Image segmentation method and device based on self-training and prior guidance
By using self-training and prior guidance methods in medical image segmentation, using pseudo-masks to generate network and boundary image training models, the problems of large amount of parameters and limited noise processing capabilities in the existing technology are solved, and fast and accurate Multimodal image segmentation improves the effect of tumor segmentation.
Patent Information
- Application Number
- CN202411074124.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-07
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-08-07
AI Technical Summary
The prior art has problems in medical image segmentation with large amounts of parameters, high computing resources and limited noise processing capabilities, making it difficult to achieve fast and accurate multimodal image segmentation.
Using the image segmentation method based on self-training and prior guidance, the real mask is generated by annotating the original image dataset, the pseudo mask is generated by using the pseudo mask generation network, and the image segmentation model is trained in combination with the boundary image to reduce the amount of training parameters and Improve the segmentation effect.
在计算资源和数据有限的情况下,提高了图像分割模型的分割效果,适用于多模态医学图像的肿瘤分割,减少了训练参数量,同时增强了分割模型的先验知识,提升了肿瘤分割 Accuracy of the results.
Smart Images

Figure CN118587443B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the cross-technical field of computer technology and multimodal images, and in particular to an image segmentation method and device based on self-training and priori guidance. Background Art
[0002] Image segmentation is the technology and process of dividing an image into homogeneous regions with their own characteristics and extracting the target objects of interest. Through image segmentation, complex images can be decomposed into multiple simple image components with clear semantics, which is convenient for subsequent analysis and processing. In medical images, image segmentation of tumor areas plays a vital role in accurate diagnosis, treatment planning and disease monitoring. Manual segmentation not only requires experienced radiologists, but also the labeling process is time-consuming and laborious, so automated tumor segmentation is very necessary.
[0003] At present, image segmentation methods are mainly divided into traditional morphological methods (such as threshold segmentation, region growing, etc.), small deep learning models (such as U-Net and its variants), and large image segmentation models (such as SAM, etc.). Due to the complex anatomical structure, blurred tissue boundaries, and small lesion areas of medical images, traditional morphological methods are simple and intuitive but have high requirements for preprocessing; small deep learning models have fast training speed and low memory requirements, but have limited image noise processing capabilities; large image segmentation models have high accuracy, but the fine-tuning cost of directly applying them in the field of medical image segmentation is high and requires a lot of computing resources. Therefore, it is necessary to find a segmentation method with small parameters, fast speed, high quality, and applicable to multimodal images to assist radiologists in improving diagnostic efficiency and accuracy. Summary of the invention
[0004] The purpose of the present invention is to provide an image segmentation method and device based on self-training and prior guidance in view of the deficiencies in the prior art.
[0005] The object of the present invention is achieved through the following technical solution: an image segmentation method based on self-training and prior guidance, comprising:
[0006] Annotate the original image dataset and generate the corresponding real mask;
[0007] The pseudo-mask generation network is trained using images containing real masks, and the trained pseudo-mask generation network is used to generate pseudo-masks based on images without masks;
[0008] Extracting a boundary from a mask image to generate a boundary image; the mask image includes a real mask image and a pseudo mask image;
[0009] A three-channel image is formed based on the original image and its corresponding mask image and boundary image;
[0010] The three-channel image is used to train an image segmentation model, and the trained image segmentation model is used to segment the image to be segmented.
[0011] Furthermore, the original image data set includes two-dimensional data and three-dimensional data; the two-dimensional data and three-dimensional data in the original image data set are distinguished by modality and marked;
[0012] The three-dimensional data in the original image data set is preprocessed at least as follows: sliced according to three views: coronal plane, sagittal plane and cross-sectional plane.
[0013] Furthermore, the original image includes a CT image and an ultrasound image, wherein the CT image is three-dimensional data and the ultrasound image is two-dimensional data.
[0014] Furthermore, the pseudo-mask generation network includes a teacher model and a student model; the main parts of the teacher model and the student model use a pre-trained SAM large model; the pre-trained SAM large model includes an image encoder, a prompt encoder and a mask decoder;
[0015] The student model is trained using images containing real masks, including: keeping the parameters of the image encoder and the prompt encoder in the student model unchanged, using images containing real masks as training samples to train only the mask decoder in the student model, and the loss function of the i-th sample is expressed as:
[0016]
[0017] Among them, H represents the length of the image, W represents the width of the image, j represents the pixel index of the image, K represents the number of segmentation categories, and k represents the category index of the segmentation mask; Represents an image with a true mask, and its corresponding true mask is
[0018] Pass the trained student model parameters to the teacher model.
[0019] Furthermore, the teacher model also includes a suspicious noise decoder, which takes the output of the image encoder and the mask decoder in the teacher model as input, predicts suspicious noise and outputs a low-noise pseudo mask;
[0020] Keep the parameters of the image encoder, hint encoder and mask decoder in the teacher model unchanged, use the images containing the real mask as training samples to train only the suspicious noise decoder in the teacher model. The loss function of the suspicious noise decoder is divided into: mask refinement loss function And the noise refinement loss function
[0021]
[0022] in, represents the predicted noise mask, represents the true noise mask, represents the segmentation logic of the image with the true mask, express After the perturbation, the segmentation logic Represents the segmentation logic of the image to be requested, represents the fast Fourier transform, represents the inverse transformation, represents the amplitude, represents phase; C represents mask.
[0023] Furthermore, it also includes: using noise mask Calculated confidence As a weighting coefficient, it is used to calculate the loss function between the predicted mask of the image to be required and the low-noise pseudo mask To train the mask decoder of the student model, its loss function for:
[0024]
[0025] in, represents a low-noise pseudo mask; if the j pixel of the pseudo mask contains noise, Otherwise 0.
[0026] Furthermore, extracting the boundary from the mask image to generate the boundary image includes: using a Canny edge detection algorithm to extract the boundary from the mask image to generate the boundary image.
[0027] Furthermore, the image segmentation model includes a U-Net model, a U-Net++ model, a SegNet model, etc.
[0028] Furthermore, it also includes: using one or more evaluation indicators to evaluate the image segmentation model; the evaluation indicators include average intersection-over-union ratio and Dice coefficient.
[0029] The present invention also provides an image segmentation device based on self-training and prior guidance, comprising:
[0030] The annotation module is used to annotate several images in the original image dataset and generate corresponding true masks;
[0031] A pseudo-mask generation module is used to train a pseudo-mask generation network using an image containing a real mask, and the trained pseudo-mask generation network is used to generate a pseudo-mask based on an image without a mask;
[0032] A boundary image generation module, used to extract the boundary from the mask image and generate a boundary image; the mask image includes a real mask image and a pseudo mask image;
[0033] The image segmentation module is used to splice the original image and its corresponding mask image and boundary image into a three-channel image; the three-channel image is used to train the image segmentation model, and the trained image segmentation model is used to segment the image to be required.
[0034] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:
[0035] 1. Based on the original image and its corresponding mask image and boundary image, a three-channel image is spliced to train the image segmentation model and improve the segmentation effect of the image segmentation model.
[0036] 2. The method of the present invention can also perform segmentation on multimodal images, such as tumor segmentation, organ segmentation, etc. on multimodal medical images.
[0037] 3. In the case of limited computing resources and data, the method of the present invention can minimize the amount of training parameters while giving full play to the advantages of existing large models. The method of the present invention can also strengthen the prior knowledge of the segmentation model and further improve the results of tumor segmentation. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.
[0039] Figure 1 A schematic diagram of a process flow of an image segmentation method based on self-training and prior guidance provided by an embodiment of the present invention;
[0040] Figure 2 A schematic diagram of a modal discrimination and preprocessing process provided by an embodiment of the present invention;
[0041] Figure 3 A schematic diagram of a low-noise pseudo-mask generation network structure based on a SAM model provided in an embodiment of the present invention;
[0042] Figure 4 A schematic diagram of suspicious noise decoding provided by an embodiment of the present invention;
[0043] Figure 5 A schematic diagram of a mask inspection process provided by an embodiment of the present invention;
[0044] Figure 6 A schematic diagram of an image segmentation model and model evaluation provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0045] The present invention is described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the features of the following embodiments and implementations can be combined with each other.
[0046] The present invention provides an image segmentation method based on self-training and prior guidance, comprising:
[0047] (1) Annotate the original image dataset and generate the corresponding real mask;
[0048] Specifically, the image is annotated according to the specific task to generate the corresponding real mask; such as annotating the tumor site in the medical image.
[0049] In one embodiment, if Figure 2 As shown, the original image data set includes two-dimensional data and three-dimensional data; the two-dimensional data and three-dimensional data in the original image data set are distinguished by modality and marked; the three-dimensional data in the original image data set is preprocessed at least as follows: sliced according to three views of coronal plane, sagittal plane and cross-section respectively; so as to achieve segmentation of multimodal images.
[0050] For multimodal images, a modality discrimination module is introduced to label the type of input images; for example, two-dimensional data is labeled as 0 and three-dimensional data is labeled as 1.
[0051] (2) Using images containing real masks to train the pseudo-mask generation network, the trained pseudo-mask generation network is used to generate pseudo-masks based on images without masks;
[0052] In one embodiment, the pseudo-mask generation network includes a teacher model and a student model; the main parts of the teacher model and the student model use a pre-trained SAM large model; the pre-trained SAM large model includes an image encoder, a prompt encoder and a mask decoder;
[0053] The student model is trained using images containing real masks, including: keeping the parameters of the image encoder and the prompt encoder in the student model unchanged, using images containing real masks as training samples to train only the mask decoder in the student model, and the loss function of the i-th sample is expressed as:
[0054]
[0055] Among them, H represents the length of the image, W represents the width of the image, j represents the pixel index of the image, K represents the number of segmentation categories, and k represents the category index of the segmentation mask; Represents an image with a true mask, and its corresponding true mask is
[0056] Pass the trained student model parameters to the teacher model.
[0057] In another embodiment, if Figure 3 As shown, the teacher model also includes a suspicious noise decoder, such as Figure 4 As shown, it takes the output of the image encoder and the mask decoder in the teacher model as input, predicts suspicious noise and outputs a low-noise pseudo mask;
[0058] Keep the parameters of the image encoder, hint encoder and mask decoder in the teacher model unchanged, use the images containing the real mask as training samples to train only the suspicious noise decoder in the teacher model. The loss function of the suspicious noise decoder is divided into: mask refinement loss function And the noise refinement loss function
[0059]
[0060] in, represents the predicted noise mask, represents the true noise mask, represents the segmentation logic of the image with the true mask, express After the perturbation, the segmentation logic Represents the segmentation logic of the image to be requested, represents the fast Fourier transform, represents the inverse transformation, represents the amplitude, Indicates the phase. C is a calculated mask. In order to interfere with the low-frequency components, the C value is set to 1 only in the center area and 0 in other areas.
[0061] The teacher model now generates a low-noise pseudo mask for the image without a mask.
[0062]
[0063] In another embodiment, in order to reduce the noise effect in the pseudo mask, the noise mask is Calculated confidence As a weighting coefficient, it is used to calculate the loss function between the predicted mask and the pseudo mask of the image to be requested Come to train Confidence The calculation process is:
[0064]
[0065] Among them, if the j pixel of the pseudo mask contains noise, then Otherwise it is 0. The loss function calculated from this for:
[0066]
[0067] It should be noted that not all images without masks can be generated into pseudo masks through the pseudo mask generation network. At this time, the mask detection module is introduced to select images without masks as new datasets without masks. The ultrasound image with the pseudo mask and the ultrasound image with the mask are used as a new dataset. Pass it to the network in step (2) and re-generate the pseudo mask until all target image data have corresponding masks.
[0068] (3) extracting a boundary from the mask image to generate a boundary image; the mask image includes a real mask image and a pseudo mask image;
[0069] Specifically, the canny edge detection algorithm can be used to extract the boundary from the mask image to generate a boundary image; for the layer without a mask in the three-dimensional image data, its boundary image is replaced by a completely black image.
[0070] (4) splicing the original image and its corresponding mask image and boundary image into a three-channel image;
[0071] Specifically, if the image is two-dimensional data, it is aligned in the order of original image, mask image, and boundary image, and spliced into a three-channel image; if the image is three-dimensional data, it is aligned in the order of original image, mask image, and boundary image according to three views, and spliced into a three-channel image.
[0072] (5) The three-channel image is used to train an image segmentation model, and the trained image segmentation model is used to segment the image to be segmented.
[0073] Specifically, if the image is two-dimensional data, the corresponding spliced three-channel image is directly transmitted to the image segmentation model for training, and the trained image segmentation model is used to segment the target image.
[0074] If the image is three-dimensional data, the three-channel images corresponding to the three spliced views are transmitted to the image segmentation model for training respectively, and three trained image segmentation models are obtained. The three trained image segmentation models are used to segment the three views corresponding to the target image respectively, and the segmentation results of each view are spliced to obtain three three-dimensional segmentation probability results; after each voxel position is softmaxed according to the three three-dimensional segmentation probability results, the category with the largest probability is the final segmentation result.
[0075] In one embodiment, if Figure 6 As shown, it also includes: using one or more evaluation indicators to evaluate the image segmentation model; the evaluation indicators include average intersection-over-union ratio and Dice coefficient.
[0076] Example 1: A tumor segmentation method for multimodal medical images, such as Figure 1 As shown;
[0077] (1) Collect and annotate CT and ultrasound data;
[0078] Specifically, the lesion site of the ovarian cancer patient is scanned by a high-speed rotating X-ray transmitter and receiver, and the original CT image of the cancerous site of the patient is obtained by the image reconstruction algorithm. The tumor site of the CT image is annotated by the radiologist, and the data format of the collected CT and its mask is nii. The ultrasound image is obtained by receiving and processing the echo containing the characteristic information of the lesion site of the ovarian cancer patient. The tumor site of the ultrasound image is annotated by the radiologist, and the data format of the collected ultrasound and its mask is dicom. The CT and ultrasound image data sets are divided into training set, validation set and test set in a ratio of 7:1:2.
[0079] (2) Introducing a modality discrimination module to mark the type of input images;
[0080] The input images of this embodiment are CT and ultrasound images. The modality is determined based on the dimension of the input image. The two-dimensional data is the ultrasound image, and the three-dimensional data is the CT image. According to the result of the modality determination, the ultrasound image is marked as 0 and the CT image is marked as 1.
[0081] (3) Preprocessing CT images and ultrasound images;
[0082] Specifically, the CT image processing method includes: adjusting the window width and window level, using the min_max normalization method to normalize the image to [0,1], and then slicing according to the coronal, sagittal and cross-sectional views;
[0083] The processing method of ultrasound images includes: using the min_max normalization method to normalize the image to between [0,1].
[0084] (4) Construct a low-noise pseudo-mask generation network based on the SAM model to generate a low-noise mask from an image without a mask;
[0085] The network consists of a teacher model and a student model. The main parts of both models use the pre-trained SAM large model, and the teacher model also includes a suspicious noise decoder. The parameters of the image encoder and prompt encoder in the student model are kept unchanged. The ovarian cancer data is only used to train the mask decoder part of the student model. The cross entropy loss function is selected as the loss function, batch_size is set to 16, the learning rate is set to 0.005, the hyperparameter τ1 is set to 0.5, and the hyperparameter τ2 is set to 0.7.
[0086] Specifically, let the original image containing the true mask be The corresponding mask is Then the dataset containing the true mask is Among them, N l is the number of data. Suppose the image without mask is The dataset without mask is Keep the parameters of the image encoder and prompt encoder in the SAM model unchanged, and only train the mask decoder part. Suppose the mask decoder of the student model is The mask decoder of the teacher model is
[0087] On images containing true masks, the student model uses the cross entropy loss function for supervised learning. The loss function of the i-th sample is expressed as:
[0088]
[0089] Wherein, H represents the length of the image, W represents the width of the image; j represents the pixel index of the image, K represents the number of segmented categories, and in this embodiment, represents the number of segmented tumor categories, and k represents the category index of the segmentation mask.
[0090] Use the student model trained above Model for teachers The parameters are updated iteratively.
[0091]
[0092] Among them, ψ is the parameter of the teacher model, is the parameter of the student model, n represents the number of iterations, and β is a hyperparameter that adjusts the iteration degree of the student model parameters.
[0093] In order to make the suspicious noise decoder focus on the image features and the segmentation logic of the noise perturbation, and focus the learning on predicting high-quality segmentation masks and pixels where the pseudo mask may be wrong, we add a perturbation strategy based on fast Fourier transform to optimize the decoding process.
[0094] The loss calculation for the image with the true mask on the suspicious noise decoder is and the image without mask Perform fast Fourier transform respectively to obtain the corresponding amplitude and phase. The phase is related to the relative position of each part of the original image structure, and the amplitude is related to the energy information on the original image. The low-frequency components in are replaced by The low-frequency components in the image are reconstructed by inverse fast Fourier transform. The segmentation logic of the image with a true mask is The segmentation logic of the image without mask is The fast Fourier transform is The inverse transformation is The amplitude is Phase is but The segmentation logic after disturbance is:
[0095]
[0096] Where C is a calculated mask. In order to interfere with the low-frequency components, the C value is set to 1 only in the center area and 0 in other areas.
[0097] because Only perturbation information is added, the semantic content of the image does not change, so its true segmentation mask can be used Train the suspicious noise decoder. The real noise mask It can be calculated by comparing the original and perturbed masks, that is, 1 if the masks are the same and 0 if they are different. For images with real masks, the loss function of the suspicious noise decoder is divided into: Mask refinement loss function And the noise refinement loss function
[0098]
[0099] in, represents the predicted noise mask.
[0100] For the loss calculation of the image without mask on the suspicious noise decoder, since the real segmentation mask cannot be obtained, the pseudo mask generated by the suspicious noise decoder is used as a reference. Since there is more noise in the pseudo mask generated in the early stage of training, a higher threshold τ1 is selected to filter the pseudo mask pixels, that is, the confidence in the early stage of training The later model has a certain ability to locate pseudo masks, and the loss is calculated in the same way as resemblance.
[0101] Therefore, the teacher model For the image to be requested Generated low noise pseudo mask for:
[0102]
[0103] In order to reduce the noise effect in the pseudo mask, the noise mask Calculated confidence As a weighting coefficient, it is used to calculate the loss function between the predicted mask and the pseudo mask of the image to be requested Come to train Confidence The calculation process is:
[0104]
[0105] Among them, if the j pixel of the pseudo mask contains noise, then Otherwise it is 0. The loss function calculated from this for:
[0106]
[0107] (5) Mask generation test;
[0108] Since CT images are multi-layered and not every layer has a tumor, the mask generation test is only performed on ultrasound data. Figure 5 As shown, the specific method is to check whether each ultrasound image has a mask, including a real mask or a pseudo mask generated in step (4), and select the images without masks as a new dataset without masks. The ultrasound image with the pseudo mask and the ultrasound image with the mask are used as a new dataset. The network is passed to step (4) to generate the low-noise pseudo mask again until all ultrasound data have corresponding masks.
[0109] (6) Mask boundary extraction and boundary image generation
[0110] The masks of CT and ultrasound images are both extracted using the canny edge detection algorithm as the boundary image. The boundary images of the layers in CT that do not contain masks are replaced by completely black images.
[0111] (7) splicing the original image and its corresponding mask image and boundary image into a three-channel image;
[0112] The ultrasound image is aligned in the order of original image, mask image, and boundary image, and then spliced into a three-channel image. The CT image is aligned in the order of original image, mask image, and boundary image according to the three views, and then spliced into a three-channel image.
[0113] (8) The three-channel image is used to train an image segmentation model, and the trained image segmentation model is used to segment the image to be segmented;
[0114] The three-channel images of the ultrasound image are directly transmitted to the U-Net model for training, and the test data is input into the trained model to obtain the segmentation results. The three-channel images of the three views of the CT image are transmitted to the U-Net model for training respectively, and the test data are transmitted to the three trained models according to the three views respectively. The segmentation results of each view in the CT image are spliced to obtain the three-dimensional segmentation probability results. After each voxel position is softmaxed according to the three-dimensional segmentation probability results, the category with the highest probability is the final segmentation result.
[0115] like Figure 6 As shown, the image segmentation model is also evaluated; the evaluation indicators include average intersection-over-union ratio and Dice coefficient.
[0116] In summary, the tumor segmentation method of multimodal medical images based on self-training and prior guidance in this embodiment 1 includes a self-training mechanism, the mainstream SAM and the prior knowledge guidance of its preliminary results, which enhances the information capture of the encoder of the commonly used medical segmentation model, improves the segmentation performance, and provides a more effective method for refining the tumor segmentation results.
[0117] The present invention also provides an image segmentation device based on self-training and prior guidance, comprising:
[0118] The annotation module is used to annotate several images in the original image dataset and generate corresponding true masks;
[0119] A pseudo-mask generation module is used to train a pseudo-mask generation network using an image containing a real mask, and the trained pseudo-mask generation network is used to generate a pseudo-mask based on an image without a mask;
[0120] A boundary image generation module, used to extract the boundary from the mask image and generate a boundary image; the mask image includes a real mask image and a pseudo mask image;
[0121] The image segmentation module is used to splice the original image and its corresponding mask image and boundary image into a three-channel image; the three-channel image is used to train the image segmentation model, and the trained image segmentation model is used to segment the image to be required.
[0122] It should be noted that the device embodiment shown in this embodiment matches the content of the above method embodiment, and the content of the above method embodiment can be referred to, which will not be repeated here.
[0123] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.
Claims
1. An image segmentation method based on self-training and prior guidance, characterized in that: include: Annotate the original image dataset and generate the corresponding real mask; The pseudo-mask generation network is trained using images containing real masks, and the trained pseudo-mask generation network is used to generate pseudo-masks based on images without masks; Extracting a boundary from a mask image to generate a boundary image; the mask image includes a real mask image and a pseudo mask image; A three-channel image is formed based on the original image and its corresponding mask image and boundary image; The three-channel image is used to train an image segmentation model, and the trained image segmentation model is used to segment the image to be segmented.
2. The method according to claim 1, characterized in that The original image data set includes two-dimensional data and three-dimensional data; the two-dimensional data and three-dimensional data in the original image data set are identified by modality and marked; The three-dimensional data in the original image data set is preprocessed at least as follows: sliced according to three views: coronal plane, sagittal plane and cross-sectional plane.
3. The method according to claim 2, characterized in that The original images include CT images and ultrasound images, wherein the CT images are three-dimensional data and the ultrasound images are two-dimensional data.
4. The method according to claim 1, characterized in that: The pseudo-mask generation network includes a teacher model and a student model; the teacher model and the student model both use a pre-trained SAM large model; the pre-trained SAM large model includes an image encoder, a prompt encoder and a mask decoder; The student model is trained using images containing real masks, including: keeping the parameters of the image encoder and the prompt encoder in the student model unchanged, using images containing real masks as training samples to train only the mask decoder in the student model, and the loss function of the i-th sample is expressed as: Among them, H represents the length of the image, W represents the width of the image, j represents the pixel index of the image, K represents the number of segmentation categories, and k represents the category index of the segmentation mask; Represents an image with a true mask, and its corresponding true mask is represents the student model; Pass the trained student model parameters to the teacher model.
5. The method according to claim 4, characterized in that The teacher model also includes a suspicious noise decoder, which takes the output of the image encoder and the mask decoder in the teacher model as input, predicts suspicious noise and outputs a low-noise pseudo mask; Keep the parameters of the image encoder, hint encoder and mask decoder in the teacher model unchanged, use the images containing the real mask as training samples to train only the suspicious noise decoder in the teacher model. The loss function of the suspicious noise decoder is divided into: mask refinement loss function And the noise refinement loss function in, represents the predicted noise mask, represents the true noise mask, represents the segmentation logic of the image with the true mask, express After the perturbation, the segmentation logic Represents the segmentation logic of the image to be requested, represents the fast Fourier transform, represents the inverse transformation, represents the amplitude, represents phase; C represents mask.
6. The method according to claim 5, characterized in that Also includes: Confidence calculated with noise mask As a weighting coefficient, it is used to calculate the loss function between the predicted mask of the image to be required and the low-noise pseudo mask To train the mask decoder of the student model, its loss function for: in, represents the image to be requested; represents a low-noise pseudo mask; if the j pixel of the pseudo mask contains noise, Otherwise 0.
7. The method according to claim 1, characterized in that The step of extracting the boundary from the mask image and generating the boundary image includes: extracting the boundary from the mask image and generating the boundary image using a Canny edge detection algorithm.
8. The method according to claim 1, characterized in that: The image segmentation models include U-Net model, U-Net++ model, and SegNet model.
9. The method according to claim 1, characterized in that: Also includes: The image segmentation model is evaluated using one or more evaluation indicators; the evaluation indicators include average intersection-over-union ratio and Dice coefficient.
10. An image segmentation device based on self-training and prior guidance, characterized in that: include: The annotation module is used to annotate several images in the original image dataset and generate corresponding true masks; A pseudo-mask generation module is used to train a pseudo-mask generation network using an image containing a real mask, and the trained pseudo-mask generation network is used to generate a pseudo-mask based on an image without a mask; A boundary image generation module, used to extract the boundary from the mask image and generate a boundary image; the mask image includes a real mask image and a pseudo mask image; The image segmentation module is used to splice the original image and its corresponding mask image and boundary image into a three-channel image; the three-channel image is used to train the image segmentation model, and the trained image segmentation model is used to segment the image to be required.
