A deep neural network medical image diagnosis method based on confidence calibration

Abstract

The present invention discloses a deep neural network medical image diagnosis method based on confidence calibration, comprising: collecting medical image samples x with annotated disease category labels y and performing data cleaning and normalization processing; training neural network models LeNet‑5 and ResNet‑18 and saving the training weights; using LeNet‑5 as a teacher network and ResNet‑18 as a student network to construct a knowledge distillation model; fixing the weight of LeNet‑5, re-inputting the medical image x with the disease category label y into the LeNet‑5 and ResNet‑18 neural network models at the same time, and obtaining the confidence score corresponding to the disease category predicted for each medical image from the softmax layer of the LeNet‑5 model; calculating the KL divergence loss between the confidence score obtained above and the confidence score output by the corresponding softmax layer of ResNet‑18, and weighting it with the cross entropy loss between the ResNet‑18 prediction result and the true label to obtain a new loss function, so as to fine-tune the network parameters of ResNet‑18, and finally obtaining the deep neural network model ResNet‑18 after confidence calibration for assisting the medical image diagnosis process.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a deep neural network medical image diagnosis method, system, and storage medium based on confidence calibration. Background Art

[0002] With the advent of the big data era, the difficulty of acquiring, storing and transmitting large amounts of data has been greatly reduced. At the same time, the rapid improvement of computer computing power has enabled many data-driven machine learning technologies to be applied in practical scenarios. Among these machine learning technologies, deep learning algorithms, especially deep neural network models, are more likely to attract the attention of researchers and the favor of practitioners due to their high accuracy. In the field of medical care, deep learning image classification technology has played an important role in the automatic diagnosis of various collected medical images: doctors often use medical diagnosis systems built by deep neural network models to automatically diagnose a large number of collected medical images, and only make manual follow-up decisions for low-confidence (low-confidence) diagnosis results output by the system, which greatly reduces unnecessary and repetitive work for doctors and improves the efficiency of the hospital's disease diagnosis process for patients in the context of an aging population.

[0003] Since medical care belongs to a high-risk field, a truly reliable medical diagnosis model should not only require high diagnostic accuracy, but also the overall grasp of the model's correct diagnosis results (overall confidence) should be equivalent to the proportion of medical image samples with correct final diagnosis (accuracy), that is, the model's confidence calibration ability is required to be strong. At present, practitioners generally use the softmax layer output value corresponding to the disease category identified by the automatic medical diagnosis system as the confidence corresponding to the diagnosis result, but do not pay attention to the true reliability of this strategy to obtain confidence. Recent studies have shown that although modern deep neural networks have high prediction accuracy in many application fields, their softmax layer is prone to "overconfident" wrong predictions with high confidence for some samples. Excessive confidence will mislead doctors to believe and adopt the final diagnosis results of the neural network, which is very dangerous. Therefore, it is necessary to adopt a confidence calibration strategy to calibrate the high confidence output by the deep neural network due to the "overconfident" characteristic, so that it can output a true reflection of the confidence of the medical diagnosis system in its own prediction results. However, the existing methods for neural network confidence calibration, including post-training processing methods and training and calibration parallel methods, have limitations in application: the post-training processing method usually divides the limited labeled samples into pre-training samples for the first step and post-processing samples for the second step, which reduces the number of pre-training samples and reduces the diagnostic accuracy of deep neural networks for diseases; the training and calibration parallel method usually means that the complexity of the model and the calculation time are doubled, and large-scale neural network medical diagnosis models are not conducive to equipment deployment and will also reduce the operating efficiency of medical diagnosis systems. Although the prediction accuracy of traditional shallow neural networks is low, the confidence score output by its softmax layer has a natural high calibration feature, that is, it can more realistically reflect the degree of grasp of the prediction results by the neural network, but this feature has been ignored for a long time. Based on this, it is necessary to improve the confidence calibration method of deep neural networks currently used in medical diagnosis systems. Summary of the invention

[0004] In order to overcome the above-mentioned deficiencies of the prior art, the present invention discloses a deep neural network medical image diagnosis method based on confidence calibration, which specifically comprises the following steps:

[0005] Collect medical image training samples x that have been labeled with disease category labels y and perform data cleaning and normalization;

[0006] Train the shallow neural network model LeNet-5 and save the training weights;

[0007] Train the deep neural network model ResNet-18 and save the training weights;

[0008] A knowledge distillation model was constructed, the weights of the LeNet-5 neural network were fixed, and the medical image x with the disease category label y was re-input into the trained LeNet-5 and ResNet-18 neural network models at the same time. The predicted disease category and corresponding confidence score of each medical image were obtained from the softmax layer output of the LeNet-5 model.

[0009] The KL divergence loss is calculated between the confidence score obtained previously and the confidence score output by the corresponding softmax layer of ResNet-18, and is weighted with the cross entropy loss between the ResNet-18 prediction result and the true label to obtain a new loss function, so as to fine-tune the network parameters of ResNet-18, and finally obtain the deep neural network model ResNet-18 after confidence calibration to assist the medical diagnosis process.

[0010] Furthermore, when training the neural network models LeNet-5 and ResNet-18 and saving the training weights, a total of N disease category labels y (y = {y1, y2, ..., y i ,…y N}, where i = 1~N) medical image data x is input into LeNet-5 and ResNet-18 after data preprocessing, and the confidence scores p (respectively p) output by the two models are calculated using formula (1). t 、p s ) and the cross entropy loss CE(p,y) between the labeled diagnostic disease category y, combined with the L2 regularization term, update the parameters θ of the two networks, and wait for the loss function Loss of the two networks LeNet5 、Loss ResNet18 After full convergence, save the respective training weights. The loss function is shown in formula (2) (3):

[0011]

[0012] Loss LeNet5 =CE(p t ,y)+λ2*θ 2 (2)

[0013] Loss ResNet18 =CE(p s ,y)+λ2*θ 2 (3)

[0014] where p i For the sample image x i (i=1~N) is the confidence obtained by inputting the neural network, and λ2 is the L2 regularization coefficient.

[0015] Furthermore, the KL divergence loss is calculated using the confidence scores of the two networks, and the new loss function is obtained by weighting it with the cross entropy loss. When fine-tuning the network parameters of ResNet-18, the information of the teacher network (p t ) is passed to the student network (p s ), the weights of the shallow neural network LeNet-5 are fixed to make it the teacher network. Then, the high confidence calibration ability of the shallow neural network is used to use the medical image x with the labeled disease category y after data preprocessing and input into the pre-trained LeNet-5 and ResNet-18, and the confidence scores p output by the two models are recalculated. t 、p s , the KL divergence loss of the two is calculated according to formula (4):

[0016]

[0017] where p i,t 、p i,s The samples x are i Confidence output by the teacher network and the student network.

[0018] By minimizing the KL loss, knowledge distillation converts the p containing high-confidence calibration ability information in the teacher network LeNet-5 into t Passed to the student network ResNet-18, so that its original output overconfidence confidence score p s Converges to p t , in order to improve the confidence calibration ability of the deep neural network. At the same time, in order to prevent the low accuracy of the shallow neural network from being passed to the deep neural network and reducing its prediction accuracy, the real disease category label y of the collected medical image x is used to construct a cross entropy loss for the student network to further constrain the confidence score p s Convergence process: The confidence p output by ResNet-18 s The cross entropy loss between the category label y and the confidence p output by LeNet-5 t The KL divergence loss and L2 regularization term between them are weighted to obtain a new loss function Loss Cal-ResNet18 , as shown in formula (5), the formula is minimized to fine-tune the network parameters of ResNet-18 until it converges completely:

[0019] Loss Cal-ResNet18 =(1-λ1)*KL(p t ,p s )+λ1*CE(p s ,y)+λ2*θ 2 (5)

[0020] Where λ1 is the weight of the classification loss and θ is the network parameter.

[0021] A system for executing the above-mentioned deep neural network medical image diagnosis method based on confidence calibration, characterized in that the system comprises:

[0022] A collection unit, used to collect medical image data with annotated disease category labels, and modify the size and data enhancement to obtain enhanced collected data;

[0023] A training unit, inputting the pre-processed and enhanced collected data into a neural network model for training, and saving the trained model parameter weights;

[0024] The decision unit reads the collected medical image data of the disease category to be labeled, and transmits it to the neural network obtained by the training unit to predict the disease category and confidence level of the data to be tested, and makes a decision on the final diagnosis result of the collected data to be tested.

[0025] Furthermore, when deciding on the final category of the sample to be tested, if the confidence of the neural network output is greater than the specified threshold, the final category is the predicted category of the neural network; if the confidence of the neural network output is less than the specified threshold, the final category is the result of the doctor's manual decision-making diagnosis.

[0026] Due to the adoption of the above-mentioned technical solution, the present invention provides a deep neural network medical image diagnosis method based on confidence calibration. This method utilizes the natural characteristics of the long-neglected shallow neural network with high confidence calibration ability, and assists the modern overconfident deep neural network to improve its calibration ability, avoiding it from making predictions with high confidence, thereby reducing the serious consequences caused by doctors blindly adopting the prediction results of the medical diagnosis system. This confidence calibration strategy that uses a shallow network with strong calibration ability to assist a poor deep network training has a simpler and more intuitive training strategy and is more interpretable. In addition, since the deep neural network trained by post-processing contains the knowledge information of both the shallow network and the deep network, the new network is actually an implicit integrated model, which enables its accuracy in predicting diagnostic results to be slightly improved. Therefore, compared with the existing distance-based confidence calculation technology and the existing neural network confidence calibration strategy, the present invention proposes a method for solving the problem of overconfident predictions with high confidence levels that are easily output by deep neural network models, based on distance-based confidence calculation, which greatly improves the reliability of prediction results output by automatic medical diagnosis systems when used to assist doctors in diagnosing patients' diseases; a post-processing calibration strategy combining two networks is used to overcome the problem that traditional confidence calibration strategies may sacrifice prediction accuracy performance; and finally a single deep neural network with a simple structure is obtained, which overcomes the defects of traditional confidence calibration strategies with high computational complexity, large models, and poor rapidity; the proposed method can be used to solve the problem that traditional automatic medical diagnosis systems are easily output with overconfident predictions that mislead doctors, thereby improving the reliability of the prediction function of the automatic medical diagnosis system. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0028] Figure 1 The flowchart of the deep neural network medical image diagnosis method based on confidence calibration is shown;

[0029] Figure 2 This is a structural diagram of a deep neural network medical image diagnosis method based on confidence calibration;

[0030] Figure 3 shows the network structure diagrams of LeNet-5 (a) and ResNet-18 (b);

[0031] Figure 4 shows the reliability graph (including ECE loss) and confusion matrix of ResNet-18 before (a) and after (b) confidence calibration;

[0032] Figure 5 shows the confidence results of three different types of patient X-ray images (a, b, c) after being misjudged before and after confidence calibration; DETAILED DESCRIPTION

[0033] In order to make the technical solutions and advantages of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention:

[0034] The labeled sample data in this application is collected based on image cases of medical diagnosis tasks, such as Figure 1 A deep neural network medical image diagnosis method based on confidence calibration is shown, which specifically includes the following steps:

[0035] Collect sample data with category labels and perform data cleaning and normalization. Here, we take the ChestXRay2017 dataset as an example, and scale the dimensions of all training samples x (located in the train folder) with label y to 224×224 to facilitate subsequent unified input into the neural network.

[0036] We use LeNet-5 (proposed in 1998) as a representative of simple shallow neural networks and ResNet-18 (proposed in 2015) as a representative of modern deep neural networks. We train the shallow neural network model LeNet-5 and the deep neural network ResNet-18 separately and save the training weights. We use a total of N disease category labels y (y = {y1, y2, ..., y i ,…y N}, where i = 1~N) medical image data x is input into LeNet-5 and ResNet-18 after data preprocessing, and the confidence scores p (respectively p) output by the two models are calculated using formula (1). t 、p s ) and the cross entropy loss CE(p,y) between the labeled diagnostic disease category y, combined with the L2 regularization term, update the parameters θ of the two networks, and wait for the loss function Loss of the two networks LeNet5 、Loss ResNet18 After full convergence, save the respective training weights. The loss function is shown in (2) and (3):

[0037]

[0038] Loss LeNet5 =CE(p t ,y)+λ2*θ 2 (2)

[0039] Loss ResNet18 =CE(p s,y)+λ2*θ 2 (3)

[0040] where p i For the sample image x i (i=1~N) is the confidence obtained by inputting the neural network, and λ2 is the L2 regularization coefficient.

[0041] like Figure 2 As shown in the figure, the shallow neural network LeNet-5 is used as the teacher network, and the deep neural network ResNet-18 is used as the student network to build a knowledge distillation model. The information of the teacher network (p t ) is passed to the student network (p s ) features, and uses the natural high calibration ability of LeNet-5 to calibrate the confidence of ResNet-18. The weights of the LeNet-5 neural network are fixed, and the sample data with category labels are re-input into the trained LeNet-5 and ResNet-18 neural network models at the same time. The confidence score corresponding to each sample is obtained from the softmax layer output of the LeNet-5 model, and the KL divergence loss is calculated between the obtained confidence score and the confidence score of the corresponding softmax layer output of ResNet-18, and the new loss function is weighted with the cross entropy loss between the ResNet-18 prediction result and the true label. In this way, the network parameters of ResNet-18 are fine-tuned, and finally the neural network model ResNet-18 after confidence calibration is obtained for auxiliary medical diagnosis process. The loss function of the confidence calibration process consists of the KL divergence of the confidence scores of the two networks, the classification loss between the confidence score of ResNet-18 and the true category label y, and the weighted L2 regularization term. The KL divergence loss is shown in formula (4), and the final loss function Loss Cal-ResNet18 As shown in formula (5):

[0042]

[0043] Loss Cal-ResNet18 =(1-λ1)*KL(p t ,p s )+λ1*CE(p s ,y)+λ2*θ 2 (5)

[0044] where p i,t 、p i,s The samples x are i The confidence output by the teacher network and the student network, λ1 is the weight of the classification loss, and θ is the network parameter.

[0045] The hierarchical structures of the shallow neural network LeNet-5 and the deep neural network ResNet-18 are shown in Figure 3(a) and Figure 3(b) respectively. Since the input image dimension is too large, the structure of the original shallow neural network LeNet-5 is improved by adding an additional set of convolutional layers and fully connected layers.

[0046] In order to more intuitively demonstrate the effectiveness of the confidence calibration based deep neural network medical image diagnosis method for ResNet-18 deep neural network confidence calibration, Figure 4 (a) and Figure 4 (b) respectively show the reliability graph (left, with ECE loss value in the title) and confusion matrix (right) of ResNet-18 prediction of verification samples in the ChestXRay2017 / test folder before and after calibration. The reliability graph is a histogram that visually reflects the confidence calibration ability of the model: it divides the confidence of the prediction result from 0 to 1 into M = 20 equally spaced intervals, and calculates the confidence values ​​that fall in each confidence interval B. m The average prediction confidence score conf(B) of all samples under (m=1~M) m )(red) and the average prediction accuracy acc(B m )(blue), the smaller the difference between the two in each interval, the better the confidence calibration ability of the model. ECE loss is based on the reliability diagram and takes into account each confidence interval B m Number of samples under |B m | is an indicator that quantitatively characterizes the confidence calibration ability of the model. The smaller it is, the better the confidence calibration effect of the model. For n validation samples, the calculation formula of ECE loss is shown in (6):

[0047]

[0048] According to Figure 4, from a qualitative perspective, the prediction accuracy of the reliability graph after calibration is significantly better than that before calibration; from a quantitative perspective, the ECE loss after calibration is also lower than that before calibration: this shows that the deep neural network medical image diagnosis method based on confidence calibration has played a good role in the confidence calibration of the ResNet-18 deep neural network. At the same time, according to the two confusion matrices on the right, it can be calculated that the prediction accuracy before and after calibration is 83.49% and 85.26% respectively, which also shows the effectiveness of the implicit integration structure in the proposed confidence calibration method to improve the accuracy of the deep neural network ResNet-18. For the hyperparameter settings in the loss functions (12), (13), and (15), take λ1=0.45, λ2=5e-4.

[0049] Corresponding to the deep neural network medical image diagnosis method based on confidence calibration provided by the present invention, the present invention also provides a deep neural network medical image diagnosis system based on confidence calibration, comprising:

[0050] A collection unit, used to collect medical image data with annotated disease category labels, and modify the size and data enhancement to obtain enhanced collected data;

[0051] A training unit, inputting the pre-processed and enhanced collected data into a neural network model for training, and saving the trained model parameter weights;

[0052] The decision unit reads the unlabeled medical image samples to be diagnosed, passes them into the neural network model obtained by the training unit, predicts and obtains the disease category and confidence information to which the samples belong, and makes a decision on the disease category to be diagnosed: if the confidence output by the neural network is greater than the specified threshold, the final diagnosis result is the predicted category of the neural network; if the confidence output by the neural network is less than the specified threshold, the final diagnosis result is the result of the doctor's manual decision-making diagnosis.

[0053] The auxiliary medical diagnosis system is further described below in conjunction with FIG5, and the confidence decision threshold in the decision unit is taken as 0.85. FIG5 (a), (b), and (c) respectively select the confidence information of the ResNet-18 output before and after calibration after one X-ray image of each of the true categories of bacterial pneumonia, health, and viral pneumonia is misdiagnosed. It can be seen that if the proposed deep neural network medical image diagnosis system based on confidence calibration is not used, the confidence of the three misclassified X-ray images is greater than 0.85, which will mislead doctors to use the results of the medical diagnosis system to make independent decisions, and may endanger the life and health of patients; after using the deep neural network medical image diagnosis system based on confidence calibration, the three misclassified X-ray images are output with a low confidence less than the threshold of 0.85, prompting doctors to use manual identification to decide the patient's condition, minimizing the risks that may be brought by the automatic medical diagnosis system.

[0054] It can be seen that the method and system of the present invention can enable the deep neural network to more accurately reflect the reliability of the model prediction results, and solve the problem of overconfidence in the deep neural network prediction.

[0055] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0056] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0057] In the several embodiments provided by the present invention, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0058] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0059] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0060] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0061] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A deep neural network medical image diagnosis method based on confidence calibration, characterized in that: include: Collect medical image training samples x that have been labeled with disease category labels y and perform data cleaning and normalization; Train the shallow neural network model LeNet-5 and save the training weights; Train the deep neural network model ResNet-18 and save the training weights; Use LeNet-5 as the teacher network and ResNet-18 as the student network to build a knowledge distillation network model; The weights of the LeNet-5 neural network are fixed, and the medical image x with the disease category label y is re-input into the trained LeNet-5 and ResNet-18 neural network models at the same time, and the predicted disease category and corresponding confidence score of each medical image are obtained from the softmax layer output of the LeNet-5 model; The KL divergence loss is calculated between the confidence score obtained previously and the confidence score output by the corresponding softmax layer of ResNet-18, and is weighted with the cross entropy loss between the ResNet-18 prediction result and the true label to obtain a new loss function, so as to fine-tune the network parameters of ResNet-18, and finally obtain the deep neural network model ResNet-18 after confidence calibration to assist the medical diagnosis process.

2. The method according to claim 1, further characterized in that: When training the neural network models LeNet-5 and ResNet-18 and saving the training weights, a total of N disease category labels y (y = {y1, y2, ..., y i ,…y N }, where i = 1~N) medical image data x is input into LeNet-5 and ResNet-18 after data preprocessing, and the confidence scores p of the two models are calculated respectively using formula (1) t 、p s The cross entropy loss CE(p,y) between the labeled diagnostic disease category y and the L2 regularization term is used to update the parameters θ of the two networks until the loss function Loss of the two networks is LeNet5 、Loss ResNet18 After full convergence, the respective training weights are saved, and the loss function is shown in formula (2) (3): Loss LeNet5 =CE(p t ,y)+λ2*θ 2 (2) Loss ResNet18 =CE(p s ,y)+λ2*θ 2 (3) where p i For the sample image x i (i=1~N) is the confidence obtained by inputting the neural network, and λ2 is the L2 regularization coefficient.

3. The method according to claim 1, further characterized in that: The KL divergence loss is calculated using the confidence scores of the two networks, and the new loss function is obtained by weighting it with the cross entropy loss. When fine-tuning the network parameters of ResNet-18, the teacher network confidence score p can be reduced by knowledge distillation. t Pass to the student network confidence score p s The weights of the shallow neural network LeNet-5 are fixed to make it a teacher network. Then, the high confidence calibration ability of the shallow neural network is used to use the medical image x with the labeled disease category y after data preprocessing and input it into the pre-trained LeNet-5 and ResNet-18, and the confidence scores p output by the two models are recalculated. t 、p s , the KL divergence loss of the two is calculated according to formula (4): where p i,t 、p i,s The samples x are i The confidence output by the teacher network and the student network is converted into p containing high confidence calibration ability information in the teacher network LeNet-5 by minimizing the KL loss. t Passed to the student network ResNet-18, so that its original output overconfidence confidence score p s Converges to p t , in order to improve the confidence calibration ability of the deep neural network; at the same time, in order to prevent the low accuracy characteristics of the shallow neural network from being transmitted to the deep neural network and reducing its prediction accuracy, the real disease category label y of the collected medical image x is used to construct the cross entropy loss for the student network to further constrain the confidence score p s Convergence process: The confidence p output by ResNet-18 s The cross entropy loss between the class label y and the confidence p output by LeNet-5 t The KL divergence loss and L2 regularization term between them are weighted to obtain a new loss function Loss Cal-ResNet18 , as shown in formula (5), the formula is minimized to fine-tune the network parameters of ResNet-18 until it converges completely: Loss Cal-ResNet18 =(1-λ1)*KL(p t ,p s )+λ1*CE(p s ,y)+λ2*θ 2 (5) Where λ1 is the weight of the classification loss and θ is the network parameter.

4. A system for executing the deep neural network medical image diagnosis method based on confidence calibration as described in any one of claims 1 to 3, characterized in that: The system comprises: The acquisition unit is used to collect medical image data with annotated disease category labels, and modify the size and perform data enhancement processing on the medical image data to obtain enhanced collected data; The training unit inputs the preprocessed and enhanced collected data into the neural network model for training, and saves the trained model parameter weights; The decision unit reads the collected medical image data of the disease category to be labeled, and transmits it to the neural network obtained by the training unit to predict the disease category and confidence level of the data to be tested, and makes a decision on the final diagnosis result of the collected data to be tested.

5. The method according to claim 4, further characterized in that: When making decisions on the final diagnosis results of the collected data to be tested, if the confidence level of the neural network output is greater than the specified threshold, the final diagnosis result is the predicted category of the neural network; if the confidence level of the neural network output is less than the specified threshold, the final diagnosis result is decided manually by the doctor.

Images