An uncertainty estimation system for medical image segmentation

Abstract

A medical image segmentation uncertainty estimation system belongs to the field of medical image segmentation technology. The present invention solves the problems of poor accuracy, versatility and flexibility of existing medical image segmentation uncertainty estimation technologies. The present invention establishes a probability space distribution of data through logical mapping of data, which can achieve better modeling of the semantic rationality and the existing model uncertainty and arbitrary uncertainty in the segmentation results, thereby improving the accuracy of the uncertainty estimation technology. Moreover, the uncertainty perception module M2 of the present invention can be flexibly integrated into the existing segmentation framework, that is, it can be connected to any deterministic medical image segmentation backbone network M1, and the two can be connected together for use to form an uncertainty-aware medical image segmentation model, thereby achieving a universal improvement to the deterministic medical image segmentation algorithm. The method of the present invention can be applied to medical image segmentation.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image segmentation, and in particular relates to a medical image segmentation uncertainty estimation system. Background Art

[0002] With the rapid development of medical image analysis technology, medical image segmentation plays an increasingly important role in clinical diagnosis and treatment. However, traditional medical image segmentation algorithms are often unable to effectively deal with fuzzy boundaries, artifacts, and subjective differences between different annotators in the image, resulting in uncertainty in the segmentation results. This uncertainty not only affects the accuracy of medical image analysis, but may also have a negative impact on subsequent clinical decisions. Therefore, effective estimation and processing of uncertainty in medical image segmentation has become a technical problem that needs to be solved urgently. Existing uncertainty estimation technologies usually rely on specific models and training data, lack versatility and flexibility, and the accuracy of uncertainty estimation is poor, which cannot meet the diverse needs of clinical applications. Therefore, proposing a new uncertainty estimation technology for medical image segmentation is a problem that needs to be solved urgently. Summary of the Invention

[0003] The purpose of the present invention is to solve the problems of poor accuracy, versatility and flexibility of existing medical image segmentation uncertainty estimation technologies, and to propose a medical image segmentation uncertainty estimation system.

[0004] The technical solution adopted by the present invention to solve the above technical problems is: a medical image segmentation uncertainty estimation system, the system includes a medical image preprocessing module and an uncertainty-aware medical image segmentation model M, the uncertainty-aware medical image segmentation model M includes a medical image segmentation backbone network module M1 and an uncertainty-aware module M2; wherein:

[0005] The medical image preprocessing module is used to preprocess the diseased medical image to obtain the preprocessed diseased medical image;

[0006] The medical image segmentation backbone network module M1 is used to process the input pre-processed diseased medical image and output a logic map; and the logic map is used as the input of the uncertainty perception module M2;

[0007] The uncertainty perception module M2 is used to process the input logic map to obtain various segmentation results and uncertainty estimation results for the diseased medical image.

[0008] Furthermore, the working process of the medical image preprocessing module is as follows:

[0009] Step S11, performing denoising processing on the diseased medical image to obtain a denoised diseased medical image;

[0010] Step S12: cropping the denoised diseased medical image, that is, cropping an image of size N×N centered on the lesion area from the denoised diseased medical image;

[0011] Step S13: normalize the pixel values ​​in the cropped image to obtain a normalized image.

[0012] Furthermore, the structure of the medical image segmentation backbone network module M1 is: all layers before the last activation function layer in the U-Net segmentation network;

[0013] The logical mapping diagram output by the medical image segmentation backbone network module M1 is the output of the previous layer of the last activation function layer in the U-Net segmentation network.

[0014] Furthermore, the uncertainty perception module M2 includes two parallel branches, and each branch includes a Gaussian distribution modeling module M LMG and uncertainty measurement module M UMS .

[0015] Furthermore, the Gaussian distribution modeling module M LMG It includes three parallel convolutional layers, Gaussian distribution modeling module M LMG It is used to perform pixel-by-pixel joint probability Gaussian distribution modeling on the logical map, and the Gaussian distribution modeling module M LMG The working process is:

[0016] Step S21: Pass the logistic map through the first convolutional layer to obtain the output μ of the first convolutional layer; pass the logistic map through the second convolutional layer to obtain the output P of the second convolutional layer; pass the logistic map through the third convolutional layer to obtain the output D of the third convolutional layer;

[0017] μ=conv1(η)

[0018] P = conv2(η)

[0019] D = conv3(η)

[0020] Where η represents the logistic map, conv1(·) represents the first convolutional layer, conv2(·) represents the second convolutional layer, and conv3(·) represents the third convolutional layer.

[0021] Step S22: Calculate the covariance matrix Σ based on P and D, that is, obtain the pixel-by-pixel joint probability Gaussian distribution N(μ,Σ) of the logical map.

[0022] Furthermore, the covariance matrix Σ is calculated based on P and D, specifically:

[0023] Σ=P×P T +D

[0024] Among them, P T is the transpose of P.

[0025] Furthermore, the uncertainty measurement module M UMS The working process is:

[0026] Step S31: Sampling the input pixel-by-pixel joint probability Gaussian distribution to obtain n / 2 sampling results, and recording the i-th sampling result as Z i ;

[0027] Step S32: The i-th sampling result Z i After the evidence layer, the Dirichlet distribution parameters when the segmentation result of the pixel (h, w) in the i-th sampling result is foreground are obtained And the Dirichlet distribution parameter when the segmentation result is the background

[0028] Step S33, use Calculate the probability that the segmentation result of pixel (h,w) is foreground use Calculate the probability that the segmentation result of pixel (h,w) is background

[0029]

[0030] Step S34: and Get the segmentation result Y of pixel (h,w) h,w :

[0031]

[0032] Where j = 0 or 1, if Y h,w = 0, then the pixel (h, w) is the foreground, if Y h,w =1, then the pixel (h, w) is the background; the segmentation results of all pixels in the i-th sampling result are used to form the segmentation result corresponding to the i-th sampling result;

[0033] Step S35: Based on the segmentation result Y of the pixel (h, w) h,w Calculate the uncertainty U of pixel (h,w) A(h,w) :

[0034]

[0035] The uncertainty of all pixels in the i-th sampling result is used to form the arbitrary uncertainty corresponding to the i-th sampling result;

[0036] Step S36: Using the method from step S32 to step S35, obtain n / 2 segmentation results of the image and the arbitrary uncertainties corresponding to the n / 2 segmentation results.

[0037] Furthermore, the specific process of step S32 is as follows:

[0038]

[0039] Among them, Softplus(·) is the Softplus activation function layer.

[0040] Furthermore, each branch of the uncertainty perception module M2 obtains n / 2 segmentation results of the image, and then processes the n segmentation results of the two branches to obtain model uncertainty;

[0041] The calculation method of the model uncertainty is:

[0042] According to the segmentation results of the pixel (h, w) in each segmentation result of the image, the variance of the segmentation result of the pixel (h, w) is calculated, and the variance of the segmentation results of all pixels together constitutes the model uncertainty result.

[0043] Furthermore, the training process of the uncertainty-aware medical image segmentation model M is as follows:

[0044] Step 1: Obtain a dataset of diseased medical images, perform denoising on each diseased medical image in the dataset, and then perform data enhancement on the denoised images to obtain an enhanced dataset of diseased medical images;

[0045] Step 2: Perform cropping processing on each diseased medical image obtained in step 1 to obtain a cropped diseased medical image dataset; and perform cropping processing on the segmentation label corresponding to each diseased medical image obtained in step 1 to obtain a cropped segmentation label corresponding to each diseased medical image;

[0046] The cropped segmentation label set corresponding to a diseased medical image is recorded as Y = {Y1, Y2, ..., Y n}, n is the number of segmentation labels corresponding to each diseased medical image, and the training set is composed of each cropped diseased medical image and the cropped segmentation labels corresponding to each cropped diseased medical image;

[0047] Step 3: Use the training set obtained in step 2 to train the uncertainty-aware medical image segmentation model M. The loss function L used in the training process is:

[0048]

[0049] Among them, L D represents the loss of the first branch, L′ D represents the loss of the second branch;

[0050]

[0051] Among them, x k represents the kth diseased medical image in the input training set, k = 1, 2, ..., K, K represents the total number of images in the training set, η k Represents a diseased medical image x k The output after the medical image segmentation backbone network module M1, P(η k |x k ) represents the diseased medical image x k After passing through the medical image segmentation backbone network module M1, the output is η k The probability of y i,j represents the j-th predicted segmentation result in the i-th branch, P(y i,j |η k ) indicates that the j-th predicted segmentation result in the i-th branch is y i,j probability;

[0052] The training is stopped until the loss function converges, and the trained uncertainty-aware medical image segmentation model M is obtained.

[0053] The beneficial effects of the present invention are:

[0054] The present invention selects the existing medical image segmentation backbone network M1 based on deep learning, establishes the probability space distribution of data through the logical mapping of data, and can achieve better modeling for the semantic rationality of the segmentation results and the existing model uncertainty and arbitrary uncertainty, thereby improving the accuracy of the uncertainty estimation technology and providing more reliable support for medical image analysis. Moreover, the uncertainty perception module M2 of the present invention can be flexibly integrated into the existing segmentation framework, that is, it can be connected to any deterministic medical image segmentation backbone network M1, and the two can be connected together for use to form an uncertainty-aware medical image segmentation model, thereby achieving a universal improvement to the deterministic medical image segmentation algorithm. Since the uncertainty perception module of the present invention can be plug-and-play after the deterministic medical image segmentation algorithm, the versatility and flexibility of the existing uncertainty estimation technology are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a flow chart of a medical image segmentation uncertainty estimation system of the present invention;

[0056] Figure 2 is a schematic diagram of the uncertainty perception module M2;

[0057] The figure takes n=6 as an example;

[0058] Figure 3 It is the logistic mapping Gaussian distribution modeling module M LMG Schematic diagram of;

[0059] Figure 4 is the uncertainty measurement module M UMS Schematic diagram of . DETAILED DESCRIPTION

[0060] Specific implementation method 1: Combination Figure 1 and Figure 2 This embodiment describes a medical image segmentation uncertainty estimation system, which includes a medical image preprocessing module and an uncertainty-aware medical image segmentation model M. The uncertainty-aware medical image segmentation model M includes a medical image segmentation backbone network module M1 and an uncertainty-aware module M2; wherein:

[0061] The medical image preprocessing module is used to preprocess the diseased medical image to obtain the preprocessed diseased medical image;

[0062] The medical image segmentation backbone network module M1 is used to process the input pre-processed diseased medical image and output a logic map; and the logic map is used as the input of the uncertainty perception module M2;

[0063] The uncertainty perception module M2 is used to process the input logic map to obtain various segmentation results and uncertainty estimation results for the diseased medical image.

[0064] Specific embodiment 2: This embodiment differs from specific embodiment 1 in that the working process of the medical image preprocessing module is as follows:

[0065] Step S11, performing denoising processing on the diseased medical image to obtain a denoised diseased medical image;

[0066] Step S12: cropping the denoised diseased medical image, that is, cropping an image of size N×N centered on the lesion area from the denoised diseased medical image;

[0067] Step S13: normalize the pixel values ​​in the cropped image to obtain a normalized image.

[0068] Other steps and parameters are the same as those in the first embodiment.

[0069] Specific embodiment three: This embodiment differs from specific embodiment one or two in that the structure of the medical image segmentation backbone network module M1 is: all layers before the last activation function layer in the U-Net segmentation network;

[0070] That is, the logical mapping diagram output by the medical image segmentation backbone network module M1 is the output of the previous layer of the last activation function layer in the U-Net segmentation network.

[0071] Other steps and parameters are the same as those in the first or second embodiment.

[0072] The medical image segmentation backbone network module M1 in this embodiment can also be all layers before the last activation function layer in other traditional segmentation networks, and the logical mapping diagram is the intermediate result obtained by the traditional segmentation network before the last activation layer.

[0073] Specific embodiment 4: This embodiment differs from any one of specific embodiments 1 to 3 in that the uncertainty perception module M2 includes two parallel branches, and each branch includes a Gaussian distribution modeling module M LMG and uncertainty measurement module M UMS .

[0074] The other steps and parameters are the same as those in the first to third embodiments.

[0075] Specific implementation method five: Combination Figure 3 The present embodiment is different from the first to fourth embodiments in that the Gaussian distribution modeling module M LMG It includes three parallel convolutional layers, Gaussian distribution modeling module M LMG It is used to perform pixel-by-pixel joint probability Gaussian distribution modeling on the logical map, and the Gaussian distribution modeling module M LMG The working process is:

[0076] Step S21: Pass the logistic map through the first convolutional layer to obtain the output μ of the first convolutional layer; pass the logistic map through the second convolutional layer to obtain the output P of the second convolutional layer; pass the logistic map through the third convolutional layer to obtain the output D of the third convolutional layer;

[0077] Among them, the output μ of the first convolution layer is used as the mean of the Gaussian distribution, the output P of the second convolution layer is used as the low-rank factor of the Gaussian distribution, and the output D of the third convolution layer is used as the diagonal element of the Gaussian distribution;

[0078] μ=conv1(η)

[0079] P = conv2(η)

[0080] D = conv3(η)

[0081] Where η represents the logistic map, conv1(·) represents the first convolutional layer, conv2(·) represents the second convolutional layer, and conv3(·) represents the third convolutional layer.

[0082] Step S22: Calculate the covariance matrix Σ based on P and D, that is, obtain the pixel-by-pixel joint probability Gaussian distribution N(μ,Σ) of the logical map.

[0083] The other steps and parameters are the same as those in the first to fourth embodiments.

[0084] This implementation can reduce the size of computing resources occupied by the covariance matrix.

[0085] Specific embodiment 6: This embodiment differs from any one of specific embodiments 1 to 5 in that the covariance matrix Σ is calculated based on P and D, specifically:

[0086] Σ=P×P T +D

[0087] Among them, P T is the transpose of P.

[0088] The other steps and parameters are the same as those in the first to fifth embodiments.

[0089] Specific implementation method seven: combination Figure 4 The difference between this embodiment and any one of the specific embodiments 1 to 6 is that the uncertainty measurement module M UMS The working process is:

[0090] Step S31: Sample the input pixel-by-pixel joint probability Gaussian distribution to obtain n / 2 sampling results, and record the i-th sampling result as Z i ;

[0091] Step S32: The i-th sampling result Z i After the evidence layer (Softplus activation function layer), the Dirichlet distribution parameters when the segmentation result of pixel (h, w) in the i-th sampling result is foreground are obtained And the Dirichlet distribution parameter when the segmentation result is the background

[0092] Step S33, use Calculate the probability that the segmentation result of pixel (h,w) is foreground use Calculate the probability that the segmentation result of pixel (h,w) is background

[0093]

[0094] Step S34: and Get the segmentation result Y of pixel (h,w) h,w :

[0095]

[0096] Where j = 0 or 1, if Y h,w = 0, then the pixel (h, w) is the foreground, if Y h,w =1, then the pixel (h, w) is the background; the segmentation results of all pixels in the i-th sampling result are used to form the segmentation result corresponding to the i-th sampling result;

[0097] Step S35: Based on the segmentation result Y of the pixel (h, w) h,w Calculate the uncertainty U of pixel (h,w) A(h,w) :

[0098]

[0099] The uncertainty of all pixels in the i-th sampling result is used to form the arbitrary uncertainty corresponding to the i-th sampling result;

[0100] Step S36: Using the method from step S32 to step S35, obtain n / 2 segmentation results of the image and the arbitrary uncertainties corresponding to the n / 2 segmentation results.

[0101] The other steps and parameters are the same as those in the first to sixth embodiments.

[0102] Specific embodiment eight: This embodiment differs from any one of specific embodiments one to seven in that the specific process of step S32 is as follows:

[0103]

[0104] Among them, Softplus(·) is the Softplus activation function layer.

[0105] The other steps and parameters are the same as those in the first to seventh embodiments.

[0106] Specific embodiment 9: This embodiment differs from any one of specific embodiments 1 to 8 in that each branch of the uncertainty perception module M2 obtains n / 2 segmentation results of the image, and then processes the n segmentation results of the two branches to obtain model uncertainty;

[0107] The calculation method of the model uncertainty is:

[0108] According to the segmentation results of pixel (h, w) in each segmentation result of the image, the variance of the segmentation result of pixel (h, w) is calculated, and the variance of the segmentation results of all pixels together constitute the model uncertainty result (the variance of the segmentation result of pixel (h, w) is used as the uncertainty of pixel (h, w) in the model uncertainty result).

[0109] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0110] In the present invention, the n segmentation results obtained by the two branches are multiple reasonable segmentation results for the lesions in the diseased image. Assuming n = 6, if the segmentation results of the first to sixth segmentation results for pixel (h, w) are 0 (that is, in the first to sixth segmentation results, pixels (h, w) are all foreground), then the variance corresponding to pixel (h, w) is 0 (calculated based on 6 zeros), that is, in the overall model uncertainty, the value corresponding to pixel (h, w) is 0. Similarly, the variance corresponding to each pixel can be calculated, that is, the value corresponding to each pixel in the overall model uncertainty is obtained. Arbitrary uncertainty and model uncertainty together constitute the uncertainty of medical image segmentation.

[0111] Specific embodiment 10: This embodiment differs from any one of specific embodiments 1 to 9 in that the training process of the uncertainty-aware medical image segmentation model M is as follows:

[0112] Step 1: Obtain a dataset of diseased medical images, perform denoising on each diseased medical image in the dataset, and then perform data enhancement on the denoised images to obtain an enhanced dataset of diseased medical images;

[0113] Step 2: Perform cropping processing on each diseased medical image obtained in step 1 to obtain a cropped diseased medical image dataset; and perform cropping processing on the segmentation label corresponding to each diseased medical image obtained in step 1 to obtain a cropped segmentation label corresponding to each diseased medical image;

[0114] The cropping method of the image and label is the same as step S12;

[0115] The cropped segmentation label set corresponding to a diseased medical image is recorded as Y = {Y1, Y2, ..., Y n}, n is the number of reasonable segmentation labels corresponding to each diseased medical image, and the training set is composed of each cropped diseased medical image and the cropped segmentation labels corresponding to each cropped diseased medical image;

[0116] Step 3: Use the training set obtained in step 2 to train the uncertainty-aware medical image segmentation model M. The loss function L used in the training process is:

[0117]

[0118] Among them, L D represents the loss of the first branch, L′ D represents the loss of the second branch;

[0119]

[0120] Among them, x k represents the kth diseased medical image in the input training set, k = 1, 2, ..., K, K represents the total number of images in the training set, η k Represents a diseased medical image x k The output after the medical image segmentation backbone network module M1, P(η k |x k ) represents the diseased medical image x k After passing through the medical image segmentation backbone network module M1, the output is η k The probability of y i,j represents the j-th predicted segmentation result in the i-th branch, P(y i,j |η k ) indicates that the j-th predicted segmentation result in the i-th branch is y i,j The probability of ,the integral is calculated by Monte Carlo sampling method;

[0121] The loss of the second branch is similar to that of the first branch, and the logarithm value can be 10;

[0122] The training is stopped until the loss function converges, and the trained uncertainty-aware medical image segmentation model M is obtained.

[0123] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0124] The above examples are merely illustrative of the calculation model and process of the present invention and are not intended to limit the embodiments of the present invention. Persons skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. This list of embodiments is not exhaustive; however, any obvious variations or modifications derived from the technical solution of the present invention remain within the scope of protection of the present invention.

Claims

1. A medical image segmentation uncertainty estimation system, characterized in that: The system includes a medical image preprocessing module and an uncertainty-aware medical image segmentation model M, wherein the uncertainty-aware medical image segmentation model M includes a medical image segmentation backbone network module M1 and an uncertainty-aware module M2; wherein: The medical image preprocessing module is used to preprocess the diseased medical image to obtain the preprocessed diseased medical image; The medical image segmentation backbone network module M1 is used to process the input pre-processed diseased medical image and output a logic map; and the logic map is used as the input of the uncertainty perception module M2; The uncertainty perception module M2 is used to process the input logic map to obtain various segmentation results and uncertainty estimation results of the diseased medical image; The uncertainty perception module M2 includes two parallel branches, and each branch includes a Gaussian distribution modeling module M LMG and uncertainty measurement module M UMS ; The Gaussian distribution modeling module M LMG It includes three parallel convolutional layers, Gaussian distribution modeling module M LMG It is used to perform pixel-by-pixel joint probability Gaussian distribution modeling on the logical map, and the Gaussian distribution modeling module M LMG The working process is: Step S21: Pass the logistic map through the first convolutional layer to obtain the output μ of the first convolutional layer; pass the logistic map through the second convolutional layer to obtain the output P of the second convolutional layer; pass the logistic map through the third convolutional layer to obtain the output D of the third convolutional layer; μ=conv1(η) P = conv2(η) D = conv3(η) Where η represents the logistic map, conv1(·) represents the first convolutional layer, conv2(·) represents the second convolutional layer, and conv3(·) represents the third convolutional layer. Step S22: Calculate the covariance matrix Σ based on P and D, that is, obtain the pixel-by-pixel joint probability Gaussian distribution N(μ,Σ) of the logical map; The uncertainty measurement module M UMS The working process is: Step S31: Sampling the input pixel-by-pixel joint probability Gaussian distribution to obtain n / 2 sampling results, and recording the i-th sampling result as Z i ; Step S32: The i-th sampling result Z i After the evidence layer, the Dirichlet distribution parameters when the segmentation result of the pixel (h, w) in the i-th sampling result is foreground are obtained And the Dirichlet distribution parameter when the segmentation result is the background Step S33, use Calculate the probability that the segmentation result of pixel (h,w) is foreground use Calculate the probability that the segmentation result of pixel (h,w) is background Step S34: and Get the segmentation result Y of pixel (h,w) h,w : Where j = 0 or 1, if Y h,w = 0, then the pixel (h, w) is the foreground, if Y h,w =1, then the pixel (h, w) is the background; the segmentation results of all pixels in the i-th sampling result are used to form the segmentation result corresponding to the i-th sampling result; Step S35: Based on the segmentation result Y of the pixel (h, w) h,w Calculate the uncertainty U of pixel (h,w) A(h,w) : The uncertainty of all pixels in the i-th sampling result is used to form the arbitrary uncertainty corresponding to the i-th sampling result; Step S36: Using the method from step S32 to step S35, obtain n / 2 segmentation results of the image and arbitrary uncertainties corresponding to the n / 2 segmentation results; Each branch of the uncertainty perception module M2 obtains n / 2 segmentation results of the image, and then processes the n segmentation results of the two branches to obtain model uncertainty; The calculation method of the model uncertainty is: According to the segmentation results of the pixel (h, w) in each segmentation result of the image, the variance of the segmentation result of the pixel (h, w) is calculated, and the variance of the segmentation results of all pixels together constitutes the model uncertainty result.

2. A medical image segmentation uncertainty estimation system according to claim 1, characterized in that: The working process of the medical image preprocessing module is as follows: Step S11, performing denoising processing on the diseased medical image to obtain a denoised diseased medical image; Step S12: cropping the denoised diseased medical image, that is, cropping an image of size N×N centered on the lesion area from the denoised diseased medical image; Step S13: normalize the pixel values ​​in the cropped image to obtain a normalized image.

3. A medical image segmentation uncertainty estimation system according to claim 2, characterized in that: The structure of the medical image segmentation backbone network module M1 is: all layers before the last activation function layer in the U-Net segmentation network; The logical mapping diagram output by the medical image segmentation backbone network module M1 is the output of the previous layer of the last activation function layer in the U-Net segmentation network.

4. A medical image segmentation uncertainty estimation system according to claim 3, characterized in that: The covariance matrix Σ is calculated based on P and D, specifically: Σ=P×P T +D Among them, P T is the transpose of P.

5. The medical image segmentation uncertainty estimation system according to claim 4, characterized in that: The specific process of step S32 is as follows: Among them, Softplus(·) is the Softplus activation function layer.

6. The medical image segmentation uncertainty estimation system according to claim 5, characterized in that: The training process of the uncertainty-aware medical image segmentation model M is as follows: Step 1: Obtain a dataset of diseased medical images, perform denoising on each diseased medical image in the dataset, and then perform data enhancement on the denoised images to obtain an enhanced dataset of diseased medical images; Step 2: Perform cropping processing on each diseased medical image obtained in step 1 to obtain a cropped diseased medical image dataset; and perform cropping processing on the segmentation label corresponding to each diseased medical image obtained in step 1 to obtain a cropped segmentation label corresponding to each diseased medical image; The cropped segmentation label set corresponding to a diseased medical image is recorded as Y = {Y1, Y2, ..., Y n }, n is the number of segmentation labels corresponding to each diseased medical image, and the training set is composed of each cropped diseased medical image and the cropped segmentation labels corresponding to each cropped diseased medical image; Step 3: Use the training set obtained in step 2 to train the uncertainty-aware medical image segmentation model M. The loss function L used in the training process is: Among them, L D represents the loss of the first branch, L′ D represents the loss of the second branch; Among them, x k represents the kth diseased medical image in the input training set, k = 1, 2, ..., K, K represents the total number of images in the training set, η k Represents a diseased medical image x k The output after the medical image segmentation backbone network module M1, P(η k |x k ) represents the diseased medical image x k After passing through the medical image segmentation backbone network module M1, the output is η k The probability of y i,j represents the j-th predicted segmentation result in the i-th branch, P(y i,j |η k ) indicates that the j-th predicted segmentation result in the i-th branch is y i,j probability; The training is stopped until the loss function converges, and the trained uncertainty-aware medical image segmentation model M is obtained.

Images