Multi-organ knowledge migration modeling and semi-supervised data iteration labeling method

Abstract

The invention discloses a multi-organ knowledge migration modeling and semi-supervised data iteration labeling method, and relates to the technical field of image processing, and the method comprises the steps: carrying out the screening of a data set, and obtaining a plurality of target plain-scan CT images; wherein the target plain-scan CT image comprises a plurality of target organs and labels corresponding to the target organs; training a knowledge migration model by using each target plain-scan CT image; using the trained knowledge migration model to determine the performance score of each target plain-scan CT image; and performing multi-round label labeling on the to-be-labeled plain-scan CT image according to the performance score to obtain a target labeling data set. According to the method, the plain-scan CT image is labeled through modeling and training of the knowledge migration model, migration knowledge in open source data can be fully mined, and labeling of the target data set is efficiently completed in an iterative optimization mode.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method. Background Art

[0002] High-accuracy training of medical image segmentation models critically relies on large, high-quality labeled datasets. Traditionally, the annotation process relies on manual, slice-by-slice annotation of 3D medical images. This is not only time-consuming and labor-intensive, but also subject to subjectivity, leading to high-quality fluctuations in annotation quality, severely hindering the widespread adoption of automated segmentation technology in clinical practice. Furthermore, different medical tasks require varying levels of annotated organ types and accuracy. This is particularly true for high-risk procedures such as puncture planning, where accurate identification of key organs and their spatial locations is crucial. However, large-scale, finely annotated datasets specifically targeting puncture-related areas are currently scarce, hindering efficient model training. To alleviate this issue, researchers have recently proposed using knowledge transfer and semi-supervised learning techniques to improve model performance on low-label data. Knowledge transfer allows for the transfer of general medical image representations learned from large, open-source datasets to the target task, thereby shortening model training time and improving segmentation accuracy. Semi-supervised learning, on the other hand, leverages the underlying structural information in unlabeled data to iteratively optimize model performance based on limited labeled samples, further reducing the annotation burden on physicians. Although some studies have attempted to combine transfer learning with semi-supervised learning for medical image segmentation, they still face many challenges when dealing with tasks involving multiple organs and different structural complexities, such as inconsistent label quality of open source datasets, mismatch of target organ types, limited generalization ability of transfer models, and large differences in the importance of labeled organs. Summary of the Invention

[0003] The main purpose of the embodiments of the present application is to propose a multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method to improve the labeling efficiency of medical images.

[0004] To achieve the above objectives, one aspect of an embodiment of the present application proposes a multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method, the method comprising the following steps:

[0005] A plurality of target plain scan CT images are obtained by screening from a data set; wherein the target plain scan CT images include a plurality of target organs and labels corresponding to the target organs;

[0006] Using each of the target plain scan CT images to train a knowledge transfer model;

[0007] Determining a performance score for each of the target plain scan CT images using the trained knowledge transfer model;

[0008] Multiple rounds of labeling are performed on the unenhanced CT images to be annotated according to the performance scores to obtain a target annotated dataset.

[0009] In some embodiments, the step of screening and obtaining a plurality of target plain scan CT images from a data set comprises the following steps:

[0010] Selecting labels corresponding to various target organs as target labels;

[0011] Selecting candidate plain CT images of the chest, abdomen, or chest and abdomen from the data set;

[0012] A plain scan CT image containing the target label and having connectivity components of various target organs meeting preset conditions is selected from the candidate plain scan CT images as the target plain scan CT image.

[0013] In some embodiments, the training of the knowledge transfer model using each of the target plain scan CT images comprises the following steps:

[0014] Cropping, resampling, and then normalizing each of the target plain scan CT images to obtain a preprocessed image;

[0015] Randomly rotating, randomly scaling, or mirroring the preprocessed image and the corresponding mask to obtain enhanced data;

[0016] Downsampling the enhanced data and then using the downsampling result to train a first-level three-dimensional U-Net to obtain a first-level output result;

[0017] Upsampling the first-level output result to a segmentation result consistent with the original voxel spacing and size, and then using the segmentation result and the enhanced data to train a second-level 3D U-Net; wherein the first-level 3D U-Net and the second-level 3D U-Net serve as the knowledge transfer model, and the resolution of the first-level 3D U-Net is lower than the resolution of the second-level 3D U-Net;

[0018] Calculating a scaling gradient according to the optimization function, and then using the scaling gradient to update parameters of the knowledge transfer model when training the knowledge transfer model;

[0019] Adjusting parameters of the knowledge transfer model inference process based on patch overlap and voting mechanism;

[0020] After the inference is completed, the predicted labels of the knowledge transfer model are analyzed. If the connected components of various target organs are greater than 1, the largest connected component is retained as the label corresponding to the target organ.

[0021] In some embodiments, the cropping, resampling, and then normalizing each of the target plain scan CT images to obtain a preprocessed image comprises the following steps:

[0022] cropping each of the target plain scan CT images to a non-zero value region to obtain a cropped image;

[0023] Counting the median of all voxel spacings in the cropped image;

[0024] Solving the matrix equation using the first endpoint condition and the data nodes to obtain a quadratic differential value, and then resampling the voxel spacing of all the cropped images to the median using a third-order spline interpolation;

[0025] The expression using third-order spline interpolation is:

[0026] f i (x) = a i +b i (xx i )+c i (xx i ) 2 ;

[0027]

[0028] Among them, (x i ,y i ) represents the pixels in the cropped image, h i represents the step length, h i =x i+1 -x i , m i represents the second differential value;

[0029] Extracting pixel intensity values ​​of all areas covered by the segmentation mask in each of the target plain scan CT images and constructing a statistical distribution;

[0030] Based on the statistical distribution of the pixel intensity values, determining the 0.5% and 99.5% quantiles as truncation thresholds, removing 1% of extreme outliers at both ends of the distribution, and retaining pixel intensity values ​​between the 0.5% and 99.5% quantiles as target pixel intensity values;

[0031] Calculating the average and standard deviation of the intensity values ​​according to the target pixel intensity values, and then performing normalization processing;

[0032] The expression of the normalization process is:

[0033]

[0034] Among them, x nrepresents the result of the normalization processing, x represents the image obtained after resampling, μ represents the average value, and δ represents the average value.

[0035] In some embodiments, calculating the scaling gradient according to the optimization function comprises the following steps:

[0036] The loss function is determined as follows:

[0037] L total =L dice +L CE ;

[0038]

[0039] Among them, L total is the loss function, L dice is the loss of DICE similarity coefficient, L CE is the cross entropy loss, K represents the total number of category sets, and each category calculates a DICE similarity coefficient loss and takes the average. is the predicted probability of class k, is the true label of category k, N is the total number of samples;

[0040] The current gradient of the knowledge transfer model is calculated as follows:

[0041]

[0042] Among them, g t is the current gradient, t represents the time step, θ t-1 is the parameter of the previous time step;

[0043] The first moment estimate of the gradient is calculated as follows:

[0044] m t =β1m t-1 +(1-β1)g t ;

[0045] Among them, m t The first-order moment estimate for the t time step, β1 represents the hyperparameter of the historical first-order moment attenuation factor, m t-1 represents the first-order moment estimate at the previous time step;

[0046] The second moment estimate of the gradient is calculated as follows:

[0047]

[0048] Among them, v t is the second-order moment estimate at time step t, β2 represents the hyperparameter of the historical first-order moment decay factor, and v t-1Represents the second-order moment estimate of the previous time step;

[0049] The scaling gradient is calculated as follows:

[0050]

[0051] Among them, g′ t is the scaling gradient, l is the initial learning rate, and ε is a decimal greater than 0.

[0052] In some embodiments, adjusting parameters of the knowledge transfer model inference process based on patch overlap and voting mechanism includes the following steps:

[0053] We perform data augmentation at test time by overlapping selected patches by choosing a 1 / 2 patch size and then mirroring all patches along the three valid axes.

[0054] After data augmentation at test time, 64 predictions are determined for each voxel.

[0055] Voting on each of the prediction results according to the weight of the voxels, wherein the closer to the center of the voxel, the higher the weight;

[0056] Complete aggregate reasoning based on voting results.

[0057] In some embodiments, performing multiple rounds of labeling on the unenhanced CT images to be annotated according to the performance scores to obtain a target annotated dataset includes the following steps:

[0058] The organ knowledge with a performance score higher than 0.95 is regarded as the first-level transfer knowledge, and the organ knowledge with a performance score lower than or equal to 0.95 is regarded as other transfer knowledge;

[0059] Using the trained knowledge transfer model to perform performance scores on the first-level transfer knowledge and the other transfer knowledge, and then taking the organ knowledge with a performance score higher than 0.90 but lower than 0.95 as the second-level transfer knowledge;

[0060] Using the trained knowledge transfer model to perform performance scoring on the first-level transfer knowledge and the second-level transfer knowledge, and then taking the organ knowledge with a performance score higher than 0.90 as the third-level transfer knowledge;

[0061] The target annotation dataset is obtained by fusing the first-level transfer knowledge, the second-level transfer knowledge, and the third-level transfer knowledge.

[0062] To achieve the above objectives, another aspect of the present application provides a multi-organ knowledge transfer modeling and semi-supervised data iterative labeling device, the device comprising:

[0063] An image screening unit, configured to screen a plurality of target plain scan CT images from a data set; wherein the target plain scan CT images include a plurality of target organs and labels corresponding to the target organs;

[0064] A model training unit, configured to train a knowledge transfer model using each of the target plain scan CT images;

[0065] a performance scoring unit, configured to determine a performance score for each of the target plain scan CT images using the trained knowledge transfer model;

[0066] The iterative labeling unit is used to perform multiple rounds of labeling on the unenhanced CT images to be labeled according to the performance scores to obtain a target labeling dataset.

[0067] To achieve the above-mentioned purpose, another aspect of an embodiment of the present application provides an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the above-mentioned method when executing the computer program.

[0068] To achieve the above-mentioned purpose, another aspect of an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program implements the above-mentioned method when executed by a processor.

[0069] The embodiments of the present application include at least the following beneficial effects:

[0070] The present application can filter multiple target plain scan CT images from a data set; wherein the target plain scan CT images include multiple target organs and labels corresponding to the various target organs; use each target plain scan CT image to train a knowledge transfer model; use the trained knowledge transfer model to determine the performance score of each target plain scan CT image; perform multiple rounds of labeling on the plain scan CT images to be labeled based on the performance score to obtain a target labeled data set. The present application uses knowledge transfer model modeling and training to label plain scan CT images, which can fully mine the transfer knowledge in open source data and efficiently complete the labeling of the target data set through iterative optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0072] Figure 1 A flowchart of a multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method provided in an embodiment of the present application;

[0073] Figure 2 A flowchart of knowledge extraction and migration of large-scale datasets provided in an embodiment of the present application;

[0074] Figure 3 A training flow chart of the knowledge transfer model provided in an embodiment of the present application;

[0075] Figure 4 Flowchart of the semi-supervised iterative labeling algorithm provided in the embodiment of the present application;

[0076] Figure 5 A schematic diagram of the structure of a multi-organ knowledge transfer modeling and semi-supervised data iterative labeling device provided in an embodiment of the present application;

[0077] Figure 6 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0078] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of the present application. They are merely examples of devices and methods consistent with some aspects of the embodiments of the present application as detailed in the appended claims.

[0079] It will be understood that the terms "first", "second", etc. used in this application may be used herein to describe various concepts, but unless otherwise specified, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the words "if" and "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0080] The terms "at least one", "plurality", "each", "any", etc. used in this application include "at least one", "two" or more, "plurality" or "each", "any" or "any one", "each" or "any one" as used herein.

[0081] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0082] The embodiment of the present application provides a multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method, which relates to the field of image processing technology. The multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method provided in the embodiment of the present application can be applied to a terminal, can also be applied to a server, and can also be software running in a terminal or a server. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, and a car terminal, etc., but is not limited to this; the server side can be configured as an independent physical server, or can be configured as a server cluster or distributed system composed of multiple physical servers, and can also be configured as a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network; the software can be an application that implements a multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method, etc., but is not limited to the above forms.

[0083] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and the like. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments in which tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0084] Reference Figure 1 The embodiment of the present application provides a multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method, which may include but is not limited to S100 to S130, as follows:

[0085] S100: Filtering and obtaining a plurality of target plain scan CT images from a data set; wherein the target plain scan CT images include a plurality of target organs and labels corresponding to the target organs;

[0086] S110: training a knowledge transfer model using each of the target plain scan CT images;

[0087] S120: Determine a performance score for each of the target plain scan CT images using the trained knowledge transfer model;

[0088] S130: Perform multiple rounds of labeling on the unenhanced CT images to be labeled according to the performance scores to obtain a target labeled dataset.

[0089] Optionally, the step of screening a plurality of target plain scan CT images from the data set comprises the following steps:

[0090] Selecting labels corresponding to various target organs as target labels;

[0091] Selecting candidate plain CT images of the chest, abdomen, or chest and abdomen from the data set;

[0092] A plain scan CT image containing the target label and having connectivity components of various target organs meeting preset conditions is selected from the candidate plain scan CT images as the target plain scan CT image.

[0093] Optionally, the training of the knowledge transfer model using each of the target plain scan CT images comprises the following steps:

[0094] Cropping, resampling, and then normalizing each of the target plain scan CT images to obtain a preprocessed image;

[0095] Randomly rotating, randomly scaling, or mirroring the preprocessed image and the corresponding mask to obtain enhanced data;

[0096] Downsampling the enhanced data and then using the downsampling result to train a first-level three-dimensional U-Net to obtain a first-level output result;

[0097] Upsampling the first-level output result to a segmentation result consistent with the original voxel spacing and size, and then using the segmentation result and the enhanced data to train a second-level 3D U-Net; wherein the first-level 3D U-Net and the second-level 3D U-Net serve as the knowledge transfer model, and the resolution of the first-level 3D U-Net is lower than the resolution of the second-level 3D U-Net;

[0098] Calculating a scaling gradient according to the optimization function, and then using the scaling gradient to update parameters of the knowledge transfer model when training the knowledge transfer model;

[0099] Adjusting parameters of the knowledge transfer model inference process based on patch overlap and voting mechanism;

[0100] After the inference is completed, the predicted labels of the knowledge transfer model are analyzed. If the connected components of various target organs are greater than 1, the largest connected component is retained as the label corresponding to the target organ.

[0101] Optionally, the cropping, resampling, and then normalizing each of the target plain scan CT images to obtain a preprocessed image comprises the following steps:

[0102] cropping each of the target plain scan CT images to a non-zero value region to obtain a cropped image;

[0103] Counting the median of all voxel spacings in the cropped image;

[0104] Solving the matrix equation using the first endpoint condition and the data nodes to obtain a quadratic differential value, and then resampling the voxel spacing of all the cropped images to the median using a third-order spline interpolation;

[0105] The expression using third-order spline interpolation is:

[0106] f i (x) = a i +b i (xx i )+c i (xx i ) 2 ;

[0107]

[0108] Among them, (x i ,y i ) represents the pixels in the cropped image, h i represents the step length, h i =x i+1 -x i , m i represents the second differential value;

[0109] Extracting pixel intensity values ​​of all areas covered by the segmentation mask in each of the target plain scan CT images and constructing a statistical distribution;

[0110] Based on the statistical distribution of the pixel intensity values, determining the 0.5% and 99.5% quantiles as truncation thresholds, removing 1% of extreme outliers at both ends of the distribution, and retaining pixel intensity values ​​between the 0.5% and 99.5% quantiles as target pixel intensity values;

[0111] Calculating the average and standard deviation of the intensity values ​​according to the target pixel intensity values, and then performing normalization processing;

[0112] The expression of the normalization process is:

[0113]

[0114] Among them, x n represents the result of the normalization processing, x represents the image obtained after resampling, μ represents the average value, and δ represents the average value.

[0115] Optionally, calculating the scaling gradient according to the optimization function comprises the following steps:

[0116] The loss function is determined as follows:

[0117] L total =L dice +L CE ;

[0118]

[0119] Among them, L total is the loss function, L dice is the loss of DICE similarity coefficient, L CE is the cross entropy loss, K represents the total number of category sets, and each category calculates a DICE similarity coefficient loss and takes the average. is the predicted probability of class k, is the true label of category k, N is the total number of samples;

[0120] The current gradient of the knowledge transfer model is calculated as follows:

[0121]

[0122] Among them, g t is the current gradient, t represents the time step, θ t-1 is the parameter of the previous time step;

[0123] The first moment estimate of the gradient is calculated as follows:

[0124] m t =β1m t-1 +(1-β1)g t ;

[0125] Among them, m t The first-order moment estimate for the t time step, β1 represents the hyperparameter of the historical first-order moment attenuation factor, m t-1 represents the first-order moment estimate at the previous time step;

[0126] The second moment estimate of the gradient is calculated as follows:

[0127]

[0128] Among them, v t is the second-order moment estimate at time step t, β2 represents the hyperparameter of the historical first-order moment decay factor, and v t-1 Represents the second-order moment estimate of the previous time step;

[0129] The scaling gradient is calculated as follows:

[0130]

[0131] Among them, g′ t is the scaling gradient, l is the initial learning rate, and ε is a decimal greater than 0.

[0132] Optionally, adjusting parameters of the knowledge transfer model reasoning process according to the patch overlap and voting mechanism comprises the following steps:

[0133] We perform data augmentation at test time by overlapping selected patches by choosing a 1 / 2 patch size and then mirroring all patches along the three valid axes.

[0134] After data augmentation at test time, 64 predictions are determined for each voxel.

[0135] Voting on each of the prediction results according to the weight of the voxels, wherein the closer to the center of the voxel, the higher the weight;

[0136] Complete aggregate reasoning based on voting results.

[0137] Optionally, performing multiple rounds of labeling on the unenhanced CT images to be annotated according to the performance scores to obtain a target annotated dataset comprises the following steps:

[0138] The organ knowledge with a performance score higher than 0.95 is regarded as the first-level transfer knowledge, and the organ knowledge with a performance score lower than or equal to 0.95 is regarded as other transfer knowledge;

[0139] Using the trained knowledge transfer model to perform performance scores on the first-level transfer knowledge and the other transfer knowledge, and then taking the organ knowledge with a performance score higher than 0.90 but lower than 0.95 as the second-level transfer knowledge;

[0140] Using the trained knowledge transfer model to perform performance scoring on the first-level transfer knowledge and the second-level transfer knowledge, and then taking the organ knowledge with a performance score higher than 0.90 as the third-level transfer knowledge;

[0141] The target annotation dataset is obtained by fusing the first-level transfer knowledge, the second-level transfer knowledge, and the third-level transfer knowledge.

[0142] Next, the solution of the embodiment of the present application will be introduced and explained in detail with reference to specific application examples.

[0143] This embodiment discloses a multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method. Autonomous segmentation of medical images relies on a large number of labeled data sets, and manual data labeling is inefficient and time-consuming. With the help of knowledge transfer of open source data sets and semi-supervised learning methods, the data labeling rate can be improved. The method of this embodiment takes into account issues such as open source data set quality optimization and migration knowledge grading. The method of this embodiment mainly consists of three parts: open source data set knowledge extraction, open source knowledge transfer model and semi-supervised iterative labeling method. By using this embodiment, the labeling efficiency of medical segmentation data sets can be improved.

[0144] Specifically, this embodiment includes the following technical solutions:

[0145] 1. Knowledge extraction from large-scale data sets.

[0146] 1.1. Large-scale dataset knowledge source.

[0147] Selecting a large-scale dataset as a knowledge source allows for rapid initial annotation through knowledge transfer, leveraging knowledge from large-scale open-source datasets, thereby improving physician annotation efficiency. A dataset suitable as a knowledge source should meet the following characteristics: 1. It should have annotations for most organs in chest and abdominal CT scans, thus meeting the requirements for multi-puncture region data; 2. The data size should be large enough to still have sufficient data to support training after screening; 3. It should include chest and abdominal plain scan CT scans.

[0148] Due to the limitations of existing open-source chest and abdomen multi-organ segmentation datasets, this example uses the large-scale dataset TotalSegmentator published by Wasserthal et al. in 2023. This dataset contains rich annotations of human anatomical structures. Although most organs are not currently required for this study, its rich data annotation resources can extract valuable knowledge. For example, the flowchart of large-scale dataset knowledge extraction and migration can be referred to Figure 2 First, the knowledge labels to be extracted need to be confirmed with a professional puncture surgeon. Then, two rounds of optimization screening are performed on the large-scale data. The initial screening is to select the appropriate CT data types, and the second round of screening is to ensure the quality of the labels. Finally, the filtered data is subjected to label mapping to obtain formatted data, and ultimately, a knowledge transfer model is obtained through training.

[0149] 1.2. Double optimization screening.

[0150] After determining the labels, rigorous data set screening is essential. To this end, this study proposed a two-step optimization screening process. The first step prioritizes CT data type, ensuring that the selected data is plain scan and within the puncture range. The second step focuses on the data label type and quality, ensuring that the acquired data labels meet the requirements.

[0151] 1.2.1. First level of screening: CT data type screening.

[0152] By reading the "research data type" label in the metadata table, CT data that does not include the chest and abdomen are eliminated, such as CT data that only contains a single neck, a single pelvis, or a head. Only CT image data containing the chest, abdomen, or chest and abdomen are retained. The specific data are as follows: there are 160 cases of data that do not include the pelvis and neck (151 of which are training data and 9 are test data); there are 230 cases of data that contain the pelvis; there are 79 cases of data that contain the neck (76 are training data and 3 are test data). Since the tip of some lungs may appear in the neck CT, some neck CT is taken into consideration. In addition, in actual applications, puncture surgery requires the use of plain CT, so the enhanced CT in the data is removed.

[0153] 1.2.2. Second level of screening: organ label quality assurance screening.

[0154] After completing the CT type screening, the label quality of the data needs to be carefully checked to ensure the validity of the extracted knowledge.

[0155] The most important factor in determining label quality is the connectivity of key organs. In the field of medical image segmentation, connectivity refers to the spatial interconnectedness of labeled pixels (2D) or voxels (3D) of the same category (organ or tissue). Generally speaking, regions belonging to the same organ should be spatially continuous and connected, and should not be split into multiple independent fragments or small blocks. Neighborhood analysis is used to obtain the connectivity components of each organ, and the connectivity of the organ is determined based on the connectivity components.

[0156] 2. Construction and training of open source knowledge transfer model.

[0157] The knowledge extracted from large-scale data sets is trained with the help of an adaptive framework. The training process of the knowledge transfer model is as follows: Figure 3As shown in the figure, it specifically includes the following five modules: (1) Data preprocessing module: Reduce the amount of model calculation by cropping invalid areas before training; and solve the problem of different voxel spacing through data resampling method; finally, solve the problem of CT intensity difference through data normalization. (2) Cascade network module: Through the architecture of cascade network, solve the problem of organ size difference and improve the segmentation accuracy of knowledge transfer model. (3) Training optimization module: Through adaptive learning rate adjustment, realize dynamic adjustment of model learning rate to ensure the gradient descent trend during model training; through various data enhancement methods, increase the amount of training data while enhancing the generalization ability of the model. (4) Inference optimization module: During the model inference process, patch overlap and mirror reversal are used to enhance the prediction data, and then the voting mechanism is used to enhance the credibility of the model. (5) Post-processing module: According to the connectivity analysis of a certain organ in the training data, the small components of the prediction results with errors are removed to ensure the rationality of the prediction results.

[0158] 2.1. Data preprocessing

[0159] Data preprocessing is a crucial part of the automatic segmentation algorithm and includes image cropping, resampling, and normalization, each of which will be analyzed below. First, all data is clipped to a non-zero region through cropping. The voxel spacing of a dataset can be affected by differences in scanner model and acquisition protocol, but convolutional neural networks cannot understand differences in voxel spacing. To address this issue and enable the knowledge transfer model to correctly learn the spatial semantics of the data, the voxel spacing needs to be resampled. The median voxel spacing of all data in the dataset is obtained through statistical knowledge extraction, and the voxel spacing of all data is then resampled to this median.

[0160] The image data is resampled by using third-order spline interpolation, assuming that the i-th slice is (x i ,yi) indicates that the calculation step length h i =x i+1 -x i . Use the first endpoint condition and data nodes to solve the matrix equation to obtain the quadratic differential value m i , and then the third-order equation can be calculated.

[0161] f i (x) = a i +b i (xx i )+c i (xx i ) 2 (1-1)

[0162]

[0163] Furthermore, the obtained data is normalized. The pixel intensity values ​​I of all the segmentation mask coverage areas in the open source dataset are extracted and the statistical distribution is constructed. Based on the extracted intensity values, the 0.5% and 99.5% quantiles are determined as the truncation thresholds, and the extreme outliers of 1% at both ends of the distribution are removed, and the percentile I of the main data range is retained. 0.5,0.95 , and based on this, the entire data set is normalized. 0.5,0.95 Calculate the mean μ and standard deviation δ of the intensity value, and perform z-score normalization to obtain the normalized result. For data x, the normalized result is as follows:

[0164]

[0165] Through the above methods, the preprocessing of the extracted data, such as cropping, resampling and normalization, is completed, preparing for the training of the knowledge transfer model.

[0166] 2.2. Knowledge transfer network architecture and data enhancement.

[0167] Considering that CT data is three-dimensional volumetric data, when processing slices of data using a two-dimensional U-Net, information along the z-axis cannot be aggregated and extracted. Therefore, a two-dimensional U-Net structure is not used. A simple three-dimensional U-Net result may not account for differences in organ size. Therefore, a cascaded U-Net structure is used to address the problem of large differences in organ labels in the extracted knowledge.

[0168] After preprocessing, the data is first downsampled to a lower resolution. This operation reduces the amount of data, enabling faster network learning. The downsampled data enters the first low-resolution training phase, which uses a low-resolution 3D U-Net structure to help the knowledge transfer model quickly acquire global information. However, because downscaling can result in the loss of some detailed information, the output of the low-resolution phase must be upsampled for the next stage of learning.

[0169] The low-resolution results are upsampled to obtain a low-resolution segmentation result that matches the original voxel spacing and size. The upsampled result is then used as an additional input channel for feature fusion with the original image, and both are then fed into a second-level, high-resolution 3D U-Net for training. At the high-resolution stage, the network can learn and converge quickly because of the initial segmentation results from the low-resolution stage, and the high-resolution network can better extract detailed knowledge from the data.

[0170] At the same time, during training, it is also necessary to enhance the training data to improve the generalization ability of the model. First, the image and mask are subjected to corresponding data enhancement operations such as random rotation, random scaling, and mirroring.

[0171] 2.3. Optimize function.

[0172] The organ segmentation error is calculated using a combination of the Dice Similarity Coefficient (DSC) and cross-entropy loss to train the network. The DICE metric is commonly used to measure the similarity between two sets and is particularly applicable to medical image segmentation scenarios. The multi-category loss function is calculated as follows:

[0173] L total =L dice +L CE (1-4)

[0174]

[0175] Among them, K represents the total number of category sets, and a Dice Loss is calculated for each category and the mean is taken. is the predicted probability of category k, is the true label of category k, and N is the total number of samples.

[0176] The optimizer uses Adam, which uses the first-order moment to simulate the momentum of the gradient, making the optimization direction more stable. The second-order moment is used to adaptively adjust the learning rate to prevent oscillation or divergence caused by excessive learning rate. The following calculation is used to achieve adaptive adjustment of the learning rate, set the initial learning rate l, and calculate the current gradient g of the model. t , the specific calculation is as follows:

[0177]

[0178] Among them, L total (θ) represents the loss function, t represents the time step, θ t-1 are the parameters of the previous time step.

[0179] Then calculate the first-order moment estimate m of the gradient t , using the moving exponential average, the specific calculation is as follows:

[0180] m t =β1m t-1 +(1-β1)g t (1-8)

[0181] Among them, β1 represents the hyperparameter of the historical first-order moment attenuation factor, m t-1 Represents the first moment estimate at the previous time step.

[0182] Then calculate the second-order moment estimate v t , the moving exponential average of the squared gradient is calculated as follows:

[0183]

[0184] Among them, β2 represents the hyperparameter of the historical first-order moment attenuation factor, v t-1 Represents the second moment estimate at the previous time step.

[0185] Finally, the scaled gradient can be calculated ε is a very small number to prevent the denominator from being 0. The scaled gradient is then used to update the parameters.

[0186] 2.4. Reasoning optimization and connectivity analysis.

[0187] The final inference process is optimized using patch overlap and voting mechanisms. Patches are selected by overlapping them with a size of 1 / 2, and data augmentation is performed during testing by mirroring all patches along the three valid axes. After data augmentation, a single voxel can generate 64 predictions, which are then used for voting. Furthermore, considering that the network's segmentation accuracy varies at different locations on a patch, with a particularly declining trend at patch edges, when aggregating predictions from different patches, the weights of voxels near the center are increased to ensure they are higher than those near the edges, before completing the aggregated prediction.

[0188] Finally, the post-processing method uses connectivity analysis. Similar to the knowledge extraction process, most organs in the data (liver, gallbladder, pancreas, left kidney, right kidney, left upper lobe, left lower lobe, right upper lobe, right middle lobe, right lower lobe, heart, aorta) should have only one connected component. Therefore, the identifier of a single connected component is added to the general attributes of these organ labels. After the inference is completed, the predicted results are analyzed. If the connected component of the above organs is greater than 1, only the largest component is retained, and the remaining predicted small components are directly removed.

[0189] 3. Iterative labeling method based on semi-supervised learning.

[0190] Traditionally, doctors use annotation tools to annotate two-dimensional slices of three-dimensional data, one slice at a time. This method is inefficient, and the long labeling process can lead to poor label quality. To address these issues and improve doctors' labeling efficiency, this study leverages a knowledge transfer model trained on large open-source datasets, combined with an iterative labeling algorithm using semi-supervised learning, to assist doctors in labeling.

[0191] 3.1. Classification of transferable knowledge.

[0192] An open-source knowledge transfer model is constructed through training on the dataset extracted through double optimization. Different organ knowledge performs differently on the transfer model, so each piece of knowledge needs to be graded based on its performance on the open-source knowledge transfer model.

[0193] Using the three-level classification annotation method, organ knowledge with a performance score above 0.95 indicates that the transfer model has better learned this knowledge and performs better. Therefore, the transferred knowledge is more reliable and requires less manual modification. Therefore, this type of organ knowledge is defined as first-level transferred knowledge. In the three-stage iterative annotation process, first-level transferred knowledge generates the initial annotations in the first stage and participates in label generation in the second stage.

[0194] For organ knowledge with performance above 0.90 but below 0.95, this indicates poor learning performance and requires further modification. Therefore, this knowledge can only be used to generate initial labels in the first stage. After the initial model is established, this performance is not included in the second stage of labeling. This type of knowledge is defined as secondary transfer knowledge.

[0195] For organ knowledge with a performance lower than 0.90, it indicates that the learning performance of this type of knowledge is poor and requires a lot of manual correction. Therefore, this type of organ knowledge is discarded.

[0196] 3.2. Three-stage iterative labeling.

[0197] After determining the available knowledge, follow Figure 4 The flowchart of the semi-supervised iterative labeling algorithm shown in the figure performs the iterative labeling process of semi-supervised learning.

[0198] In the first phase, initial annotations were generated using the primary and secondary labels from an open-source knowledge transfer model, starting with 25% of the data. Professional doctors reviewed and modified this data, adding labels for the remaining unlabeled organs. For example, the knowledge transfer model generated initial labels for the lung lobes, heart, aorta, and bony organs. Doctors were responsible for revising these initial labels and labeling the pulmonary vessels and trachea. Once these revisions were complete, the initial model for the first phase was trained on the 25% data.

[0199] In the second phase, based on this initial model, initial annotations are generated for an additional 35% of the data, including the primary organ labels corresponding to the open source model and labels for other organs. Because the primary organ labels from the open source model performed well, they can still be used in the second phase. The primary organ labels obtained by the initial model are merged with those obtained by the open source knowledge transfer model to obtain auxiliary primary organ annotations for this data. The remaining organ labels obtained by the initial model are directly used as auxiliary organ annotations for the remaining organs in this data. Doctors make simple modifications to the primary organs in this 35% of data, while more extensive modifications are required for the remaining organs. The resulting data is merged with the initial 25% of data, resulting in a total of 60% of data in the second phase.

[0200] In the third stage, an advanced model is trained based on the 60% of the data obtained in the first two stages. Due to the increased amount of training data, the advanced model performs better in organ segmentation. Therefore, in the third stage, there is no need to use the knowledge transfer model. Instead, the advanced model is directly used to complete the auxiliary labeling of the remaining 40% of the data. Professional doctors can directly check, confirm, and slightly modify the newly generated labels. This process uses knowledge transfer from open source datasets and a semi-supervised learning iterative labeling algorithm to generate auxiliary labels. Compared to direct labeling by doctors, auxiliary labels can greatly improve data labeling efficiency.

[0201] Beneficial effects:

[0202] This example proposes a knowledge transfer modeling and semi-supervised iterative labeling method for multi-organ segmentation tasks. This method leverages knowledge extraction and structural optimization from large-scale open-source datasets to construct a transfer model. This transfer knowledge is then used to perform three-level classification and three-stage iterative learning within the target data. This significantly improves the efficiency and accuracy of medical image dataset construction, providing powerful data support for high-precision medical scenarios such as puncture surgery planning.

[0203] Reference Figure 5 The embodiment of the present application further provides a multi-organ knowledge transfer modeling and semi-supervised data iterative labeling device, which can implement the above-mentioned multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method, and the device includes:

[0204] An image screening unit, configured to screen a plurality of target plain scan CT images from a data set; wherein the target plain scan CT images include a plurality of target organs and labels corresponding to the target organs;

[0205] A model training unit, configured to train a knowledge transfer model using each of the target plain scan CT images;

[0206] a performance scoring unit, configured to determine a performance score for each of the target plain scan CT images using the trained knowledge transfer model;

[0207] The iterative labeling unit is used to perform multiple rounds of labeling on the unenhanced CT images to be labeled according to the performance scores to obtain a target labeling dataset.

[0208] It can be understood that the contents of the above method embodiments are all applicable to the present device embodiments, the functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0209] The present application also provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method of the present application. The electronic device can be any smart terminal, such as a tablet computer or an in-vehicle computer.

[0210] It can be understood that the contents of the above method embodiments are all applicable to the embodiments of the present device, the functions specifically implemented by the embodiments of the present device are the same as those of the method of the present application, and the beneficial effects achieved are also the same as those achieved by the method of the present application.

[0211] See also Figure 6 , Figure 6 The hardware structure of an electronic device according to another embodiment is shown. The electronic device includes:

[0212] The processor 601 may be implemented as a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is configured to execute relevant programs to implement the technical solutions provided in the embodiments of the present application.

[0213] The memory 602 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 602 can store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 602 and is called by the processor 601 to execute the methods of the embodiments of this application.

[0214] Input / output interface 603, used to implement information input and output;

[0215] Communication interface 604, used to implement communication interaction between this device and other devices, which can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WiFi, Bluetooth, etc.);

[0216] Bus 605 , which transmits information between various components of the device (e.g., processor 601 , memory 602 , input / output interface 603 , and communication interface 604 );

[0217] The processor 601 , the memory 602 , the input / output interface 603 and the communication interface 604 are connected to each other in communication within the device via a bus 605 .

[0218] An embodiment of the present application further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the method of the present application is implemented.

[0219] It can be understood that the contents of the above method embodiments are all applicable to the present storage medium embodiment, the functions specifically implemented by the present storage medium embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0220] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0221] The embodiments described in the embodiments of this application are intended to more clearly illustrate the technical solutions of the embodiments of this application and do not constitute a limitation on the technical solutions provided by the embodiments of this application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0222] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0223] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, i.e., they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0224] Those skilled in the art will appreciate that all or some of the steps in the methods, systems, and functional modules / units in the devices disclosed above may be implemented as software, firmware, hardware, or appropriate combinations thereof.

[0225] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0226] It should be understood that in this application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0227] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the above-mentioned units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0228] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0229] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0230] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes multiple instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present application. The aforementioned storage medium includes: various media that can store programs, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0231] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and essence of the present invention should be within the scope of the present invention.

Claims

1. A multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method, characterized by: The method comprises the following steps: A plurality of target plain scan CT images are obtained by screening from a data set; wherein the target plain scan CT images include a plurality of target organs and labels corresponding to the target organs; Using each of the target plain scan CT images to train a knowledge transfer model; Determining a performance score for each of the target plain scan CT images using the trained knowledge transfer model; Multiple rounds of labeling are performed on the unenhanced CT images to be annotated according to the performance scores to obtain a target annotated dataset.

2. A multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method according to claim 1, characterized in that: The method of screening and obtaining a plurality of target plain scan CT images from the data set comprises the following steps: Selecting labels corresponding to various target organs as target labels; Selecting candidate plain CT images of the chest, abdomen, or chest and abdomen from the data set; A plain scan CT image containing the target label and having connectivity components of various target organs meeting preset conditions is selected from the candidate plain scan CT images as the target plain scan CT image.

3. A multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method according to claim 1, characterized in that: The method of training the knowledge transfer model using each of the target plain scan CT images comprises the following steps: Cropping, resampling, and then normalizing each of the target plain scan CT images to obtain a preprocessed image; Randomly rotating, randomly scaling, or mirroring the preprocessed image and the corresponding mask to obtain enhanced data; Downsampling the enhanced data and then using the downsampling result to train a first-level three-dimensional U-Net to obtain a first-level output result; Upsampling the first-level output result to a segmentation result consistent with the original voxel spacing and size, and then using the segmentation result and the enhanced data to train a second-level 3D U-Net; wherein the first-level 3D U-Net and the second-level 3D U-Net serve as the knowledge transfer model, and the resolution of the first-level 3D U-Net is lower than the resolution of the second-level 3D U-Net; Calculating a scaling gradient according to the optimization function, and then using the scaling gradient to update parameters of the knowledge transfer model when training the knowledge transfer model; Adjusting parameters of the knowledge transfer model inference process based on patch overlap and voting mechanism; After the inference is completed, the predicted labels of the knowledge transfer model are analyzed. If the connected components of various target organs are greater than 1, the largest connected component is retained as the label corresponding to the target organ.

4. A multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method according to claim 3, characterized in that: The step of cropping, resampling, and then normalizing each of the target plain scan CT images to obtain a preprocessed image comprises the following steps: cropping each of the target plain scan CT images to a non-zero value region to obtain a cropped image; Counting the median of all voxel spacings in the cropped image; Solving the matrix equation using the first endpoint condition and the data nodes to obtain a quadratic differential value, and then resampling the voxel spacing of all the cropped images to the median using a third-order spline interpolation; The expression using third-order spline interpolation is: f i (x)=a i +b i (x-x i )+c i (x-x i ) 2 ; Among them, (x i ,y i ) represents the pixels in the cropped image, h i represents the step length, h i =x i+1 -x i , m i represents the second differential value; Extracting pixel intensity values ​​of all areas covered by the segmentation mask in each of the target plain scan CT images and constructing a statistical distribution; Based on the statistical distribution of the pixel intensity values, determining the 0.5% and 99.5% quantiles as truncation thresholds, removing 1% of extreme outliers at both ends of the distribution, and retaining pixel intensity values ​​between the 0.5% and 99.5% quantiles as target pixel intensity values; Calculating the average and standard deviation of the intensity values ​​according to the target pixel intensity values, and then performing normalization processing; The expression of the normalization process is: Among them, x n represents the result of the normalization processing, x represents the image obtained after resampling, μ represents the average value, and δ represents the average value.

5. A multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method according to claim 3, characterized in that: The step of calculating the scaling gradient according to the optimization function comprises the following steps: The loss function is determined as follows: L total =L dice +L CE ; Among them, L total is the loss function, L dice is the loss of DICE similarity coefficient, L CE is the cross entropy loss, K represents the total number of category sets, and each category calculates a DICE similarity coefficient loss and takes the average. is the predicted probability of class k, is the true label of category k, N is the total number of samples; The current gradient of the knowledge transfer model is calculated as follows: Among them, g t is the current gradient, t represents the time step, θ t-1 is the parameter of the previous time step; The first moment estimate of the gradient is calculated as follows: m t =β1m t-1 +(1-β1)g t ; Among them, m t The first-order moment estimate for the t time step, β1 represents the hyperparameter of the historical first-order moment attenuation factor, m t-1 represents the first-order moment estimate at the previous time step; The second moment estimate of the gradient is calculated as follows: Among them, v t is the second-order moment estimate at time step t, β2 represents the hyperparameter of the historical first-order moment decay factor, and v t-1 Represents the second-order moment estimate of the previous time step; The scaling gradient is calculated as follows: Among them, g′ t is the scaling gradient, l is the initial learning rate, and ε is a decimal greater than 0.

6. A multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method according to claim 3, characterized in that: The step of adjusting the parameters of the knowledge transfer model reasoning process according to the patch overlap and voting mechanism comprises the following steps: We perform data augmentation at test time by overlapping selected patches by choosing a 1 / 2 patch size and then mirroring all patches along the three valid axes. After data augmentation at test time, 64 predictions are determined for each voxel. Voting on each of the prediction results according to the weight of the voxels, wherein the closer to the center of the voxel, the higher the weight; Complete aggregate reasoning based on voting results.

7. A multi-organ knowledge transfer modeling and semi-supervised data iterative labeling method according to any one of claims 1 to 6, characterized in that: The step of performing multiple rounds of labeling on the unenhanced CT images to be annotated according to the performance scores to obtain a target annotated dataset comprises the following steps: The organ knowledge with a performance score higher than 0.95 is regarded as the first-level transfer knowledge, and the organ knowledge with a performance score lower than or equal to 0.95 is regarded as other transfer knowledge; Using the trained knowledge transfer model to perform performance scores on the first-level transfer knowledge and the other transfer knowledge, and then taking the organ knowledge with a performance score higher than 0.90 but lower than 0.95 as the second-level transfer knowledge; Using the trained knowledge transfer model to perform performance scoring on the first-level transfer knowledge and the second-level transfer knowledge, and then taking the organ knowledge with a performance score higher than 0.90 as the third-level transfer knowledge; The target annotation dataset is obtained by fusing the first-level transfer knowledge, the second-level transfer knowledge, and the third-level transfer knowledge.

8. A multi-organ knowledge transfer modeling and semi-supervised data iterative labeling device, characterized by: The device comprises: An image screening unit, configured to screen a plurality of target plain scan CT images from a data set; wherein the target plain scan CT images include a plurality of target organs and labels corresponding to the target organs; A model training unit, configured to train a knowledge transfer model using each of the target plain scan CT images; a performance scoring unit, configured to determine a performance score for each of the target plain scan CT images using the trained knowledge transfer model; The iterative labeling unit is used to perform multiple rounds of labeling on the unenhanced CT images to be labeled according to the performance scores to obtain a target labeling dataset.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images