Pseudo-label selection method, system and equipment for computed tomography image segmentation based on multi-modal model and medium

Abstract

The invention discloses a pseudo label selection method, system and device for computed tomography image segmentation based on a multi-modal model and a medium, and the method comprises the steps: marking unmarked planing organ types in an original CT image data set through employing a Totalsegment, and obtaining a noise pseudo label; carrying out degradation on the manually labeled label; forming a training set by the original CT image, the manually-labeled label and the degraded manually-labeled label, and inputting the training set into a BiomedCLIP image encoder to extract visual features; after the prompt words are initialized, the prompt words are input into a text encoder of the BiomedCLIP to extract text features; constructing a joint loss function to supervise and train cue words, superposing the category information of the noise pseudo-tag to the trained cue words, inputting the cue words into a text encoder to obtain text features, transforming the original CT image and the noise pseudo-tag, and inputting the transformed original CT image and noise pseudo-tag into an image encoder to extract visual features; and judging the quality of the noise pseudo tag based on the similarity of the text feature and the visual feature. According to the invention, the efficiency of screening high-quality pseudo labels is improved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and computer vision, and specifically relates to a pseudo-label selection method, system, device and medium for computed tomography image segmentation based on a multimodal model. Background Art

[0002] Applying deep learning to medical imaging, especially using image segmentation technology to assist doctors in diagnosis, has become a current research hotspot. However, there are noise, artifacts and low contrast in computed tomography (CT) images. In addition, the anatomical structure is complex and there is a lack of labeled data, which leads to many challenges in CT image segmentation.

[0003] Although major hospitals generate a large amount of CT data every day, it is expensive to obtain a large number of accurately labeled medical images because medical images involve patient privacy and require professional knowledge to interpret medical images. They must be labeled by professional doctors. In order to solve the problem of insufficient labeled data, pseudo labels are widely used in semi-supervised learning and unsupervised learning. The prediction results generated by the model are used as labels for training data to expand the data scale. However, the uneven quality of pseudo labels has a significant impact on the training effect of the model, especially in CT images, where noise and inconsistent pseudo labels will reduce the model's ability to recognize medical lesion features. Therefore, efficiently screening out high-quality pseudo labels has become a key issue that needs to be solved urgently.

[0004] The development of artificial intelligence technology has triggered profound changes in the field of medical image processing. Automated methods can reduce subjective differences between observers and achieve quantification of anatomical features. The paper (“Multi-modal contrastive mutual learning and pseudo-label re-learning for semi-supervised medical imagesegmentation”, Medical Image Analysis, 2023.) proposed a multi-modal contrastive mutual learning and pseudo-label re-learning method for semi-supervised medical image segmentation. In another study, the paper (“Pseudo-labelguided contrastive learning for semi-supervised medical image segmentation”, Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition, 2023.) proposed a pseudo-label-guided contrastive learning framework for semi-supervised medical image segmentation. In addition, the paper (“Semi-supervised medical image segmentation via cross teaching between CNN and Transformer”, Proceedings of the IEEE / CVF International Conference on Computer Vision, 2021.) proposed a cross-learning method between CNN and Transformer for semi-supervised medical image segmentation. However, these studies mainly focus on the fusion of multimodal data and the generation of pseudo labels, and have not yet fully solved the problem of uneven pseudo label quality. It can be seen from the above work that there are still few studies on the selection of pseudo labels using multimodal models. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a pseudo-label selection method, system, device and medium for computed tomography image segmentation based on a multimodal model to explicitly measure the quality of labels, so as to expand as many reliable labels as possible without introducing new training data.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A pseudo label selection method for computed tomography image segmentation based on a multimodal model comprises the following steps:

[0008] S1, use the original CT image dataset with manually annotated labels to train the segmentation model Totalsegmentator;

[0009] S2. Use the trained segmentation model Totalsegmentator to label the unlabeled anatomical organ types in the original CT image dataset. The labels labeled by the segmentation model Totalsegmentator are noise pseudo labels.

[0010] S3, degenerating the manually annotated labels to obtain degraded manually annotated labels;

[0011] S4, preparing a pre-trained multimodal model BiomedCLIP, wherein the multimodal model BiomedCLIP comprises an image encoder and a text encoder, and the original CT image, the manually annotated labels and the degraded manually annotated labels form a training data set, and the training data set is transformed and input into the image encoder of the multimodal model BiomedCLIP to extract visual features;

[0012] S5, construct a learnable prompt word, initialize the prompt word and input it into the text encoder of the multimodal model BiomedCLIP to extract text features;

[0013] S6. Construct a joint loss function using visual features and text features to supervise and train learnable prompt words, with the training goal of minimizing the joint loss. After the training is completed, freeze the prompt word parameters and the parameters of the multimodal model BiomedCLIP.

[0014] S7, superimposing the category information of the noise pseudo-label generated by the segmentation model Totalsegmentator onto the trained prompt word, and then inputting it into the text encoder in the multimodal model BiomedCLIP to obtain text features, transforming the original CT image and its corresponding noise pseudo-label and inputting it into the image encoder in the multimodal model BiomedCLIP to extract visual features;

[0015] S8. Calculate the similarity between the text features and the visual features obtained in S7, and judge the quality of the noise pseudo-label based on the similarity between the text features and the visual features.

[0016] To optimize the above technical solutions, the specific measures taken also include:

[0017] Furthermore, in S3, the step of degrading the manually annotated labels is as follows:

[0018] Multiple degradation operations are randomly selected from the following degradation operations to degrade the manually annotated labels: adding random noise, Gaussian blur, randomly moving pixel positions, random affine transformation, nonlinear deformation, simulating missing areas, adding random shapes and random morphological transformation.

[0019] Further, in S4, the training data set is represented as:

[0020] ;

[0021] Where D T is the training data set, is the i-th CT image, is the manually annotated label corresponding to the i-th CT image, is the degraded manual annotation label corresponding to the i-th CT image, and N is the number of samples in the training dataset;

[0022] The transformation of the training data set is specifically as follows:

[0023] Will Corresponding or First perform maximum projection to obtain , then Corresponding and Converting from 3D to 2D images, the resulting 2D images are uniformly recorded as , then Perform edge extraction operation to obtain edge information ,Will , and Stacked together to form three channels of data;

[0024] The image encoder of the multimodal model BiomedCLIP is expressed as follows:

[0025] ;

[0026] In the formula, is the mapping function of the image encoder, is the input image set, Represents a D-dimensional feature space.

[0027] Furthermore, in S5, the text encoder of the multimodal model BiomedCLIP is expressed as follows:

[0028] ;

[0029] In the formula, is the mapping function of the text encoder, is the input text set, Represents a D-dimensional feature space.

[0030] Further, in S6, the training objective is to minimize the joint loss and is expressed as:

[0031]

[0032] In the formula, is the joint loss function, is the cross entropy loss, is the weight of the contrast loss, is the contrast loss;

[0033] The calculation formula for cross entropy loss is as follows:

[0034] ;

[0035] Where N is the number of samples in the training data set, is the indicator function, is the i-th CT image, is the manually annotated label corresponding to the i-th CT image, is the degraded manual annotation label corresponding to the i-th CT image, represents the probability that the label of the i-th CT image is a manually annotated label, represents the probability that the label of the i-th CT image is a degraded manual annotation label;

[0036] The calculation formula of contrast loss is as follows:

[0037] ;

[0038] In the formula, is the visual feature of the i-th manually annotated label, is the text feature of the i-th manually annotated label, and m represents the lower limit of the distance between the visual feature of the manually annotated label and the text feature of the degraded manually annotated label; Represents the text features of the i-th degenerate human annotation label.

[0039] Furthermore, in S8, the similarity between the text feature and the visual feature is specifically measured using cosine similarity.

[0040] The present invention also proposes a pseudo label selection system for computed tomography image segmentation based on a multimodal model, comprising:

[0041] The noise pseudo-label generation module is used to use the trained segmentation model Totalsegmentator to annotate the unlabeled anatomical organ types in the original CT image dataset and generate noise pseudo-labels;

[0042] A degradation module is used to degrade the manually annotated labels to obtain degraded manually annotated labels;

[0043] A multimodal model BiomedCLIP, wherein the multimodal model BiomedCLIP comprises an image encoder and a text encoder;

[0044] The multi-view construction module is used to form a training data set from the original CT images, manually annotated labels, and degraded manually annotated labels, and then transform the training data set and input it into the image encoder of the multimodal model BiomedCLIP;

[0045] The image encoder is used to extract visual features; the text encoder is used to extract text features;

[0046] The prompt word training module is used to initialize the prompt words and then input them into the text encoder of the multimodal model BiomedCLIP to extract text features; a joint loss function is constructed using visual features and text features to supervise and train the learnable prompt words, and the training goal is to minimize the joint loss; after the training is completed, the prompt word parameters and the parameters of the multimodal model BiomedCLIP are frozen.

[0047] The noise pseudo-label quality judgment module is used to superimpose the category information of the noise pseudo-label generated by the segmentation model Totalsegmentator on the trained prompt word, and then input it into the text encoder in the multimodal model BiomedCLIP to obtain text features, transform the original CT image and its corresponding noise pseudo-label and input it into the image encoder of the multimodal model BiomedCLIP to extract visual features; the quality of the noise pseudo-label is judged based on the similarity between text features and visual features.

[0048] The present invention also proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the pseudo-label selection method for computed tomography image segmentation based on a multimodal model as described above is implemented.

[0049] The present invention also proposes a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the pseudo-label selection method for computed tomography image segmentation based on a multimodal model as described above.

[0050] The beneficial effects of the present invention are:

[0051] The present invention can make explicit judgments on a large number of pseudo labels, and can directly screen out high-quality pseudo labels without any manual intervention, thus avoiding the problems of low efficiency and high subjectivity caused by artificially setting thresholds or judging label quality by human eyes. The screened high-quality labels are used to train the segmentation model of the computed tomography images, and the trained segmentation model of the computed tomography images is used to segment the organs of the computed tomography images. The method can automatically generate high-quality training data to guide the segmentation model to generate high-precision organ segmentation results without adding manual annotations. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is an overall framework diagram of the pseudo-label selection method for computed tomography image segmentation based on a multimodal model proposed in the present invention. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0054] Embodiment 1

[0055] The present invention proposes a pseudo label selection method for computed tomography image segmentation based on a multimodal model. The overall framework of the method is as follows: Figure 1 As shown, the following steps are included:

[0056] S1. Use the original CT image dataset with manually annotated labels to train the segmentation model Totalsegmentator; Totalsegmentator can segment 117 organ categories. The original CT image dataset contains N types of manually annotated labels. In this embodiment, the manually annotated labels are also referred to as high-quality labels.

[0057] S2. Expand the data set used for downstream tasks. Use the trained segmentation model Totalsegmentator to label the unlabeled anatomical organ types in the original CT image data set. The labels annotated by the segmentation model Totalsegmentator are noise pseudo labels. The N types of labeled data contained in the original data are expanded to M types of labeled data. At this time, the original CT image data set contains not only N types of manually labeled high-quality labeled data, but also MN types of noise pseudo-label data generated by Totalsegmentator. express;

[0058] S3, degrade the manually annotated labels to obtain degraded manually annotated labels. This part of the data is used Indicates; the degraded manually annotated labels are also referred to as low-quality labels in this embodiment; S3 is specifically:

[0059] Design the degradation module. The degradation module contains three types of transformations: 1) noise and interference, 2) geometric transformation, and 3) morphological change. The three categories specifically include the following transformations:

[0060] 1) Noise and interference

[0061] Add Random Noise: Simulates degradation of the signal by adding random noise.

[0062] Gaussian Blur: Reduces image detail through blurring, simulating reduced image resolution or loss of focus.

[0063] 2) Geometric transformation

[0064] Random displacement: Randomly move the pixel positions in the label image to simulate image registration errors.

[0065] Random affine transformation: Perform random scaling, rotation, and translation on the label image to simulate the geometric distortion during label annotation.

[0066] Elastic deformation: nonlinear deformation of the label image to simulate the natural changes of tissue structure.

[0067] 3) Morphological processing

[0068] Simulate missing areas: Randomly delete some areas in the label image to simulate the situation where some parts of the label are occluded or missing.

[0069] Add random shapes: Randomly add some irrelevant shapes to the label image to simulate mislabeling or interference in the label.

[0070] Random morphological transformation: Perform random morphological operations such as dilation and erosion on the label image to simulate the irregular changes of the image edge.

[0071] Each manually annotated label will select at least 2-5 of these transformations, and the selection is completely random. After three rounds of artificial degradation, all the data were obtained. .

[0072] S4. Prepare a pre-trained multimodal model BiomedCLIP, which includes an image encoder and a text encoder. The original CT image, the manually annotated labels, and the degraded manually annotated labels form a training data set. The training data set is represented as:

[0073] ;

[0074] Where D T is the training data set, is the i-th CT image, is the manually annotated label corresponding to the i-th CT image, is the degraded manual annotation label corresponding to the i-th CT image, and N is the number of samples in the training dataset.

[0075] After the training data set is transformed, it is input into the image encoder of the multimodal model BiomedCLIP to extract visual features; the specific process of transforming the training data set is: Corresponding or First perform maximum projection to obtain , then Corresponding and Converting from 3D to 2D images, the resulting 2D images are uniformly recorded as , then Perform edge extraction operation to obtain edge information ,Will , and Stacked together to form three channels of data.

[0076] The image encoder of the multimodal model BiomedCLIP is expressed as follows:

[0077] ;

[0078] In the formula, is the mapping function of the image encoder, is the input image set, Represents a D-dimensional feature space.

[0079] The text encoder of the multimodal model BiomedCLIP is expressed as follows:

[0080] ;

[0081] In the formula, is the mapping function of the text encoder, is the input text set, Represents a D-dimensional feature space.

[0082] S5. Construct a learnable prompt word (prompt). The prompt word is set to 128 tokens. A token is the smallest unit after text segmentation in the field of natural language processing. It can be a word, a subword, or a special symbol. After initializing the prompt word, input it into the text encoder of the multimodal model BiomedCLIP to extract text features;

[0083] S6. Use visual features and text features to construct a joint loss function to supervise and train the learnable prompt. After several rounds of training, the prompt can effectively express high-quality labels and low-quality labels. The training goal is to minimize the joint loss; it can be expressed as:

[0084]

[0085] In the formula, is the joint loss function, is the cross entropy loss, is the weight of the contrast loss, is the contrast loss;

[0086] The cross entropy loss is used to ensure that the features output by the image encoder are aligned with the features (labels) output by the correct text encoder. The calculation formula of the cross entropy loss is as follows:

[0087] ;

[0088] Where N is the number of samples in the training data set, is the indicator function, is the i-th CT image, is the manually annotated label corresponding to the i-th CT image, is the degraded manual annotation label corresponding to the i-th CT image, represents the probability that the label of the i-th CT image is a manually annotated label, represents the probability that the label of the i-th CT image is a degraded manual annotation label;

[0089] In order to enhance the prompt's ability to judge high-quality and low-quality labels, ensure that the image features of high-quality labels are close to their corresponding text features, and keep a distance from low-quality text features, a contrast loss is designed to further supervise the training of the learnable prompt. The calculation formula of the contrast loss is as follows:

[0090] ;

[0091] In the formula, is the visual feature of the i-th manually annotated label, is the text feature of the i-th manually annotated label, and m represents the lower limit of the distance between the visual feature of the manually annotated label and the text feature of the degraded manually annotated label; Represents the text features of the i-th degenerate human annotation label.

[0092] After training is completed, freeze the cue word parameters and the parameters of the multimodal model BiomedCLIP.

[0093] S7. The noise pseudo-labels generated by the large segmentation model Totalsegmentator actually have category information. In order to further improve the robustness of the model's label judgment, the category information of the noise pseudo-labels generated by the large segmentation model Totalsegmentator is superimposed on the trained prompt words, and then input into the text encoder in the multimodal model BiomedCLIP to obtain text features. The original CT image and its corresponding noise pseudo-label are transformed and input into the image encoder of the multimodal model BiomedCLIP to extract visual features.

[0094] S8, calculating the similarity between the text features and the visual features obtained in S7, and judging the quality of the noise pseudo-label based on the similarity between the text features and the visual features. In this embodiment, the similarity between the text features and the visual features is measured by cosine similarity, and the expression of cosine similarity is as follows:

[0095] ;

[0096] In the formula, Represents the cosine similarity between text features and visual features, and For the input image and text, Represents visual features, Represents text features.

[0097] Embodiment 2

[0098] The present invention proposes a pseudo label selection system for computed tomography image segmentation based on a multimodal model corresponding to the method of embodiment 1, comprising:

[0099] The noise pseudo-label generation module is used to use the trained segmentation model Totalsegmentator to annotate the unlabeled anatomical organ types in the original CT image dataset and generate noise pseudo-labels;

[0100] A degradation module is used to degrade the manually annotated labels to obtain degraded manually annotated labels;

[0101] A multimodal model BiomedCLIP, wherein the multimodal model BiomedCLIP comprises an image encoder and a text encoder;

[0102] The multi-view construction module is used to form a training data set from the original CT images, manually annotated labels, and degraded manually annotated labels, and then transform the training data set and input it into the image encoder of the multimodal model BiomedCLIP;

[0103] The image encoder is used to extract visual features; the text encoder is used to extract text features;

[0104] The prompt word training module is used to initialize the prompt words and then input them into the text encoder of the multimodal model BiomedCLIP to extract text features; a joint loss function is constructed using visual features and text features to supervise and train the learnable prompt words, and the training goal is to minimize the joint loss; after the training is completed, the prompt word parameters and the parameters of the multimodal model BiomedCLIP are frozen.

[0105] The noise pseudo-label quality judgment module is used to superimpose the category information of the noise pseudo-label generated by the segmentation model Totalsegmentator on the trained prompt word, and then input it into the text encoder in the multimodal model BiomedCLIP to obtain text features, transform the original CT image and its corresponding noise pseudo-label and input it into the image encoder of the multimodal model BiomedCLIP to extract visual features; the quality of the noise pseudo-label is judged based on the similarity between text features and visual features.

[0106] The implementation methods of each module and module function in the system are completely consistent with the steps of the method in Example 1, so they will not be repeated here.

[0107] Embodiment 3

[0108] The present invention proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the pseudo-label selection method for computed tomography image segmentation based on a multimodal model as described in Example 1 is implemented.

[0109] Embodiment 4

[0110] The present invention proposes a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the pseudo-label selection method for computed tomography image segmentation based on a multimodal model as described in the first embodiment.

[0111] The present invention can make explicit judgments on a large number of pseudo labels, and can directly screen out high-quality pseudo labels without any manual intervention, thus avoiding the problems of low efficiency and high subjectivity caused by artificially setting thresholds or judging label quality by human eyes. The screened high-quality labels are used to train the segmentation model of the computed tomography images, and the trained segmentation model of the computed tomography images is used to segment the organs of the computed tomography images. The present invention can automatically generate high-quality training data to guide the segmentation model to generate high-precision organ segmentation results without adding manual annotations.

[0112] In the embodiments disclosed in the present application, the computer storage medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or apparatus. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. More specific examples of computer storage media may include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0113] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0114] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A pseudo-label selection method for computed tomography image segmentation based on a multimodal model, characterized in that: The following steps are involved: S1, use the original CT image dataset with manually annotated labels to train the segmentation model Totalsegmentator; S2. Use the trained segmentation model Totalsegmentator to label the unlabeled anatomical organ types in the original CT image dataset. The labels labeled by the segmentation model Totalsegmentator are noise pseudo labels. S3, degenerating the manually annotated labels to obtain degraded manually annotated labels; S4, preparing a pre-trained multimodal model BiomedCLIP, wherein the multimodal model BiomedCLIP comprises an image encoder and a text encoder, and the original CT image, the manually annotated labels and the degraded manually annotated labels form a training data set, and the training data set is transformed and input into the image encoder of the multimodal model BiomedCLIP to extract visual features; S5, construct a learnable prompt word, initialize the prompt word and input it into the text encoder of the multimodal model BiomedCLIP to extract text features; S6. Construct a joint loss function using visual features and text features to supervise and train learnable prompt words, with the training goal of minimizing the joint loss. After the training is completed, freeze the prompt word parameters and the parameters of the multimodal model BiomedCLIP. S7, superimposing the category information of the noise pseudo-label generated by the segmentation model Totalsegmentator onto the trained prompt word, and then inputting it into the text encoder in the multimodal model BiomedCLIP to obtain text features, transforming the original CT image and its corresponding noise pseudo-label and inputting it into the image encoder in the multimodal model BiomedCLIP to extract visual features; S8. Calculate the similarity between the text features and the visual features obtained in S7, and judge the quality of the noise pseudo-label based on the similarity between the text features and the visual features.

2. The pseudo label selection method for computed tomography image segmentation based on a multimodal model according to claim 1, characterized in that: In S3, the degradation of the manually annotated labels is specifically as follows: Multiple degradation operations are randomly selected from the following degradation operations to degrade the manually annotated labels: adding random noise, Gaussian blur, randomly moving pixel positions, random affine transformation, nonlinear deformation, simulating missing areas, adding random shapes and random morphological transformation.

3. The pseudo label selection method for computed tomography image segmentation based on a multimodal model according to claim 1, characterized in that: In S4, the training data set is represented as: ； Where D T is the training data set, is the i-th CT image, is the manually annotated label corresponding to the i-th CT image, is the degraded manual annotation label corresponding to the i-th CT image, and N is the number of samples in the training dataset; The transformation of the training data set is specifically as follows: Will Corresponding or First perform maximum projection to obtain , then Corresponding and Converting from 3D to 2D images, the resulting 2D images are uniformly recorded as , then Perform edge extraction operation to obtain edge information ,Will , and Stacked together to form three channels of data; The image encoder of the multimodal model BiomedCLIP is expressed as follows: ； In the formula, is the mapping function of the image encoder, is the input image set, Represents a D-dimensional feature space.

4. The pseudo label selection method for computed tomography image segmentation based on a multimodal model according to claim 1, characterized in that: In S5, the text encoder of the multimodal model BiomedCLIP is expressed as follows: ； In the formula, is the mapping function of the text encoder, is the input text set, Represents a D-dimensional feature space.

5. The pseudo label selection method for computed tomography image segmentation based on a multimodal model according to claim 1, characterized in that: In S6, the training objective is to minimize the joint loss and is expressed as: ； In the formula, is the joint loss function, is the cross entropy loss, is the weight of the contrast loss, is the contrast loss; The calculation formula for cross entropy loss is as follows: ； Where N is the number of samples in the training data set, is the indicator function, is the i-th CT image, is the manually annotated label corresponding to the i-th CT image, is the degraded manual annotation label corresponding to the i-th CT image, represents the probability that the label of the i-th CT image is a manually annotated label, represents the probability that the label of the i-th CT image is a degraded manual annotation label; The calculation formula of contrast loss is as follows: ； In the formula, is the visual feature of the i-th manually annotated label, is the text feature of the i-th manually annotated label, and m represents the lower limit of the distance between the visual feature of the manually annotated label and the text feature of the degraded manually annotated label; Represents the text features of the i-th degenerate human annotation label.

6. The pseudo label selection method for computed tomography image segmentation based on a multimodal model according to claim 1, characterized in that: In S8, the similarity between the text feature and the visual feature is specifically measured using cosine similarity.

7. A pseudo-label selection system for computed tomography image segmentation based on a multimodal model, characterized in that: include: The noise pseudo-label generation module is used to use the trained segmentation model Totalsegmentator to annotate the unlabeled anatomical organ types in the original CT image dataset and generate noise pseudo-labels; A degradation module is used to degrade the manually annotated labels to obtain degraded manually annotated labels; A multimodal model BiomedCLIP, wherein the multimodal model BiomedCLIP comprises an image encoder and a text encoder; The multi-view construction module is used to form a training data set from the original CT images, manually annotated labels, and degraded manually annotated labels, and then transform the training data set and input it into the image encoder of the multimodal model BiomedCLIP; The image encoder is used to extract visual features; The text encoder is used to extract text features; The prompt word training module is used to initialize the prompt words and then input them into the text encoder of the multimodal model BiomedCLIP to extract text features; a joint loss function is constructed using visual features and text features to supervise and train the learnable prompt words, and the training goal is to minimize the joint loss; after the training is completed, the prompt word parameters and the parameters of the multimodal model BiomedCLIP are frozen. The noise pseudo-label quality judgment module is used to superimpose the category information of the noise pseudo-label generated by the segmentation model Totalsegmentator onto the trained prompt word, and then input it into the text encoder in the multimodal model BiomedCLIP to obtain text features. The original CT image and its corresponding noise pseudo-label are transformed and input into the image encoder of the multimodal model BiomedCLIP to extract visual features. The quality of noisy pseudo-labels is judged based on the similarity between text features and visual features.

8. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the pseudo label selection method for computed tomography image segmentation based on a multimodal model as described in any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables a computer to execute the pseudo label selection method for computed tomography image segmentation based on a multimodal model as described in any one of claims 1 to 6.

Images