Semi-supervised medical image segmentation method based on multiple networks
Through the semi-supervised medical image segmentation method based on multi-network, the deep learning network is trained using pseudo-label samples, which solves the problem of difficulty in obtaining samples in medical image segmentation networks, and improves the accuracy and efficiency of image segmentation.
Patent Information
- Application Number
- CN202510367005.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-26
- Publication Date
- 2025-07-11
AI Technical Summary
In the field of medical image analysis, data annotation is expensive and challenging, resulting in high cost and small number of training samples acquisitions, and low image segmentation accuracy of existing medical image segmentation networks.
Using a semi-supervised medical image segmentation method based on multiple networks, by training multiple deep learning networks, the first and second image segmentation networks are screened out using the image segmentation loss function, pseudo-label samples are generated using the first label-free medical image samples, the second image segmentation network is trained, and the segmentation processing of the medical image samples to be segmented is combined with the first and third image segmentation networks.
It improves the accuracy of medical image segmentation, increases the number of training samples, reduces costs, and achieves efficient image segmentation effect.
Smart Images
Figure CN120298429A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of medical image segmentation, and particularly to a semi-supervised medical image segmentation method based on multiple networks. Background Art
[0002] In the field of computer vision, the accuracy of semantic segmentation has been significantly improved due to a large amount of labeled data. However, in the field of medical image analysis, data annotation is expensive and challenging, which requires a large number of medical expert annotators. Therefore, in the existing medical image segmentation technologies, there are technical defects such as high cost and small quantity of training samples obtained, resulting in low image segmentation accuracy of the trained medical image segmentation network. Summary of the Invention
[0003] Based on this, the purpose of the present application is to provide a semi-supervised medical image segmentation method based on multiple networks, which can overcome the deficiencies of the prior art.
[0004] In order to achieve the above purpose, the technical solution adopted by the present application is as follows:
[0005] A semi-supervised medical image segmentation method based on multiple networks, comprising:
[0006] Training a plurality of deep learning networks according to a plurality of initial medical image samples to obtain a plurality of image segmentation networks;
[0007] According to the image segmentation loss functions of each of the image segmentation networks, obtaining a plurality of first image segmentation networks and a plurality of second image segmentation networks; the image segmentation loss function of the first image segmentation network is less than that of the second image segmentation network;
[0008] Inputting a plurality of first unlabeled medical image samples into the plurality of first image segmentation networks respectively to obtain pseudo-labeled medical image samples output by the plurality of first image segmentation networks; the pseudo-labeled medical image samples include medical images corresponding to the predicted segmentation objects of the first image segmentation networks;
[0009] Training the plurality of second image segmentation networks according to the pseudo-labeled medical image samples to obtain a plurality of third image segmentation networks;
[0010] Performing segmentation processing on the medical image samples to be segmented through the first image segmentation networks and the third image segmentation networks to obtain target segmentation images.
[0011] In one embodiment, the step of obtaining a plurality of first image segmentation networks and a plurality of second image segmentation networks according to the image segmentation loss functions of each of the image segmentation networks includes:
[0012] Obtain the image segmentation loss functions of each of the image segmentation networks;
[0013] Arrange all the image segmentation networks from low to high according to the image segmentation loss functions, determine several of the image segmentation networks arranged in the front as the first image segmentation networks, and determine the remaining image segmentation networks as the second image segmentation networks.
[0014] In one embodiment, the step of obtaining the image segmentation loss functions of each of the image segmentation networks includes:
[0015] Input a plurality of first annotated medical image samples into each of the image segmentation networks to obtain the predicted segmentation images output by each of the image segmentation networks;
[0016] According to the predicted segmentation images of a plurality of the same image segmentation network and the segmentation objects annotated in the plurality of first annotated medical image samples, obtain the image segmentation loss function of this image segmentation network.
[0017] In one embodiment, the step of obtaining the image segmentation loss function of this image segmentation network according to the predicted segmentation images of a plurality of the same image segmentation network and the segmentation objects annotated in the plurality of first annotated medical image samples includes:
[0018] Obtain the image segmentation loss function through the following formula:
[0019]
[0020] where Dice i represents the image segmentation loss function of the i-th image segmentation network, represents the predicted segmentation image, T = y L represents the segmentation object, and ∈ represents the smoothing term.
[0021] In one embodiment, the step of inputting a plurality of first unannotated medical image samples into the several first image segmentation networks respectively to obtain the pseudo-labeled medical image samples output by the several first image segmentation networks includes:
[0022] Input the plurality of first unannotated medical image samples into the several first image segmentation networks to obtain a plurality of first predicted medical image samples;
[0023] According to the first voxel prediction values of the plurality of first predicted medical image samples, obtain the second voxel prediction values of the second predicted medical image samples corresponding to each of the first unannotated medical image samples;
[0024] Normalize the second voxel prediction values of each voxel in the second predicted medical image sample to obtain the pseudo-label medical image sample and the corresponding labeled voxel prediction values.
[0025] In one embodiment, the step of obtaining the second voxel prediction values of each of the first unlabeled medical image samples corresponding to the second predicted medical image samples according to the first voxel prediction values of the plurality of first predicted medical image samples includes:
[0026] Obtain the second voxel prediction values through the following formula:
[0027]
[0028] where represents the second voxel prediction value of the voxel point with coordinates (x, y, z), represents the first voxel prediction value of the voxel point with coordinates (x, y, z) in the first predicted medical image sample No. 1, represents the first voxel prediction value of the voxel point with coordinates (x, y, z) in the second predicted medical image sample No. 2, represents the first voxel prediction value of the voxel point with coordinates (x, y, z) in the I-th predicted medical image sample, and I is the last first predicted medical image sample.
[0029] In one embodiment, the step of normalizing the second voxel prediction values of each voxel in the second predicted medical image sample to obtain the pseudo-label medical image sample and the corresponding labeled voxel prediction values includes:
[0030] Obtain the labeled voxel prediction values through the following formula:
[0031]
[0032] where represents the labeled voxel prediction value of the voxel point with coordinates (x, y, z), τ represents a preset value, [k, x, y, z] represents the prediction value of the voxel point with coordinates (x, y, z) in the second predicted medical image sample being class k, and C represents the total number of voxel point classes, represents the second voxel prediction value of the voxel point with coordinates (x, y, z) being class c.
[0033] In one embodiment, the step of segmenting the medical image sample to be segmented through the first image segmentation network and the third image segmentation network to obtain the target segmentation image includes:
[0034] Input the medical image samples to be segmented into each of the first image segmentation networks and the third image segmentation network respectively to obtain a plurality of candidate segmentation images;
[0035] Generate the target segmentation image according to the voxel means of the voxels corresponding to the candidate segmentation images.
[0036] In one embodiment, the plurality of initial medical image samples include a plurality of second annotated medical image samples;
[0037] The step of training a plurality of deep learning networks according to a plurality of initial medical image samples to obtain a plurality of image segmentation networks includes:
[0038] Input the plurality of second annotated medical image samples into the plurality of deep learning networks respectively to obtain a plurality of first model output images;
[0039] Obtain the confusion region of the second annotated medical image sample according to the differences of the plurality of first model output images corresponding to the same second annotated medical image sample;
[0040] Construct the review loss function of each deep learning network according to the plurality of second annotated medical image samples, the plurality of first model output images and the confusion region;
[0041] Train the plurality of deep learning networks according to the review loss function until the plurality of image segmentation networks are obtained.
[0042] In one embodiment, the plurality of initial medical image samples include a plurality of second unannotated medical image samples;
[0043] The step of training a plurality of deep learning networks according to a plurality of initial medical image samples to obtain a plurality of image segmentation networks includes:
[0044] Perform image transformation on the plurality of second unannotated medical image samples to obtain a plurality of third unannotated image samples;
[0045] Input the plurality of second unannotated medical image samples and the plurality of third unannotated image samples into the plurality of deep learning networks respectively to obtain a plurality of second model output images and a plurality of third model output images;
[0046] Construct the consistency constraint function of each deep learning network according to the plurality of second model output images and the plurality of third model output images;
[0047] Train the plurality of deep learning networks according to the consistency constraint function until the plurality of image segmentation networks are obtained.
[0048] Compared with the traditional technology, the beneficial effects of the present application are as follows:
[0049] The semi-supervised medical image segmentation method based on multiple networks of the present application trains multiple deep learning networks according to multiple initial medical image samples. After obtaining multiple image segmentation networks, according to the image segmentation loss functions of each image segmentation network, a number of first image segmentation networks and a number of second image segmentation networks are obtained. Then, multiple first unlabeled medical image samples are respectively input into the number of first image segmentation networks to obtain pseudo-labeled medical image samples output by the number of first image segmentation networks. Then, according to the pseudo-labeled medical image samples, a number of second image segmentation networks are trained to obtain a number of third image segmentation networks, so as to obtain a number of first image segmentation networks and a number of third image segmentation networks with good network performance, and the medical image samples to be segmented are segmented through the first image segmentation networks and the third image segmentation networks to obtain the target segmentation image. The number of training samples can be increased through the image segmentation networks with lower image segmentation loss functions and better network performance for training the image segmentation networks with higher image segmentation loss functions and worse network performance. The number of training samples of the image segmentation networks with higher image segmentation loss functions and worse network performance can be increased efficiently and at low cost, so as to obtain the third image segmentation networks with lower image segmentation loss functions and better network performance. Combining with the first image segmentation networks with lower image segmentation loss functions and better network performance to segment the medical image samples to be segmented, a target segmentation image with high segmentation accuracy can be obtained.
[0050] For better understanding and implementation, the present application will be described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 It is a step diagram of the semi-supervised medical image segmentation method based on multiple networks according to an embodiment of the present application;
[0052] Figure 2 It is a flow chart of the semi-supervised medical image segmentation method based on multiple networks according to an embodiment of the present application; DETAILED DESCRIPTION OF THE EMBODIMENTS
[0053] To make the objectives, technical solutions and advantages of the present application clearer, the embodiments of the present application will be further described in detail below with reference to the accompanying drawings.
[0054] It should be clear that the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts belong to the scope protected by the embodiments of the present application.
[0055] When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. In the description of the present application, it should be understood that the terms "first", "second", "third", etc. are only used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to specific circumstances. The singular forms of "a", "the", and "said" used in the present application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. The words "if" / "when" used herein can be interpreted as "when...", "when...", or "in response to a determination".
[0056] In addition, in the description of the present application, unless otherwise specified, "a plurality" means two or more. "And / or" describes the association relationship of associated objects and indicates that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally represents an "or" relationship between the associated objects before and after.
[0057] Please refer to Figure 1-2 , Figure 1 is a step diagram of a multi-network-based semi-supervised medical image segmentation method according to the first embodiment of the present application, Figure 2 is a flowchart of a multi-network-based semi-supervised medical image segmentation method according to the first embodiment of the present application, including:
[0058] S1: Train multiple deep learning networks according to multiple initial medical image samples to obtain multiple image segmentation networks.
[0059] Among them, the initial medical image samples can be 3D images, such as two image datasets of LA MRI and Pancreas-CT, which contain left atrial MRI data and pancreas CT data.
[0060] Image transformation includes cropping, rotating, flipping, etc. of the image.
[0061] S2: Obtain a number of first image segmentation networks and a number of second image segmentation networks according to the image segmentation loss functions of each of the image segmentation networks; the image segmentation loss function of the first image segmentation network is less than that of the second image segmentation network.
[0062] S3: Input the multiple first unlabeled medical image samples into the several first image segmentation networks respectively to obtain the pseudo-labeled medical image samples output by the several first image segmentation networks; the pseudo-labeled medical image samples include the medical images corresponding to the predicted segmentation objects of the first image segmentation networks.
[0063] S4: Train the several second image segmentation networks according to the pseudo-labeled medical image samples to obtain several third image segmentation networks.
[0064] Among them, the training of the several second image segmentation networks can be guided by the contrast loss function of the several second image segmentation networks. For example, the number of the deep learning networks is 3, the first image segmentation networks are the first 2 deep learning networks, and the second image segmentation network is the 3rd deep learning network. At this time, the contrast loss function is shown in the following formula:
[0065]
[0066] Among them, represents the contrast loss function of the second image segmentation network, N represents the total number of voxel points of the pseudo-labeled medical image samples, and a voxel point refers to the smallest unit in a 3D image. represents the predicted value of the voxel point of the pseudo-labeled medical image sample. represents the predicted value of the voxel point output by the second image segmentation network.
[0067] S5: Perform segmentation processing on the medical image sample to be segmented through the first image segmentation network and the third image segmentation network to obtain the target segmentation image.
[0068] In a feasible embodiment, the step of S2: obtaining several first image segmentation networks and several second image segmentation networks according to the image segmentation loss functions of the respective image segmentation networks includes:
[0069] S21: Obtain the image segmentation loss functions of the respective image segmentation networks.
[0070] S22: Arrange all the image segmentation networks from low to high according to the image segmentation loss functions, determine the several image segmentation networks ranked in the front as the first image segmentation networks, and determine the remaining image segmentation networks as the second image segmentation networks.
[0071] In a feasible embodiment, the step of S21: obtaining the image segmentation loss functions of the respective image segmentation networks includes:
[0072] S211: Input multiple first labeled medical image samples into each of the image segmentation networks to obtain the predicted segmentation images output by each of the image segmentation networks.
[0073] S212: Based on the multiple predicted segmentation images of the same image segmentation network and the segmented objects labeled in the multiple first labeled medical image samples, obtain the image segmentation loss function of this image segmentation network.
[0074] In a feasible embodiment, the step S22: Based on the multiple predicted segmentation images of the same image segmentation network and the segmented objects labeled in the multiple first labeled medical image samples, obtaining the image segmentation loss function of this image segmentation network includes:
[0075] Obtain the image segmentation loss function through the following formula:
[0076]
[0077] where Dice i represents the image segmentation loss function of the i-th image segmentation network, represents the predicted segmentation image, T = y L represents the segmented object, and ∈ represents the smoothing term.
[0078] In a feasible embodiment, the step S3: Input multiple first unlabeled medical image samples into the several first image segmentation networks respectively to obtain the pseudo-labeled medical image samples output by the several first image segmentation networks includes:
[0079] S31: Input the multiple first unlabeled medical image samples into the several first image segmentation networks to obtain multiple first predicted medical image samples.
[0080] S32: Based on the first voxel prediction values of the multiple first predicted medical image samples, obtain the second voxel prediction values of the second predicted medical image samples corresponding to each of the first unlabeled medical image samples.
[0081] S33: Normalize the second voxel prediction values of each voxel of the second predicted medical image samples to obtain the pseudo-labeled medical image samples and the corresponding labeled voxel prediction values.
[0082] In a feasible embodiment, the step S32: Based on the first voxel prediction values of the multiple first predicted medical image samples, obtaining the second voxel prediction values of the second predicted medical image samples corresponding to each of the first unlabeled medical image samples includes:
[0083] Obtain the second voxel prediction value through the following formula:
[0084]
[0085] wherein, represents the second voxel prediction value of class c for the voxel point with coordinates (x, y, z), represents the first voxel prediction value of class c for the voxel point with coordinates (x, y, z) in the first predicted medical image sample of the first one, represents the first voxel prediction value of class c for the voxel point with coordinates (x, y, z) in the first predicted medical image sample of the second one, represents the first voxel prediction value of class c for the voxel point with coordinates (x, y, z) in the first predicted medical image sample of the I-th one, where I is the last first predicted medical image sample.
[0086] In a feasible embodiment, the step S33: normalizing the second voxel prediction values of each voxel of the second predicted medical image sample to obtain the pseudo-labeled medical image sample and the corresponding labeled voxel prediction values includes:
[0087] Obtaining the labeled voxel prediction values through the following formula:
[0088]
[0089] wherein, represents the labeled voxel prediction value of the voxel point with coordinates (x, y, z), τ represents a preset value, [k, x, y, z] represents the prediction value of class k for the voxel point with coordinates (x, y, z) in the second predicted medical image sample, C represents the total number of voxel point classes, represents the second voxel prediction value of class c for the voxel point with coordinates (x, y, z).
[0090] In a feasible embodiment, the step S5: segmenting the medical image sample to be segmented through the first image segmentation network and the third image segmentation network to obtain the target segmentation image includes:
[0091] S51: Inputting the medical image sample to be segmented into each of the first image segmentation network and the third image segmentation network respectively to obtain a plurality of candidate segmentation images.
[0092] S52: Generating the target segmentation image according to the voxel mean values of the same voxel corresponding to the candidate segmentation images.
[0093] In a feasible embodiment, the plurality of initial medical image samples include a plurality of second labeled medical image samples;
[0094] Step S1: Training multiple deep learning networks based on multiple initial medical image samples to obtain multiple image segmentation networks, including:
[0095] S111: Inputting the multiple second labeled medical image samples into the multiple deep learning networks respectively to obtain multiple first model output images.
[0096] S112: Obtaining the confusion region of the second labeled medical image sample according to the differences between multiple first model output images corresponding to the same second labeled medical image sample.
[0097] For example, the number of the deep learning networks is 3, including 2 VNets and 1 3D-ResVnet. Among them, VNet is a convolutional segmentation model for 3D images, and the deep learning network obtained by replacing the encoder of VNet with 3D convolutional ResNet34 is 3D-ResVnet.
[0098] The confusion region is obtained through the following formula:
[0099]
[0100] where Δ represents the confusion region, represents the first training segmentation image in binary form output by the first deep learning network, represents the first training segmentation image in binary form output by the second deep learning network, represents the first training segmentation image in binary form output by the third deep learning network, represents the first training segmentation image output by the i-th deep learning network, and σ(·) represents the binary conversion function.
[0101] S113: Constructing the review loss function of each deep learning network according to the multiple second labeled medical image samples, the multiple first model output images and the confusion region.
[0102] Among them, the review loss function of each deep learning network is constructed through the following formula:
[0103]
[0104] represents the review loss function of the i-th deep learning network, N represents the total number of voxel points of the second labeled medical image sample, Δ represents the confusion region, y L represents the second labeled medical image sample, represents the first model output image of the i-th deep learning network.
[0105] S114: Train the multiple deep learning networks according to the review loss function until the multiple image segmentation networks are obtained.
[0106] Optionally, when the function value of the review loss function is less than a preset review loss threshold, or the function values of the loss functions of each deep learning network are all less than a preset loss threshold, determine the multiple deep learning networks as multiple image segmentation networks.
[0107] In a feasible embodiment, the multiple initial medical image samples include multiple second unlabeled medical image samples;
[0108] The step S1 of training the multiple deep learning networks according to the multiple initial medical image samples to obtain multiple image segmentation networks includes:
[0109] S121: Perform image transformation on the multiple second unlabeled medical image samples to obtain multiple third unlabeled image samples.
[0110] For example, perform transformation processing on the second unlabeled medical image samples through the image transformation T(·) to obtain multiple third unlabeled image samples.
[0111] S122: Input the multiple second unlabeled medical image samples and the multiple third unlabeled image samples into the multiple deep learning networks respectively to obtain multiple second model output images and multiple third model output images.
[0112] S123: Construct a consistency constraint function for each of the deep learning networks according to the multiple second model output images and the multiple third model output images.
[0113] Among them, since the third unlabeled image samples are obtained by performing image transformation on the second unlabeled medical image samples, the result after performing inverse image transformation processing on the third model output images should be the same as that of the second model output images. That is, the consistency constraint function can be expressed as:
[0114]
[0115] Among them, represents the consistency constraint function, N represents the total number of voxel points of the second model output image, represents the second model output image, represents the third model output image, and T -1 (·) represents inverse transformation processing.
[0116] Among them, the inverse image transformation is the reverse transformation process of the image transformation. For example, if the image transformation is that the output image of the second model is rotated 5 degrees clockwise along the x-axis to obtain the third unlabeled image sample, then the inverse image transformation is that the output image of the third model is rotated 5 degrees counterclockwise.
[0117] S124: Train the multiple deep learning networks according to the consistency constraint function until the multiple image segmentation networks are obtained.
[0118] S53: Train each of the second image segmentation networks according to each of the loss functions until the several third image segmentation networks are obtained.
[0119] Among them, for the unlabeled medical image samples, the original image and the transformed image can also be obtained through the image transformation T(·), and then input into the first image segmentation network and the third image segmentation network to obtain and and After the inverse image transformation T -1 (·), the output of the transformed data should be consistent with the original output To ensure the transformation consistency regularization, the mean square error (MSE) loss between these two outputs is minimized:
[0120]
[0121] In summary, the overall loss function of the first image segmentation network of the present application is shown in the following formula:
[0122]
[0123] The overall loss function of the third image segmentation network of the present application is shown in the following formula:
[0124]
[0125] Among them, represents the overall loss function of the i-th image segmentation network, β and λ respectively represent preset weights, represents the review loss function, represents the supervised segmentation loss, including the Dice loss and the CE loss, that is represents the mean square error loss, represents the contrast loss function.
[0126] The semi-supervised medical image segmentation method based on multiple networks of the present application trains multiple deep learning networks according to multiple initial medical image samples. After obtaining multiple image segmentation networks, according to the image segmentation loss functions of each image segmentation network, a number of first image segmentation networks and a number of second image segmentation networks are obtained. Then, multiple first unlabeled medical image samples are respectively input into the number of first image segmentation networks to obtain pseudo-labeled medical image samples output by the number of first image segmentation networks. Then, according to the pseudo-labeled medical image samples, the number of second image segmentation networks are trained to obtain a number of third image segmentation networks, so as to obtain a number of first image segmentation networks and a number of third image segmentation networks with good network performance, and the medical image sample to be segmented is segmented by the first image segmentation network and the third image segmentation network to obtain the target segmentation image. The number of training samples can be increased by the image segmentation network with a lower image segmentation loss function and better network performance for training the image segmentation network with a higher image segmentation loss function and worse network performance. The number of training samples of the image segmentation network with a higher image segmentation loss function and worse network performance can be increased efficiently and at low cost, so as to obtain a third image segmentation network with a lower image segmentation loss function and better network performance. Then, combined with the first image segmentation network with a lower image segmentation loss function and better network performance, the medical image sample to be segmented is segmented to obtain a target segmentation image with high segmentation accuracy.
[0127] The device embodiments described above are merely illustrative. The components described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present application. Those of ordinary skill in the art can understand and implement it without creative efforts.
[0128] Persons skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0129] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to produce a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices produce means for implementing the selected functions in the process Figure 1 in one process or multiple processes and / or blocks Figure 1 or means for implementing the selected functions in one block or multiple blocks. These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means for implementing the selected functions in the process Figure 1 in one process or multiple processes and / or blocks Figure 1 or means for implementing the selected functions in one block or multiple blocks.
[0130] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the selected functions in the process Figure 1 in one process or multiple processes and / or blocks Figure 1 or means for implementing the selected functions in one block or multiple blocks.
[0131] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and a memory.
[0132] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM), and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of computer-readable media.
[0133] A computer-readable medium includes both permanent and non-permanent, removable and non-removable media that can implement information storage by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic tape disk storage or other magnetic storage devices, or any other non-transitory medium that can be used to store information that can be accessed by a computing device. As defined herein, a computer-readable medium does not include transitory computer-readable media such as modulated data signals and carrier waves.
[0134] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or apparatus comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or apparatus. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising the element.
[0135] The above are only embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.
Claims
1. A semi-supervised medical image segmentation method based on multiple networks, characterized in that, Including: Training multiple deep learning networks based on multiple initial medical image samples to obtain multiple image segmentation networks; Obtaining a number of first image segmentation networks and a number of second image segmentation networks according to the image segmentation loss functions of the respective image segmentation networks; the image segmentation loss function of the first image segmentation network is less than that of the second image segmentation network; Inputting multiple first unlabeled medical image samples into the number of first image segmentation networks respectively to obtain pseudo-labeled medical image samples output by the number of first image segmentation networks; the pseudo-labeled medical image samples include medical images corresponding to the predicted segmentation objects of the first image segmentation network; Training the number of second image segmentation networks according to the pseudo-labeled medical image samples to obtain a number of third image segmentation networks; Performing segmentation processing on the medical image sample to be segmented through the first image segmentation network and the third image segmentation network to obtain a target segmentation image.
2. The semi-supervised medical image segmentation method based on multiple networks according to claim 1, characterized in that The step of obtaining a number of first image segmentation networks and a number of second image segmentation networks according to the image segmentation loss functions of the respective image segmentation networks includes: Obtaining the image segmentation loss functions of the respective image segmentation networks; Arranging all the image segmentation networks from low to high according to the image segmentation loss function, determining the number of image segmentation networks arranged in the front as the first image segmentation networks, and the remaining image segmentation networks as the second image segmentation networks.
3. The semi-supervised medical image segmentation method based on multiple networks according to claim 2, characterized in that The step of obtaining the image segmentation loss functions of the respective image segmentation networks includes: Inputting multiple first labeled medical image samples into the respective image segmentation networks to obtain predicted segmentation images output by the respective image segmentation networks; Obtaining the image segmentation loss function of the image segmentation network according to the multiple predicted segmentation images of the same image segmentation network and the segmentation objects labeled by the multiple first labeled medical image samples.
4. The semi-supervised medical image segmentation method based on multiple networks according to claim 3, characterized in that, The step of obtaining the image segmentation loss function of the image segmentation network according to the multiple predicted segmentation images of the same image segmentation network and the segmentation objects labeled by the multiple first labeled medical image samples includes: Obtaining the image segmentation loss function through the following formula: Among them, Dice i represents the image segmentation loss function of the i-th image segmentation network, represents the predicted segmented image, T = y L represents the segmentation object, and ∈ represents the smoothing term.
5. The semi-supervised medical image segmentation method based on multiple networks according to claim 1, characterized in that, The step of inputting multiple first unlabeled medical image samples into the number of first image segmentation networks respectively to obtain pseudo-labeled medical image samples output by the number of first image segmentation networks includes: Inputting the multiple first unlabeled medical image samples into the number of first image segmentation networks to obtain multiple first predicted medical image samples; Obtaining the second voxel prediction values of the second predicted medical image samples corresponding to the respective first unlabeled medical image samples according to the first voxel prediction values of the multiple first predicted medical image samples; Performing normalization processing on the second voxel prediction values of each voxel of the second predicted medical image samples to obtain the pseudo-labeled medical image samples and the corresponding labeled voxel prediction values.
6. The semi-supervised medical image segmentation method based on multiple networks according to claim 5, wherein The step of obtaining the second voxel prediction value of each of the first unlabeled medical image samples according to the first voxel prediction value of the multiple first predictive medical image samples includes: Obtaining the second voxel prediction value through the following formula: Among them, indicates that the voxel point with coordinates (x, y, z) is the second voxel prediction value of class c. indicates the first voxel prediction value of the voxel point with coordinates (x, y, z) of the first predicted medical image sample. [c, x, y, z] indicates the first voxel prediction value of the voxel point with coordinates (x, y, z) of the second predicted medical image sample. indicates the first voxel prediction value of the voxel point with coordinates (x, y, z) of the I-th predicted medical image sample, where I is the last predicted medical image sample.
7. The semi-supervised medical image segmentation method based on multiple networks according to claim 5, characterized in that, The step of normalizing the second voxel prediction value of each voxel of the second predictive medical image sample to obtain the pseudo-label medical image sample and the corresponding labeled voxel prediction value includes: Obtaining the labeled voxel prediction value through the following formula: Among them, represents the predicted value of the label voxel of the voxel point with coordinates (x, y, z), and τ represents a preset value. represents the predicted value that the voxel point with coordinates (x, y, z) of the second predicted medical image sample is of class k, and C represents the total number of voxel point classes. represents the second voxel prediction value that the voxel point with coordinates (x, y, z) is of class c.
8. The semi-supervised medical image segmentation method based on multiple networks according to claim 1, characterized in that The step of performing segmentation processing on the medical image sample to be segmented through the first image segmentation network and the third image segmentation network to obtain a target segmentation image includes: Inputting the medical image sample to be segmented into each of the first image segmentation network and the third image segmentation network respectively to obtain a plurality of candidate segmentation images; Generating the target segmentation image according to the voxel mean value of the same voxel corresponding to the candidate segmentation images.
9. The semi-supervised medical image segmentation method based on multiple networks according to claim 1, wherein The plurality of initial medical image samples include a plurality of second labeled medical image samples; The step of training a plurality of deep learning networks according to a plurality of initial medical image samples to obtain a plurality of image segmentation networks includes: Inputting the plurality of second labeled medical image samples into the plurality of deep learning networks respectively to obtain a plurality of first model output images; Obtaining the confusion region of the second labeled medical image sample according to the differences between the plurality of first model output images corresponding to the same second labeled medical image sample; Constructing a review loss function for each of the deep learning networks according to the plurality of second labeled medical image samples, the plurality of first model output images, and the confusion region; Training the plurality of deep learning networks according to the review loss function until the plurality of image segmentation networks are obtained.
10. The semi-supervised medical image segmentation method based on multiple networks according to claim 1, wherein The plurality of initial medical image samples include a plurality of second unlabeled medical image samples; The step of training a plurality of deep learning networks according to a plurality of initial medical image samples to obtain a plurality of image segmentation networks includes: Performing image transformation on the plurality of second unlabeled medical image samples to obtain a plurality of third unlabeled image samples; Inputting the plurality of second unlabeled medical image samples and the plurality of third unlabeled image samples into the plurality of deep learning networks respectively to obtain a plurality of second model output images and a plurality of third model output images; Constructing a consistency constraint function for each of the deep learning networks according to the plurality of second model output images and the plurality of third model output images; Training the plurality of deep learning networks according to the consistency constraint function until the plurality of image segmentation networks are obtained.