A method and system for cavity organ image segmentation

Abstract

This invention discloses a method and system for segmenting images of hollow organs, relating to the field of image segmentation technology. The method includes the following steps: collecting an image dataset of hollow organs containing intracavitary material; using the U-net algorithm to segment the images in the dataset into cavities and the material contained within them, projecting the cavities and material onto polar coordinates, and obtaining the cavity wall region by the difference between the polar coordinates of the cavity and the material; introducing a self-supervised BYOL architecture, using contrast and reconstruction losses to strengthen the cavity wall region, and obtaining the weights of the cavity wall region; using images with label quality higher than a threshold and transfer weights to supervise the training of a deep learning model, constructing a panoramic segmentation model of the image based on the training results; and using the panoramic segmentation model to segment the image containing intracavitary material, obtaining an image of the hollow organ after segmentation of the intracavitary material. In this invention, the U-net algorithm not only improves the network's ability to recognize the boundaries of hollow organs and eliminates the dependence of deep models on large amounts of data, but also achieves good results in image segmentation.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and in particular to a method and system for segmenting images of cavity organs. Background Art

[0002] Hollow organs include the nose, pharynx, trachea, esophagus, gastrointestinal tract, bladder, uterus and vagina. The characteristics of lesions in hollow organs are erosion, ulceration and bleeding of the organ lining, which causes cells to fall off and be discharged from the cavity.

[0003] Clinically, medical images are usually processed by experienced radiologists. Such manual operations are cumbersome and time-consuming, and the results may be biased due to human subjective consciousness. Since the location of lesions in hollow organs is on the organ lining, such as the bladder, it is difficult to distinguish the mixed area of ​​lesions and bladder walls due to its thin lining wall, which greatly increases the difficulty of medical staff. Therefore, applying artificial intelligence to it can greatly reduce the workload of doctors. Medical image segmentation technology based on deep learning can not only improve processing efficiency, but also have a certain guarantee of its accuracy, which can provide good auxiliary role for doctors in diagnosis and treatment. At present, there are still many shortcomings in the segmentation of lesion images of hollow organs. Take bladder tumors as an example. So far, there are two main research methods for segmenting bladder walls and tumors: model-driven methods and data-driven methods. In model-driven methods, segmentation is mainly achieved by making various improvements to level sets or by training classifiers with manual features. Data-driven methods tend to build a deep network trained by a large number of images. Compared with model-driven methods, data-driven methods can extract more targeted features according to tasks. However, data-driven methods with better performance require more data for training.

[0004] This requires professional radiologists to accurately depict the walls of the cavity organs before network training. As mentioned above, it is difficult for people to distinguish the mixed area between the cavity lining and the lesion area. Even professional radiologists may not provide completely accurate labels for each image. In addition, the labels of small data sets cannot meet the training requirements of deep neural networks. The high difficulty, large workload and high cost of labeling have led to a very limited size of the corresponding data set. In the case of inconsistent data set quality, the stability of the deep neural network is affected, and the image segmentation accuracy is affected. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for cavity organ image segmentation in view of the deficiencies of the above-mentioned prior art, so as to solve the problem that the labels of small data sets in the prior art cannot meet the training requirements of deep neural networks, and when the quality of data sets is inconsistent, the labeling difficulty is high, the workload is large, and the cost is too high, which leads to the corresponding data set size being very limited. When the quality of data sets is inconsistent, the stability of the deep neural network is affected, and the image segmentation accuracy is affected.

[0006] The present invention specifically provides the following technical solution: a method for segmenting a cavity organ image, comprising the following steps:

[0007] Collecting a dataset of images of hollow organs with intracavitary substances; wherein the images in the dataset of images of hollow organs have labels for distinguishing between the cavity wall and the intracavitary substances;

[0008] Using a U-net algorithm to segment the image into a cavity and a substance in the cavity, and projecting the cavity and the substance into polar coordinates, obtaining a cavity wall region by the difference between the cavity polar coordinates and the substance polar coordinates, and removing images with label errors greater than an error threshold by using the obtained cavity wall region;

[0009] Using a BYOL architecture of self-supervised training, the cavity wall region is subjected to comparison and reconstruction loss training to obtain an optimized cavity wall region and a weight of the cavity wall region;

[0010] The weights are transferred to a deep learning model, and images whose label errors are less than an error threshold are input into the deep learning model for supervised training, and a panoramic segmentation model of the image is constructed based on the training results;

[0011] The panoramic segmentation model is used to segment the image to be segmented with the intracavitary material, so as to obtain the cavity organ image after the intracavitary material is segmented.

[0012] Preferably, the step of eliminating images with label errors greater than an error threshold from the obtained cavity wall region comprises the following steps:

[0013] Comparing the area of ​​the cavity wall region with the area of ​​the label of the image data set projected to polar coordinates;

[0014] The threshold discrimination method is used for the compared areas to eliminate images whose labels differ by more than the error threshold.

[0015] Preferably, the step of transferring the weights to a deep learning model, inputting images whose label errors are less than an error threshold into the deep learning model for supervised training, and constructing a panoramic segmentation model of the image through the training results comprises the following steps:

[0016] Use a deep learning model to batch normalize input images whose label quality is above a threshold;

[0017] Introducing a leaky relu layer into the deep learning model, eliminating the neuron death scenario when the input transfer weight is negative; where death means that the input weight cannot be updated;

[0018] Increasing the number of filters in each convolutional layer in the deep learning model, and performing convolution operations on the input image through the leaky relu layer and the transfer weight input convolutional layer to capture the characteristics of the data;

[0019] The deep learning model is evaluated using independent data, and is optimized according to the model evaluation results to obtain a panoramic segmentation model; wherein the optimization includes adjusting model parameters, loss function, and learning rate.

[0020] Preferably, the deep learning model is used to perform batch normalization on input images whose label errors are less than an error threshold. The specific expression is:

[0021]

[0022]

[0023]

[0024] Y i =γ×X i ′+β

[0025] Among them, ε is a constant, γ and β are learned and optimized according to the training data, μ is the mean, σ is the standard deviation, Xi′ is the normalized result, and Xi is the input data set {X 1 ,X 2 ,...,X m}The i-th input data in .

[0026] Preferably, DiceFocal loss and VAE loss are used as the loss function of the panoramic segmentation model. Under the condition of data category imbalance, the panoramic segmentation model focuses on categories whose number is lower than a threshold to suppress sample imbalance.

[0027] Preferably, the specific expression of the DiceFocal loss is:

[0028]

[0029]

[0030] DFL=FL+DL

[0031] Among them, x is the prediction result, y is the label, FL is the Focal Loss loss function, DL is the Dice Loss loss function, DFL is the DiceFocal loss loss function, P t is the predicted probability of the true category of the sample, γ * To adjust the hyperparameter of the weight factor, α t For each weight.

[0032] Preferably, the specific expression of the VAE loss is:

[0033]

[0034] Among them, MSE is the reconstruction loss, KL is the divergence regularization term, and VL is the VAE loss function.

[0035] Preferably, when the self-supervised training BYOL architecture is introduced, noise is added to the images filtered out during the data screening process, and the original image is predicted using self-supervised learning.

[0036] Preferably, after the image to be segmented with intracavitary material is segmented using the panoramic segmentation model, data enhancement is performed on the segmentation result, and the data enhancement includes obtaining the maximum connected component and filling holes.

[0037] The present invention provides a system for segmenting images of cavity organs, comprising:

[0038] An acquisition module, for collecting a data set of images of hollow organs with intracavitary substances; wherein the images in the data set of images of the hollow organs have labels for distinguishing between the cavity wall and the intracavitary substances;

[0039] a cavity wall acquisition module, for segmenting the image into a cavity and a substance in the cavity by using a U-net algorithm, and projecting the cavity and the substance into polar coordinates, obtaining a cavity wall region by the difference between the cavity polar coordinates and the substance polar coordinates, and removing images whose label errors are greater than an error threshold by using the obtained cavity wall region;

[0040] A weight acquisition module, for performing comparison and reconstruction loss training on the cavity wall region using a BYOL architecture of self-supervised training, to obtain an optimized cavity wall region and a weight of the cavity wall region;

[0041] A task construction module, used to transfer the weights to a deep learning model, and input images whose label errors are less than an error threshold into the deep learning model for supervised training, and construct a panoramic segmentation model of the image through the training results;

[0042] The segmentation module is used to segment the image to be segmented with the intracavitary material using the panoramic segmentation model to obtain the cavity organ image after the intracavitary material is segmented.

[0043] Compared with the prior art, the present invention has the following significant advantages:

[0044] The present invention uses the U-net algorithm to segment the image into a cavity and the material contained in the cavity, and obtains the cavity wall by the difference of polar coordinates. It can not only improve the network's recognition ability of the cavity organ boundary, but also solve the dependence of the deep model on large amounts of data, strengthen the model's perception of details, and improve the network's recognition ability of the organ boundary. At the same time, the cavity wall area is strengthened through contrast and reconstruction loss, and the weight of the cavity wall area is obtained. The weight is transferred, and an image with a label quality higher than a threshold is used as input to further reduce the difficulty of segmenting the organ wall and the cavity material. While getting rid of the dependence of the deep model on a large amount of data, it also achieves a better effect on image segmentation. Moreover, under the premise of using a data set with different label qualities, a better segmentation effect can be achieved through training. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a general schematic diagram of the present invention;

[0046] Figure 2 Filter structure graphs for data;

[0047] Figure 3 This is the network structure diagram for self-supervised learning;

[0048] Figure 4 This is a diagram of the supervised learning network and post-processing structure;

[0049] Figure 5 This is the overall structure diagram;

[0050] Figure 6 The following are some examples of self-supervised segmentation results; (a), (b), (c), (d) are original images of different organs, (e), (f), (g), (h) are images after adding noise to the original images of different organs, (i), (j), (k), (l) are self-supervised segmentation results;

[0051] Figure 7 The performance comparison report diagram of the embodiment of the present invention and other studies; wherein (a) is the first group of comparative data, and (b) is the second group of comparative data. DETAILED DESCRIPTION

[0052] The following is a clear and complete description of the technical solutions of the embodiments of the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0053] An embodiment of the present invention provides a method for segmenting a cavity organ image, comprising the following steps:

[0054] Step S1: Collecting a dataset of images of hollow organs with intracavitary substances; wherein the images in the dataset of images of hollow organs have labels for distinguishing between the cavity wall and the intracavitary substances.

[0055] Step S2: Use the U-net algorithm to segment the image into the cavity and the material in the cavity, and project the cavity and the material into polar coordinates. Obtain the cavity wall area by the difference between the cavity polar coordinates and the material polar coordinates, and remove images with label errors greater than the error threshold by using the obtained cavity wall area. Use the designed discrimination rules to filter out data with labeling problems.

[0056] The discrimination rule is as follows: by comparing the area of ​​the cavity wall with the area of ​​the image dataset's label projected to the polar coordinates, a threshold discrimination method is used to eliminate the image datasets whose labels differ by more than the threshold.

[0057] Step S3: Using the BYOL architecture of self-supervised training, the cavity wall region is trained with comparison and reconstruction losses to obtain an optimized cavity wall region and a weight of the cavity wall region.

[0058] When the self-supervised training BYOL architecture is introduced, noise is added to the images filtered out during the data screening process, and the original image is predicted through self-supervised learning.

[0059] Step S4: Transfer the weights to the deep learning model, and input the images whose label errors are less than the error threshold into the deep learning model for supervised training, and build a panoramic segmentation model of the image based on the training results.

[0060] This step specifically includes:

[0061] Use high-quality labeled images for supervised learning. Use DiceFocal loss and VAE loss as loss functions. Under the condition of imbalanced data categories, let the model focus on a small number of categories to suppress the imbalance of samples.

[0062] The SegNet deep learning model is selected for model training, and the input images are classified and labeled at the pixel level. First, batch normalization (BN) is performed on the input images whose label quality is higher than the threshold (label error is less than the error threshold). Here, the input data set is given as {X 1 ,X 2 ,...,X m}, the specific formula is as follows:

[0063] Mean:

[0064] Standard deviation σ:

[0065] Normalization:

[0066] Y i =γ×X i ′+β

[0067] Among them, ε is a small number to avoid the denominator being 0, and γ and β will be learned and optimized according to the training data. BN processing is used to improve the speed and stability of neural network training convergence.

[0068] In addition, a leaky relu layer is introduced into the deep learning model to eliminate the neuron death scenario when the input transfer weight is negative; death means that the input weight cannot be updated. The specific formula is:

[0069]

[0070] This layer is used to solve the problem of neuron "death" that may occur in traditional ReLU when negative input is applied. The sparsity and expression capabilities are improved by introducing this layer, thereby increasing the nonlinear characteristics of the network.

[0071] Increase the number of filters in each convolutional layer in the deep learning model, perform convolution operations on the input image through the leaky relu layer and the transfer weight input convolutional layer, and increase the network width to enhance the representation ability of the model to better capture the complex features of the data, improve the model segmentation performance, and reduce overfitting.

[0072] Afterwards, the model is evaluated using independent data and optimized based on the model evaluation results to obtain a panoramic segmentation model. The optimization includes adjusting model parameters, loss function, learning rate, etc.

[0073] Using DiceFocal loss and VAE loss as the loss function of the panoramic segmentation model, under the condition of data category imbalance, the panoramic segmentation model focuses on categories with a number below the threshold to suppress the imbalance of samples.

[0074] Among them, the specific expression of DiceFocal loss is:

[0075] FL(P t )=-α t (1-P t ) γ log(P t )

[0076]

[0077] DFL=FL+DL

[0078] Among them, x is the prediction result, y is the label, FL is the Focal Loss loss function, DL is the Dice Loss loss function, DFL is the DiceFocal loss loss function, P t is the predicted probability of the true category of the sample, γ is the hyperparameter for adjusting the weight factor, α t For each weight.

[0079] The specific expression of VAE loss is:

[0080]

[0081] Among them, MSE is the reconstruction loss, KL is the divergence regularization term, and VL is the VAE loss function.

[0082] Step S5: using the panoramic segmentation model to segment the image to be segmented with the intracavitary material, to obtain the cavity organ image after segmenting the intracavitary material.

[0083] After the image to be segmented with intracavitary material is segmented using the panoramic segmentation model, data enhancement is performed on the segmentation result. The data enhancement includes obtaining the maximum connected component and filling holes.

[0084] Example:

[0085] The present invention provides a method for segmenting a cavity organ image, using bladder image data as a segmentation object, comprising:

[0086] The overall framework of the present invention is as follows Figure 1 shown.

[0087] The dataset used for bladder cancer segmentation is usually a mixture of high-quality and low-quality labels, so it is necessary to screen the dataset to eliminate the incorrect annotations that radiologists may make. The present invention uses U-net to segment the image into urine and bladder. The segmentation result is projected into polar coordinates, and the urine area is subtracted from the bladder area to obtain the bladder wall area a priori. The obtained bladder wall area is then compared with the area of ​​the label projected into the polar coordinates, and the threshold discrimination method is used to eliminate the label datasets that differ too much, such as Figure 2 shown.

[0088] Secondly, for the remaining high-quality data set after filtering, we cropped it into images of 256*256 in size. Then we introduced the Bootstrap Your Own Latent (BYOL) architecture for self-supervised training, added noise to random areas of the image, and used a self-supervised learning method based on random crops for training, as shown in the manual. Figure 3 As shown in the figure, the model is allowed to predict the original image, which strengthens the model's perception of details and improves the network's ability to recognize the bladder boundary. The training results are shown in Figure 6 shown.

[0089] On this basis, if Figure 4 As shown, the weights obtained from self-supervised learning are transferred, images with high-quality labels are used as input, and a panoramic segmentation model is constructed to further reduce the difficulty of segmenting the bladder wall and tumor. In addition, the present invention uses DiceFocal loss.

[0090] In the Focal loss formula, α for urine is set to 0.1, α for bladder wall is set to 0.6, and α for tumor is set to 0.8; in the Dice loss formula, x is the predicted result and y is the label. Under the condition of imbalanced data categories, this loss function will focus the model's attention on rare categories, thereby suppressing sample imbalance.

[0091] At the same time, the present invention also uses VAE Loss, including reconstruction loss and kl divergence regularization term, which correspond to the two goals that the model training process hopes to achieve. The specific formula is:

[0092] The reconstruction loss (i.e., reconstruction loss) calculates the MSE loss between the decoded vector and the input vector, which reflects the difference between the result generated by VAE and the input. The corresponding goal is to make the result generated by VAE as similar to the input as possible. The kl divergence regularization term is more complicated to understand than the MSE loss. It makes the latent variables generated by the encoder conform to the standard normal distribution as much as possible.

[0093] In supervised training, the present invention adopts the SegNet segmentation network, which is mainly composed of two parts: encoder and decoder. The encoder is a network model that follows VGG16 and mainly parses object information. The decoder converts the parsed information into the final image form, that is, each pixel is represented by the color corresponding to its object information.

[0094] Each encoder layer in VGG16 has a decoder layer, so the decoder also has 13 convolutional layers. The encoder process includes convolution operations (convolution + batch normalization + ReLU) and maximum pooling operations; the decoder process includes upsampling operations and softmax classifier operations. With the help of this segmentation network, boundaries can be effectively depicted.

[0095] The model obtained by supervised training is transferred to post-processing, the data set after data augmentation (TTA) is segmented, and the results are subjected to data augmentation processing, including obtaining the maximum connected component and filling holes, to achieve better segmentation effect.

[0096] Finally, the data obtained by the network model constructed by the present invention is analyzed, such as Figure 6 (a)-(l), the performance comparison report of the present invention and other models is as follows Figure 7 (a) and (b) are sufficient to prove the superiority and effectiveness of the present invention in the segmentation of hollow organ images.

[0097] Based on the above methods and statements, the present invention provides a system for cavity organ image segmentation, including: an acquisition module, a cavity wall acquisition module, a weight acquisition module, a task construction module and a segmentation module.

[0098] Among them, the acquisition module is used for a cavity organ imaging dataset with intracavitary substances; wherein the images in the cavity organ imaging dataset have labels for distinguishing the cavity wall and the cavity substance; the cavity wall acquisition module is used to use the U-net algorithm to segment the image into the cavity and the substance in the cavity, and project the cavity and the substance into polar coordinates, obtain the cavity wall area by the difference between the cavity polar coordinates and the substance polar coordinates, and remove the image with label error greater than the error threshold through the obtained cavity wall area; the weight acquisition module is used to use the BYOL architecture of self-supervised training to perform contrast and reconstruction loss training on the cavity wall area to obtain the optimized cavity wall area and the weight of the cavity wall area; the task construction module is used to transfer the weights to the deep learning model, and input the images with label errors less than the error threshold into the deep learning model for supervised training, and construct a panoramic segmentation model of the image through the training results; the segmentation module is used to use the panoramic segmentation model to segment the image to be segmented with intracavitary substances, and obtain the cavity organ image after the intracavitary substances are segmented.

[0099] The present invention also provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of a method for segmenting a cavity organ image.

[0100] The present invention also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of a method for segmenting a cavity organ image are implemented.

[0101] The above content is a further detailed description of the present invention in combination with a specific preferred embodiment. For technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for segmenting a cavity organ image, characterized in that: The steps include: Collecting a dataset of images of hollow organs with intracavitary substances; wherein the images in the dataset of images of hollow organs have labels for distinguishing between the cavity wall and the intracavitary substances; Using a U-net algorithm to segment the image into a cavity and a substance in the cavity, and projecting the cavity and the substance into polar coordinates, obtaining a cavity wall region by the difference between the cavity polar coordinates and the substance polar coordinates, and removing images with label errors greater than an error threshold by using the obtained cavity wall region; The BYOL architecture of self-supervised training is adopted to perform contrast loss and reconstruction loss training on the cavity wall region to obtain an optimized cavity wall region and a weight of the cavity wall region; The weights are transferred to a deep learning model, and images whose label errors are less than an error threshold are input into the deep learning model for supervised training, and a panoramic segmentation model of the image is constructed based on the training results; Using the panoramic segmentation model to segment the image to be segmented with the intracavitary material, to obtain the cavity organ image after segmenting the intracavitary material; The weights are transferred to the deep learning model, and images whose label errors are less than the error threshold are input into the deep learning model for supervised training, and a panoramic segmentation model of the image is constructed based on the training results, including the following steps: Use the deep learning model to batch normalize input images whose label error is less than the error threshold; Introducing a leaky relu layer into the deep learning model to eliminate the neuron death scenario that occurs when the transfer weight is negative; Among them, death means that the input weight cannot be updated; Increasing the number of filters in each convolutional layer in the deep learning model, and performing convolution operations on the input image through the leaky relu layer and the transfer weight input convolutional layer to capture the characteristics of the data; The deep learning model is evaluated using independent data, and is optimized according to the model evaluation results to obtain a panoramic segmentation model; wherein the optimization includes adjusting model parameters, loss function, and learning rate.

2. The method for segmenting a cavity organ image according to claim 1, characterized in that: The method of eliminating images whose label errors are greater than an error threshold from the obtained cavity wall region comprises the following steps: Comparing the area of ​​the cavity wall region with the area of ​​the label in the image data set projected onto polar coordinates; The threshold discrimination method is used for the compared area to eliminate labels with a difference greater than the error threshold and the image to which the label belongs.

3. The method for segmenting a cavity organ image according to claim 1, characterized in that: The deep learning model is used to batch normalize input images whose label errors are less than the error threshold. The specific expression is: AND i =γ×X i ′+β Among them, ε is a constant, γ and β are learned and optimized according to the training data, μ is the mean, σ is the standard deviation, Xi′ is the normalized result, and Xi is the input data set {X1,X2,...,X m }The i-th input data in .

4. The method for segmenting a cavity organ image according to claim 3, characterized in that: DiceFocalloss and VAE loss are used as the loss functions of the panoramic segmentation model. Under the condition of data category imbalance, the panoramic segmentation model focuses on categories whose number is lower than a threshold to suppress the imbalance of samples.

5. The method for segmenting a cavity organ image according to claim 4, characterized in that: The specific expression of the DiceFocalloss is: FL(P t )=-a t (1-P t ) γ log(P t ) DFL=FL+DL Among them, x is the prediction result, y is the label, FL is the Focal Loss loss function, DL is the DiceLoss loss function, DFL is the DiceFocal loss loss function, P t is the predicted probability of the true category of the sample, γ is the hyperparameter for adjusting the weight factor, α t are the weights of various types of cavity wall areas.

6. The method for segmenting a cavity organ image according to claim 4, characterized in that: The specific expression of the VAE loss is: VL=MSE(X,X′)+KL(N(μ1,σ1 2 ),N(0,1)) Among them, MSE is the reconstruction loss, KL is the divergence regularization term, and VL is the VAE loss function.

7. The method for segmenting a cavity organ image according to claim 1, characterized in that: When the BYOL architecture for self-supervised training is introduced, noise is added to the images filtered out during the data screening process, and self-supervised learning is used to predict the original image.

8. The method for segmenting a cavity organ image according to claim 1, characterized in that: After the image to be segmented with the intracavitary material is segmented by using the panoramic segmentation model, data enhancement is performed on the segmentation result, and the data enhancement includes obtaining the maximum connected component and filling the hole processing.

9. A system for segmenting images of hollow organs, characterized in that: include: An acquisition module, for collecting a data set of images of hollow organs with intracavitary substances; wherein the images in the data set of images of the hollow organs have labels for distinguishing between the cavity wall and the intracavitary substances; a cavity wall acquisition module, for segmenting the image into a cavity and a substance in the cavity by using a U-net algorithm, and projecting the cavity and the substance into polar coordinates, obtaining a cavity wall region by the difference between the cavity polar coordinates and the substance polar coordinates, and removing images whose label errors are greater than an error threshold by using the obtained cavity wall region; A weight acquisition module, for adopting a BYOL architecture of self-supervised training, performing contrast loss and reconstruction loss training on the cavity wall region, and obtaining an optimized cavity wall region and a weight of the cavity wall region; A task construction module, used to transfer the weights to a deep learning model, and input images whose label errors are less than an error threshold into the deep learning model for supervised training, and construct a panoramic segmentation model of the image through the training results; A segmentation module, used to segment the image to be segmented with the intracavitary material using the panoramic segmentation model, to obtain the image of the cavity organ after the intracavitary material is segmented; The task construction module transfers the weights to the deep learning model, and inputs images whose label errors are less than the error threshold into the deep learning model for supervised training, and constructs a panoramic segmentation model of the image through the training results, including the following steps: Use the deep learning model to batch normalize input images whose label error is less than the error threshold; Introducing a leaky relu layer into the deep learning model to eliminate the neuron death scenario that occurs when the transfer weight is negative; Among them, death means that the input weight cannot be updated; Increasing the number of filters in each convolutional layer in the deep learning model, and performing convolution operations on the input image through the leaky relu layer and the transfer weight input convolutional layer to capture the characteristics of the data; The deep learning model is evaluated using independent data, and is optimized according to the model evaluation results to obtain a panoramic segmentation model; wherein the optimization includes adjusting model parameters, loss function, and learning rate.

Images