Evidence uncertainty progressive guidance based medical image segmentation method and device

Abstract

The application discloses a medical image segmentation method and device based on evidence uncertainty progressive guidance, and the method comprises the following steps: S1, obtaining training samples and corresponding labels; S2, training a medical image segmentation model according to the training samples and the corresponding labels; S3, obtaining test samples; S4, inputting the test samples into the trained medical image segmentation model for prediction, calculating the parameters of the Dirichlet distribution, and calculating the expected tensor of the Dirichlet distribution; S5, calculating a segmentation evidence uncertainty matrix; S6, taking the class with the maximum class probability as the predicted class, and then the prediction segmentation result output by the medical image segmentation model based on the test samples is: prediction i = argmax p i ; S7, visualizing the prediction segmentation result of the test samples and the segmentation evidence uncertainty matrix. The application improves the segmentation precision, greatly reduces the calculation cost, and provides more accurate uncertainty estimation.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image segmentation, and in particular to a credible medical image segmentation method and device guided by progressive evidence uncertainty. Background Art

[0002] Existing medical image segmentation methods typically utilize neural network models, leveraging trained neural network models to accurately identify and segment specific anatomical structures or pathological features of interest within medical images. Current deep learning neural network-based models often interpret the class probabilities output by the Softmax function as classification confidence, treating the semantic segmentation task as a pixel-level classification task focused on the class of each pixel. However, the confidence obtained in this way is uncalibrated, and the confidence estimates provided are inaccurate when faced with out-of-distribution samples.

[0003] In the practical application of medical image segmentation technology, due to the interference of various factors during medical image acquisition, the generated medical images often contain noise and artifacts, which affects the accurate segmentation of specific anatomical structures. In addition, the internal structure of the neural network is difficult to explain, and the results it produces often have low credibility. Doctors find it difficult to understand and explain how the neural network model makes segmentation decisions. In order to avoid clinical risks, further human analysis is required. In order to improve these shortcomings and increase the credibility of the neural network segmentation results, uncertainty metrics are usually introduced for auxiliary analysis. Existing medical image segmentation methods often use algorithms such as integrated networks, Monte Carlo Dropout, calibration, and evidence deep learning to evaluate uncertainty, but these medical image segmentation methods still have some shortcomings, as follows:

[0004] (1) Medical images in real scenes have large scale differences. The feature extraction modules of neural networks in existing medical image segmentation methods are mostly based on 3x3 convolution kernels, which cannot achieve adaptive extraction of image features.

[0005] (2) The integrated network requires multiple retraining of the model, which will result in high computational costs for complex models. At the same time, the integrated model stores these trained models for inference, which requires a large storage space, which is a major problem for some medical institutions with limited resources.

[0006] (3) Introducing Monte Carlo Dropout to estimate Bayesian uncertainty in deep neural networks may lead to inconsistent outputs. At the same time, the inference process of Bayesian neural networks usually requires multiple samplings, which requires large computing resources and time.

[0007] (4) Although calibration-based methods can produce highly calibrated segmentation results, their effectiveness in uncertainty measurement is usually lower than that of other uncertainty measurement methods;

[0008] (5) Although the current mainstream evidence deep learning method can model the uncertainty of segmentation, it will significantly reduce the accuracy of model segmentation;

[0009] (6) Existing uncertainty measurement methods often only produce uncertainty results, thus ignoring the guiding role of uncertainty results in model segmentation, resulting in poor model segmentation effects;

[0010] (7) Medical image segmentation tasks usually focus more on the segmentation accuracy and uncertainty of the segmentation area edges, but existing models based on the UNet network architecture generally have the defect of paying little attention to the segmentation edges;

[0011] In summary, although these methods have been widely used in the field of medical image segmentation, their shortcomings still lead to significant limitations in processing actual medical images. Summary of the Invention

[0012] The present invention provides a reliable medical image segmentation method and device with progressive guidance of evidence uncertainty, which has low dependence on data quality, low computing and storage resource overhead, and high application of segmentation uncertainty, and can solve at least one of the above technical problems.

[0013] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0014] The reliable medical image segmentation method guided by evidence uncertainty step by step includes the following steps:

[0015] S1, obtain training samples and corresponding labels;

[0016] S2. training a credible medical image segmentation model based on the training samples and corresponding labels;

[0017] S3. Obtain test samples;

[0018] S4, inputting the test sample into the trained credible medical image segmentation model for prediction, and calculating the parameter α of the Dirichlet distribution i , calculate the expected tensor p of the Dirichlet distribution i =α i / S i , S i is the parameter α i sum;

[0019] S5. Calculate the segmentation evidence uncertainty matrix U = C / S i , C is the total number of segmentation categories;

[0020] S6. Let the category with the largest category probability be the predicted category, and the predicted segmentation result output by the credible medical image segmentation model based on the test sample is: prediction i =argmax p i ;

[0021] S7, predict the segmentation result of the test sample i and the segmentation evidence uncertainty matrix U for visualization.

[0022] Furthermore, in S2, the training of the credible medical image segmentation model further includes:

[0023] S21, input the training sample x i , the size of the two-dimensional tensor is m×n;

[0024] S22. Initialize the evidence uncertainty matrix to be a full-1 matrix U, and the size of the two-dimensional tensor is the same as the training sample x. i The two-dimensional tensors of are the same size;

[0025] S23, the evidence matrix e obtained based on the credible medical image segmentation model and the evidence uncertainty matrix prediction i =f(x i ,w), calculate the parameter α of Dirichlet distribution according to subjective logic theory i , and use this to calculate the evidence uncertainty matrix u for this segmentation i :

[0026] α i =e i +1,u i =C / S i

[0027] in, C is the total number of categories included in all sample pixels, i represents the sample index position, f(·) is the function form of the neural network, and w is the parameter of the neural network;

[0028] S24. Update the evidence uncertainty matrix U to u i ;

[0029] S25. Calculate the updated evidence uncertainty matrix u i The difference between the original evidence uncertainty matrix U is used to determine whether the following conditions are met:

[0030]

[0031] Among them, ε is a given sufficiently small positive number;

[0032] If the condition is satisfied, the credible medical image segmentation model is considered to have converged, and the subsequent steps are continued. If the condition is not satisfied, the process returns to S23 until the above condition is satisfied, at which point the calculated evidence uncertainty tends to be stable.

[0033] S26. Calculate the parameter α of the Dirichlet distribution based on the prediction of the credible medical image segmentation model. i , calculate the expected tensor p of the Dirichlet distribution i =α i / S i ;

[0034] S27, let the category with the largest category probability be the predicted category, then the predicted segmentation result output by the credible medical image segmentation model based on the training sample is: prediction i =argmax p i ;

[0035] S28, calculating the loss based on the predicted segmentation result obtained in S27 and the evidence uncertainty obtained in S25, and performing back propagation to update the weight of the credible medical image segmentation model;

[0036] S29. Repeat the above training steps until the trusted medical image segmentation model reaches the maximum number of iterations or the loss no longer decreases.

[0037] Furthermore, in S2, the credible medical image segmentation model is optimized To learn, that is, To optimize the goal;

[0038] in, is the loss function, is the evidence cross entropy loss function, is the KL divergence regularization loss function, is the uncertainty regularization loss function, λ1 is the attenuation coefficient, and λ2 is the empirical hyperparameter.

[0039] Furthermore, in S2, if the optimization objective is explained by the loss function of a single medical image sample, the medical image segmentation task is regarded as a pixel-level classification task. For a basic classification task, the likelihood loss function is written as follows:

[0040]

[0041] Among them, C is the total number of segmentation categories, x is the xth sample, One-hot encoding of the Ground Truth representing the data, represents the predicted probability of the credible medical image segmentation model for category C;

[0042] In the framework of evidence theory, by associating Dirichlet distribution with confidence distribution, the above formula (1) can be rewritten as the following formula (2):

[0043]

[0044] in, is the evidence cross entropy loss function, represents the Dirichlet distribution parameter of the Cth type of sample, and S is the sum of the Dirichlet distribution parameters;

[0045] In order to ensure that samples that cannot be correctly classified generate less evidence, the KL divergence regularization term is introduced for constraint. The calculation formula is as follows (3):

[0046]

[0047] in, is the KL divergence regularization loss function, Γ(α) is the Gamma function, ψ(·) is the Digamma function, Represents the adjusted parameters to ensure that the Ground Truth evidence does not mistakenly become 0;

[0048] In order to constrain uncertainty and balance the prediction probability and uncertainty range, the uncertainty regularization term is introduced, and the calculation formula is as follows (4):

[0049]

[0050] in, is the uncertainty regularization loss function;

[0051] Combining equations (1) to (4), we get the final loss function as follows (5):

[0052]

[0053] Wherein, λ1 is a decay coefficient used to prevent the trusted medical image segmentation model from converging prematurely and thus preventing the trusted medical image segmentation model from insufficiently learning. λ1 = min(1,10×epoch / total_epochs, epoch is the current number of training rounds, total_epochs is the total number of training rounds, and λ2 is an empirical hyperparameter.

[0054] The device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the above-mentioned reliable medical image segmentation method guided by progressive evidence uncertainty.

[0055] The beneficial effects of the present invention are embodied in:

[0056] 1. Class uncertainty is modeled based on subjective logic. Compared with the traditional method of using first-order probability to express uncertainty, subjective logic explicitly models uncertainty by including the quality of uncertainty in subjective opinions, thus having the ability to reflect whether the given probability is reliable, solving the defect of traditional methods that cannot reflect the credibility of the given probability.

[0057] 2. Express subjective opinions through Dirichlet distribution, and then generate subjective opinions through neural network learning evidence in observed samples.

[0058] 3. Use evidence uncertainty to gradually guide the segmentation process, improve segmentation accuracy, make full use of the guiding role of uncertainty in segmentation, and solve the problem of decreased segmentation accuracy caused by evidence uncertainty modeling.

[0059] 4. The present invention also includes a micro-step convolution and dilated convolution module fused through a gating mechanism. This module utilizes the characteristics of smaller receptive field of micro-step convolution and larger receptive field of dilated convolution, and weightedly fuses the two through a gating mechanism, so that the model can adaptively capture the detail information and global information in medical images, so that the two compensate each other to achieve better feature extraction effect.

[0060] 5. The present invention also includes an uncertainty attention module, which uses the attention mechanism to take the input medical image as a query for the segmentation uncertainty map, calculates the attention matrix, and then jumps it with the input through a multi-layer perceptron, so that the model can pay attention to the content of interest in the segmentation uncertainty map, and better play the role of progressive guidance of evidence uncertainty. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The drawings described herein are used to provide further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0062] Figure 1 This is an overall flow chart of the trusted medical image segmentation method according to an embodiment of the present invention.

[0063] Figure 2 This is a model reasoning flow chart of the trusted medical image segmentation method according to an embodiment of the present invention.

[0064] Figure 3 Schematic diagram of the model structure of the credible medical image segmentation method according to an embodiment of the present invention.

[0065] Figure 43 is a comparison chart of the medical image segmentation method according to an embodiment of the present invention and other medical image segmentation methods.

[0066] Figure 5 It is a structural block diagram of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. In the absence of conflict, the embodiments in this application and the features in the embodiments can be combined with each other. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0068] It should be noted that the meaning of "and / or" appearing throughout the text includes three parallel solutions. Taking "A and / or B" as an example, it includes solution A, solution B, or solutions in which both A and B are satisfied. In addition, "multiple" refers to more than two. In addition, the technical solutions between the various embodiments can be combined with each other, but this must be based on the fact that ordinary technicians in this field can implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0069] In order to make the solution of the embodiment of the present invention easier to understand and better reflect the difference from the existing implementation scheme, the following first briefly describes the existing reliable medical image segmentation method in the prior art:

[0070] Existing credible medical image segmentation methods are mainly divided into two categories: credible medical image segmentation methods that use deep neural networks with Monte Carlo Dropout or ensemble methods to estimate epistemic uncertainty, and credible medical image segmentation methods that use direct modeling methods such as calibration or evidential deep learning to estimate aleatoric uncertainty.

[0071] In addition, existing credible medical image segmentation models often ignore the progressive guiding role of segmentation uncertainty in the segmentation process, and only use the segmentation uncertainty map as an auxiliary tool to assist doctors in further diagnosis, resulting in room for improvement in segmentation accuracy. In particular, the evidence deep learning quantification uncertainty method, which is the same as the present invention, will cause problems such as decreased segmentation accuracy when it does not utilize the progressive guiding role of evidence uncertainty in the segmentation process.

[0072] Specifically, the Bayesian neural network defines that the network weights obey the Gaussian distribution, and the posterior distribution of the network weights is approximately inferred through variational inference and the Monte Carlo method, which brings about a high computational cost. Although the computational cost of estimating uncertainty using a deep neural network with the Monte Carlo Dropout method or an integrated method is lower than that of the Bayesian neural network, the former requires running the neural network multiple times in the inference phase, and the latter requires training and storing multiple deep neural networks, so there are also problems such as high computational and storage costs. Moreover, compared with non-Bayesian neural networks, both of the above methods cannot achieve the goal of estimating cognitive uncertainty through a single forward propagation through a single neural network. The method provided in the embodiment of the present invention is to model uncertainty through the Dirichlet distribution, define the learning process of the deep neural network as a process of evidence acquisition, and use the uncertainty of the evidence of segmentation for the progressive guidance of the segmentation process, thereby achieving full utilization of uncertainty.

[0073] Existing reliable medical image segmentation methods often use direct modeling to estimate the aleatory uncertainty in medical image segmentation using first-order probabilities. The probability estimates obtained by this approach are often uncalibrated and can lead to overconfidence when dealing with out-of-distribution data. The embodiments of the present invention model the Dirichlet distribution to obtain evidence from the data. Therefore, when dealing with out-of-distribution data, higher uncertainty estimates can be obtained due to the limited evidence collected.

[0074] To clearly illustrate the credible medical image segmentation method proposed by the present invention, the following will be explained in detail in conjunction with specific embodiments:

[0075] Since the baseline model of the embodiment of the present invention is the UNet segmentation model, it is necessary to briefly describe the U-shaped neural network segmentation model. The UNet segmentation model is a deep learning model widely used in the field of image segmentation. It is named because its structure resembles the letter "U". UNet is an "encoder-decoder" structure. The role of the encoder is to gradually extract the features of the input image through a series of convolutional layers and pooling layers, reduce the spatial resolution of the image, and increase the abstract level of the features. After each pooling operation, the size of the image is halved, and the number of feature channels is increased, thereby capturing richer image feature information. The decoder is the opposite of the encoder. It gradually restores the spatial resolution of the image through a series of upsampling operations, and at the same time combines the feature maps from the corresponding levels of the contraction path to achieve accurate segmentation of the image. In this process, the features of different levels in the encoder are fused with the features of the corresponding levels in the decoder through jump connections, so that the feature information of different levels can be fully utilized, which includes both low-level detail information and high-level semantic information, which helps to improve the accuracy of segmentation. As Figure 3As shown, the embodiment of the present invention adopts a modified UNet segmentation model.

[0076] See also Figure 3 The modified UNet segmentation model specifically modifies the network structure of the downsampling part of the UNet segmentation model, replacing the original multi-layer convolution and pooling layers with a multi-layer stack of micro-step convolution and dilated convolution modules fused through a gating mechanism. This micro-step convolution and dilated convolution module fused through a gating mechanism divides the input into four branches after passing through a 1*1 convolution and a maximum pooling layer, and passes through different micro-step convolution layers or dilated convolution layers respectively, and finally fuses them through a gating mechanism. Micro-step convolution has a smaller receptive field than ordinary convolution and can better process detailed information; dilated convolution has a larger receptive field than ordinary convolution and has a good effect in sensing a wide range of information. The gating mechanism fuses the two and makes them compensate for each other, so that the module can adaptively sense features of interest of different sizes and ranges in actual medical images, achieving better feature extraction effects.

[0077] See also Figure 3 The modified UNet segmentation model also introduces an uncertainty attention module. The specific calculation method is as follows: Input medical image tensor X i And uncertainty map U, perform 1×1 convolution operation on the medical image to obtain matrix Q, the uncertainty map is divided into two branches after passing through the maximum pooling layer and 1×1 convolution, and perform 1×1 convolution operation on the matrix K and the matrix V respectively. Then the attention matrix A is calculated as A=Softmax(K·V), matrix Q and attention matrix A are matrix multiplied, and after passing through the multi-layer perceptron, they are combined with the input medical image x i Perform a jump connection to get the output Output, the process can be expressed as

[0078] For the sake of brevity, some symbols of the formula of the embodiment of the present invention are given here: i represents the sample index position; C represents that all samples contain pixels of C categories in total.

[0079] The embodiment of the present invention first trains the neural network. The specific training process is as follows: each sample is defined as a sample obtained by sampling from the population independently and identically distributed, and the category y corresponding to each pixel of the sample is defined. i Obeying the category distribution, Dirichlet distribution is used as the conjugate prior of the category distribution likelihood function. By training the neural network through the intermediate output of the evidence head function, the evidence matrix corresponding to each category of the sample is output, that is:

[0080] e i =f(x i ;w)

[0081] Among them, x i is the i-th sample input, f(·) is the function form of the neural network, and w is the parameter of the neural network;

[0082] Calculate the parameter α corresponding to the Dirichlet distribution through the evidence matrix i =e i +1, by computing the expected tensor p of the Dirichlet distribution i =α i / S i ;

[0083] Among them, S i pass By calculating, we can get an estimate of the category probability.

[0084] Based on subjective logic theory, the category uncertainty estimate u can be calculated by the formula i =C / S i The category with the highest probability is considered the predicted category, and the predicted segmentation result is: prediction i =argmax p i .

[0085] The medical image segmentation task can be regarded as a pixel-level classification task. For a basic classification task, the likelihood loss function can be written as follows:

[0086]

[0087] in, One-hot encoding of the Ground Truth representing the data, represents the model's predicted probability for category c. In the framework of evidence theory, by associating the Dirichlet distribution with the confidence distribution, the above formula (1) can be rewritten as the following formula (2):

[0088]

[0089] In order to ensure that samples that cannot be correctly classified generate less evidence, the KL divergence regularization term is introduced for constraint, and its calculation formula is as follows (3):

[0090]

[0091] Where Γ(α) is the Gamma function, Represents the adjusted parameters to ensure that the Ground Truth evidence does not mistakenly become 0.

[0092] Subsequently, in order to constrain uncertainty and balance the prediction probability and uncertainty range, the present invention proposes The regularization term is calculated as follows (4):

[0093]

[0094] The final loss function can be expressed as:

[0095]

[0096] Among them, λ1 = min(1,10×epoch / total_epochs) is the decay coefficient, whose purpose is to prevent the model from converging too early and causing insufficient model learning. epoch is the current training round number, and total_epochs is the total number of training rounds. λ2 is an empirical hyperparameter, and the recommended value is λ2 = 0.5.

[0097] In summary, for a single medical image sample, the trusted medical image segmentation model proposed in the embodiment of the present invention uses the following function as the optimization target:

[0098]

[0099] In the embodiment of the present invention, based on the optimization goal expressed in the above formula (6), the Adam optimizer is used to train a credible medical image segmentation model.

[0100] In traditional evidence deep learning, since it is necessary to ensure that all the output evidence is not less than 0, a Relu function or a Softplus function is usually added to the last layer of the neural network as the evidence head of the neural network so that f(x i ; w) ≥ 0. In order to enable the model to estimate greater uncertainty, it is usually necessary to enable it to express a smaller amount of evidence. Therefore, this paper proposes a new activation function as the evidence head of the neural network, which is calculated as follows (7):

[0101] activation(x)=e -Relu(x) +Relu(x)-1(7)

[0102] The calculation formula of the ReLU function is as follows (8):

[0103]

[0104] The new activation function proposed in this paper has some favorable properties compared to the ReLU function: when x is not greater than 0, the function value is 0; when x is greater than 0, the value of the activation function is always smaller than the ReLU function, and when x is large, the function approaches the ReLU function. Therefore, the activation function is good for both high and low evidence samples.

[0105] In addition, simply modeling and quantifying uncertainty in this way will lead to a decrease in segmentation accuracy. Therefore, this paper introduces an algorithm that uses evidence uncertainty to gradually guide the segmentation process of a credible medical image segmentation model. Figure 1-Figure 2 The embodiment of the present invention proposes a trustworthy medical image segmentation method based on progressive guidance of evidence uncertainty, which mainly includes the following execution steps:

[0106] S1, obtain training samples and corresponding labels;

[0107] S2. training a credible medical image segmentation model based on the training samples and corresponding labels;

[0108] S3. Obtain test samples;

[0109] S4, inputting the test sample into the trained credible medical image segmentation model for prediction, and calculating the parameter α of the Dirichlet distribution i , calculate the expected tensor p of the Dirichlet distribution i =α i / S i , S i is the parameter α i sum;

[0110] S5. Calculate the segmentation evidence uncertainty matrix U = C / S i , C is the total number of segmentation categories;

[0111] S6. Let the category with the largest category probability be the predicted category, and the predicted segmentation result output by the credible medical image segmentation model based on the test sample is: prediction i =argmax p i ;

[0112] S7, predict the segmentation result of the test sample i and the segmentation evidence uncertainty matrix U for visualization.

[0113] Wherein, in S2, the training of the credible medical image segmentation model further includes:

[0114] S21, input the training sample x i , the size of the two-dimensional tensor is m×n;

[0115] S22. Initialize the evidence uncertainty matrix to be a full-1 matrix U, and the size of the two-dimensional tensor is the same as the training sample x. i The two-dimensional tensors of are the same size;

[0116] S23, the evidence matrix e obtained based on the credible medical image segmentation model and the evidence uncertainty matrix predictioni =f(x i ,w), calculate the parameter α of Dirichlet distribution according to subjective logic theory i , and use this to calculate the evidence uncertainty matrix u for this segmentation i :

[0117] α i =e i +1,u i =C / S i

[0118] in, C is the total number of categories included in all sample pixels, i represents the sample index position, f(·) is the function form of the neural network, and w is the parameter of the neural network;

[0119] S24. Update the evidence uncertainty matrix U to u i ;

[0120] S25. Calculate the updated evidence uncertainty matrix u i The difference between the original evidence uncertainty matrix U is used to determine whether the following conditions are met:

[0121]

[0122] Among them, ε is a given sufficiently small positive number;

[0123] If the condition is satisfied, the credible medical image segmentation model is considered to have converged, and the subsequent steps are continued. If the condition is not satisfied, the process returns to S23 until the above condition is satisfied, at which point the calculated evidence uncertainty tends to be stable.

[0124] S26. Calculate the parameter α of the Dirichlet distribution based on the prediction of the credible medical image segmentation model. i , calculate the expected tensor p of the Dirichlet distribution i =α i / S i ;

[0125] S27, let the category with the largest category probability be the predicted category, then the predicted segmentation result output by the credible medical image segmentation model based on the training sample is: prediction i =argmax p i ;

[0126] S28, calculating the loss based on the predicted segmentation result obtained in S27 and the evidence uncertainty obtained in S25, and performing back propagation to update the weight of the credible medical image segmentation model;

[0127] S29. Repeat the above training steps until the trusted medical image segmentation model reaches the maximum number of iterations or the loss no longer decreases.

[0128] See also Figure 4 ,Comparing this medical image segmentation method with several current mainstream ,medical image segmentation methods, it can be seen that the method proposed in ,the present invention has better segmentation accuracy and more accurate ,estimation of the uncertainty of the prediction results.

[0129] In summary, the embodiment of the present invention models the category uncertainty in the medical image segmentation task based on the subjective logic theory, and explicitly models the uncertainty by including the uncertainty quality in the subjective opinion, thereby having the ability to reflect whether the given probability is reliable. At the same time, the embodiment of the present invention proposes an evidence uncertainty progressively guided segmentation algorithm, which fully utilizes the guiding role of evidence uncertainty in the segmentation process and improves the segmentation accuracy. The method adopted by the embodiment of the present invention not only greatly reduces the computational cost, but also provides a more accurate uncertainty estimate. The solution provided by the embodiment of the present invention can not only complete the uncertainty estimation through a single forward propagation through a single neural network, but also provide a robust and calibrated uncertainty estimate when facing out-of-distribution data. Therefore, the embodiment of the present invention has the advantage of reducing computational costs compared to Bayesian neural networks, deep neural networks using Monte Carlo Dropout, and integration methods, and has better segmentation accuracy compared to calibration methods and evidence deep learning methods.

[0130] An embodiment of the present invention further provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the above-mentioned method for credible medical image segmentation guided by progressive evidence uncertainty.

[0131] See also Figure 5 An embodiment of the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the above-mentioned method for credible medical image segmentation guided by progressive evidence uncertainty.

[0132] An embodiment of the present invention further provides a computer program product comprising instructions, which, when executed on a computer, enables the computer to execute the steps of the above-mentioned method for trustworthy medical image segmentation guided by progressive evidence uncertainty.

[0133] It is understandable that the system, device and storage medium provided in the embodiments of the present invention correspond to the method provided in the embodiments of the present invention. The explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts of the above-mentioned reliable medical image segmentation method with progressive guidance of evidence uncertainty.

[0134] It should be noted that those skilled in the art will understand that all or part of the steps implemented in the embodiments of the present invention can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using hardware, it can be implemented in whole or in part in the form of purchased standard parts or modified parts. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiments of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media integrated therein. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a DVD), or a semiconductor medium (eg, a solid state disk (SSD)).

[0135] It should be understood that the examples and implementation methods described herein are for illustrative purposes only and are not intended to limit the present invention. Those skilled in the art may make various modifications or changes based on them. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A trustworthy medical image segmentation method guided by progressive evidence uncertainty, characterized by: The following steps are involved: S1, obtain training samples and corresponding labels; S2. training a credible medical image segmentation model based on the training samples and corresponding labels; S3. Obtain test samples; S4, inputting the test sample into the trained credible medical image segmentation model for prediction, and calculating the parameters of the Dirichlet distribution , calculate the expected tensor of the Dirichlet distribution , S i For parameters sum; S5. Calculate the segmentation evidence uncertainty matrix , C is the total number of segmentation categories; S6. Let the category with the largest category probability be the predicted category, and the predicted segmentation result output by the credible medical image segmentation model based on the test sample: prediction i =argmax p i ; S7, the predicted segmentation result of the test sample and the segmentation evidence uncertainty matrix Perform visualization; In S2, the training of the credible medical image segmentation model further includes: S21. Input the training sample , the size of the two-dimensional tensor is ; S22. Initialize the evidence uncertainty matrix to an all-1 matrix , a two-dimensional tensor with the same size as the training sample The two-dimensional tensors of are the same size; S23, an evidence matrix predicted based on the credible medical image segmentation model and the evidence uncertainty matrix , calculate the parameter α of the Dirichlet distribution according to the subjective logic theory i , and use this to calculate the evidence uncertainty matrix u for this segmentation i : ， in, , C is the total number of categories included in all sample pixels, i represents the sample index position, is the functional form of the neural network, are the parameters of the neural network; S24. Update the evidence uncertainty matrix for ; S25. Calculate the updated evidence uncertainty matrix u i The difference between the original evidence uncertainty matrix U is used to determine whether the following conditions are met: in, is a given sufficiently small positive number; If the condition is satisfied, the credible medical image segmentation model is considered to have converged, and the subsequent steps are continued. If the condition is not satisfied, the process returns to S23 until the above condition is satisfied, at which point the calculated evidence uncertainty tends to be stable. S26. Calculate the parameters of the Dirichlet distribution based on the prediction of the credible medical image segmentation model. , calculate the expected tensor of the Dirichlet distribution ; S27, let the category with the largest category probability be the predicted category, then the predicted segmentation result output by the credible medical image segmentation model based on the training sample is: prediction i =argmax p i ; S28, calculating the loss based on the predicted segmentation result obtained in S27 and the evidence uncertainty obtained in S25, and performing back propagation to update the weight of the credible medical image segmentation model; S29. Repeat the above training steps until the trusted medical image segmentation model reaches the maximum number of iterations or the loss no longer decreases.

2. The method for credible medical image segmentation guided by progressive evidence uncertainty according to claim 1, characterized in that: In S2, the credible medical image segmentation model is optimized To learn, that is, To optimize the goal; in, is the loss function, is the evidence cross entropy loss function, is the KL divergence regularization loss function, is the uncertainty regularization loss function, λ1 is the attenuation coefficient, and λ2 is the empirical hyperparameter.

3. The method for credible medical image segmentation guided by progressive evidence uncertainty according to claim 2, wherein: In S2, if the optimization objective is explained by the loss function of a single medical image sample, the medical image segmentation task is regarded as a pixel-level classification task. For a basic classification task, the likelihood loss function is written as follows: （1） Among them, C is the total number of segmentation categories, x is the xth sample, One-hot encoding of the Ground Truth representing the data, represents the predicted probability of the credible medical image segmentation model for category C; In the framework of evidence theory, by associating Dirichlet distribution with confidence distribution, the above formula (1) can be rewritten as the following formula (2): （2） in, is the evidence cross entropy loss function, represents the Dirichlet distribution parameter of the Cth type of sample, and S is the sum of the Dirichlet distribution parameters; In order to ensure that samples that cannot be correctly classified generate less evidence, the KL divergence regularization term is introduced for constraint. The calculation formula is as follows (3): （3） in, is the KL divergence regularization loss function, is the Gamma function, is the Digamma function, , represents the adjusted parameter, which is used to ensure that the evidence of Ground Truth does not mistakenly become 0; In order to constrain uncertainty and balance the prediction probability and uncertainty range, the uncertainty regularization term is introduced, and the calculation formula is as follows (4): （4） in, is the uncertainty regularization loss function; Combining the above equations (1) to (4), we get the final loss function as follows (5): （5） Among them, λ1 is the attenuation coefficient, which is used to prevent the credible medical image segmentation model from converging prematurely, thereby avoiding insufficient learning of the credible medical image segmentation model. , epoch is the current number of training rounds, total_epochs is the total number of training rounds, and λ2 is the empirical hyperparameter.

4. The device is characterized in that The invention comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the credible medical image segmentation method guided by progressive evidence uncertainty as described in any one of claims 1 to 3.

Images