Multi-organ segmentation method and system based on multi-source integrated distillation
Through the multi-source integrated distillation method, a highly efficient multi-organ segmentation model is trained using pre-trained single-organ segmentation model and label-free data sets, solving the problems of lack of multi-organ labeling data and data privacy protection, and achieving efficient multi-organ segmentation effect.
Patent Information
- Application Number
- CN202210268374.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-18
- Publication Date
- 2025-03-28
- Estimated Expiration
- 2042-03-18
AI Technical Summary
In the prior art, medical image multi-organ segmentation models are lacking in exploration, and due to the lack of multi-organ labeling data and data privacy protection, it is very difficult to train an efficient multi-organ segmentation model.
Using a multi-organ segmentation method based on multi-source integrated distillation, using multiple pre-trained single-organ segmentation models and label-free target domain datasets, an efficient multi-organ segmentation model is obtained through the prediction and transformation of teacher and student models.
This method reduces the need for labeling training data, avoids the use of original training data, improves the accuracy of multi-organ segmentation, and saves calculation time consumption.
Smart Images
Figure CN114612451B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical fields of computer vision, image processing and computer-aided medical diagnosis, and in particular to a multi-organ segmentation method and system based on multi-source integrated distillation. Background Art
[0002] Organ segmentation is one of the basic tasks of computer-aided medical diagnosis and plays a vital role in disease discovery and subsequent treatment. The purpose of medical image segmentation is to extract anatomical regions of interest from medical images in an automatic or semi-automatic manner. Depending on the clinical application, the specific regions of interest can range from tumors to bones to blood vessels. In recent years, the booming development of deep learning has also brought new ideas and breakthroughs to medical image processing. Although deep learning methods have been widely used in this task, their success depends largely on high-quality annotations, which are expensive and difficult to obtain in practice. In addition, doctors usually only annotate a certain organ that they are interested in and are good at. Therefore, the main existing research focuses on single-organ segmentation models, and there is a lack of exploration of multi-organ segmentation models. Multi-organ segmentation models can reduce computational and storage overheads and can also utilize prior knowledge between organs. In order to create a multi-organ dataset, multiple organs in a single image need to be annotated, which requires the joint cooperation of experts in multiple different fields, greatly increasing the difficulty and cost of annotation, resulting in a lack of multi-organ annotated datasets and limiting the implementation of multi-organ segmentation models.
[0003] The success of deep learning-based semantic segmentation of medical images also depends on the availability of medical image data. However, institutions are reluctant to share imaging data for various reasons, the most important of which is privacy protection. Strict health information privacy regulations require the removal of protected health information before sharing data outside the home institution, a process that is both expensive and time-consuming in the context of medical image data. Institutions that intentionally or unintentionally release health information may suffer serious consequences, including huge fines and criminal prosecution. In addition, in many cases, when imaging data are publicly accessible (e.g., cancer imaging data), the corresponding labels are rarely available due to the lack of expert annotations of the target area and the lack of information on the equipment and standards for annotation. In practice, there are many other reasons why such data are unavailable, including data storage and transmission, limited number of cases, and difficulties in data cleaning and annotation. Therefore, it is very difficult to train an efficient medical model in this case. The above-mentioned issues of missing multi-organ annotations and data privacy have put forward new requirements for modeling methods.
[0004] In view of the defects in the prior art, a new technical solution needs to be proposed. Summary of the invention
[0005] In view of the defects in the prior art, the object of the present invention is to provide a multi-organ segmentation method and system based on multi-source integrated distillation.
[0006] According to a multi-organ segmentation method based on multi-source integrated distillation provided by the present invention, the method comprises the following steps:
[0007] Step S1: The input image is predicted by the teacher and student models to obtain multiple sets of segmentation maps;
[0008] Step S2: convert the teacher image segmentation map to output and expand it to multi-category output;
[0009] Step S3: converting the output into organ and background regions according to the region-based mask;
[0010] Step S4: supervised learning is performed on different organ regions and background regions respectively, and the transformed prediction results of the teacher and student models are aligned to obtain a student multi-organ segmentation model;
[0011] Step S5: Input the predicted image and pass it through the student multi-organ segmentation model to obtain the organ segmentation prediction result.
[0012] Preferably, step S1 uses a teacher single-organ neural network segmentation model composed of a convolutional layer and a fully connected layer to predict the input image, and the formula is as follows:
[0013]
[0014] Where x is the input image, the superscript j indicates the jth dimension, and f sm is the mth teacher neural network model, δ k represents the softmax prediction output in the kth dimension, is the predicted probability of the kth dimension obtained by the mth teacher neural network model;
[0015] The student multi-organ neural network segmentation model composed of convolutional layers and fully connected layers is used to predict the input image. The formula is as follows:
[0016]
[0017] Among them, f t is the student neural network model, is the predicted probability of the kth dimension obtained by the student neural network model.
[0018] Preferably, step S2 comprises:
[0019] For organ regions, keep the background predictions of the teacher segmentation model The prediction of the organ numbered m remains unchanged. Shift to the mth dimension and then set the values of all other dimensions to 0. The formula is as follows:
[0020]
[0021] in, is the output of the teacher model after the output conversion mapping to the organ numbered m;
[0022] For the background area, the prediction values of all M teacher single-organ segmentation models are averaged, and the formula is as follows:
[0023]
[0024] in, It is the output of the teacher model to the background after the output conversion mapping.
[0025] Preferably, step S3 comprises:
[0026] Get the predicted mask for organ m, the formula is as follows:
[0027]
[0028] The prediction of the organ region by the converted teacher single organ segmentation model is as follows:
[0029]
[0030] The prediction of the organ region by converting the student multi-organ segmentation model is as follows:
[0031]
[0032] Get the mth teacher single organ segmentation model For the background mask, the formula is as follows:
[0033]
[0034] Aggregate all M teacher single organ segmentation models for the background mask, the formula is as follows:
[0035]
[0036] The conversion teacher single organ segmentation model predicts the background area, and the formula is as follows:
[0037]
[0038] The prediction of the background area for the student multi-organ segmentation model is converted as follows:
[0039]
[0040] Preferably, the step S4 sets different weights for different organ regions and background regions, aligning the transformed prediction results of the teacher and student models, and the formula is as follows:
[0041]
[0042] Among them, λ is a hyperparameter used to balance the weights of organ regions and background regions.
[0043] Preferably, the prediction image is input in step S5 and predicted by the student multi-organ segmentation model, and the formula is as follows:
[0044] p=δ(f t (x)) (13)
[0045] Among them, p is the obtained multi-organ segmentation result.
[0046] The present invention also provides a multi-organ segmentation system based on multi-source integrated distillation, the system comprising the following modules:
[0047] Module M1: The input image is predicted by the teacher and student models to obtain multiple sets of segmentation maps;
[0048] Module M2: Output conversion of the teacher image segmentation map and expansion to multi-category output;
[0049] Module M3: Convert the output to organ and background regions based on region-based masks;
[0050] Module M4: supervised learning is performed on different organ regions and background regions respectively, and the transformed prediction results of the teacher and student models are aligned to obtain a student multi-organ segmentation model;
[0051] Module M5: Input the predicted image and pass it through the student multi-organ segmentation model to obtain the organ segmentation prediction result.
[0052] Preferably, the module M1 uses a teacher single-organ neural network segmentation model composed of a convolutional layer and a fully connected layer to predict the input image, and the formula is as follows:
[0053]
[0054] Where x is the input image, and the superscript j indicates the jth dimension. is the mth teacher neural network model, δ k represents the softmax prediction output in the kth dimension, is the predicted probability of the kth dimension obtained by the mth teacher neural network model;
[0055] The student multi-organ neural network segmentation model composed of convolutional layers and fully connected layers is used to predict the input image. The formula is as follows:
[0056]
[0057] Among them, f t is the student neural network model, is the predicted probability of the kth dimension obtained by the student neural network model;
[0058] The module M2 comprises:
[0059] For organ regions, keep the background predictions of the teacher segmentation model The prediction of the organ numbered m remains unchanged. Shift to the mth dimension and then set the values of all other dimensions to 0. The formula is as follows:
[0060]
[0061] in, is the output of the teacher model after the output conversion mapping to the organ numbered m;
[0062] For the background area, the prediction values of all M teacher single-organ segmentation models are averaged, and the formula is as follows:
[0063]
[0064] in, It is the output of the teacher model to the background after the output conversion mapping.
[0065] Preferably, the module M3 comprises:
[0066] Get the predicted mask for organ m, the formula is as follows:
[0067]
[0068] The prediction of the organ region by the converted teacher single organ segmentation model is as follows:
[0069]
[0070] The prediction of the organ region by converting the student multi-organ segmentation model is as follows:
[0071]
[0072] Get the mth teacher single organ segmentation model For the background mask, the formula is as follows:
[0073]
[0074] Aggregate all M teacher single organ segmentation models for the background mask, the formula is as follows:
[0075]
[0076] The conversion teacher single organ segmentation model predicts the background area, and the formula is as follows:
[0077]
[0078] The prediction of the background area for the student multi-organ segmentation model is converted as follows:
[0079]
[0080] Preferably, the module M4 sets different weights for different organ regions and background regions, aligning the transformed prediction results of the teacher and student models, and the formula is as follows:
[0081]
[0082] Among them, λ is a hyperparameter used to balance the weights of organ regions and background regions;
[0083] The module M5 inputs the prediction image and predicts it through the student multi-organ segmentation model. The formula is as follows:
[0084] p=δ(f t (x)) (13)
[0085] Among them, p is the obtained multi-organ segmentation result.
[0086] Compared with the prior art, the present invention has the following beneficial effects:
[0087] 1. The present invention provides a privacy-friendly multi-organ segmentation method and system based on multi-source integrated distillation, which can alleviate the problems of missing multi-organ annotations and data privacy to a certain extent. In the absence of original training data, the present invention uses multiple pre-trained single-organ segmentation models for different organs and an unlabeled target domain dataset to train an efficient multi-organ segmentation model, reducing the need for annotation of training data and eliminating the need for original training data;
[0088] 2. The multi-organ segmentation model trained by the present invention aggregates the knowledge of multiple pre-trained single-organ segmentation models. Compared with the single-organ segmentation model and other multi-organ segmentation models, it further improves the segmentation accuracy of each organ and saves a lot of computing time compared with the single-organ segmentation model.
[0089] 3. The present invention conducts learning and training for different organs and background regions respectively, and controls the ratio of the two through hyperparameters, so that the multi-organ segmentation model can pay attention to both organs and background regions at the same time, and the learning of organs and background regions assists and constrains each other, thereby jointly improving the segmentation performance of multiple organs;
[0090] 4. The multi-organ segmentation model trained by the present invention utilizes the input of the target domain for training, so that the model can adapt to the predicted distribution and environment and has better performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0091] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:
[0092] Figure 1 A flow chart of a method in an embodiment of the present invention;
[0093] Figure 2 Schematic diagram of the system in the embodiment of the present invention. DETAILED DESCRIPTION
[0094] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.
[0095] The purpose of the present invention is to provide a multi-organ segmentation method and system based on multi-source integrated distillation. In the scenario where there is only a single-organ pre-training model but no original training data, a model that can accurately segment multiple organs is trained to help data modeling in real medical scenarios.
[0096] The present invention provides an embodiment, a multi-organ segmentation method based on multi-source integrated distillation, comprising the following steps:
[0097] Step S1: The input image is predicted by the teacher and student models to obtain multiple sets of segmentation maps. The input image is predicted by the teacher and student models to obtain multiple sets of segmentation maps. The teacher single-organ neural network segmentation model composed of convolutional layers and fully connected layers is used to predict the input image. The formula is as follows:
[0098]
[0099] Where x is the input image, and the superscript j indicates the jth dimension. is the mth teacher neural network model, δ k represents the softmax prediction output in the kth dimension, is the predicted probability of the kth dimension obtained by the mth teacher neural network model.
[0100] The student multi-organ neural network segmentation model composed of convolutional layers and fully connected layers is used to predict the input image. The formula is as follows:
[0101]
[0102] Among them, f t is the student neural network model, is the predicted probability of the kth dimension obtained by the student neural network model.
[0103] Step S2: Output conversion of the teacher image segmentation map and expansion to multi-category output. Output conversion of the teacher image segmentation map and expansion to multi-category output include: For organ regions, keep the background prediction of the teacher segmentation model The prediction of the organ numbered m remains unchanged. Shift to the mth dimension and then set the values of all other dimensions to 0. The formula is as follows:
[0104]
[0105] in, It is the output of the teacher model for the organ numbered m after the output transformation mapping.
[0106] For the background area, the prediction values of all M teacher single-organ segmentation models are averaged, and the formula is as follows:
[0107]
[0108] in, It is the output of the teacher model to the background after the output conversion mapping.
[0109] Step S3: Convert the organ and background regions according to the region-based mask. Converting the organ and background regions according to the region-based mask includes: obtaining a predicted mask for the organ m, and the formula is as follows:
[0110]
[0111] The prediction of the organ region by the converted teacher single organ segmentation model is as follows:
[0112]
[0113] The prediction of the organ region by converting the student multi-organ segmentation model is as follows:
[0114]
[0115] Get the mth teacher single organ segmentation model For the background mask, the formula is as follows:
[0116]
[0117] Aggregate all M teacher single organ segmentation models for the background mask, the formula is as follows:
[0118]
[0119] The conversion teacher single organ segmentation model predicts the background area, and the formula is as follows:
[0120]
[0121] The prediction of the background area for the student multi-organ segmentation model is converted as follows:
[0122]
[0123] Step S4: Perform supervised learning on different organ regions and background regions respectively, so that the prediction results after conversion of the teacher and student models are aligned, and an efficient student multi-organ segmentation model is obtained. Perform supervised learning on different organ regions and background regions respectively, so that the prediction results after conversion of the teacher and student models are aligned, and an efficient student multi-organ segmentation model is obtained, including: setting different weights for different organ regions and background regions, aligning the prediction results after conversion of the teacher and student models, and the formula is as follows:
[0124]
[0125] Among them, λ is a hyperparameter used to balance the weights of organ regions and background regions.
[0126] Step S5: Input the prediction image and pass it through the student multi-organ segmentation model to obtain the organ segmentation prediction result. Inputting the prediction image and passing it through the student multi-organ segmentation model to obtain the organ segmentation prediction result includes: inputting the prediction image and passing it through the student multi-organ segmentation model for prediction, the formula is as follows:
[0127] p=δ(f t (x)) (13)
[0128] Among them, p is the obtained multi-organ segmentation result.
[0129] The above-mentioned embodiment of the present invention utilizes the aggregation of single-organ knowledge of multiple teacher models and a learning method based on organs and background areas to realize the training of multi-organ models, and has achieved good results in multi-organ prediction.
[0130] Based on the same concept of the above embodiment, in another embodiment of the present invention, a privacy-friendly multi-organ segmentation system based on multi-source integrated distillation is provided. Figure 2 The schematic diagram of the system of this embodiment is shown in FIG. It includes:
[0131] Organ segmentation model module: The input image is predicted by the teacher and student models to obtain multiple groups of segmentation maps.
[0132] Single organ segmentation output conversion module: The teacher image segmentation map is converted into output and expanded to multi-category output.
[0133] Mask Definition and Transformation Module: Transforms the output into organ and background regions according to region-based masks.
[0134] Supervised alignment module: supervised learning is performed on different organ regions and background regions respectively, so that the converted prediction results of the teacher and student models are aligned to obtain an efficient student multi-organ segmentation model.
[0135] Prediction module: The input prediction image is passed through the student multi-organ segmentation model to obtain the organ segmentation prediction result.
[0136] The present invention also provides a multi-organ segmentation system based on multi-source integrated distillation, the system comprising the following modules:
[0137] Module M1: The input image is predicted by the teacher and student models to obtain multiple sets of segmentation maps; the teacher single-organ neural network segmentation model composed of convolutional layers and fully connected layers is used to predict the input image. The formula is as follows:
[0138]
[0139] Where x is the input image, and the superscript j indicates the jth dimension. is the mth teacher neural network model, δ k represents the softmax prediction output in the kth dimension, is the predicted probability of the kth dimension obtained by the mth teacher neural network model;
[0140] The student multi-organ neural network segmentation model composed of convolutional layers and fully connected layers is used to predict the input image. The formula is as follows:
[0141]
[0142] Among them, f t is the student neural network model, is the predicted probability of the kth dimension obtained by the student neural network model.
[0143] Module M2: Convert the output of the teacher image segmentation map to expand it to multi-category output; for the organ area, keep the background prediction of the teacher segmentation model The prediction of the organ numbered m remains unchanged. Shift to the mth dimension and then set the values of all other dimensions to 0. The formula is as follows:
[0144]
[0145] in, is the output of the teacher model after the output conversion mapping to the organ numbered m;
[0146] For the background area, the prediction values of all M teacher single-organ segmentation models are averaged, and the formula is as follows:
[0147]
[0148] in, It is the output of the teacher model to the background after the output conversion mapping.
[0149] Module M3: Convert the output to organ and background regions based on the region-based mask; obtain the predicted mask for organ m, the formula is as follows:
[0150]
[0151] The prediction of the organ region by the converted teacher single organ segmentation model is as follows:
[0152]
[0153] The prediction of the organ region by converting the student multi-organ segmentation model is as follows:
[0154]
[0155] Get the mth teacher single organ segmentation model For the background mask, the formula is as follows:
[0156]
[0157] Aggregate all M teacher single organ segmentation models for the background mask, the formula is as follows:
[0158]
[0159] The conversion teacher single organ segmentation model predicts the background area, and the formula is as follows:
[0160]
[0161] The prediction of the background area for the student multi-organ segmentation model is converted as follows:
[0162]
[0163] Module M4: supervised learning is performed on different organ regions and background regions respectively, and the prediction results after conversion of the teacher and student models are aligned to obtain a student multi-organ segmentation model; different weights are set for different organ regions and background regions, and the prediction results after conversion of the teacher and student models are aligned. The formula is as follows:
[0164]
[0165] Among them, λ is a hyperparameter used to balance the weights of organ regions and background regions.
[0166] Module M5: Input the prediction image and pass it through the student multi-organ segmentation model to obtain the organ segmentation prediction result; Input the prediction image and pass it through the student multi-organ segmentation model for prediction, the formula is as follows:
[0167] p=δ(f t (x)) (13)
[0168] Among them, p is the obtained multi-organ segmentation result.
[0169] In summary, the above embodiment utilizes the coarse super-resolution result of a learnable coarse super-resolution model, utilizes a high-frequency discrete learnable dictionary to discretely encode the high-frequency part of the image, utilizes a local autoregressive model to complete the generation from the coarse super-resolution result to the high-frequency dictionary encoding, and utilizes a high-definition image generation module to generate a high-definition image corresponding to the final low-definition image, thereby alleviating the problems of missing multi-organ annotations and data privacy, while improving the segmentation accuracy of each organ, and saving a lot of computing time compared to the single-organ segmentation model.
[0170] The present invention can utilize a publicly available pre-trained single-organ segmentation model and an unlabeled multi-organ dataset to perform multi-organ segmentation and achieve good organ segmentation effects.
[0171] It should be noted that the steps in the method provided by the present invention can be implemented by using corresponding modules, devices, units, etc. in the system. Those skilled in the art can refer to the technical solution of the method to realize the composition of the system, that is, the embodiments in the method can be understood as preferred examples for constructing the system, which will not be elaborated here.
[0172] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various modifications or variations within the scope of the claims, which does not affect the essence of the present invention. The above preferred features can be used in any combination without conflicting with each other.
[0173] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.
[0174] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.
Claims
1. A multi-organ segmentation method based on multi-source integrated distillation, characterized in that: The method comprises the following steps: Step S1: The input image is predicted by the teacher and student models to obtain multiple sets of segmentation maps; Step S2: convert the teacher image segmentation map to output and expand it to multi-category output; Step S3: converting the output into organ and background regions according to the region-based mask; Step S4: supervised learning is performed on different organ regions and background regions respectively, and the transformed prediction results of the teacher and student models are aligned to obtain a student multi-organ segmentation model; Step S5: input the prediction image and pass it through the student multi-organ segmentation model to obtain the organ segmentation prediction result; The step S2 comprises: For organ regions, keep the background predictions of the teacher segmentation model The prediction of the organ numbered m remains unchanged. Shift to the mth dimension and then set the values of all other dimensions to 0. The formula is as follows: in, is the output of the teacher model after the output conversion mapping to the organ numbered m; For the background area, the prediction values of all M teacher single-organ segmentation models are averaged, and the formula is as follows: The step S3 comprises: Get the predicted mask for organ m, the formula is as follows: The prediction of the organ region by the converted teacher single organ segmentation model is as follows: The prediction of the organ region by converting the student multi-organ segmentation model is as follows: Get the mth teacher single organ segmentation model For the background mask, the formula is as follows: Aggregate all M teacher single organ segmentation models for the background mask, the formula is as follows: The conversion teacher single organ segmentation model predicts the background area, and the formula is as follows: The prediction of the background area for the student multi-organ segmentation model is converted as follows: The step S4 sets different weights for different organ regions and background regions, and aligns the transformed prediction results of the teacher and student models. The formula is as follows: Among them, λ is a hyperparameter used to balance the weights of organ regions and background regions.
2. The multi-organ segmentation method based on multi-source integrated distillation according to claim 1, characterized in that: The step S1 uses the teacher single organ neural network segmentation model composed of convolutional layers and fully connected layers to predict the input image. The formula is as follows: Where x is the input image, and the superscript j indicates the jth dimension. is the mth teacher neural network model, δ k represents the softmax prediction output in the kth dimension, is the predicted probability of the kth dimension obtained by the mth teacher neural network model; The student multi-organ neural network segmentation model composed of convolutional layers and fully connected layers is used to predict the input image. The formula is as follows: Among them, f t is the student neural network model, is the predicted probability of the kth dimension obtained by the student neural network model.
3. The multi-organ segmentation method based on multi-source integrated distillation according to claim 1, characterized in that: In step S5, the prediction image is input and predicted by the student multi-organ segmentation model, and the formula is as follows: p=δ(f t (x))(13) Among them, p is the obtained multi-organ segmentation result.
4. A multi-organ segmentation system based on multi-source integrated distillation, characterized in that: The system includes the following modules: Module M1: The input image is predicted by the teacher and student models to obtain multiple sets of segmentation maps; Module M2: Output conversion of the teacher image segmentation map and expansion to multi-category output; Module M3: Convert the output to organ and background regions based on region-based masks; Module M4: supervised learning is performed on different organ regions and background regions respectively, and the transformed prediction results of the teacher and student models are aligned to obtain a student multi-organ segmentation model; Module M5: Input the predicted image and pass it through the student multi-organ segmentation model to obtain the organ segmentation prediction result; The module M2 comprises: For organ regions, keep the background predictions of the teacher segmentation model The prediction of the organ numbered m remains unchanged. Shift to the mth dimension and then set the values of all other dimensions to 0. The formula is as follows: in, is the output of the teacher model after the output conversion mapping to the organ numbered m; For the background area, the prediction values of all M teacher single-organ segmentation models are averaged, and the formula is as follows: in, It is the output of the teacher model to the background after the output conversion mapping; The module M3 comprises: Get the predicted mask for organ m, the formula is as follows: The prediction of the organ region by the converted teacher single organ segmentation model is as follows: The prediction of the organ region by converting the student multi-organ segmentation model is as follows: Get the mth teacher single organ segmentation model For the background mask, the formula is as follows: Aggregate all M teacher single organ segmentation models for the background mask, the formula is as follows: The conversion teacher single organ segmentation model predicts the background area, and the formula is as follows: The prediction of the background area for the student multi-organ segmentation model is converted as follows: The module M4 sets different weights for different organ regions and background regions, aligning the transformed prediction results of the teacher and student models, and the formula is as follows: Among them, λ is a hyperparameter used to balance the weights of organ regions and background regions.
5. The multi-organ segmentation system based on multi-source integrated distillation according to claim 4, characterized in that: The module M1 uses the teacher single-organ neural network segmentation model composed of convolutional layers and fully connected layers to predict the input image. The formula is as follows: Where x is the input image, and the superscript j indicates the jth dimension. is the mth teacher neural network model, δ k represents the softmax prediction output in the kth dimension, is the predicted probability of the kth dimension obtained by the mth teacher neural network model; The student multi-organ neural network segmentation model composed of convolutional layers and fully connected layers is used to predict the input image. The formula is as follows: Among them, f t is the student neural network model, is the predicted probability of the kth dimension obtained by the student neural network model.
6. The multi-organ segmentation system based on multi-source integrated distillation according to claim 4, characterized in that: The module M5 inputs the prediction image and predicts it through the student multi-organ segmentation model. The formula is as follows: p=δ(f t (x))(13) Among them, p is the obtained multi-organ segmentation result.