Multi-organ segmentation method, training method, medium and electronic device

Abstract

The application provides a multi-organ segmentation method, a training method, a medium and an electronic device. The multi-organ segmentation method comprises: acquiring a medical image, wherein the medical image comprises multiple organs; and processing the medical image by using a multi-organ segmentation model based on contrast learning to obtain a multi-organ segmentation result of the medical image, wherein the multi-organ segmentation model comprises an encoder module, a feature projection module and a decoder module, the output ends of the encoder module are connected with the feature projection module and the decoder module respectively, and the feature projection module is used to obtain features required by contrast learning. The multi-organ segmentation method has the advantages of high efficiency, high speed and the like, and can avoid segmentation errors caused by human subjective errors.

Description

Technical Field

[0001] This invention relates to a medical image processing method, and more particularly to a multi-organ segmentation method, training method, medium, and electronic device. Background Technology

[0002] Organ segmentation refers to identifying the regions belonging to various organs in medical images and recognizing the organs within those images. Multi-organ segmentation of abdominal CT images can identify multiple abdominal organs in medical images. The human abdominal cavity contains several important organs, including the liver, gallbladder, pancreas, spleen, and kidneys. Multi-organ segmentation refers to segmenting the regions encompassed by each target organ in the image; there are multiple target organs. Abdominal CT images have low contrast, and the boundaries of various organs are blurred and diverse. Due to abdominal pressure, some organs in the human body (such as the stomach, pancreas, and liver) exhibit significant deformation, and there is structural overlap between different organs. Manual segmentation is extremely time-consuming and labor-intensive; therefore, the use of multi-organ segmentation technology to achieve automatic segmentation of multiple abdominal organs is of great significance. Summary of the Invention

[0003] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a multi-organ segmentation method, training method, medium and electronic device for realizing automatic segmentation of multiple organs.

[0004] To achieve the above and other related objectives, a first aspect of the present invention provides a multi-organ segmentation method. The multi-organ segmentation method includes: acquiring a medical image, the medical image including multiple organs; processing the medical image using a contrastive learning-based multi-organ segmentation model to obtain multi-organ segmentation results of the medical image, wherein the multi-organ segmentation model includes an encoder module, a feature projection module, and a decoder module, the output of the encoder module being connected to the feature projection module and the decoder module respectively, and the feature projection module being used to acquire features required for contrastive learning.

[0005] In one embodiment of the first aspect, the training method of the multi-organ segmentation model includes: acquiring a dataset, the dataset including single-organ labeled data and / or multi-organ labeled data; acquiring a training dataset, a test dataset, and a validation dataset based on the dataset; constructing the multi-organ segmentation model; training the multi-organ segmentation model using the training dataset, and using the validation dataset to verify the state and convergence of the multi-organ segmentation model during training; and testing the performance of the multi-organ segmentation model using the test dataset.

[0006] In one embodiment of the first aspect, the dataset includes CT sequence images. Obtaining a training dataset, a test dataset, and a validation dataset based on the dataset includes: setting the voxel spacing of the three dimensions of the CT sequence images to the same value; setting the orientation of the CT sequence images to the same value; cropping the CT values ​​of the CT sequence images; normalizing the CT sequence images; and dividing the CT sequence images into the training dataset, the test dataset, and the validation dataset according to a preset ratio.

[0007] In one embodiment of the first aspect, training the multi-organ segmentation model using the training dataset includes: using supervised learning to update the network parameters of the multi-organ segmentation model for labeled regions in the training dataset, and using contrastive learning to update the network parameters of the multi-organ segmentation model for unlabeled regions in the training dataset.

[0008] In one embodiment of the first aspect, training the multi-organ segmentation model using any image data in the training dataset includes: sampling the image data and its labeled data to obtain two overlapping 3D image slices; inputting the two overlapping 3D image slices into the multi-organ segmentation model to obtain two segmentation probability maps; updating the network parameters of the multi-organ segmentation model using supervised learning for labeled regions in the image data; and updating the network parameters of the multi-organ segmentation model using contrastive learning for unlabeled regions in the image data; wherein the loss function used in the supervised learning and the contrastive learning is obtained from the two segmentation probability maps.

[0009] In one embodiment of the first aspect, the method for obtaining the first loss function used in the supervised learning includes: selecting a set of labeled voxels from the two segmentation probability maps based on the labeled data of the image data; obtaining the set of real labels corresponding to the set of labeled voxels from the labeled data of the image data; and obtaining the first loss function based on the set of labeled voxels and its corresponding set of real labels.

[0010] In one embodiment of the first aspect, the method for obtaining the second loss function used in the contrastive learning includes: filtering the overlapping regions of the two segmentation probability maps to obtain the predicted values ​​of unlabeled voxels; using the feature projection module to obtain the feature set corresponding to the unlabeled voxels; obtaining the confidence set corresponding to the unlabeled voxels based on the predicted values ​​of the unlabeled voxels; and obtaining the second loss function based on the feature set and confidence set corresponding to the unlabeled voxels.

[0011] A second aspect of the present invention provides a training method for a multi-organ segmentation model, the training method comprising: acquiring a dataset, the dataset including single-organ labeled data and / or multi-organ labeled data; acquiring a training dataset, a test dataset, and a validation dataset based on the dataset; constructing the multi-organ segmentation model; training the multi-organ segmentation model using the training dataset, and using the validation dataset to verify the state and convergence of the multi-organ segmentation model during training; and testing the performance of the multi-organ segmentation model using the test dataset.

[0012] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-organ segmentation method described in any one of the first aspects of the present invention, and / or the training method for the multi-organ segmentation model described in the second aspect of the present invention.

[0013] A fourth aspect of the present invention also provides an electronic device, the electronic device comprising: a memory storing a computer program; and a processor communicatively connected to the memory, wherein, when the computer program is invoked, the processor executes the multi-organ segmentation method described in any one of the first aspects of the present invention, and / or the training method for the multi-organ segmentation model described in the second aspect of the present invention.

[0014] As described above, the multi-organ segmentation method, training method, medium, and electronic device provided in one or more embodiments of the present invention have the following beneficial effects:

[0015] The described multi-organ segmentation method utilizes a contrastive learning-based multi-organ segmentation model to segment medical images and obtain multi-organ segmentation results. This process requires minimal human intervention, thus offering advantages such as high efficiency and speed, while also avoiding segmentation errors caused by human subjectivity.

[0016] Furthermore, the multi-organ segmentation model can effectively utilize single-organ labeled data to increase the amount of training data during training, thereby improving performance. Moreover, the multi-organ segmentation model can utilize not only labeled image regions but also unlabeled image regions during training, which further enhances training effectiveness. Attached Figure Description

[0017] Figure 1 The flowchart shown is a multi-organ segmentation method in an embodiment of the present invention.

[0018] Figure 2 The flowchart shown is a training method for a multi-organ segmentation model in an embodiment of the present invention.

[0019] Figure 3The flowchart shown is a process for obtaining the training dataset, test dataset, and validation dataset in an embodiment of the present invention.

[0020] Figure 4 The diagram shows a detailed flowchart of the training method for the multi-organ segmentation model in an embodiment of the present invention.

[0021] Figure 5 The flowchart shown is a method for obtaining the first loss function in an embodiment of the present invention.

[0022] Figure 6 The flowchart shown is a method for obtaining the second loss function in an embodiment of the present invention.

[0023] Figure 7 The diagram shown is a structural schematic of an electronic device according to an embodiment of the present invention.

[0024] Component designation explanation

[0025] 700 electronic devices

[0026] 710 Memory

[0027] 720 processor

[0028] 730 monitor

[0029] S11~S12

[0030] S21~S25

[0031] Steps S31 to S35

[0032] Steps S41 to S44

[0033] Steps S51 to S53 Detailed Implementation

[0034] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0035] It should be noted that the illustrations provided in the following embodiments are merely schematic representations of the basic concept of the present invention. The illustrations only show components relevant to the present invention and are not drawn according to the actual number, shape, and size of the components in implementation. In actual implementation, the form, quantity, and proportion of each component can be arbitrarily changed, and the component layout may also be more complex. Furthermore, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0036] Abdominal CT images have low contrast, and the boundaries of various organs in the images are blurred and diverse. Due to the influence of abdominal pressure, some organs in the human body (such as the stomach, pancreas, liver, etc.) are greatly deformed and there is structural overlap between different organs. Manual segmentation is very time-consuming and laborious. Therefore, it is of great significance to realize the automatic segmentation of multiple organs in the abdomen using multi-organ segmentation technology.

[0037] To address at least the aforementioned problems, this invention provides a multi-organ segmentation method. This method utilizes a contrastive learning-based multi-organ segmentation model to segment medical images, yielding multi-organ segmentation results. This process requires minimal human intervention, thus offering advantages such as high efficiency and speed, while also avoiding segmentation errors caused by human subjectivity.

[0038] The multi-organ segmentation method provided by the present invention will now be described in detail through specific embodiments and accompanying drawings.

[0039] Figure 1 The flowchart shown is a multi-organ segmentation method according to an embodiment of the present invention. Figure 1 As shown, the multi-organ segmentation method provided in this embodiment includes the following steps S11 and S12.

[0040] S11, Acquire medical images, which include multiple organs. In some embodiments, the medical images may be abdominal CT images including organs such as the liver, gallbladder, pancreas, spleen, and kidneys, but the present invention is not limited thereto.

[0041] S12, a multi-organ segmentation model based on contrastive learning is used to process medical images to obtain multi-organ segmentation results. The multi-organ segmentation model includes an encoder module, a feature projection module, and a decoder module. The output of the encoder module is connected to both the feature projection module and the decoder module. The feature projection module is used to acquire the features required for contrastive learning.

[0042] As described above, the multi-organ segmentation method provided in this embodiment utilizes a contrastive learning-based multi-organ segmentation model to segment medical images and obtain multi-organ segmentation results. This process requires virtually no human intervention, thus offering advantages such as high efficiency and speed, while also avoiding segmentation errors caused by human subjective mistakes.

[0043] In one embodiment of the present invention, the encoder module includes four downsampling layers for encoding medical images to obtain high-level semantic features of the images. Two branches are connected at the end of the encoder module: a feature projection module and a decoder module. The feature projection module includes a first three-dimensional convolutional layer, a ReLU activation function layer, and a second three-dimensional convolutional layer connected in sequence. The decoder module includes four upsampling layers. Each upsampling layer in the decoder module is skipped to a corresponding downsampling layer and receives features directly passed from the corresponding downsampling layer as well as features passed from the previous upsampling layer.

[0044] It should be noted that the structure of the encoder module, feature projection module and decoder module described above is only one feasible way in this embodiment, but the present invention is not limited thereto.

[0045] Figure 2 This is shown as a training method for a multi-organ segmentation model in one embodiment of the present invention. For example... Figure 2 As shown, the training method for the multi-organ segmentation model in this embodiment includes the following steps S21 to S25.

[0046] S21, Obtain the dataset, which includes single-organ annotation data and / or multi-organ annotation data. For any image data in the dataset, if the image data contains 5 organs, the multi-organ annotation data can be the annotation data of all 5 organs in the image data, or it can be the annotation data of 2 or 3 organs in the image data. It should be noted that although the single-organ annotation data only contains the annotation of a specific organ and does not contain the annotation of other organs, the single-organ annotation data still contains some useful information. Therefore, the multi-organ segmentation model can improve the performance of single-organ segmentation based on this useful information.

[0047] Optionally, if this embodiment requires segmenting C organs, for a series of datasets {D0, ..., D...} K}, D k ={(X k ;Y k )},k∈[0,1,...,K],D k A collection containing input images and the organ annotations corresponding to these input images. in If D kIf there is no label for organ c, then In this embodiment, a C-dimensional one-hot vector T can be used. k ∈{0,1} C D represents k Does it contain a label for a specific organ?

[0048] S22, obtain the training dataset, test dataset, and validation dataset based on the dataset, wherein the training dataset, test dataset, and validation dataset have no overlap with each other.

[0049] S23, Construct a multi-organ segmentation model. In some feasible implementations, a multi-organ segmentation model can be constructed based on U-Net, but this invention is not limited thereto.

[0050] S24. The multi-organ segmentation model is trained using the training dataset, and the state and convergence of the multi-organ segmentation model are checked using the validation dataset during training.

[0051] S25, the performance of the multi-organ segmentation model is tested using the test dataset.

[0052] Optionally, the dataset may include CT sequence images. Figure 3 This embodiment demonstrates the method used to train the multi-organ segmentation model using the training dataset. For example... Figure 3 As shown, in this embodiment, training the multi-organ segmentation model using the training dataset includes the following steps S31 to S35.

[0053] S31, set the voxel spacing of the three dimensions of the CT sequence image to the same value, which can be, for example, a common nominal value.

[0054] S32, set the orientation of the CT sequence images to be the same. For example, the orientation of the CT sequence images can be set to the same RAI coordinate.

[0055] S33, truncate the CT values ​​of the CT sequence image. For example, the CT values ​​of the CT sequence image can be truncated to the range of [-68, 200].

[0056] S34, Normalize the CT sequence images. For example, normalization can be achieved by subtracting the image mean from the CT value of each pixel in the CT sequence image and dividing by the variance.

[0057] S35 divides the CT sequence images into training dataset, test dataset, and validation dataset according to a preset ratio.

[0058] In one embodiment of the present invention, training a multi-organ segmentation model using a training dataset includes: using supervised learning to update the network parameters of the multi-organ segmentation model for labeled regions in the training dataset, and using contrastive learning to update the network parameters of the multi-organ segmentation model for unlabeled regions in the training dataset.

[0059] Optionally, for each dataset D k Any image data in Using this image data The process of training a multi-organ segmentation model is as follows: Figure 4 As shown, the process includes steps S41 to S44.

[0060] S41, for image data and its labeled data Sampling is performed to obtain two overlapping 3D image slices. and in and The dimensions are the same, and their corresponding annotations are respectively... and

[0061] S42, slice two overlapping 3D images and Input the multi-organ segmentation model to obtain two segmentation probability maps and in Corresponding to Corresponding to according to and You can obtain their respective pseudo-tags. and

[0062] S43. For the labeled regions in the image data, the network parameters of the multi-organ segmentation model are updated using supervised learning.

[0063] S44. For unlabeled regions in the image data, the network parameters of the multi-organ segmentation model are updated using a contrastive learning approach.

[0064] In supervised learning and contrastive learning, the loss function used is obtained from two segmentation probability maps.

[0065] Optionally, Figure 5 The flowchart shown in this embodiment illustrates the process of obtaining the first loss function, where the first loss function is the loss function used in step S43. For example... Figure 5 As shown, the method for obtaining the first loss function includes the following steps S51 to S53.

[0066] S51, based on the labeled data of the image data, select the set of labeled voxels from the two segmentation probability maps. Where N l This represents the total number of labeled voxels.

[0067] S52, Obtain the labeled voxel set from the labeled data of the image data. Corresponding real annotation set

[0068] S53, based on the labeled voxel set and its corresponding set of real annotations Obtain the first loss function. In some feasible implementations, the first loss function L Sup It can be expressed using the following formula (1).

[0069]

[0070] Among them, T k,c Represents dataset D k Is there any annotation information for organ c? If so, then T k,c If it is 1, then T. k,c =0, Represents dataset D k If the i-th data contains labeled information, does the n-th voxel in the voxel set belong to organ c? If so, then... =1, otherwise =0, The model represents the dataset D. k The probability that the nth voxel in the set of labeled voxels in the i-th data is predicted to be organ c ranges from 0 to 1.

[0071] Optionally, Figure 6 The flowchart shown is for obtaining the second loss function in an embodiment of the present invention, wherein the second loss function is the loss function used in step S44. Figure 6 As shown, the method for obtaining the second loss function in this embodiment includes the following steps S61 to S64.

[0072] S61, for the two segmentation probability maps and The overlapping regions are filtered to obtain the predicted values ​​of unlabeled voxel points. in, express The corresponding predicted values ​​for unlabeled voxels, express The predicted value of the corresponding unlabeled voxel.

[0073] S62, use the feature projection module to obtain the feature set corresponding to the unlabeled voxel points. in, Corresponding to Corresponding to

[0074] S63, Predicted values ​​for unlabeled voxels Downsampling is performed to obtain the voxel prediction set corresponding to the feature set. Then obtain the confidence set corresponding to the unlabeled voxel points. Where, N uo This represents the total number of unlabeled voxels after downsampling. express The predicted probability of the corresponding voxel. express The corresponding voxel prediction probability.

[0075] S64, obtain the second loss function based on the feature set and confidence set corresponding to the unlabeled voxels. For example, in some feasible implementations, directional contrastive learning can be used, allowing learning to proceed from features with low confidence to features with high confidence. In this case, the second loss function L... CL It can be expressed using the following formula (2).

[0076]

[0077] in:

[0078]

[0079]

[0080] in, φ represents the cosine similarity between two corresponding features. ne F represents the feature corresponding to a negative sample voxel with a different pseudo-label from voxel n. u This represents the set of negative sample voxels. This indicates whether the confidence level of voxel n predicted in the first 3D slice is less than the confidence level predicted in the second 3D slice. If it is less, the value is 1; otherwise, the value is 0. This indicates whether the confidence level of voxel n predicted in the second 3D slice is greater than a certain set threshold γ. If it is greater, the value is 1; otherwise, the value is 0. γ represents a pre-given threshold, which can be, for example, 0.6.

[0081] Optionally, based on the first and second loss functions mentioned above, the overall loss function L used in the training of the multi-organ segmentation model can be represented by the following equation (5).

[0082] L = L sup +λL CL Equation (5);

[0083] Where λ is a coefficient, the value of which can be set according to actual needs or experience.

[0084] As described above, in this embodiment, two overlapping 3D volumetric slices are sampled from the CT sequence images for training each time, making the network more robust to changes in the background. Furthermore, this embodiment uses contrastive learning to learn more robust voxel features, which can conveniently utilize unlabeled voxel information in single-organ labeled data, thus improving segmentation accuracy.

[0085] In one embodiment of the present invention, when processing medical images using the multi-organ segmentation model, for each medical image, a sliding window along three axes can be used to traverse all regions of the image. The image slices sampled each time are of the same size, and the number of overlapping pixels between two adjacent image slices is also the same.

[0086] Based on the above description of the multi-organ segmentation method and the training method of the multi-organ segmentation model, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements... Figure 1 The multi-organ segmentation method and / or shown Figure 2 The training method for the multi-organ segmentation model is shown.

[0087] In this invention, any combination of one or more storage media can be used. The storage medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, RAM, ROM, an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0088] The present invention also provides an electronic device. Figure 7 The diagram shown is a structural schematic of an electronic device 700 according to an embodiment of the present invention. Figure 7 As shown, in this embodiment, the electronic device 700 includes a memory 710 and a processor 720.

[0089] The memory 710 is used to store computer programs; preferably, the memory 710 includes various media that can store program code, such as ROM, RAM, magnetic disk, USB flash drive, memory card or optical disk.

[0090] Specifically, memory 710 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. Electronic device 700 may further include other removable / non-removable, volatile / non-volatile computer system storage media. Memory 710 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present invention.

[0091] The processor 720 is connected to the memory 710 and is used to execute the computer program stored in the memory 710, so that the electronic device 700 performs... Figure 1 The multi-organ segmentation method and / or shown Figure 2 The training method for the multi-organ segmentation model is shown.

[0092] Optionally, the processor 720 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0093] Optionally, in this embodiment, the electronic device 700 may further include a display 730. The display 730 is communicatively connected to the memory 710 and the processor 720, and is used to display the relevant GUI interactive interface of the multi-organ segmentation method and / or the training method of the multi-organ segmentation model.

[0094] The scope of protection for the multi-organ segmentation method and the training method for the multi-organ segmentation model described in this invention is not limited to the execution order of the steps listed in this embodiment. Any solution implemented by adding, subtracting, or replacing steps in the prior art based on the principles of this invention is included within the scope of protection of this invention.

[0095] In summary, the multi-organ segmentation method provided in this embodiment of the invention utilizes a contrastive learning-based multi-organ segmentation model to segment medical images and obtain multi-organ segmentation results. This process requires virtually no human intervention, thus offering advantages such as high efficiency and speed, while also avoiding segmentation errors caused by human subjectivity.

[0096] Furthermore, the multi-organ segmentation model can effectively utilize single-organ labeled data to increase the amount of training data during training, thereby improving performance. Moreover, the multi-organ segmentation model can utilize not only labeled image regions but also unlabeled image regions during training, which further enhances training effectiveness.

[0097] Therefore, this invention effectively overcomes the various shortcomings of the prior art and has high industrial application value.

[0098] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for multi-organ segmentation, characterized in that, The multi-organ segmentation method includes: Acquire medical images, which include multiple organs; The medical image is processed using a multi-organ segmentation model based on contrastive learning to obtain the multi-organ segmentation result of the medical image. The multi-organ segmentation model includes an encoder module, a feature projection module, and a decoder module. The output of the encoder module is connected to the feature projection module and the decoder module, respectively. The feature projection module is used to obtain the features required for contrastive learning. The training method for the multi-organ segmentation model includes: acquiring a dataset, which includes single-organ labeled data and multi-organ labeled data; acquiring a training dataset, a test dataset, and a validation dataset based on the dataset; constructing the multi-organ segmentation model; training the multi-organ segmentation model using the training dataset, and using the validation dataset to test the state and convergence of the multi-organ segmentation model during training; and testing the performance of the multi-organ segmentation model using the test dataset; wherein the number of organs to be segmented is C, and the number of organ labels in some labeled data in the dataset is less than C but greater than 1; Training the multi-organ segmentation model using the training dataset includes: using supervised learning to update the network parameters of the multi-organ segmentation model for labeled regions in the training dataset, and using contrastive learning to update the network parameters of the multi-organ segmentation model for unlabeled regions in the training dataset. For any image data in the training dataset, training the multi-organ segmentation model using the image data includes: sampling the image data and its labeled data to obtain two overlapping 3D image slices; inputting the two overlapping 3D image slices into the multi-organ segmentation model to obtain two segmentation probability maps; updating the network parameters of the multi-organ segmentation model using supervised learning for labeled regions in the image data; and updating the network parameters of the multi-organ segmentation model using contrastive learning for unlabeled regions in the image data; wherein the loss function used in the supervised learning and the contrastive learning is obtained from the two segmentation probability maps. The method for obtaining the first loss function used in the supervised learning includes: selecting a set of labeled voxels from the two segmentation probability maps based on the labeled data of the image data; obtaining the set of ground truth labels corresponding to the set of labeled voxels from the labeled data of the image data; and obtaining the first loss function based on the set of labeled voxels and its corresponding set of ground truth labels. The first loss function is as follows: Where, N l T represents the number of labeled voxels. k,c Represents dataset D k Is there any annotation information for organ C? Represents dataset D k Does the i-th data contain labeled information about whether the n-th voxel in the voxel set is organ c? The model represents the dataset D. k The probability that the nth voxel in the set of voxels with labeled information in the i-th data is predicted to be organ c; The method for obtaining the second loss function used in the contrastive learning includes: filtering the overlapping regions of the two segmentation probability maps to obtain the predicted values ​​of unlabeled voxels; using the feature projection module to obtain the feature set corresponding to the unlabeled voxels; obtaining the confidence set corresponding to the unlabeled voxels based on the predicted values ​​of the unlabeled voxels; and obtaining the second loss function based on the feature set and confidence set corresponding to the unlabeled voxels. The second loss function is as follows: in: Where, N uo This represents the total number of unlabeled voxels after downsampling. φ represents the cosine similarity between two corresponding features. ne F represents the feature corresponding to a negative sample voxel with a different pseudo-label from voxel n. u This represents the set of negative sample voxels. This indicates whether the confidence level of the prediction for voxel n in the first 3D slice is less than the confidence level of the prediction in the second 3D slice. This indicates whether the confidence level of the prediction of voxel n in the second 3D slice is greater than a certain set threshold γ, where γ represents the threshold. The total loss function used in training the multi-organ segmentation model, based on the first and second loss functions, is: L = L sup +λL CL , where λ is the coefficient.

2. The multi-organ segmentation method according to claim 1, characterized in that, The dataset includes CT sequence images, and the training dataset, test dataset, and validation dataset are obtained based on the dataset, including: Set the voxel spacing in the three dimensions of the CT sequence image to the same value; Set the orientation of the CT sequence images to be the same; The CT values ​​of the CT sequence images are extracted; The CT sequence images are normalized. The CT sequence images are divided into the training dataset, the test dataset, and the validation dataset according to a preset ratio.

3. A training method for a multi-organ segmentation model, characterized in that, The training method for the multi-organ segmentation model includes: acquiring a dataset, which includes single-organ labeled data and / or multi-organ labeled data; Based on the dataset, obtain the training dataset, test dataset, and validation dataset; Construct the multi-organ segmentation model; The multi-organ segmentation model is trained using the training dataset, and the state and convergence of the multi-organ segmentation model are tested using the validation dataset during training. The performance of the multi-organ segmentation model was tested using the test dataset. The number of organs to be segmented is C, and the number of organ annotations in some labeled data in the dataset is less than C but greater than 1. Training the multi-organ segmentation model using the training dataset includes: using supervised learning to update the network parameters of the multi-organ segmentation model for labeled regions in the training dataset, and using contrastive learning to update the network parameters of the multi-organ segmentation model for unlabeled regions in the training dataset. For any image data in the training dataset, training the multi-organ segmentation model using the image data includes: sampling the image data and its labeled data to obtain two overlapping 3D image slices; inputting the two overlapping 3D image slices into the multi-organ segmentation model to obtain two segmentation probability maps; updating the network parameters of the multi-organ segmentation model using supervised learning for labeled regions in the image data; and updating the network parameters of the multi-organ segmentation model using contrastive learning for unlabeled regions in the image data; wherein the loss function used in the supervised learning and the contrastive learning is obtained from the two segmentation probability maps. The method for obtaining the first loss function used in the supervised learning includes: selecting a set of labeled voxels from the two segmentation probability maps based on the labeled data of the image data; obtaining the set of ground truth labels corresponding to the set of labeled voxels from the labeled data of the image data; and obtaining the first loss function based on the set of labeled voxels and its corresponding set of ground truth labels. The first loss function is as follows: Where, N l T represents the number of labeled voxels. k,c Represents dataset D k Is there any annotation information for organ C? Represents dataset D k Does the i-th data contain labeled information about whether the n-th voxel in the voxel set is organ c? The model represents the dataset D. k The probability that the nth voxel in the set of voxels with labeled information in the i-th data is predicted to be organ c; The method for obtaining the second loss function used in the contrastive learning includes: filtering the overlapping regions of the two segmentation probability maps to obtain the predicted values ​​of unlabeled voxels; using the feature projection module to obtain the feature set corresponding to the unlabeled voxels; obtaining the confidence set corresponding to the unlabeled voxels based on the predicted values ​​of the unlabeled voxels; and obtaining the second loss function based on the feature set and confidence set corresponding to the unlabeled voxels. The second loss function is as follows: in: Where, N uo This represents the total number of unlabeled voxels after downsampling. φ represents the cosine similarity between two corresponding features. ne F represents the feature corresponding to a negative sample voxel with a different pseudo-label from voxel n. u This represents the set of negative sample voxels. This indicates whether the confidence level of the prediction for voxel n in the first 3D slice is less than the confidence level of the prediction in the second 3D slice. This indicates whether the confidence level of the prediction of voxel n in the second 3D slice is greater than a certain set threshold γ, where γ represents the threshold. The total loss function used in training the multi-organ segmentation model, based on the first and second loss functions, is: L = L sup +λL CL , where λ is the coefficient.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the multi-organ segmentation method of any one of claims 1 to 2, and / or the training method of the multi-organ segmentation model of claim 3.

5. An electronic device, characterized in that, The electronic device includes: A memory that stores a computer program; The processor, which is communicatively connected to the memory, executes the multi-organ segmentation method according to any one of claims 1 to 2 and / or the training method of the multi-organ segmentation model according to claim 3 when calling the computer program.

Images