Slice segmentation model training method and device, slice segmentation method and device

Abstract

The application provides a slice segmentation model training method and device, and a slice segmentation method and device. The slice segmentation model training method comprises the following steps: inputting original slice images in a first sample set into an initial slice segmentation model to obtain intermediate feature maps and inference values; determining a first loss value of the original slice images with respect to the initial slice segmentation model based on the inference values, true values and the intermediate feature maps; determining a target original slice image based on the first loss value and iteratively training the initial slice segmentation model until convergence; inputting each slice image in a second sample set into the converged initial slice segmentation model to determine a value quantity corresponding to each slice image; determining a new slice image from the second sample set based on the value quantities and adding the new slice image to the first sample set; and iteratively training the converged initial slice segmentation model using the updated first sample set. Through the above method, the target slice segmentation model can achieve the expected performance by using only a small number of samples for training.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a slice segmentation model training method and apparatus, and a slice segmentation method and apparatus. Background Technology

[0002] Serial slice microscopy facilitates the parallel operation of multiple microscopic instruments and is widely used in the study of samples at both the microscopic and mesoscopic scales. With the continuous development of microscopic imaging technology and the deepening of research, researchers have an increasing demand for observing microscopic and mesoscopic structures, and the size of the sample slices being observed is also increasing, with the number of slices now reaching tens of thousands. Automatic and precise segmentation of the sample slices can effectively accelerate the microscopic imaging process, saving manpower and time.

[0003] In related technologies, most deep learning-based slice segmentation models rely on data-driven approaches and depend on the support of large datasets specific to the application domain. However, in continuous slice microscopy, due to significant morphological differences between continuous slices made from different samples, slice segmentation models trained on existing continuous slice datasets may not be applicable to new continuous slice segmentation tasks. When faced with new continuous slice segmentation tasks, directly using existing slice segmentation models is unlikely to achieve satisfactory results, and it is highly likely that a new training set will need to be created, leading to a waste of human and material resources.

[0004] Therefore, for slice segmentation models, how to achieve the expected performance with only a small number of samples for training is an urgent problem to be solved. Summary of the Invention

[0005] To address the problems existing in the prior art, the present invention provides a slice segmentation model training method and apparatus, and a slice segmentation method and apparatus.

[0006] This invention provides a method for training a slice segmentation model, comprising:

[0007] The original slice images in the first sample set are input into the initial slice segmentation model to obtain at least one intermediate feature map and inference value corresponding to the original slice images output by the initial slice segmentation model. The inference value is used to characterize the probability that the segmentation result output by the initial slice segmentation model belongs to the true value; the true value is used to characterize that the segmentation result is correct.

[0008] Based on the inference value, the true value, and each of the intermediate feature maps, a first loss value for the original slice image is determined for the initial slice segmentation model.

[0009] Based on the first loss value, the original target slice image is determined from the first sample set; the initial slice segmentation model is iteratively trained using the original target slice image until the initial slice segmentation model converges.

[0010] Each slice image in the second sample set is input into the converged initial slice segmentation model to obtain the segmentation results output by the converged initial slice segmentation model; based on each slice segmentation result, the value corresponding to each slice image is determined.

[0011] Based on the value values, new slice images are determined from the second sample set and added to the first sample set; the converged initial slice segmentation model is iteratively trained using the updated first sample set until the training stopping condition is met, and the target slice segmentation model is obtained.

[0012] Optionally, the initial slice segmentation model includes a global average pooling layer and a fully connected layer;

[0013] The step of determining the first loss value of the original slice image for the initial slice segmentation model based on the inferred value, the ground truth value, and each of the intermediate feature maps includes:

[0014] Based on the inference value and the true value, a second loss value for the original slice image is determined for the initial slice segmentation model;

[0015] Each of the intermediate feature maps is input into the global average pooling layer to obtain the feature vector corresponding to each of the intermediate feature maps;

[0016] Each of the aforementioned feature vectors is input into the fully connected layer to obtain the predicted loss value of the original slice image for the initial slice segmentation model;

[0017] The first loss value is determined based on the predicted loss value and the second loss value.

[0018] Optionally, determining the first loss value based on the predicted loss value and the second loss value includes:

[0019] Based on the predicted loss value and the second loss value, the first loss value is calculated using the first loss function;

[0020] The first loss function is expressed by the following formula (1):

[0021]

[0022] in, This is the first loss value; is the predicted loss value; l is the second loss value.

[0023] Optionally, the slice segmentation result includes N segmented sub-slices, the posterior probability of each segmented sub-slice belonging to the target category, and the segmentation mask of each segmented sub-slice. The segmentation mask is a binary image, and the pixel value of each pixel in the binary image represents the posterior probability of the pixel belonging to the target category.

[0024] The step of determining the value corresponding to each slice image based on the slice segmentation results includes:

[0025] For each of the sliced ​​images, the uncertainty of the segmentation mask is determined based on the segmentation mask;

[0026] Based on the uncertainty of the segmentation mask and the posterior probability, the uncertainty of the slice image with respect to the initial slice segmentation model is determined;

[0027] The value is determined based on the uncertainty of the initial slice segmentation model and the predicted loss value.

[0028] Optionally, determining new slice images from the second sample set and adding them to the first sample set based on each of the value values ​​includes:

[0029] The slice images whose value is greater than the first threshold are identified as the newly added slice images;

[0030] The newly added sliced ​​image is added to the first sample set.

[0031] Optionally, the training stopping condition is that the third loss value of the initial slice segmentation model reaches the second threshold, and / or the number of newly added slice images reaches the third threshold; the third loss value is calculated using the following formula (2):

[0032]

[0033] Where L is the third loss value; This is the second loss value; y is the inference value; y is the truth value; λ is the hyperparameter; This is the first loss value.

[0034] The present invention also provides a slicing method, comprising:

[0035] Obtain the slice image to be segmented;

[0036] The slice image to be segmented is input into the target slice segmentation model to obtain the slice segmentation result output by the target slice segmentation model;

[0037] The target slice segmentation model is obtained by training the slice segmentation model training method.

[0038] The present invention also provides a slice segmentation model training device, comprising:

[0039] The first input module is used to input the original slice images from the first sample set into the initial slice segmentation model to obtain at least one intermediate feature map and inference value output by the initial slice segmentation model corresponding to the original slice images. The inference value is used to characterize the probability that the segmentation result output by the initial slice segmentation model belongs to the true value; the true value is used to characterize that the segmentation result is correct.

[0040] The determination module is used to determine a first loss value of the original slice image for the initial slice segmentation model based on the inference value, the true value and each of the intermediate feature maps;

[0041] The first training module is used to determine the target original slice image from the first sample set based on the first loss value; and to iteratively train the initial slice segmentation model using the target original slice image until the initial slice segmentation model converges.

[0042] The second determining module is used to input each slice image in the second sample set into the converged initial slice segmentation model to obtain the segmentation results output by the converged initial slice segmentation model; and to determine the value corresponding to each slice image based on each slice segmentation result.

[0043] The second training module is used to determine new slice images from the second sample set based on the value values ​​and add them to the first sample set; and to iteratively train the converged initial slice segmentation model using the updated first sample set until the training stopping condition is met, thereby obtaining the target slice segmentation model.

[0044] The present invention also provides a slicing and dividing device, comprising:

[0045] The acquisition module is used to acquire the slice image to be segmented;

[0046] The second input module is used to input the slice image to be segmented into the target slice segmentation model to obtain the slice segmentation result output by the target slice segmentation model;

[0047] The target slice segmentation model is obtained by training the slice segmentation model training method.

[0048] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the slice segmentation model training method or slice segmentation method as described above.

[0049] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the slice segmentation model training method or slice segmentation method as described above.

[0050] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the slice segmentation model training method or slice segmentation method as described above.

[0051] The slice segmentation model training method provided by this invention involves inputting the original slice images from a first sample set into an initial slice segmentation model to obtain at least one intermediate feature map and inference value corresponding to the original slice images output by the initial slice segmentation model. Based on the inference value, the ground truth value, and each intermediate feature map, a first loss value for the original slice images relative to the initial slice segmentation model is determined. Based on the first loss value, target original slice images that the current initial slice segmentation model has not yet fully learned can be identified from the first sample set. That is, the initial slice segmentation model is iteratively trained using the target original slice images until the initial slice segmentation model converges, allowing the initial slice segmentation model to fully learn from the original slice images in the first sample set. Simultaneously, each slice image from a second sample set is input into the converged initial slice segmentation model to obtain the segmentation results output by the initial slice segmentation model. Based on each slice segmentation result, the value corresponding to each slice image is determined. Based on the value corresponding to each slice image, samples with greater value to the current converged initial slice segmentation model can be selected from the second sample set and added to the first sample set for iterative training to obtain the target slice segmentation model. This achieves the target slice segmentation model with the expected performance using only a small number of samples, reducing the consumption of human and material resources. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0053] Figure 1 This is one of the flowcharts illustrating the slice segmentation model training method provided by the present invention;

[0054] Figure 2 This is a schematic diagram illustrating the process of determining the first loss value provided by the present invention;

[0055] Figure 3 This is the second flowchart illustrating the slice segmentation model training method provided by the present invention;

[0056] Figure 4 This is a flowchart illustrating the slicing and segmentation method provided by the present invention;

[0057] Figure 5 This is a schematic diagram of the structure of the slice segmentation model training device provided by the present invention;

[0058] Figure 6 This is a schematic diagram of the slicing and dividing device provided by the present invention;

[0059] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0061] The terminology used in one or more embodiments of the present invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of the invention. The singular forms “a,” “the,” and “the” used in one or more embodiments of the invention and in the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” used in one or more embodiments of the invention refers to and includes any or all possible combinations of one or more associated listed items.

[0062] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of the present invention, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of the present invention, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0063] To facilitate a clearer understanding of the various embodiments of the present invention, some relevant background knowledge will be introduced as follows.

[0064] Serial slice microscopy, which facilitates parallel operation of multiple microscopic instruments, is widely used in the study of samples at both the microscopic and mesoscopic scales. The serial slice microscopy method first involves attaching collected slices to a wafer for subsequent automatic imaging by the microscopic imaging equipment. Each wafer can carry hundreds of sequential slices. During microscopy, the wafer with slices is first imaged to obtain a navigation map. The position of each slice is then marked on this navigation map. By calculating the mapping relationship between the pixel coordinates of the wafer navigation map and the physical coordinates of the stage of the microscopic imaging equipment using key points, a correspondence between the navigation map pixels and the stage's physical coordinate system is established, thus obtaining the true coordinates of each slice on the wafer on the stage. After obtaining the true coordinates of each slice, the microscopic imaging equipment can then image each slice.

[0065] Precise labeling of slices on wafer navigation maps is fundamental for subsequent imaging by microscopic equipment. Wafers are on the decimeter scale, while the coordinates of slices captured by microscopic equipment are accurate to the nanometer or micrometer level. Therefore, accurate slice labeling significantly improves the speed and quality of microscopic imaging. With the continuous development of microscopic imaging technology and the deepening of research, researchers have an increasing demand for observing micro- and mesoscopic structures, and the size of the samples being observed is also increasing, with the number of slices reaching tens of thousands. Automatic and precise segmentation of slices on wafer navigation maps can effectively accelerate the microscopic imaging process, saving manpower and time.

[0066] In recent years, with the rapid development of deep learning, segmentation methods have made great progress. However, the algorithms in these technologies are mainly geared towards instance segmentation problems in natural scenes. During the segmentation of continuous slices, the high similarity between slices—even abnormal slices such as wrinkles and damage—can be highly similar to normal slices, making it difficult for algorithms to distinguish between normal and abnormal slices in a sequence. Furthermore, due to the high similarity between slices, the slice boundaries are difficult to distinguish in cases of adhesion, further complicating the algorithm's ability to differentiate between different slices.

[0067] Most deep learning-based slice segmentation neural networks are data-driven, relying on massive datasets specific to their application domains. This also limits the widespread application of deep learning neural networks: first, creating such large datasets requires significant manpower, resources, and time, making them very expensive; second, the generalization ability of neural networks is highly limited by the training dataset, and their performance suffers greatly when faced with data not included in the training dataset. In continuous slice microscopy, because continuous slices from different samples vary considerably in morphology, slice segmentation networks trained on existing continuous slice datasets may not be suitable for segmenting new samples. When faced with new automatic continuous slice detection tasks, directly using existing neural segmentation networks is unlikely to achieve satisfactory results, and it may be necessary to recreate the dataset, further increasing the waste of manpower and resources.

[0068] Therefore, in order to address the aforementioned technical problems, and to avoid recreating the training set and reduce the consumption of manpower and resources when training the slice segmentation model, this invention provides a slice segmentation model training method and apparatus, and a slice segmentation method and apparatus.

[0069] The following is combined with Figure 1 The slice segmentation model training method provided by this invention will be described in detail. Figure 1 This is one of the flowcharts illustrating the slice segmentation model training method provided by this invention. See [link / reference]. Figure 1 As shown, the method includes steps 101-105, wherein:

[0070] Step 101: Input the original slice images from the first sample set into the initial slice segmentation model to obtain at least one intermediate feature map and inference value output by the initial slice segmentation model corresponding to the original slice images. The inference value is used to characterize the probability that the segmentation result output by the initial slice segmentation model belongs to the true value; the true value is used to characterize that the segmentation result is correct.

[0071] It should be noted that the subject of this invention can be any electronic device with the function of training a slice segmentation model, such as a smartphone, smartwatch, desktop computer, laptop, etc. The embodiments of this invention can be applied to scenarios involving the training of slice segmentation models.

[0072] Among them, the slice segmentation model is used to segment continuous slices. It can be any network model that has the function of segmenting continuous slices, such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short Term Memory (LSTM), etc.

[0073] In this embodiment, the first sample set includes the original slice image and the ground truth value corresponding to the original slice image. The ground truth value is used to characterize the correct segmentation result that the initial slice segmentation model should theoretically output after the original slice image is input into the initial slice segmentation model.

[0074] Specifically, the original slice images from the first sample set are first input into the initial slice segmentation model to obtain at least one intermediate feature map and an inference value corresponding to the original slice images output by the initial slice segmentation model. The inference value is used to characterize the probability that the segmentation result output by the initial slice segmentation model belongs to the true value. The closer the inference value is to the true value, the more accurate the result output by the slice segmentation model is. It should be noted that the intermediate feature map is generated after the initial slice segmentation model processes the original slice images after inputting them into the initial slice segmentation model.

[0075] Step 102: Based on the inference value, the true value, and each of the intermediate feature maps, determine the first loss value of the original slice image for the initial slice segmentation model.

[0076] After obtaining at least one intermediate feature map and inference value corresponding to the original slice image output by the initial slice segmentation model, it is necessary to determine the first loss value of the original slice image for the initial slice segmentation model based on the inference value, the ground truth value corresponding to the original slice image in the first sample set, and each intermediate feature map.

[0077] The first loss value is the final predicted loss value for the current initial slice segmentation model, which is used to filter the target original slice images from the first sample set that have not been fully learned by the current initial slice segmentation model.

[0078] Step 103: Based on the first loss value, determine the target original slice image from the first sample set; use the target original slice image to iteratively train the initial slice segmentation model until the initial slice segmentation model converges.

[0079] In this embodiment, after determining the first loss value of the original slice image for the initial slice segmentation model, it is necessary to determine the target original slice image from the first sample set based on the first loss value, wherein the target original slice image is the original slice image in the first sample set that the current initial slice segmentation model has not fully learned.

[0080] By iteratively training the initial slice segmentation model using the original target slice image until the initial slice segmentation model converges, the initial slice segmentation model can be fully learned from the original slice images in the first sample set.

[0081] Step 104: Input each slice image in the second sample set into the converged initial slice segmentation model to obtain the segmentation results output by the converged initial slice segmentation model; based on each slice segmentation result, determine the value corresponding to each slice image.

[0082] In this embodiment, after obtaining the converged initial slice segmentation model, it is necessary to input each slice image in the second sample set into the converged initial slice segmentation model to obtain the segmentation results output by the converged initial slice segmentation model, and determine the value corresponding to each slice image based on each slice segmentation result.

[0083] The value is used to characterize whether the slice image samples in the second sample set have been fully learned by the converged initial slice segmentation model. The higher the value, the less fully the current converged initial slice segmentation model has learned the slice image samples corresponding to that value.

[0084] Step 105: Based on the value values, determine the new slice images from the second sample set and add them to the first sample set; use the updated first sample set to iteratively train the converged initial slice segmentation model until the training stopping condition is met, and obtain the target slice segmentation model.

[0085] In this embodiment, after determining the value of each slice image in the second sample set, new slice images are determined from the second sample set based on each value and added to the first sample set; then, the updated first sample set is used to iteratively train the converged initial slice segmentation model until the training stopping condition is met, and the target slice segmentation model is obtained.

[0086] The process of iteratively training the converged initial slice segmentation model using the updated first sample set is as follows:

[0087] The updated first sample set is input into the converged initial slice segmentation model for the first round of training, so that the converged initial slice segmentation model can learn from the updated first sample set.

[0088] Then, the updated first sample set is input into the initial slice segmentation model trained in the first round for the second round of training; the above method is repeated until the training stops after N rounds of iteration, and the target slice segmentation model is obtained.

[0089] It should be noted that each newly added slice image in the second sample set can be an unlabeled slice image or a labeled slice image; if the newly added slice image is an unlabeled slice image, it needs to be labeled before adding the newly added slice image to the first sample set.

[0090] The slice segmentation model training method provided by this invention involves inputting the original slice images from a first sample set into an initial slice segmentation model to obtain at least one intermediate feature map and inference value corresponding to the original slice images output by the initial slice segmentation model. Based on the inference value, the ground truth value, and each intermediate feature map, a first loss value for the original slice images relative to the initial slice segmentation model is determined. Based on the first loss value, target original slice images that the current initial slice segmentation model has not yet fully learned can be identified from the first sample set. That is, the initial slice segmentation model is iteratively trained using the target original slice images until the initial slice segmentation model converges, allowing the initial slice segmentation model to fully learn from the original slice images in the first sample set. Simultaneously, each slice image from a second sample set is input into the converged initial slice segmentation model to obtain the segmentation results output by the initial slice segmentation model. Based on each slice segmentation result, the value corresponding to each slice image is determined. Based on the value corresponding to each slice image, samples with greater value to the current converged initial slice segmentation model can be selected from the second sample set and added to the first sample set for iterative training to obtain the target slice segmentation model. This achieves the target slice segmentation model with the expected performance using only a small number of samples, reducing the consumption of human and material resources.

[0091] Optionally, in one possible implementation of the present invention, the initial slice segmentation model includes a global average pooling layer and a fully connected layer;

[0092] The determination of the first loss value of the original slice image for the initial slice segmentation model based on the inferred value, the true value, and each of the intermediate feature maps can be achieved in the following manner, specifically including steps 1)-4):

[0093] Step 1) Based on the inference value and the true value, determine the second loss value of the original slice image for the initial slice segmentation model;

[0094] Step 2) Input each of the intermediate feature maps into the global average pooling layer to obtain the feature vectors corresponding to each of the intermediate feature maps;

[0095] Step 3) Input each of the feature vectors into the fully connected layer to obtain the prediction loss value of the original slice image for the initial slice segmentation model;

[0096] Step 4) Determine the first loss value based on the predicted loss value and the second loss value.

[0097] In this embodiment, the inferred value and the true value need to be substituted into the loss function corresponding to the initial slice segmentation model for calculation, so as to obtain the second loss value, that is, the true loss value of the original slice image for the initial slice segmentation model.

[0098] It should be noted that different initial slice segmentation models correspond to different loss functions, such as log-likelihood loss function and cross-entropy loss function. This invention does not limit the loss function corresponding to the initial slice segmentation model.

[0099] In practical applications, the global average pooling layer and the fully connected layer are attached to the initial slice segmentation model. In practice, the global average pooling layer and the fully connected layer can be regarded as loss prediction modules, which are used to predict the loss value of the initial slice segmentation model.

[0100] First, each intermediate feature map is input into a global average pooling layer for global average pooling operation to obtain a one-dimensional vector (i.e., feature vector) corresponding to each intermediate feature map.

[0101] After inputting all the one-dimensional vectors into the shared fully connected layer, they are concatenated together to obtain the total feature vector (i.e., the prediction loss value), which can be expressed by the following formula (3):

[0102]

[0103] in, To predict the loss value; Θ loss is the loss prediction module; h is the intermediate feature map.

[0104] Finally, based on the predicted loss value and the second loss value, the first loss value (i.e., the final predicted loss value for the current initial slice segmentation model) is determined.

[0105] Optionally, determining the first loss value based on the predicted loss value and the second loss value can be achieved in the following ways:

[0106] Based on the predicted loss value and the second loss value, the first loss value is calculated using the first loss function;

[0107] The first loss function is expressed by the following formula (1):

[0108]

[0109] in, This is the first loss value; is the predicted loss value; l is the second loss value.

[0110] The following is combined with Figure 2 The process of determining the first loss value will be further explained. Figure 2 This is a schematic diagram illustrating the process of determining the first loss value provided by the present invention.

[0111] like Figure 2 As shown, the original slice images in the first sample set are first input into the initial slice segmentation model to obtain at least one intermediate feature map and inference value corresponding to the original slice images output by the initial slice segmentation model.

[0112] Secondly, the inferred value and the true value are substituted into the loss function corresponding to the initial slice segmentation model (i.e., the main network loss function) to obtain the main network loss value (i.e., the second loss value); multiple intermediate feature maps are input into the loss prediction module to obtain the predicted loss value; the predicted loss value and the main network loss value are substituted into the loss function corresponding to the loss prediction module (i.e., the first loss function) to obtain the module loss value (i.e., the first loss value).

[0113] It should be noted here that, while determining the target original slice image from the first sample set based on the first loss value, and then iteratively training the initial slice segmentation model using the target original slice image, it is also necessary to iteratively train the loss prediction module using the target original slice image until the initial slice segmentation model and the loss prediction module converge.

[0114] In other words, if the initial slice segmentation model includes a global average pooling layer and a fully connected layer, it is necessary to iteratively train the initial slice segmentation model, which includes a global average pooling layer and a fully connected layer, using the original target slice image until the initial slice segmentation model converges.

[0115] In the above implementation, based on the inferred value and the ground truth value, a second loss value for the original slice image relative to the initial slice segmentation model is determined. Each intermediate feature map is input into a global average pooling layer to obtain the feature vector corresponding to each intermediate feature map. Then, each feature vector is input into a fully connected layer to obtain the predicted loss value of the original slice image relative to the initial slice segmentation model. Finally, based on the predicted loss value and the second loss value, a first loss value is determined. Through the above method, the loss value of the current initial slice segmentation model can be predicted. Based on the first loss value, the target original slice image that the current initial slice segmentation model has not fully learned can be identified from the first sample set. This allows the identification of samples that are more helpful to the current initial slice segmentation model, thereby enabling the target slice segmentation model to achieve the expected performance with only a small number of samples for training, reducing the consumption of human and material resources.

[0116] Optionally, in one possible implementation of the present invention, the slice segmentation result includes N segmented sub-slices, the posterior probability of each segmented sub-slice belonging to the target category, and the segmentation mask of each segmented sub-slice. The segmentation mask is a binary image, and the pixel value of each pixel in the binary image represents the posterior probability of the pixel belonging to the target category.

[0117] The determination of the value corresponding to each slice image based on the slice segmentation results can be achieved through the following methods, specifically including steps a)-c):

[0118] Step a) For each slice image, determine the uncertainty of the segmentation mask based on the segmentation mask;

[0119] Step b) Based on the uncertainty of the segmentation mask and the posterior probability, determine the uncertainty of the slice image for the initial slice segmentation model;

[0120] Step c) Determine the value based on the uncertainty of the initial slice segmentation model and the predicted loss value.

[0121] In this embodiment, the degree to which the initial slice segmentation model learns each slice image in the second sample set is also reflected in the value of the initial slice segmentation model for each slice image in the second sample set.

[0122] The value is reflected by the uncertainty of the slice segmentation result output by the initial slice segmentation model. The uncertainty is used to characterize the degree to which the initial slice segmentation model has learned the slice images in the second sample set. The higher the uncertainty, the more the initial slice segmentation model has not yet learned the slice image sufficiently.

[0123] In practical applications, given a slice image x in the second sample set, the slice segmentation result output by the initial slice segmentation model can be expressed by the following formula (4):

[0124]

[0125] Where n represents the initial slice segmentation model that detected n segmented sub-slices in slice image x; y' represents the slice segmentation result; The category described for the i-th segmented sub-slice, such as whether the segmented sub-slice is complete, incomplete, wrinkled, or contaminated; Let be the posterior probability that the i-th segment belongs to the target category; The bounding box of the i-th segmented sub-slice; Let be the segmentation mask for the i-th segmented sub-slice. It should be noted that the segmentation mask is a binary image where all pixel values ​​are between 0 and 1. The pixel value of each pixel in the binary image represents the posterior probability that the pixel belongs to the target category. For ease of subsequent explanation, Use p i Replace; will Use m i replace.

[0126] For each slice image in the second sample set, the uncertainty of the segmentation mask needs to be determined based on the segmentation mask of that slice image. Specifically, it can be calculated using the following formula (5):

[0127]

[0128] Among them, U i The segmentation mask m corresponding to the i-th segmented sub-slice i The uncertainty; h, w are the segmentation masks m corresponding to the i-th segmented sub-slice, respectively. i Height and width; The segmentation mask m corresponding to the i-th segmented sub-slice i The pixel value at coordinates (u,v); the function H is the entropy calculation formula, denoted by p. The entropy calculation formula can be expressed as: H(p)=-plog p-(1-p)log(1-p).

[0129] After determining the uncertainty of the segmentation mask, it is necessary to determine the uncertainty of the slice image relative to the initial slice segmentation model based on the uncertainty of the segmentation mask and the posterior probability. Specifically, it can be calculated using the following formula (6):

[0130]

[0131] Among them, U yThe uncertainty of the sliced ​​image with respect to the initial slice segmentation model; p i U represents the posterior probability that the i-th segment belongs to the target category; i Let be the uncertainty of the segmentation mask corresponding to the i-th segmented sub-slice.

[0132] Then, based on the uncertainty of the initial slice segmentation model and the predicted loss value, the value is determined, which can be calculated using the following formula (7):

[0133]

[0134] Among them, V x The value corresponding to the slice image; η is the predicted loss value; η is a hyperparameter that balances the uncertainty of the loss prediction module and the segmentation mask, and is generally a positive number.

[0135] In the above implementation, for each slice image, the uncertainty of the segmentation mask is determined based on the segmentation mask in the slice segmentation result. Then, based on the uncertainty of the segmentation mask and the posterior probability, the uncertainty of the slice image for the initial slice segmentation model is determined. Based on the uncertainty of the initial slice segmentation model and the predicted loss value, the value is determined. Through the above method, based on the value corresponding to each slice image, samples with greater value for the current initial slice segmentation model can be selected from the second sample set. That is, based on the value, samples that the current initial slice segmentation model has not yet fully learned can be determined from the second sample set and added to the first sample set for iterative training to obtain the target slice segmentation model. This achieves the expected performance of the target slice segmentation model with only a small number of samples for training, reducing the consumption of human and material resources.

[0136] Optionally, in one possible implementation of the present invention, the step of determining new slice images from the second sample set and adding them to the first sample set based on each of the value quantities can be achieved in the following way, specifically including steps [1]-[2]:

[0137] Step [1]: Identify the slice images whose value is greater than the first threshold as the newly added slice images;

[0138] Step [2]: Add the newly added sliced ​​image to the first sample set.

[0139] Specifically, after determining the value of each slice image in the second sample set, slice images with a value greater than the first threshold in the second sample set are identified as new slice images, and these new slice images are added to the first sample set to train the initial slice segmentation model.

[0140] In the above implementation, based on the value of each slice image, samples with greater value to the current converged initial slice segmentation model can be selected from the second sample set and added to the first sample set for training to obtain the target slice segmentation model. This achieves the expected performance of the target slice segmentation model with only a small number of samples for training, reducing the consumption of human and material resources.

[0141] Optionally, in one possible implementation of this invention, the training stopping condition is that the third loss value of the initial slice segmentation model reaches a second threshold, and / or the number of newly added slice images reaches a third threshold; the third loss value is calculated using the following formula (2):

[0142]

[0143] Where L is the third loss value; This is the second loss value; y is the inference value; y is the truth value; λ is the hyperparameter; This is the first loss value.

[0144] In practical applications, when the slice images in the second sample set are not labeled, it is necessary to label the newly added slice images at the same time as identifying them. Training stops when the cost of labeling the newly added slice images reaches the budget, that is, when the number of newly added slice images reaches the third threshold.

[0145] Figure 3 This is the second flowchart illustrating the slice segmentation model training method provided by this invention. See also... Figure 3 As shown, the method includes steps 301-311, wherein:

[0146] Step 301: Input the original slice images from the first sample set into the initial slice segmentation model to obtain at least one intermediate feature map and inference value corresponding to the original slice images output by the initial slice segmentation model.

[0147] Specifically, the inference value is used to characterize the probability that the segmentation result output by the initial slice segmentation model belongs to the true value; the true value is used to characterize the correctness of the segmentation result.

[0148] Step 302: Based on the inference value and the true value, determine the second loss value of the original slice image for the initial slice segmentation model.

[0149] Step 303: Input each intermediate feature map into the global average pooling layer in the initial slice segmentation model to obtain the feature vector corresponding to each intermediate feature map.

[0150] Step 304: Input each feature vector into the fully connected layer of the initial slice segmentation model to obtain the prediction loss value of the original slice image for the initial slice segmentation model.

[0151] Step 305: Based on the predicted loss value and the second loss value, determine the first loss value of the original slice image for the initial slice segmentation model.

[0152] Step 306: Based on the first loss value, determine the original target slice image from the first sample set; use the original target slice image to iteratively train the initial slice segmentation model until the initial slice segmentation model converges.

[0153] Step 307: Input each slice image in the second sample set into the converged initial slice segmentation model to obtain the segmentation results output by the converged initial slice segmentation model; for each slice image, determine the uncertainty of the segmentation mask based on the segmentation mask in the segmentation results.

[0154] Step 308: Based on the uncertainty of the segmentation mask and the posterior probability, determine the uncertainty of the slice image for the converged initial slice segmentation model.

[0155] Step 309: Based on the uncertainty and prediction loss value of the converged initial slice segmentation model, determine the value corresponding to the slice image.

[0156] Step 310: Determine the slice images with a value greater than the first threshold as new slice images.

[0157] Step 311: Add the newly added slice image to the first sample set, and use the updated first sample set to iteratively train the converged initial slice segmentation model until the training stopping condition is met, and obtain the target slice segmentation model.

[0158] Figure 4 This is a flowchart illustrating the slicing and segmentation method provided by the present invention. See [link / reference]. Figure 4 As shown, the method includes steps 401-402, wherein:

[0159] Step 401: Obtain the slice image to be segmented.

[0160] Step 402: Input the slice image to be segmented into the target slice segmentation model to obtain the slice segmentation result output by the target slice segmentation model;

[0161] The target slice segmentation model is obtained by training the slice segmentation model training method.

[0162] The slice segmentation method provided by this invention inputs the slice image to be segmented into the target slice segmentation model, which can realize the automatic segmentation of continuous slice images, ensure the segmentation accuracy, and achieve accurate automatic segmentation of continuous slice images.

[0163] The slice segmentation model training device provided by the present invention is described below. The slice segmentation model training device described below can be referred to in correspondence with the slice segmentation model training method described above. Figure 5 This is a schematic diagram of the structure of the slice segmentation model training device provided by the present invention, as shown below. Figure 5 As shown, the slice segmentation model training device 500 includes: a first input module 501, a first determination module 502, a first training module 503, a second determination module 504, and a second training module 505, wherein:

[0164] The first input module 501 is used to input the original slice image from the first sample set into the initial slice segmentation model to obtain at least one intermediate feature map and inference value output by the initial slice segmentation model corresponding to the original slice image. The inference value is used to characterize the probability that the segmentation result output by the initial slice segmentation model belongs to the true value. The true value is used to characterize that the segmentation result is correct.

[0165] The first determining module 502 is used to determine a first loss value of the original slice image for the initial slice segmentation model based on the inference value, the true value and each of the intermediate feature maps;

[0166] The first training module 503 is used to determine the target original slice image from the first sample set based on the first loss value; and to train the initial slice segmentation model using the target original slice image until the initial slice segmentation model converges.

[0167] The second determining module 504 is used to input each slice image in the second sample set into the converged initial slice segmentation model to obtain the segmentation results output by the converged initial slice segmentation model; and to determine the value corresponding to each slice image based on each slice segmentation result.

[0168] The second training module 505 is used to determine new slice images from the second sample set based on the value values ​​and add them to the first sample set; and to iteratively train the converged initial slice segmentation model using the updated first sample set until the training stopping condition is met, thereby obtaining the target slice segmentation model.

[0169] The slice segmentation model training device provided by this invention inputs the original slice images from a first sample set into an initial slice segmentation model to obtain at least one intermediate feature map and inference value corresponding to the original slice images output by the initial slice segmentation model. Based on the inference value, the ground truth value, and each intermediate feature map, a first loss value for the original slice image with respect to the initial slice segmentation model is determined. Based on the first loss value, target original slice images that the current initial slice segmentation model has not yet fully learned can be determined from the first sample set. That is, the initial slice segmentation model is trained using the target original slice images until the initial slice segmentation model converges, allowing the initial slice segmentation model to fully learn from the original slice images in the first sample set. Simultaneously, each slice image from a second sample set is input into the converged initial slice segmentation model to obtain each segmentation result output by the initial slice segmentation model. Based on each slice segmentation result, the value corresponding to each slice image is determined. Based on the value corresponding to each slice image, samples with greater value to the current converged initial slice segmentation model can be selected from the second sample set and added to the first sample set for training to obtain the target slice segmentation model. This achieves the target slice segmentation model with the expected performance using only a small number of samples for training, reducing the consumption of human and material resources.

[0170] Optionally, the initial slice segmentation model includes a global average pooling layer and a fully connected layer;

[0171] The first determining module 502 is further used for:

[0172] Based on the inference value and the true value, a second loss value for the original slice image is determined for the initial slice segmentation model;

[0173] Each of the intermediate feature maps is input into the global average pooling layer to obtain the feature vector corresponding to each of the intermediate feature maps;

[0174] Each of the aforementioned feature vectors is input into the fully connected layer to obtain the predicted loss value of the original slice image for the initial slice segmentation model;

[0175] The first loss value is determined based on the predicted loss value and the second loss value.

[0176] Optionally, the first determining module 502 is further configured to:

[0177] Based on the predicted loss value and the second loss value, the first loss value is calculated using the first loss function;

[0178] The first loss function is expressed by the following formula (1):

[0179]

[0180] in, This is the first loss value; is the predicted loss value; l is the second loss value.

[0181] Optionally, the slice segmentation result includes N segmented sub-slices, the posterior probability of each segmented sub-slice belonging to the target category, and the segmentation mask of each segmented sub-slice. The segmentation mask is a binary image, and the pixel value of each pixel in the binary image represents the posterior probability of the pixel belonging to the target category.

[0182] The second determining module 504 is further used for:

[0183] For each of the sliced ​​images, the uncertainty of the segmentation mask is determined based on the segmentation mask;

[0184] Based on the uncertainty of the segmentation mask and the posterior probability, the uncertainty of the slice image with respect to the initial slice segmentation model is determined;

[0185] The value is determined based on the uncertainty of the initial slice segmentation model and the predicted loss value.

[0186] Optionally, the second training module 505 is further used for:

[0187] The slice images whose value is greater than the first threshold are identified as the newly added slice images;

[0188] The newly added sliced ​​image is added to the first sample set.

[0189] Optionally, the training stopping condition is that the third loss value of the initial slice segmentation model reaches the second threshold, and / or the number of newly added slice images reaches the third threshold; the third loss value is calculated using the following formula (2):

[0190]

[0191] Where L is the third loss value; This is the second loss value; y is the inference value; y is the truth value; λ is the hyperparameter; This is the first loss value.

[0192] The slicing device provided by the present invention is described below. The slicing device described below and the slicing method described above can be referred to in correspondence. Figure 6 This is a schematic diagram of the slicing and dividing device provided by the present invention, as shown below. Figure 6 As shown, the slicing and segmentation device 600 includes: an acquisition module 601 and an input module 602, wherein:

[0193] The acquisition module 601 is used to acquire the slice image to be segmented;

[0194] The second input module 602 is used to input the slice image to be segmented into the target slice segmentation model to obtain the slice segmentation result output by the target slice segmentation model;

[0195] The target slice segmentation model is obtained by training the slice segmentation model training method.

[0196] The slicing segmentation device provided by the present invention inputs the slice image to be segmented into the target slice segmentation model, which can realize the automatic segmentation of continuous slice images, ensure the segmentation accuracy, and achieve accurate automatic segmentation effect of continuous slice images.

[0197] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other through the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a slice segmentation model training method. This method includes: inputting original slice images from a first sample set into an initial slice segmentation model to obtain at least one intermediate feature map and an inference value output by the initial slice segmentation model corresponding to the original slice image; the inference value characterizing the probability that the segmentation result output by the initial slice segmentation model belongs to the true value; the true value characterizing the correctness of the segmentation result; determining a first loss value for the original slice image against the initial slice segmentation model based on the inference value, the true value, and each intermediate feature map; and, based on the first loss value, selecting from the first sample set... The process involves: identifying the target original slice image; iteratively training the initial slice segmentation model using the target original slice image until the initial slice segmentation model converges; inputting each slice image from the second sample set into the converged initial slice segmentation model to obtain the segmentation results output by the converged initial slice segmentation model; determining the value corresponding to each slice image based on each slice segmentation result; identifying new slice images from the second sample set based on each value and adding them to the first sample set; and training the converged initial slice segmentation model using the updated first sample set until the training stopping condition is met to obtain the target slice segmentation model.

[0198] Alternatively, a slice segmentation method can be executed, which includes:

[0199] Obtain the slice image to be segmented; input the slice image to be segmented into the target slice segmentation model to obtain the slice segmentation result output by the target slice segmentation model; wherein, the target slice segmentation model is trained by the slice segmentation model training method.

[0200] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0201] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the slice segmentation model training method provided by the above methods. The method includes: inputting the original slice images in a first sample set into an initial slice segmentation model to obtain at least one intermediate feature map and an inference value output by the initial slice segmentation model corresponding to the original slice images; the inference value is used to characterize the probability that the segmentation result output by the initial slice segmentation model belongs to the true value; the true value is used to characterize that the segmentation result is correct; and based on the inference value, the true value, and each of the intermediate feature maps, determining that the original slice image is for the... The initial slice segmentation model is given a first loss value. Based on the first loss value, a target original slice image is determined from the first sample set. The initial slice segmentation model is trained using the target original slice image until it converges. Each slice image from the second sample set is input into the converged initial slice segmentation model to obtain the segmentation results output by the converged initial slice segmentation model. Based on each slice segmentation result, the value corresponding to each slice image is determined. Based on each value, a new slice image is determined from the second sample set and added to the first sample set. The converged initial slice segmentation model is iteratively trained using the updated first sample set until the training stopping condition is met to obtain the target slice segmentation model.

[0202] Alternatively, a slice segmentation method can be executed, which includes:

[0203] Obtain the slice image to be segmented; input the slice image to be segmented into the target slice segmentation model to obtain the slice segmentation result output by the target slice segmentation model; wherein, the target slice segmentation model is trained by the slice segmentation model training method.

[0204] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a slice segmentation model training method provided by the above methods. The method includes: inputting original slice images from a first sample set into an initial slice segmentation model to obtain at least one intermediate feature map and an inference value output by the initial slice segmentation model corresponding to the original slice image; the inference value being used to characterize the probability that the segmentation result output by the initial slice segmentation model belongs to the true value; the true value being used to characterize that the segmentation result is correct; and determining a first loss value for the original slice image relative to the initial slice segmentation model based on the inference value, the true value, and each of the intermediate feature maps. Based on the first loss value, the target original slice image is determined from the first sample set; the initial slice segmentation model is iteratively trained using the target original slice image until the initial slice segmentation model converges; each slice image from the second sample set is input into the converged initial slice segmentation model to obtain the segmentation results output by the converged initial slice segmentation model; based on each slice segmentation result, the value corresponding to each slice image is determined; based on each value, new slice images are determined from the second sample set and added to the first sample set; the converged initial slice segmentation model is iteratively trained using the updated first sample set until the training stopping condition is met to obtain the target slice segmentation model;

[0205] Alternatively, a slice segmentation method can be executed, which includes:

[0206] Obtain the slice image to be segmented; input the slice image to be segmented into the target slice segmentation model to obtain the slice segmentation result output by the target slice segmentation model; wherein, the target slice segmentation model is trained by the slice segmentation model training method.

[0207] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0208] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0209] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for training a slice segmentation model, characterized in that, include: The original slice image in the first sample set is input into the initial slice segmentation model to obtain at least one intermediate feature map and inference value corresponding to the original slice image output by the initial slice segmentation model. The inference value is used to characterize the probability that the first segmentation result output by the initial slice segmentation model belongs to the true value. The truth value is used to characterize that the first segmentation result is correct; Based on the inference value, the true value, and each of the intermediate feature maps, a first loss value for the original slice image is determined for the initial slice segmentation model. Based on the first loss value, the original target slice image is determined from the first sample set; the initial slice segmentation model is iteratively trained using the original target slice image until the initial slice segmentation model converges. Each slice image in the second sample set is input into the converged initial slice segmentation model to obtain the second segmentation results output by the converged initial slice segmentation model; based on each second segmentation result, the value corresponding to each slice image is determined. Based on the value values, new slice images are determined from the second sample set and added to the first sample set; The converged initial slice segmentation model is iteratively trained using the updated first sample set until the training stopping condition is met, thus obtaining the target slice segmentation model. The second segmentation result includes N segmented sub-slices, the posterior probability of each segmented sub-slice belonging to the target category, and the segmentation mask of each segmented sub-slice; The step of determining the value corresponding to each slice image based on each of the second segmentation results includes: Based on the uncertainty of the segmentation mask and the posterior probability, the uncertainty of the slice image with respect to the initial slice segmentation model is determined; ； in, The uncertainty of the sliced ​​image with respect to the initial slice segmentation model; Let be the posterior probability that the i-th segment belongs to the target category; Let n be the uncertainty of the segmentation mask corresponding to the i-th segmented sub-slice, and n be the number of segmented sub-slices detected by the initial slice segmentation model in the slice image. The value is determined based on the uncertainty and predicted loss value of the initial slice segmentation model.

2. The slice segmentation model training method according to claim 1, characterized in that, The initial slice segmentation model includes a global average pooling layer and a fully connected layer; The step of determining the first loss value of the original slice image for the initial slice segmentation model based on the inferred value, the ground truth value, and each of the intermediate feature maps includes: Based on the inference value and the true value, a second loss value for the original slice image is determined for the initial slice segmentation model; Each of the intermediate feature maps is input into the global average pooling layer to obtain the feature vector corresponding to each of the intermediate feature maps; Each of the aforementioned feature vectors is input into the fully connected layer to obtain the predicted loss value of the original slice image for the initial slice segmentation model; The first loss value is determined based on the predicted loss value and the second loss value.

3. The slice segmentation model training method according to claim 2, characterized in that, Determining the first loss value based on the predicted loss value and the second loss value includes: Based on the predicted loss value and the second loss value, the first loss value is calculated using the first loss function; The first loss function is expressed by the following formula (1): （1）； in, This is the first loss value; The predicted loss value; This is the second loss value.

4. The slice segmentation model training method according to claim 2 or 3, characterized in that, The segmentation mask is a binary image, and the pixel value of each pixel in the binary image represents the posterior probability that the pixel belongs to the target category; The step of determining the value corresponding to each slice image based on each of the second segmentation results further includes: For each of the sliced ​​images, the uncertainty of the segmentation mask is determined based on the segmentation mask.

5. The slice segmentation model training method according to claim 1, characterized in that, The step of determining new slice images from the second sample set and adding them to the first sample set based on each of the aforementioned value values ​​includes: The slice images whose value is greater than the first threshold are identified as the newly added slice images; The newly added sliced ​​image is added to the first sample set.

6. The slice segmentation model training method according to claim 3, characterized in that, The training stopping condition is that the third loss value of the initial slice segmentation model reaches the second threshold, and / or the number of newly added slice images reaches the third threshold; the third loss value is calculated by the following formula (2): （2）； in, This is the third loss value; This is the second loss value; The inference value; The truth value is stated; For hyperparameters; This is the first loss value.

7. A slicing method, characterized in that, include: Obtain the slice image to be segmented; The slice image to be segmented is input into the target slice segmentation model to obtain the slice segmentation result output by the target slice segmentation model; The target slice segmentation model is obtained by training the slice segmentation model training method according to any one of claims 1-6.

8. A slice segmentation model training device, characterized in that, include: The first input module is used to input the original slice image in the first sample set into the initial slice segmentation model to obtain at least one intermediate feature map and inference value output by the initial slice segmentation model corresponding to the original slice image. The inference value is used to characterize the probability that the first segmentation result output by the initial slice segmentation model belongs to the true value. The truth value is used to characterize that the first segmentation result is correct; The first determining module is used to determine a first loss value of the original slice image for the initial slice segmentation model based on the inference value, the true value and each of the intermediate feature maps; A first training module is used to determine the original slice image of the target from the first sample set based on the first loss value; The initial slice segmentation model is iteratively trained using the original target slice image until the initial slice segmentation model converges. The second determining module is used to input each slice image in the second sample set into the converged initial slice segmentation model to obtain each second segmentation result output by the converged initial slice segmentation model. Based on each of the second segmentation results, the value corresponding to each slice image is determined; The second training module is used to determine new slice images from the second sample set and add them to the first sample set based on each of the value quantities; The initial slice segmentation model is iteratively trained using the updated first sample set until the training stopping condition is met, thereby obtaining the target slice segmentation model. The second segmentation result includes N segmented sub-slices, the posterior probability of each segmented sub-slice belonging to the target category, and the segmentation mask of each segmented sub-slice; The step of determining the value corresponding to each slice image based on each of the second segmentation results includes: Based on the uncertainty of the segmentation mask and the posterior probability, the uncertainty of the slice image with respect to the initial slice segmentation model is determined; ； in, The uncertainty of the sliced ​​image with respect to the initial slice segmentation model; Let be the posterior probability that the i-th segment belongs to the target category; Let n be the uncertainty of the segmentation mask corresponding to the i-th segmented sub-slice, and n be the number of segmented sub-slices detected by the initial slice segmentation model in the slice image. The value is determined based on the uncertainty and predicted loss value of the initial slice segmentation model.

9. A slicing and dividing device, characterized in that, include: The acquisition module is used to acquire the slice image to be segmented; The second input module is used to input the slice image to be segmented into the target slice segmentation model to obtain the slice segmentation result output by the target slice segmentation model; The target slice segmentation model is obtained by training the slice segmentation model training method according to any one of claims 1-6.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the slice segmentation model training method as described in any one of claims 1 to 6, or the slice segmentation method as described in claim 7.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the slice segmentation model training method as described in any one of claims 1 to 6, or the slice segmentation method as described in claim 7.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the slice segmentation model training method as described in any one of claims 1 to 6, or the slice segmentation method as described in claim 7.

Images