Liver blood vessel segmentation method, device, equipment and medium based on weakly supervised learning

Abstract

The present invention discloses a liver blood vessel segmentation method, device, equipment and medium based on weak supervised learning, and the method includes: training two liver blood vessel segmentation network models with differences through noisy labels; using the two models to segment respectively to obtain predicted labels; assigning weights to predicted labels and noisy labels through a label optimization module to obtain weighted optimized labels; using the optimized labels to iteratively train the two models respectively to obtain two iterative models; using the two iterative models to continue segmenting to obtain new predicted labels; when the segmentation effect of the iterative model is improved, continue to call the label optimization module to obtain new optimized labels, and continue to iterate the model until convergence; segment the input CT image through the converged model to obtain the segmentation result. The present invention can effectively reduce the adverse effects of noisy labels on the segmentation results, thereby reducing the manpower, material resources and time consumed by label annotation.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a liver blood vessel segmentation method, device, equipment and medium based on weakly supervised learning. Background Art

[0002] Primary liver malignant tumor is one of the most common malignant tumors, with high mortality and high incidence. Partial liver resection is one of the most commonly used treatments for liver malignant tumors. The positional information between liver tumors and liver blood vessels and other information as well as accurate modeling of liver vascular structure can greatly improve the success rate of the entire liver segmentation. Therefore, the algorithm for accurately obtaining liver vascular annotation is an important part of improving the accuracy of liver segmentation. At the same time, high-quality liver vascular annotation is also an indispensable part of other liver surgeries such as liver transplantation. For example, computer-assisted liver surgery (e.g., ablation and embolization) is used to treat liver tumors, but before performing these surgeries, doctors need to understand the liver contour and its venous system, which can be used for preoperative planning and intraoperative navigation. During liver embolization surgery, it is necessary to accurately locate the liver vessels that supply the tumor, and understanding the detailed location of the tumor between the liver vessels and the diameter of the vessels can determine the results of the ablation surgery.

[0003] Due to the complex structure, low contrast, large amount of noise and various pathological changes of liver blood vessels, obtaining a complete, accurate and high-quality liver blood vessel annotation dataset requires professional radiologists to spend a lot of time and energy. Currently, most of the publicly available datasets are noisy liver blood vessel labels. Noisy labels provide very limited guidance for liver blood vessel segmentation and may even provide incorrect guidance for liver blood vessel segmentation.

[0004] Document 1 (Chinese invention patent with publication number CN105741251A) discloses a method for vascular segmentation of liver CTA sequence images, which first performs contrast enhancement and smooth noise preprocessing on the input three-dimensional liver sequence images, then uses OOF and OFA algorithms to enhance the liver blood vessels and their boundaries and refine the blood vessel center; according to the geometric structure of the blood vessels, the seed points of the blood vessel centerline are automatically searched, and the centerline of the liver blood vessels is extracted and the liver blood vessel tree is constructed; finally, the liver blood vessels are preliminarily segmented by combining the fast marching method and the corresponding blood vessel and background grayscale histograms are calculated, and the graph cut algorithm is used to achieve accurate segmentation of the liver blood vessels. However, the traditional liver blood vessel segmentation method cannot be compared with the deep learning algorithm in terms of accuracy and computational efficiency.

[0005] Document 2 (Chinese invention patent with publication number CN113344940A) discloses a deep learning-based liver vascular image segmentation method, which first obtains an original computed tomography image, manually marks a liver vascular region mask image, and trains a first 3D image segmentation model, uses the first 3D image segmentation model to obtain a first 3D image segmentation model segmentation result, and based on the result, marks the mask image of the missed segmentation region and the mask image of the mis-segmentation region, then channels the original computed tomography image, the segmentation result of the first 3D image segmentation model, the mask image of the missed segmentation region, and the mask image of the mis-segmentation region to obtain a first fused image, and uses the first fused image to train a second 3D image segmentation model until the second 3D image segmentation model is trained, and uses the trained second 3D image segmentation model to obtain a liver vascular segmentation result. However, this method requires accurate and complete manual annotation, and considering the difficulty of liver vascular annotation, this method consumes a very high cost of manpower, material resources, and time.

[0006] At the same time, since most deep learning segmentation methods adopt a supervised learning model, in practical applications, due to the complex structure of liver blood vessels, the susceptibility to pathological changes, and the low contrast and high noise defects of liver blood vessels after angiography, high-quality liver blood vessel labeling usually requires very solid professional knowledge and a large amount of manpower and material resources.

[0007] The semi-supervised learning algorithm can make full use of a small number of high-quality labels and low-quality noisy labels, thereby effectively reducing the errors in vascular segmentation, improving the accuracy and robustness of liver vascular segmentation, and providing high-quality segmentation results. Therefore, the development of a liver vascular segmentation method based on a semi-supervised learning algorithm can better assist doctors in labeling blood vessels. Summary of the invention

[0008] In view of the problems that traditional liver blood vessel segmentation methods in the prior art have low accuracy and are greatly affected by noise near the liver blood vessels, and that liver blood vessel segmentation methods based on deep learning require accurate and complete manual annotation, resulting in very high consumption of manpower, material resources and time costs, the purpose of the present invention is to provide a liver blood vessel segmentation method, device, equipment and medium based on weak supervised learning.

[0009] To achieve the above object, the technical solution of the present invention is:

[0010] In a first aspect, the present invention provides a liver blood vessel segmentation method based on weakly supervised learning, comprising the following steps:

[0011] S1. obtaining two liver vascular segmentation network models having differences from each other by training liver region images with noisy labels, wherein the noisy labels are used to mark liver vessels;

[0012] S2. Using the two liver blood vessel segmentation network models obtained in S1, respectively segment the liver region image to obtain predicted labels for marking liver blood vessels;

[0013] S3, assigning weights to each of the predicted labels and each of the noisy labels through a label optimization module to obtain optimized labels with weights;

[0014] S4, using the optimized labels to replace the noisy labels in the liver region image to obtain an optimized image;

[0015] S5, iteratively training the two liver vascular segmentation network models in S1 respectively using the optimized image to obtain two iterative models;

[0016] S6, using the two iterative models obtained in S5 to segment the optimized image to obtain a new prediction label;

[0017] S7, comparing the new predicted label in S6 with the predicted label in S2, and judging whether the segmentation effect of the iterative model is improved, if yes, proceeding to the next step, otherwise convergence;

[0018] S8, optimizing each of the new prediction labels and the optimized labels by the label optimization module to obtain a new optimized label;

[0019] S9, repeating the process of S4-S8 until the segmentation effect of the iterative model is no longer improved and converges, thereby obtaining a converged liver vascular segmentation network model;

[0020] S10. Segment the input CT image using the converged liver vascular segmentation network model to obtain a segmentation result.

[0021] Preferably, in S1, each of the noisy labels constitutes a noisy label set; in S2, after the two liver vascular segmentation network models segment the liver region image respectively, two corresponding prediction label sets with the prediction labels are obtained;

[0022] Then in S3, the step of assigning weights to each of the predicted labels and each of the noisy labels through the label optimization module to obtain optimized labels with weights includes:

[0023] Combining the two predicted label sets with the noisy label set to obtain the intersection between the sets;

[0024] According to the intersection between the sets, the confidence levels are assigned to the intersection and difference between the sets respectively.

[0025] The labels contained in each intersection and each difference are weighted according to the confidence.

[0026] Preferably, tags belonging to the same intersection or difference set have the same weight.

[0027] Furthermore, before S1, the steps include: training a liver segmentation network model using a CT image set carrying a liver label;

[0028] The method before S10 also includes the step of segmenting a liver region image from the input CT image by using the liver segmentation network model, so that the converged liver vascular segmentation network model obtains a segmentation result by segmenting the liver region image.

[0029] Furthermore, the method further includes performing an expansion operation on the segmentation result of the liver segmentation network model so that the liver region image includes a liver with a complete vascular structure.

[0030] Preferably, in S7, the step of determining whether the segmentation effect of the iterative model is improved comprises:

[0031] Compare whether the difference between the new predicted label in S6 and the predicted label in S2 is higher than a preset threshold. If yes, the segmentation effect is improved, otherwise, it is not improved.

[0032] Preferably, in S1, the two liver vascular segmentation network models that are different from each other are SE-UNet focusing on accuracy and Res-UNet focusing on recall.

[0033] In a second aspect, the present invention further provides a liver blood vessel segmentation device based on weakly supervised learning, comprising:

[0034] An image acquisition module, used for acquiring a liver region image with a noisy label;

[0035] A model training module, used for training two liver vascular segmentation network models having differences from each other using the liver region image;

[0036] A blood vessel segmentation module, used to segment the liver region image using the liver blood vessel segmentation network model, so as to obtain a predicted label for marking the liver blood vessels;

[0037] A label optimization module, used to optimize the predicted label and the noisy label to obtain an optimized label with a weight;

[0038] an iterative training module, used to replace the noisy labels in the liver region image with the optimized labels to obtain an optimized image, and then iteratively train the two liver vascular segmentation network models respectively through the optimized image to obtain two iterative models, and the two iterative models are used to respectively segment the optimized image to obtain new prediction labels;

[0039] An effect judgment module, used to compare the new predicted label with the predicted label, and judge whether the segmentation effect of the iterative model is improved according to the comparison result;

[0040] and a circulation module, which is used to call the functions of the label optimization module, the iterative training module and the effect judgment module when the segmentation effect is improved, until the segmentation effect is no longer improved and converges to obtain a converged liver vascular segmentation network model.

[0041] In a third aspect, the present invention further provides an electronic device, comprising a memory storing executable program code and a processor coupled to the memory; wherein the processor calls the executable program code stored in the memory to execute the method as described above.

[0042] In a fourth aspect, the present invention further provides a computer-readable storage medium storing a computer program, wherein the computer program executes the method described above when executed by a processor.

[0043] By adopting the above technical solution, the beneficial effects of the present invention are:

[0044] 1. Due to the construction of two different liver vascular segmentation network models, the two segmentation network models are combined to identify the liver vessels as comprehensively as possible;

[0045] 2. Due to the setting of the label optimization module, optimized labels can be obtained through the label optimization module, thereby reducing the erroneous guidance of the model caused by noisy labels, more effectively extracting the effective information in the noisy labels, and being able to use the optimized labels to continue training the liver vascular segmentation network model. At the same time, during the training process, as the performance of the liver vascular segmentation network model continues to improve, the liver vascular segmentation network model will obtain more effective segmentation labels. The use of more effective segmentation labels can continue to iteratively optimize the liver vascular segmentation network model until convergence. The converged liver vascular segmentation network model can obtain better segmentation results after being applied. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of a method according to Embodiment 1 of the present invention.

[0047] Figure 2It is a schematic diagram of the optimization process of the label optimization module in the present invention.

[0048] Figure 3 This is a schematic diagram of the structure of the third embodiment of the present invention.

[0049] Figure 4 It is a structural diagram of embodiment 4 of the present invention. DETAILED DESCRIPTION

[0050] The specific embodiments of the present invention are further described below in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0051] It should be noted that, in the description of the present invention, the directions or positional relationships indicated by the terms "up", "down", "left", "right", "front", "back", etc. are descriptions of the structure of the present invention based on the accompanying drawings, and are only for the convenience of describing the present invention, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention.

[0052] The "first" and "second" in this technical solution are only used to distinguish the names of the same or similar structures, or corresponding structures with similar functions, and are not an arrangement of the importance of these structures, nor do they have a ranking, comparison of size, or other meanings.

[0053] In addition, unless otherwise clearly specified and limited, the terms "installation" and "connection" should be understood in a broad sense. For example, the connection can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be a connection between the two structures. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood based on the overall idea of ​​the present invention and the specific context of the present solution.

[0054] Embodiment 1

[0055] A liver vascular segmentation method based on weakly supervised learning, such as Figure 1 As shown, the following steps are included:

[0056] S1. Two liver vascular segmentation network models with differences from each other are obtained by training with liver region images with noisy labels, wherein the noisy labels are used to mark the liver vessels, and the noisy labels together constitute a noisy label set.

[0057] Among them, noisy labels are usually added by medical staff. This process requires medical staff to have relatively comprehensive and rich relevant knowledge, which makes the addition of labels inevitably have certain noise. The result analysis of the fully supervised algorithm found that the noise in the noisy labels not only cannot give correct guidance to the training of the model, but also brings some wrong guidance, so it needs to be optimized.

[0058] In this embodiment, the two liver blood vessel segmentation network models that are different from each other are SE-UNet focusing on accuracy and Res-UNet focusing on recall.

[0059] S2. Use the two liver vascular segmentation network models obtained in S1 to segment the liver region images respectively, and obtain predicted labels for marking liver blood vessels.

[0060] Among them, after the two liver vascular segmentation network models (i.e., SE-UNet and Res-UNet) segment the liver region image respectively, two corresponding prediction label sets with the above prediction labels will be obtained. That is, there are two prediction label sets, corresponding to the segmentation results of the liver region image by SE-UNet and Res-UNet respectively.

[0061] S3. Assign weights to each predicted label and each noisy label through the label optimization module to obtain optimized labels with weights.

[0062] So far, there are three label sets, namely, two predicted label sets and one noisy label set. In order to better utilize the potential information in the noisy labels and reduce the erroneous guidance of the noisy labels on the predicted labels, this embodiment uses a soft constraint-based method to optimize each label, and the steps include:

[0063] Combine the two predicted label sets with the noisy label set to obtain the intersection between the sets;

[0064] According to the intersection between the sets, the confidence levels are assigned to the intersection and difference between the sets respectively.

[0065] Weights are assigned to labels contained in each intersection and each difference according to the confidence level. Generally, labels belonging to the same intersection or difference are preferably given the same weight.

[0066] like Figure 2As shown in the figure, the two predicted label sets are respectively recorded as segmentation result A and segmentation result B, and the noise label set is recorded as noise label C. The intersection of the three sets is shown in the right part of the figure. The three sets have intersections with each other. Generally speaking, the part where the two sets intersect with each other has a higher confidence, and the corresponding labels have a higher weight. For example, the weight of the intersection of segmentation result A and segmentation result B, segmentation result A and noise label C, and segmentation result B and noise label C are all 0.8, the weight of the intersection of segmentation result A, segmentation result B and noise label C is 1, the weight of the other parts in segmentation result A and segmentation result B is 0.4, and the weight of the other parts in noise label C is 0.6.

[0067] S4. Use optimized labels to replace noisy labels in the liver region image to obtain an optimized image.

[0068] S5. Perform iterative training on the two liver vascular segmentation network models in S1 respectively using the above-mentioned optimized images to obtain two iterative models.

[0069] It is understandable that after optimization, the influence of noise in noisy labels will be weakened, and the positive effects of other labels will be strengthened. Therefore, after re-training the two liver vascular segmentation network models separately using optimized images with optimized labels, a model with better segmentation effect can be obtained.

[0070] S6. Use the two iterative models obtained in S5 to segment the above-mentioned optimized image to obtain a new prediction label;

[0071] S7. Compare the new predicted labels in S6 with the predicted labels in S2 to determine whether the segmentation effect of the iterative model has been improved. If so, proceed to the next step, otherwise converge.

[0072] Whether the segmentation effect of the iterative model is better needs to be verified. If so, it means that the optimization processing of the label optimization module in S3 is effective. At this time, the iterative model obtained in S5 can be used as a better segmentation tool.

[0073] Specifically, the steps for determining whether the segmentation effect of the iterative model in S5 is improved include:

[0074] Compare the difference between the new predicted label in S6 and the predicted label in S2 to see if it is higher than the preset threshold. If so, the segmentation effect is improved, otherwise it is not improved. For example, the difference between the two is within ±5%.

[0075] It is understandable that when it is determined that the segmentation effect of the iterative model in the above S5 has been improved, it is still necessary to determine whether it is the optimal model. After proving that the label optimization module is effective, it is repeatedly applied to obtain new optimized labels, and based on the new optimized labels, an iterative model with better segmentation effect can be obtained, that is, there are steps S8 and S9.

[0076] S8. Optimize each new prediction label and optimization label through a label optimization module to obtain a new optimization label.

[0077] S9. Repeat the process of S4-S8 until the segmentation effect of the iterative model is no longer improved and converges, thereby obtaining a converged liver vascular segmentation network model.

[0078] S10. Segment the input CT image using the converged liver vascular segmentation network model to obtain a segmentation result.

[0079] Embodiment 2

[0080] Liver region images are usually abdominal CT images, which include not only the liver but also images of other organs and tissues. Noisy labels are directly annotated on the CT images by medical personnel. The public dataset MSD8 (Medical Segmentation Decathlon Task 8) is used as the input image of this embodiment. This dataset has a large number of under-labeling (60% of blood vessels are not labeled) and over-labeling (5% of non-vascular parts are labeled) problems.

[0081] In order to solve the problem of imbalanced liver blood vessel categories in abdominal CT images and reduce the influence of external liver blood vessels on the segmentation of internal liver blood vessels, in this embodiment, the following steps are further included before S1:

[0082] A liver segmentation network model is trained using a CT image set carrying a liver label; then before S10, the method also includes the step of segmenting a liver region image from the input CT image using the above-mentioned liver segmentation network model, so that the liver vascular segmentation network model after convergence in S9 obtains a segmentation result by segmenting the liver region image.

[0083] It is understandable that the segmentation result of the liver segmentation network model is usually further expanded manually so that the liver region image includes a liver with a complete vascular structure. In addition, the liver region image is processed by data enhancement to obtain the data set required by the liver vascular segmentation network model after convergence in S9.

[0084] Embodiment 3

[0085] A liver vascular segmentation device based on weakly supervised learning, such as Figure 3As shown, including:

[0086] An image acquisition module, used for acquiring a liver region image with a noisy label;

[0087] A model training module, used for training two liver vascular segmentation network models having differences from each other using liver region images;

[0088] A blood vessel segmentation module, used to segment the liver region image using a liver blood vessel segmentation network model, thereby obtaining a predicted label for marking the liver blood vessels;

[0089] A label optimization module, used to optimize the predicted labels and the noisy labels to obtain optimized labels with weights;

[0090] An iterative training module, used to replace noisy labels in a liver region image with optimized labels to obtain an optimized image, and then iteratively train two liver vascular segmentation network models using the optimized image to obtain two iterative models, and the two iterative models are used to segment the optimized image to obtain new prediction labels;

[0091] The effect judgment module is used to compare the new predicted labels with the predicted labels, and judge whether the segmentation effect of the iterative model is improved based on the comparison results;

[0092] And a circulation module is used to call the functions of the label optimization module, the iterative training module and the effect judgment module when the segmentation effect is improved, until the segmentation effect is no longer improved and converges to obtain a converged liver vascular segmentation network model.

[0093] Embodiment 4

[0094] An electronic device, such as Figure 4 As shown, it includes a memory storing executable program code and a processor coupled to the memory; wherein the processor calls the executable program code stored in the memory to execute the method steps in the above embodiment.

[0095] Embodiment 5

[0096] A computer storage medium stores a computer program, and when the computer program is executed by a processor, the method steps in the above embodiment are executed.

[0097] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0098] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0099] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0100] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A liver blood vessel segmentation method based on weakly supervised learning, characterized in that: The following steps are involved: S1. obtaining two liver vascular segmentation network models having differences from each other by training liver region images with noisy labels, wherein the noisy labels are used to mark liver vessels; S2. Using the two liver blood vessel segmentation network models obtained in S1, respectively segment the liver region image to obtain predicted labels for marking liver blood vessels; S3, assigning weights to each of the predicted labels and each of the noisy labels through a label optimization module to obtain optimized labels with weights; S4, using the optimized labels to replace the noisy labels in the liver region image to obtain an optimized image; S5, iteratively training the two liver vascular segmentation network models in S1 respectively using the optimized image to obtain two iterative models; S6, using the two iterative models obtained in S5 to segment the optimized image to obtain a new prediction label; S7, comparing the new predicted label in S6 with the predicted label in S2, and judging whether the segmentation effect of the iterative model is improved, if yes, proceeding to the next step, otherwise convergence; S8, optimizing each of the new prediction labels and the optimized labels by the label optimization module to obtain a new optimized label; S9, repeating the process of S4-S8 until the segmentation effect of the iterative model is no longer improved and converges, thereby obtaining a converged liver vascular segmentation network model; S10, segmenting the input CT image using the converged liver vascular segmentation network model to obtain a segmentation result; In S1, each of the noisy labels constitutes a noisy label set; in S2, after the two liver vascular segmentation network models segment the liver region image respectively, two corresponding prediction label sets with the prediction labels are obtained; Then in S3, the step of assigning weights to each of the predicted labels and each of the noisy labels through the label optimization module to obtain optimized labels with weights includes: Combining the two predicted label sets with the noisy label set to obtain the intersection between the sets; According to the intersection between the sets, the confidence levels are assigned to the intersection and difference between the sets respectively. Assigning weights to the labels contained in each intersection and each difference according to the confidence level; Among them, labels belonging to the same intersection or difference set have the same weight.

2. The method according to claim 1, characterized in that: S1 also includes the steps of: training a liver segmentation network model using a CT image set carrying a liver label; The method before S10 also includes the step of segmenting a liver region image from the input CT image by using the liver segmentation network model, so that the converged liver vascular segmentation network model obtains a segmentation result by segmenting the liver region image.

3. The method according to claim 2, characterized in that: The method also includes performing an expansion operation on the segmentation result of the liver segmentation network model so that the liver region image includes a liver with a complete vascular structure.

4. The method according to claim 1, characterized in that: In S7, the step of determining whether the segmentation effect of the iterative model is improved includes: Compare whether the difference between the new predicted label in S6 and the predicted label in S2 is higher than a preset threshold. If yes, the segmentation effect is improved, otherwise, it is not improved.

5. The method according to claim 1, characterized in that: In S1, the two liver vascular segmentation network models that are different from each other are SE-UNet focusing on accuracy and Res-UNet focusing on recall.

6. A liver blood vessel segmentation device based on weakly supervised learning, characterized in that: include: An image acquisition module, used for acquiring a liver region image with a noisy label; A model training module, used for training two liver vascular segmentation network models having differences from each other using the liver region image; A blood vessel segmentation module, used to segment the liver region image using the liver blood vessel segmentation network model, so as to obtain a predicted label for marking the liver blood vessels; A label optimization module, used to optimize the predicted label and the noisy label to obtain an optimized label with a weight; an iterative training module, used to replace the noisy labels in the liver region image with the optimized labels to obtain an optimized image, and then iteratively train the two liver vascular segmentation network models respectively through the optimized image to obtain two iterative models, and the two iterative models are used to respectively segment the optimized image to obtain new prediction labels; An effect judgment module, used to compare the new predicted label with the predicted label, and judge whether the segmentation effect of the iterative model is improved according to the comparison result; and a circulation module, used for calling the functions of the label optimization module, the iterative training module and the effect judgment module when the segmentation effect is improved, until the segmentation effect is no longer improved and converges to obtain a converged liver vascular segmentation network model; Wherein, each of the noisy labels constitutes a noisy label set; after the two liver vascular segmentation network models segment the liver region image respectively, two corresponding prediction label sets with the prediction labels are obtained; The step of assigning weights to each of the predicted labels and each of the noisy labels through a label optimization module to obtain optimized labels with weights includes: Combining the two predicted label sets with the noisy label set to obtain the intersection between the sets; According to the intersection between the sets, the confidence levels are assigned to the intersection and difference between the sets respectively. Assigning weights to the labels contained in each intersection and each difference according to the confidence level; Among them, labels belonging to the same intersection or difference set have the same weight.

7. An electronic device, characterized in that: It comprises a memory storing executable program code and a processor coupled to the memory; wherein the processor calls the executable program code stored in the memory to execute the method according to any one of claims 1-5.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is performed.

Images