progress reaches a set progress, such as when the learning progress reaches 70%, a specific operation will be performed.

[0128] At this time, the first sample case diagnosis heat map is updated to the preset second case diagnosis heat map to obtain an updated first sample case diagnosis heat map, and the second sample case diagnosis heat map is updated to the preset second case diagnosis heat map to obtain an updated second sample case diagnosis heat map. Assuming that the preset second case diagnosis heat map is a heat map that integrates more recent animal medical research results and more extensive case data statistical analysis. For the first sample case diagnosis heat map corresponding to the first training case feedback data of the complex disease of the dog, the importance of some symptoms may not be accurately marked originally or need to be adjusted with new research findings. By updating to the preset second case diagnosis heat map, the first sample case diagnosis heat map can be more in line with current medical cognition. Similarly, for the second sample case diagnosis heat map corresponding to the second training case feedback data of the disease of the sheep, after updating to the preset second case diagnosis heat map, it can better reflect the key factors of disease diagnosis.

[0129] Finally, according to the first training case feedback data, the updated first sample case diagnosis heat map, the second training case feedback data, and the updated second sample case diagnosis heat map, the network parameter learning of the animal case diagnosis decision network is performed to generate an animal case diagnosis decision network that has completed network parameter learning. The processed first training case feedback data and the second training case feedback data are input into the network again, and the network propagation, calculation of the prediction result, calculation of the error, and backward propagation to update the weight are performed according to the steps described above. Since the updated first sample case diagnosis heat map and the second sample case diagnosis heat map more accurately reflect the key factors of disease diagnosis, the animal case diagnosis decision network can better learn and adjust the parameters. With continuous iterative learning, the error between the prediction result of the animal case diagnosis decision network and the standard diagnosis result gradually decreases, and when the error reaches a minimum value or the performance of the network no longer improves significantly, it is considered that the network has converged. At this time, the generated is the animal case diagnosis decision network that has completed network parameter learning, which can more accurately diagnose and decide the animal case.

[0130] In one possible implementation, step S120 includes:

[0131] Determining an initialization neural network corresponding to the animal case diagnosis decision network.

[0132] The internet remote case feedback data is loaded into the initialization neural network to generate an initialization diagnosis result of the internet remote case feedback data.

[0133] The initialization diagnosis result is loaded into the animal case diagnosis decision network to generate an animal case diagnosis decision result of the internet remote case feedback data.

[0134] In this embodiment, first, the initialization neural network corresponding to the animal case diagnosis decision network is determined. This initialization neural network is the result obtained after a series of training steps, which has completed the network parameter learning based on the basic training case feedback data, the first training case feedback data, the second training case feedback data, and the corresponding sample case diagnosis heat map, and has reached a convergence state. This network structure includes an input layer, multiple hidden layers, and an output layer. The number and type of nodes in the input layer are determined according to various information features in the animal case feedback data, such as animal species, age, gender, symptom performance, and other information. The number of hidden layers, the number of neurons in each hidden layer, and the connection weights between neurons and other parameters are determined in the previous network training process. The output layer corresponds to possible animal case diagnosis results, such as classification of different diseases or severity of diseases.

[0135] Then, the internet remote case feedback data is loaded into the initialization neural network to generate an initialization diagnosis result of the internet remote case feedback data. Taking an internet remote case feedback data about a dog from an animal hospital as an example, this case feedback data contains detailed information about the dog, such as the dog is a 3-year-old male Labrador, has recently shown symptoms such as cough, lethargy, decreased appetite, has no obvious past medical history, and has no recent changes in living environment. The server processes this information according to the input format of the initialization neural network. For example, the dog's age "3 years old" is converted to a specific numerical code, the dog's breed "Labrador" is also converted to a corresponding numerical value or vector through a predefined encoding method, and the symptom performance is quantitatively encoded according to the classification and severity of different symptoms. Then, these encoded data are loaded into the input layer of the initialization neural network.

[0136] In the initialization of the neural network, data starts propagating in the network. From the input layer to the first hidden layer, each neuron performs a weighted sum operation on the input data according to the connection weights with the input layer nodes, and then performs a nonlinear transformation through an activation function (such as the ReLU function) to obtain the output result of the first hidden layer. This result is used as the input of the next hidden layer, and the above-mentioned weighted sum and nonlinear transformation operations are repeated until the data propagates to the output layer. In the output layer, according to the structure and parameters of the network, a preliminary result about the disease diagnosis of the dog is obtained, which is the initialization diagnosis result of the internet remote case feedback data. For example, the output layer may give the probability of the dog suffering from respiratory tract infection disease as 60%, and the probability of suffering from other diseases (such as digestive system diseases) as 10% and other probability distributions of different diseases, which is the initialization diagnosis result based on the initialization neural network.

[0137] Finally, this initialization diagnosis result is loaded into the animal case diagnosis decision network to generate the animal case diagnosis decision result of the internet remote case feedback data. The animal case diagnosis decision network has learned more in-depth animal case diagnosis knowledge in the previous construction process, and it not only relies on the preliminary result of the initialization neural network. When the initialization diagnosis result is input into the animal case diagnosis decision network, the network will further analyze and adjust the initialization diagnosis result. For example, the animal case diagnosis decision network may consider some factors that are not fully considered in the initialization neural network, or correct the initialization diagnosis result according to more accurate disease diagnosis logic.

[0138] Continuing with the above example of the dog's case, although the initialization diagnosis result indicates that the dog has a 60% probability of suffering from respiratory tract infection disease, the animal case diagnosis decision network may further analyze the relationship between the dog's symptoms, combine more animal disease pattern knowledge and past similar case experience, etc. If it is found that the distribution of the dog's cough symptoms over time is more similar to a certain specific respiratory disease, and the probability of this disease in the initialization diagnosis result is underestimated, the animal case diagnosis decision network will adjust this probability. At the same time, it may also re-evaluate other possible diseases according to the dog's age, gender, etc. Finally, the animal case diagnosis decision network will give a more accurate animal case diagnosis decision result, such as determining that the dog has a certain specific respiratory disease, and giving suggestions for further examination (such as blood test, X-ray examination, etc.) and possible treatment plan (such as use of specific drugs, treatment cycle, etc.) and other detailed information, which is the final animal case diagnosis decision result for the internet remote case feedback data.

[0139] Further, Figure 2A hardware structure diagram of an Internet feedback mining system 100 applied to an animal remote diagnosis platform for implementing the method provided by the embodiments of the present application is shown. As shown in Figure 2 The Internet feedback mining system 100 applied to the animal remote diagnosis platform can include at least one processor 102 (the processor 102 can include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, a transmission device 106 for communication function, and a controller 108. Those skilled in the art can understand that, Figure 2 The structure shown is only schematic, and it does not limit the structure of the Internet feedback mining system 100 applied to the animal remote diagnosis platform. For example, the Internet feedback mining system 100 applied to the animal remote diagnosis platform can include more or less components than those shown in Figure 2 or have a different configuration than that shown in Figure 2 .

[0140] The memory 104 can be used to store software programs and modules of application software, such as the program instructions corresponding to the method embodiments described above in the embodiments of the present application. The processor 102 performs various function applications and data processing by running the software programs and modules stored in the memory 104, that is, implements the above-mentioned Internet feedback mining method applied to the animal remote diagnosis platform. The memory 104 can include a high-speed random access memory, and can also include a non-volatile memory such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some examples, the memory 104 can further include a memory remotely arranged with respect to the processor 102, and these remote memories can be connected to the Internet feedback mining system 100 applied to the animal remote diagnosis platform through a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0141] The transmission device 106 is used to obtain or send data via a network. The specific examples of the above-mentioned network can include a wireless network provided by a communication provider of the Internet feedback mining system 100 applied to the animal remote diagnosis platform. In one example, the transmission device 106 includes a network adapter which can be connected to other network devices through a base station so as to be able to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency module which is used to communicate with the Internet in a wireless manner.

[0142] It should be noted that the above-mentioned embodiments of the present application are merely for the purpose of description, and do not represent the advantages and disadvantages of the embodiments. The above-mentioned embodiments of the present application are described. Other embodiments are within the scope of the appended claims. In some cases, the exceptions or steps recited in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or advantageous.

[0143] Each of the embodiments in the embodiments of the present application is described in a progressive manner, and the consistent and similar parts between each embodiment can be referred to each other, and each embodiment focuses on the difference from other embodiments. In particular, for the above different embodiments, since they are basically similar to the method embodiments, they are described more simply, and the relevant parts refer to the part of the method embodiment.

[0144] A person of ordinary skill in the art can understand that all or part of the steps of the above-mentioned embodiments can be completed by hardware, or by program to instruct relevant hardware to complete. The above-mentioned program can be stored in a computer readable storage medium, and the above-mentioned storage medium can be a read-only memory, a magnetic disk or an optical disk.

Claims

1. An internet feedback mining method applied to an animal remote diagnosis platform, characterized in that, The method comprises: acquiring internet remote case feedback data uploaded by an animal remote diagnosis platform; loading the internet remote case feedback data into an animal case diagnosis decision network to generate an animal case diagnosis decision result of the internet remote case feedback data, wherein the animal case diagnosis decision network is generated by joint model learning based on first training case feedback data, a first sample case diagnosis heat map corresponding to the first training case feedback data, second training case feedback data, and a second sample case diagnosis heat map corresponding to the second training case feedback data, the first training case feedback data and the second training case feedback data are distinguished based on an initial network learning result, the second sample case diagnosis heat map is updated based on the initial network learning result, and the initial network learning result is generated by joint model learning using basic training case feedback data of the animal case diagnosis decision network; pushing information to the animal remote diagnosis platform based on the animal case diagnosis decision result of the internet remote case feedback data; Before acquiring the internet remote case feedback data uploaded by the animal remote diagnosis platform, the method further comprises: determining a basic training case feedback data sequence of an animal case diagnosis decision network, wherein the basic training case feedback data sequence comprises a plurality of basic training case feedback data, and each basic training case feedback data has a corresponding sample case diagnosis heat map; loading the plurality of basic training case feedback data into the animal case diagnosis decision network based on the sample case diagnosis heat map to generate an initial network learning result; determining the initial network learning result corresponding to each basic training case feedback data and determining an average learning result corresponding to the initial network learning result; distinguishing first training case feedback data and second training case feedback data from the plurality of basic training case feedback data according to the average learning result; taking the sample case diagnosis heat map corresponding to the first training case feedback data as a first sample case diagnosis heat map and updating the sample case diagnosis heat map corresponding to the second training case feedback data to a second sample case diagnosis heat map; performing network parameter learning on the animal case diagnosis decision network according to the first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data, and the second sample case diagnosis heat map to generate an animal case diagnosis decision network with completed network parameter learning; the initial network learning result is a network learning error parameter, and the average learning result is a network learning average error parameter; Accordingly, the distinguishing first training case feedback data and second training case feedback data from the plurality of basic training case feedback data according to the average learning result comprises: determining the network learning error parameter corresponding to each basic training case feedback data; determining the basic training case feedback data with the network learning error parameter not less than the network learning average error parameter as the first training case feedback data; determining the basic training case feedback data corresponding to the network learning error parameter being less than the network learning average error parameter as second training case feedback data; the second training case feedback data corresponding to the sample case diagnosis heat map is updated to a second sample case diagnosis heat map, including: determining the sample case diagnosis heat map corresponding to the second training case feedback data; updating the sample case diagnosis heat map corresponding to the second training case feedback data according to a preset heat value update strategy to generate a second sample case diagnosis heat map, wherein the heat value of the sample case diagnosis heat map corresponding to the second training case feedback data is greater than the second sample case diagnosis heat map; the second training case feedback data corresponding to the sample case diagnosis heat map is updated to a preset first case diagnosis heat map, and the preset first case diagnosis heat map is used as a second sample case diagnosis heat map; or obtaining an updated sample case diagnosis heat map according to the sample case diagnosis heat map corresponding to the second training case feedback data and the initial network learning result corresponding to the second training case feedback data, and updating the sample case diagnosis heat map corresponding to the second training case feedback data using the updated sample case diagnosis heat map to generate a second sample case diagnosis heat map; the animal case diagnosis decision network is trained according to the first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data, and the second sample case diagnosis heat map, and an animal case diagnosis decision network with completed network parameter learning is generated, including: selecting first target training case feedback data from the first training case feedback data according to the first sample case diagnosis heat map, and selecting second target training case feedback data from the second training case feedback data according to the second sample case diagnosis heat map; network parameter learning is performed on the animal case diagnosis decision network according to the first target training case feedback data and the second target training case feedback data, and a network parameter learning result is generated; the steps of determining the first training case feedback data and the second training case feedback data from the plurality of basic training case feedback data, and updating the first sample case diagnosis heat map corresponding to the first training case feedback data and the second sample case diagnosis heat map corresponding to the second training case feedback data are iteratively performed using the network parameter learning result, and the animal case diagnosis decision network is continuously trained until the animal case diagnosis decision network converges, and an animal case diagnosis decision network with completed network parameter learning is generated.

2. The Internet feedback mining method applied to the animal remote diagnosis platform according to claim 1, further comprising: ​ determine the network learning weight corresponding to the second training case feedback data, and update the network learning weight based on the second sample case diagnosis heat map, to generate an updated network learning weight; wherein the network parameter learning of the animal case diagnosis decision network based on the first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data, and the second sample case diagnosis heat map to generate the animal case diagnosis decision network with completed network parameter learning comprises: the network parameter learning of the animal case diagnosis decision network based on the first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data, the second sample case diagnosis heat map, and the updated network learning weight to generate the animal case diagnosis decision network with completed network parameter learning.

3. The Internet feedback mining method applied to the animal remote diagnosis platform according to claim 1, wherein the network parameter learning of the animal case diagnosis decision network based on the first training case feedback data, the first sample case diagnosis heat map, the second training case feedback data, and the second sample case diagnosis heat map to generate the animal case diagnosis decision network with completed network parameter learning comprises: determining the current network learning progress of the animal case diagnosis decision network; when the current network learning progress reaches a set progress, updating the first sample case diagnosis heat map to a preset second case diagnosis heat map to obtain an updated first sample case diagnosis heat map; and updating the second sample case diagnosis heat map to the preset second case diagnosis heat map to obtain an updated second sample case diagnosis heat map; the network parameter learning of the animal case diagnosis decision network based on the first training case feedback data, the updated first sample case diagnosis heat map, the second training case feedback data, and the updated second sample case diagnosis heat map to generate the animal case diagnosis decision network with completed network parameter learning.

4. The Internet feedback mining method applied to the animal remote diagnosis platform according to claim 1, wherein the loading of the Internet remote case feedback data into the animal case diagnosis decision network to generate the animal case diagnosis decision result of the Internet remote case feedback data comprises: determining an initialization neural network corresponding to the animal case diagnosis decision network; loading the Internet remote case feedback data into the initialization neural network to generate an initialization diagnosis result of the Internet remote case feedback data; loading the initialization diagnosis result into the animal case diagnosis decision network to generate the animal case diagnosis decision result of the Internet remote case feedback data.

5. An internet feedback mining system applied to an animal remote diagnosis platform, characterized in that, The Internet feedback mining system applied to the animal remote diagnosis platform comprises a processor and a readable storage medium, and the readable storage medium stores a program which, when executed by the processor, implements the Internet feedback mining method applied to the animal remote diagnosis platform according to any one of claims 1-4.

Images