WO2020143309A1 - Segmentation model training method, oct image segmentation method and apparatus, device and medium - Google Patents

Abstract

A segmentation model training method, an OCT image segmentation method and apparatus, a device and a medium. The segmentation model training method comprises: acquiring a training sample image set, wherein the training sample image set comprises an original OCT image and a gold standard image (S10); inputting the original OCT image into a pre-set generator model for segmentation processing to obtain a first segmented image (S20); comparing the first segmented image with the gold standard image by means of a pre-set discriminator model to obtain a comparison result, calculating a loss function of the generator model according to the comparison result, and updating the generator model according to the loss function (S30); using the updated generator model to convert the first segmented image into a second segmented image, inputting the second segmented image and the gold standard image into the pre-set discriminator model, and updating the discriminator model according to a binary cross entropy to obtain an updated discriminator model (S40); and carrying out iterative training on the updated generator model and the updated discriminator model until a loss function of the updated discriminator model converges, then stopping the iterative training, and determining the updated generator model after the iterative training is stopped as an image lesion segmentation model (S50). The segmentation model training method improves the performance of a segmentation model by means of adversarial training, and improves the accuracy of the segmentation model.

Description

Segmentation model training method, OCT image segmentation method, device, equipment and medium

This application is based on the Chinese invention patent application with the application number 201910019566.1 filed on January 9, 2019 and titled "Segmentation Model Training Method, OCT Image Segmentation Method, Device, Equipment, and Medium", and claims its priority.

Technical field

The present application relates to the field of image detection, in particular to a segmentation model training method, an OCT image segmentation method, device, equipment, and medium.

Background technique

"OCT image" is short for optical coherence tomography (Optical Coherence Tomography). It mainly uses the basic principle of weak coherent light interferometer to detect the reflection and scattering signals of biological tissue at different depths of the incident weak coherent light to reconstruct biological tissue The internal structure image is a non-contact, non-invasive tomography of biological tissue. OCT imaging equipment is used in ophthalmology to assist doctors in observing the normal tissue structure of the posterior segment of the eye (such as the macula, optic disc, or retinal nerve fiber layer, etc.) and pathological changes. In order to provide a more accurate imaging basis for the diagnosis of related ophthalmic diseases, OCT The images are segmented to provide technical assistance for related medical procedures.

Traditionally, segmentation technologies include histogram-based, boundary-based segmentation, or region-based segmentation technologies, most of which are based on image processing algorithms. However, due to factors such as more noise in the OCT image, different lesion sizes and irregular border contours, the traditional segmentation technology often fails to take into account the integrity and accuracy of the lesions in the OCT image, that is, some lesion details or may be missing The extraneous and useless information leads to unsatisfactory segmentation results.

Summary of the invention

Embodiments of the present application provide a segmentation model training method, device, computer equipment, and storage medium to solve the problem of low segmentation model segmentation performance.

In addition, the embodiments of the present application provide an OCT image segmentation method, device, computer equipment, and storage medium to solve the problem of low lesion segmentation accuracy.

A segmentation model training method, including:

Acquiring a training sample image set, the training sample image set including the original OCT image and the gold standard image;

Input the original OCT image into a preset generator model for segmentation processing to obtain a first segmented image;

Comparing the first segmented image with the gold standard image through a preset discriminator model to obtain a comparison result, calculating a loss function of the generator model according to the comparison result, and according to the loss The function updates the generator model;

Use the updated generator model to convert the first segmented image into a second segmented image, input the second segmented image and the gold standard image into a preset discriminator model, and update according to the binary cross entropy The preset discriminator model to obtain an updated discriminator model;

Iteratively training the updated generator model and the updated discriminator model until the loss function of the updated discriminator model converges, then iterative training is stopped, and the updated after the iterative training is stopped The generator model is determined as the image lesion segmentation model.

A segmentation model training device, including:

A sample image set acquisition module, for acquiring a training sample image set, the training sample image set including the original OCT image and the gold standard image;

A segmented image acquisition module, configured to input the original OCT image into a preset generator model for segmentation processing to obtain a first segmented image;

The generator update module is used to compare the first segmented image with the gold standard image through a preset discriminator model to obtain a comparison result, and calculate the loss of the generator model according to the comparison result Function, and update the generator model according to the loss function;

The discriminator update module is used to convert the first segmented image into a second segmented image using the updated generator model, and input the second segmented image and the gold standard image into a preset discriminator model , Updating the preset discriminator model according to the binary cross entropy to obtain the updated discriminator model;

The lesion segmentation model training module is used to iteratively train the updated generator model and the updated discriminator model until the loss function of the updated discriminator model converges, then iterative training is stopped and will be stopped. The updated generator model after iterative training is determined to be an image lesion segmentation model.

An OCT image segmentation method, including:

Obtain the OCT image to be processed;

The OCT image to be processed is input into an image lesion segmentation model for segmentation to obtain a lesion image, wherein the image lesion segmentation model is obtained by training using a segmentation model training method.

An OCT image segmentation device, including:

To-be-processed image acquisition module for acquiring to-be-processed OCT images;

The lesion image acquisition module is configured to input the to-be-processed OCT image into an image lesion segmentation model for segmentation to obtain a lesion image, wherein the image lesion segmentation model is obtained by training using a segmentation model training method.

A computer device includes a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, and the processor implements the computer-readable instructions to implement the above segmentation model training method Or, the processor implements the above-mentioned OCT image segmentation method when executing the computer-readable instructions.

One or more readable storage media storing computer-readable instructions, which when executed by one or more processors, cause the one or more processors to perform the above-mentioned OCT image segmentation method.

The details of one or more embodiments of the present application are set forth in the following drawings and description, and other features and advantages of the present application will become apparent from the description, drawings, and claims.

BRIEF DESCRIPTION

In order to more clearly explain the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings required in the description of the embodiments of the present application. Obviously, the drawings in the following description are only some embodiments of the present application For those of ordinary skill in the art, without paying creative labor, other drawings can also be obtained based on these drawings.

1 is a schematic diagram of an application environment of a segmentation model training method or an OCT image segmentation method provided by an embodiment of the present application;

2 is an example diagram of a method for training a segmentation model provided by an embodiment of the present application;

FIG. 3 is another example diagram of a segmentation model training method provided by an embodiment of the present application;

4 is a schematic block diagram of a segmentation model training device provided by an embodiment of the present application;

FIG. 5 is another schematic block diagram of a segmentation model training device provided by an embodiment of the present application;

6 is an example diagram of an OCT image segmentation method provided by an embodiment of the present application;

7 is another example diagram of an OCT image segmentation method provided by an embodiment of the present application;

8 is a schematic block diagram of an OCT image segmentation device provided by an embodiment of the present application;

9 is a schematic diagram of a computer device provided by an embodiment of the present application.

detailed description

The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in this application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of protection of this application.

The segmentation model training method provided in this application can be applied in the application environment as shown in FIG. 1, in which the client communicates with the server through the network, the server receives the training sample image set sent by the client, and then the training sample image The original OCT image is input into the preset generator model for conversion processing to obtain the first segmented image; the first segmented image is compared with the gold standard image through the preset discriminator model to obtain the comparison result, according to the comparison Calculate the loss function of the generator model for the result, and update the generator model according to the loss function; then, use the updated generator model to convert the first segmented image to the second segmented image, and input the second segmented image and the gold standard image To the preset discriminator model, update the preset discriminator model according to the binary cross entropy to obtain the updated discriminator model; finally, iteratively train the updated generator model model and the updated discriminator model Until the loss function of the updated discriminator model converges, then iterative training is stopped, and the updated discriminator model after the iterative training is stopped is determined as the image lesion segmentation model. Among them, the client may be, but not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices. The server can be implemented with an independent server or a server cluster composed of multiple servers.

In one embodiment, as shown in FIG. 2, the method is applied to the server in FIG. 1 as an example for illustration, including the following steps:

S10: Obtain the training sample image set, which includes the original OCT image and the gold standard image.

Among them, the training sample image set is a set of sample images used for deep learning, including the original OCT image and the gold standard image. The original OCT image refers to an unprocessed OCT image, and the original OCT image can be acquired after being scanned by an OCT scanner. The gold standard image refers to the pre-segmented lesion image. For example, the expert outlines the location of the lesion to be segmented from the unprocessed OCT image based on professional medical knowledge, that is, the gold standard image is pre-marked for the lesion. The gold standard image can be obtained by annotating the location of each lesion in the unprocessed OCT image by experts. Optionally, by selecting a preset number of images from the public fundus retina data set (such as DRIVE or STARE) as the training sample image set. It should be noted that the size of the gold standard image is the same as the original OCT image, and the pixels of the lesion area are the preset pixel values, and the pixel value of the non-focal area is 0, so as to enhance the difference between the lesion area and the non-focal area of the OCT image. Differentiate.

S20: Input the original OCT image into a preset generator model for segmentation processing to obtain a first segmented image.

Among them, the generator model is a model for segmenting the image. The model can be a convolutional neural network model, such as a U-shaped convolutional neural network (U-Net). Specifically, the model can be trained by convolutional nerves in advance The network gets. The first segmented image refers to the result graph output by the preset generator model, that is, the image obtained by segmenting the image input to the generator model. The generator model includes a down-sampling stage and an up-sampling stage, where the down-sampling stage is composed of multiple convolutional layers and pooling layers. The down-sampling stage is used for feature extraction of the image input to the generator model. The sampling stage is composed of multiple deconvolution layers. The up-sampling stage is used to gradually restore the image details. At the same time, a jump structure is added between feature layers of the same resolution to achieve the segmentation of the target object in the image.

Specifically, the original OCT image is input into the generator model, and images of different scales are formed by sampling, and connected to the corresponding scale convolutional layer of the downsampling stage. The output end of the generator model passes the output results of the descaled convolutional layers of different scales through After upsampling, stitching is performed to obtain the first segmented image. Understandably, since the preset generator model fully considers the effect of different scale images on image segmentation, the performance of the convolutional neural network model is improved, and thus the accuracy of the original OCT image segmentation is improved.

S30: Compare the first segmented image with the gold standard image through a preset discriminator model to obtain a comparison result, calculate a loss function of the generator model according to the comparison result, and update the generator model according to the loss function.

Among them, the discriminator model is a classification network used to judge whether the image output by the generator model is consistent with the marked gold standard image. The discriminator model includes a convolutional layer, a ReLU (Rectified Linear Unit) activation layer, and a batch normalization layer (Batch Normalization). In each convolution layer, a nonlinear activation function is used to classify the obtained output to realize the image. Comparison. The comparison result is used to reflect how close the first divided image is to the gold standard image. The loss function is used to measure the degree of inconsistency between the predicted value and the true value of the model. It is a non-negative real value function. The smaller the value of the loss function, the higher the accuracy of the discriminator model.

Specifically, the first segmented image and the gold standard image are multiplied with the original OCT image respectively to obtain the processed segmented image and the processed gold standard image; then the preset number of neural network modules of different scales are used for the processing The segmented image and the processed gold standard image are characterized to obtain feature maps of different scales; because the discriminator model includes convolution layers, ReLU activation layers, and batch normalization layers, and the discriminator model uses pooling Layers, which makes the feature maps of different scales change from large to small. Finally, the feature maps of different scales are converted into fully connected layers by splicing operation, and connected to a single neuron as the final output layer, and the comparison result is obtained. Optionally, the comparison result takes a value between 0-1, if the value result is 1, it is determined that the first segmented image output by the generator model is completely consistent with the gold standard image, and if the comparison result is 0, it is determined The first split image and the gold standard image are completely inconsistent. Preferably, in the embodiment of the present application, if the comparison result is greater than 0.5, the first divided image and the gold standard image are determined to be the same; if the comparison result is less than or equal to 0.5, the first divided image and the gold standard image are determined. Inconsistent, it is necessary to update the loss function at this time, and update the generator model according to the loss function.

It should be noted that the discriminator model in this step adds convolutional layers of different scales during the classification and discrimination process. Due to the full consideration of the spatial dependence of image pixels on long and short distances, that is, large scale Images represent short-distance spatial dependencies, and small-scale images represent long-distance spatial dependencies, thereby improving the performance of the discriminator model.

S40: Use the updated generator model to convert the first segmented image into the second segmented image, input the second segmented image and the gold standard image into the preset discriminator model, and update the preset discriminant based on the binary cross entropy Model to get the updated discriminator model.

The second divided image refers to a result graph output by the updated generator model, that is, an image obtained by dividing the first divided image input at the input end. Binary Cross Entropy is a way to measure the difference between the predicted value and the actual value of the discriminator model. Specifically, a second result image is generated according to the updated generator model, the second segmented image and the gold standard image are input into a preset discriminator model, and the second segmented image and the gold standard image are discriminated according to The discriminator model outputs the binary cross-entropy update discriminator model, that is, the loss function of the preset discriminator model is defined in advance, and the loss function loss includes two parts, one part is the segmentation network (updated generator model ) loss function loss _1, a portion of the network is classified (predetermined discriminant model) loss function loss _2, both the weighted sum, i.e. _{loss = λ 1 * loss 1 +} λ 2 * loss 2, wherein, λ ₁ And λ ₂ are the weights of loss ₁ and loss ₂ , respectively, and loss is the loss function of the entire network. The preset discriminator model is updated according to the calculation result of the binary cross entropy of the loss function, and the updated discriminator model is obtained. The calculation formula of Binary Cross Entropy is as follows:

In the formula,

It is expressed as the probability that the lesion is correctly segmented in the second segmented image, y _j is the probability that the lesion segmentation of the preset second segmented image is consistent with the gold standard image, and J(y) is an expression of binary cross entropy.

S50: Iteratively train the updated generator model and the updated discriminator model until the loss function of the updated discriminator model converges, then stop iterative training, and stop the updated generator model after iterative training Determined as the image lesion segmentation model.

Among them, iterative training is a model training method in deep learning, which is used to optimize the model. The implementation process of iterative training in this step is as follows: firstly build the objective loss function of the generator model and discriminator model, and use the optimization algorithm for loop training, such as the optimization algorithm SGD (stochastic gradient descent); in each loop During the training process, all training samples are read in sequence and the current loss function of the discriminator model is calculated. Based on the optimization algorithm, the gradient descent direction is determined, so that the target loss function is gradually reduced and reaches a stable state, and the parameters of the constructed network model are realized. Optimization.

Among them, the loss function convergence means that the loss function is close to 0, for example, less than 0.1, that is, the value of the discriminator model output for a given sample (positive sample or negative sample) is close to 0.5, then the discriminator is unable to distinguish between positive and negative samples, and That is, the output of the discriminator is converged, that is, the training is stopped, and the model parameters of the last training are used as the parameters of the generator model, and then the lesion segmentation model is obtained.

Specifically, by inputting the original OCT image x into a preset generator model (G), the preset generator model learns the mapping relationship between the original OCT image x and the gold standard image y, that is, G: x- >y, output the original OCT image after segmentation, and the discriminator model (D) learns the distribution difference between the input image pair {x,y} and {G(x,y)}, so as to evaluate the updated generator model The update is performed until the parameters of the generator model reach the optimum, that is, the loss function of the updated discriminator model converges, and then the updated generator model is determined as the image lesion segmentation model. Understandably, during the model training process, the accuracy of the model segmentation is improved, and no additional post-processing steps are required, and an end-to-end OCT image lesion segmentation algorithm is implemented, thereby improving the accuracy of the lesion segmentation model.

In this embodiment, first, the training sample image set is obtained, and the training sample image set includes the original OCT image and the gold standard image; then, the original OCT image is input into a preset generator model for segmentation processing to obtain a first segmented image , Because the preset generator model fully considers the impact of different scale images on image segmentation, which improves the performance of the convolutional neural network model, and thus improves the accuracy of the original OCT image segmentation; then, through the preset discrimination The generator model compares the first segmented image with the gold standard image to obtain a comparison result, calculates the loss function of the generator model according to the comparison result, and updates the generator model according to the loss function; next, the updated The generator model converts the first segmented image into the second segmented image, inputs the second segmented image and the gold standard image into the preset discriminator model, and updates the preset discriminator model according to the binary cross entropy, after the update Discriminator model; finally, iteratively train the updated generator model and the updated discriminator model until the loss function of the updated discriminator model converges, then iterative training is stopped, and the update after iterative training is stopped After the generator model is determined as an image lesion segmentation model, the accuracy of the model segmentation is improved, and no additional post-processing steps are required, and an end-to-end OCT image lesion segmentation algorithm is implemented, thereby improving the accuracy of the lesion segmentation model .

In an embodiment, as shown in FIG. 3, in step S20, the original OCT image is input into a preset generator model for conversion processing to obtain a first segmented image, which specifically includes the following steps:

S21: Input the original OCT image into a set of down-sampling blocks of a preset generator model to obtain a feature map corresponding to the original OCT image, where the down-sampling block set is composed of N down-sampling blocks connected in sequence, and N is a positive integer.

Among them, the down-sampling block is the first convolution layer in the preset generator model, which is used to extract the basic features (such as edges, textures, etc.) of the original OCT image through convolution, because the set of down-sampling blocks is composed of N The down-sampling blocks are connected in sequence. Therefore, the extracted N basic features are fused to obtain the feature map corresponding to the original OCT image. N is a positive integer, and the size of N can be selected according to actual needs, such as N=5.

S22: Input the feature map into the abstract arrangement block to obtain the first segmented image, where the abstract arrangement block is composed of M arrangement combination units connected in sequence, M is a positive integer.

The permutation and combination unit refers to the second convolutional layer in the preset generator model. The abstract permutation block is composed of M permutation and combination units connected in sequence, and is used to permutate and combine feature maps through convolution operations to obtain more It is abstract and has the characteristics of semantic information, so as to obtain a more accurate first segmented image. M is a positive integer, and the size of M can be selected according to actual needs, such as M=4. At the same time, the activation layer in the preset generator model can increase the nonlinearity of the convolutional neural network, which is conducive to the convergence of the convolutional neural network. The activation layer can use rectified linear unit, sigmoid function, etc. as the activation function. Preferably, the activation layer may use a rectified linear unit as an activation function to accelerate the convergence speed of the convolutional neural network. The pooling layer is used to reduce the length and width of the input feature map, reducing the connection parameters and calculation amount of the preset generator model, in order to comply with displacement invariance and obtain more global information; due to the shrinking of the pooling layer A constant-size filter is used on the graph, so the relative local receptive field of each neuron will become larger, so that each neuron of the next convolutional layer can extract more global features, thus making the first A segmented image is more accurate and enhances the sensitivity of segmentation.

In this embodiment, the original OCT image is input into the set of down-sampled blocks of the preset generator model to obtain the feature map corresponding to the original OCT image, and the feature map is input into the abstract arrangement block to obtain the first segmented image, so that The first segmented image is more accurate and enhances the sensitivity of image segmentation.

In an embodiment, in step S50, iteratively training the updated generator model and the updated discriminator model, including:

Use the updated discriminator model to reverse adjust the updated generator model.

Among them, back regulation is a training method for back propagation of model parameters. Specifically, after determining the network structure of the updated discriminator model and the updated generator model, the network is trained. In the course of several training iterations, the weights and deviations of the updated generator model and the updated discriminator model are trained by back propagation. The updated discriminator model learns to find the real lesion image from the training samples. At the same time, the updated generator model feedback learns how to generate an image that is close to the gold standard image, preventing it from being recognized by the updated discriminator model. Finally, the optimal updated generator model and updated discriminator model are obtained, so as to realize the segmentation of the OCT image, so as to subsequently improve the accuracy of the model for image segmentation.

In this embodiment, the updated discriminator model is used to reversely adjust the updated generator model to realize the segmentation of the OCT image, so as to subsequently improve the accuracy of the model in segmenting the image.

It should be understood that the size of the sequence numbers of the steps in the above embodiments does not mean the order of execution, and the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

In one embodiment, a segmentation model training device is provided, and the segmentation model training device corresponds to the segmentation model training method in the above embodiment in one-to-one correspondence. As shown in FIG. 4, the segmentation model training device includes a sample image set acquisition module 10, a segmentation image acquisition module 20, a generator update module 30, a discriminator update module 40, and a lesion segmentation model training module 50. The detailed description of each functional module is as follows:

The sample image set acquisition module 10 is used to obtain a training sample image set, the training sample image set includes the original OCT image and the gold standard image;

The segmented image acquisition module 20 is used to input the original OCT image into a preset generator model for segmentation processing to obtain a first segmented image;

The generator update module 30 is used to compare the first segmented image with the gold standard image through the preset discriminator model to obtain the comparison result, calculate the loss function of the generator model according to the comparison result, and update according to the loss function Generator model;

The discriminator update module 40 is used to convert the first segmented image to the second segmented image using the updated generator model, input the second segmented image and the gold standard image into the preset discriminator model, and cross according to the binary value Entropy updates the preset discriminator model to obtain the updated discriminator model;

The lesion segmentation model training module 50 is used to iteratively train the updated generator model and the updated discriminator model until the loss function of the updated discriminator model converges, then iterative training is stopped, and after iterative training is stopped The updated generator model is determined as the image lesion segmentation model.

Preferably, as shown in FIG. 5, the divided image acquisition module 20 includes a feature map acquisition unit 21 and a divided image acquisition unit 22.

The feature map acquisition unit 21 is used to input the original OCT image into a set of down-sampling blocks of a preset generator model to obtain a feature map corresponding to the original OCT image, where the down-sampling block set is composed of N down-sampling blocks connected in sequence , N is a positive integer;

The divided image acquisition unit 22 is configured to input a feature map into an abstract arrangement block to obtain a first divided image, where the abstract arrangement block is composed of M arrangement combination units connected in sequence, where M is a positive integer.

Preferably, the lesion segmentation model training module includes an iterative training unit for reversely adjusting the updated generator model using the updated discriminator model.

In one embodiment, an OCT image segmentation method is provided. The OCT image segmentation method can also be applied in the application environment as shown in FIG. 1, in which the client communicates with the server through the network. The server receives the to-be-processed OCT image sent by the client, and then inputs the to-be-processed OCT image into an image lesion segmentation model for segmentation to obtain a lesion image. Among them, the client may be, but not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices. The server can be implemented with an independent server or a server cluster composed of multiple servers.

In one embodiment, as shown in FIG. 6, the method is applied to the server in FIG. 1 as an example for illustration, including the following steps:

S60: Obtain the OCT image to be processed.

Among them, the OCT image to be processed refers to the OCT image that needs to be segmented. The OCT image to be processed can be obtained from the database of the client by the server, or directly obtained from the system database of the client, or can be obtained by the client The third-party image acquisition tool at the terminal obtains the OCT image to be processed from the system data interface.

S70: Input the to-be-processed OCT image into an image lesion segmentation model for segmentation to obtain a lesion image, wherein the image lesion segmentation model is obtained by training using a segmentation model training method.

Specifically, the OCT image to be processed is input into the image lesion segmentation model, and the output of the lesion segmentation model is the lesion image. Understandably, since the training method of the lesion image segmentation model has higher segmentation accuracy, the accuracy of the lesion image output by the image lesion segmentation model can be improved.

In this embodiment, first, the to-be-processed OCT image is obtained; then, the to-be-processed OCT image is input into an image lesion segmentation model for segmentation to obtain a lesion image. Because the lesion image segmentation model training method has a higher segmentation accuracy, which makes The accuracy of the lesion image output by the image lesion segmentation model is improved.

In an embodiment, as shown in FIG. 7, after inputting the OCT image to be processed into an image lesion segmentation model for segmentation to obtain a lesion image, the OCT image segmentation method further includes:

S81: Calculate the area of each lesion image to obtain the area of the lesion area.

The area of the lesion area refers to the area of the area where the lesion is located in the lesion image. Specifically, the area of the lesion area can be calculated according to the position parameter of the lesion. For example, an image of a diseased area is circular and the radius is 1.5 mm, and the area of the imaged area of the diseased image is 7.07 mm ² .

S82: Use the weighted sum calculation for the area of each lesion area to obtain the lesion area.

Among them, weighted summation refers to a calculation method in which each parameter is given a corresponding weight, and then the parameter and the weight are multiplied and added. Understandably, the impact of lesions in different parts is different. Therefore, the weighted sum calculation is used for the area of each lesion area to obtain the lesion area, which makes the calculation of the lesion area more accurate, so as to provide a reference for the subsequent assessment of the disease condition according to the lesion area.

In this embodiment, first, the area of each lesion image is calculated to obtain the area of the lesion area; then, the area of each lesion area is calculated by weighted sum to obtain the area of the lesion, so that the calculation of the area of the lesion is more accurate for subsequent Provide a reference for disease evaluation according to the area of the lesion.

In an embodiment, an OCT image segmentation device is provided, and the OCT image segmentation device corresponds to the OCT image segmentation method in the above embodiment in one-to-one correspondence. As shown in FIG. 8, the segmentation model training device includes a to-be-processed image acquisition module 60 and a lesion image acquisition module 70. The detailed description of each functional module is as follows:

The to-be-processed image acquisition module 60 is used to acquire to-be-processed OCT images;

The lesion image acquisition module 70 is configured to input the to-be-processed OCT image into an image lesion segmentation model for segmentation to obtain a lesion image, wherein the image lesion segmentation model is obtained by training using a segmentation model training method.

Preferably, the OCT image segmentation device further includes an area area calculation module and a lesion area acquisition module.

The area area calculation module is used to calculate the area of each lesion image to obtain the area of the lesion area;

The focus area acquisition module is used to calculate the weighted sum of the area of each focus area to obtain the focus area.

For the specific definition of the OCT image segmentation device, please refer to the above definition of the OCT image segmentation method, which will not be repeated here. Each module in the above OCT image segmentation device may be implemented in whole or in part by software, hardware, and a combination thereof. The above modules may be embedded in the hardware form or independent of the processor in the computer device, or may be stored in the memory in the computer device in the form of software so that the processor can call and execute the operations corresponding to the above modules.

In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure may be as shown in FIG. 9. The computer device includes a processor, memory, network interface, and database connected by a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, computer-readable instructions, and a database. The internal memory provides an environment for the operation of the operating system and computer-readable instructions in the non-volatile storage medium. The database of the computer device is used to store data used by the OCT image segmentation method. The network interface of the computer device is used to communicate with external terminals through a network connection. When the computer-readable instructions are executed by the processor, a method for training a segmentation model is realized.

In one embodiment, a computer device is provided, which includes a memory, a processor, and computer-readable instructions stored on the memory and executable on the processor. The processor implements the computer-readable instructions to implement the The segmentation model training method, or the processor implements the computer-readable instructions to implement the OCT image segmentation method in the above embodiments.

In one embodiment, one or more readable storage media storing computer-readable instructions are provided, which when executed by one or more processors cause the one or more processors to execute The segmentation model training method in the above embodiment, or when the computer readable instructions are executed by one or more processors, causes the one or more processors to execute the OCT image segmentation method in the above embodiments. Wherein, the readable storage medium includes a non-volatile readable storage medium and a volatile readable storage medium.

A person of ordinary skill in the art may understand that all or part of the process in the method of the foregoing embodiments may be completed by instructing relevant hardware through computer-readable instructions, and the computer-readable instructions may be stored in a non-volatile computer In a readable storage medium, when the computer-readable instructions are executed, they may include the processes of the foregoing method embodiments. Wherein, any reference to the memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and/or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

Those skilled in the art can clearly understand that, for convenience and conciseness of description, only the above-mentioned division of each functional unit and module is used as an example for illustration. In practical applications, the above-mentioned functions may be allocated by different functional units, Module completion means that the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

The above-mentioned embodiments are only used to illustrate the technical solutions of the present application, not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, persons of ordinary skill in the art should understand that they can still implement the foregoing The technical solutions described in the examples are modified, or some of the technical features are equivalently replaced; and these modifications or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present application. Within the scope of protection of this application.

Claims (20)

1. A segmentation model training method, characterized in that the segmentation model training method includes:

   Acquiring a training sample image set, the training sample image set including the original OCT image and the gold standard image;

   Input the original OCT image into a preset generator model for segmentation processing to obtain a first segmented image;

   Comparing the first segmented image with the gold standard image through a preset discriminator model to obtain a comparison result, calculating a loss function of the generator model according to the comparison result, and according to the loss The function updates the generator model;

   Use the updated generator model to convert the first segmented image into a second segmented image, input the second segmented image and the gold standard image into a preset discriminator model, and update according to the binary cross entropy The preset discriminator model to obtain an updated discriminator model;

   Iteratively training the updated generator model and the updated discriminator model until the loss function of the updated discriminator model converges, then iterative training is stopped, and the updated after the iterative training is stopped The generator model is determined as the image lesion segmentation model.
2. The method for training a segmentation model according to claim 1, wherein the inputting the original OCT image into a preset generator model for segmentation processing to obtain a first segmented image includes:

   Input the original OCT image into a set of down-sampled blocks of the preset generator model to obtain a feature map corresponding to the original OCT image, wherein the set of down-sampled blocks is composed of N down-sampled blocks connected in sequence , N is a positive integer;

   The feature map is input into an abstract arrangement block to obtain the first segmented image, wherein the abstract arrangement block is composed of M arrangement combination units connected in sequence, and M is a positive integer.
3. The method for split model training according to claim 1, wherein the iterative training of the updated generator model and the updated discriminator model includes:

   Use the updated discriminator model to reverse adjust the updated generator model.
4. An OCT image segmentation method, characterized in that the OCT image segmentation method includes:

   Obtain the OCT image to be processed;

   Input the to-be-processed OCT image into an image lesion segmentation model for segmentation to obtain a lesion image, wherein the image lesion segmentation model is obtained by training using the segmentation model training method according to any one of claims 1 to 3. of.
5. The OCT image segmentation method according to claim 4, wherein after the inputting of the OCT image to be processed into an image lesion segmentation model for segmentation to obtain a lesion image, the OCT image segmentation method further comprises:

   Calculating the area of each lesion image to obtain the area of the lesion area;

   The weighted sum is used to calculate the area of each lesion area to obtain the lesion area.
6. A segmentation model training device, characterized in that the segmentation model training device includes:

   A sample image set acquisition module, for acquiring a training sample image set, the training sample image set including the original OCT image and the gold standard image;

   A segmented image acquisition module, configured to input the original OCT image into a preset generator model for segmentation processing to obtain a first segmented image;

   The generator update module is used to compare the first segmented image with the gold standard image through a preset discriminator model to obtain a comparison result, and calculate the loss of the generator model according to the comparison result Function, and update the generator model according to the loss function;

   The discriminator update module is used to convert the first segmented image into a second segmented image using the updated generator model, and input the second segmented image and the gold standard image into a preset discriminator model , Updating the preset discriminator model according to the binary cross entropy to obtain the updated discriminator model;

   The lesion segmentation model training module is used to iteratively train the updated generator model and the updated discriminator model until the loss function of the updated discriminator model converges, then iterative training is stopped and will be stopped. The updated generator model after iterative training is determined to be an image lesion segmentation model.
7. The segmentation model training device according to claim 6, wherein the segmented image acquisition module includes:

   A feature map acquisition unit, configured to input the original OCT image into the set of down-sampled blocks of the preset generator model to obtain a feature map corresponding to the original OCT image, wherein the set of down-sampled blocks is composed of N The downsampling blocks are connected in sequence, N is a positive integer;

   A divided image acquisition unit is used to input the feature map into an abstract arrangement block to obtain the first divided image, wherein the abstract arrangement block is composed of M arrangement combination units connected in sequence, M is a positive integer.
8. The segmentation model training device according to claim 6, wherein the lesion segmentation model training module includes an iterative training unit for reversely adjusting the updated generator model using the updated discriminator model
9. An OCT image segmentation device, characterized in that the OCT image segmentation device includes:

   To-be-processed image acquisition module for acquiring to-be-processed OCT images;

   A lesion image acquisition module, configured to input the to-be-processed OCT image into an image lesion segmentation model for segmentation to obtain a lesion image, wherein the image lesion segmentation model adopts any one of claims 1 to 3 Partial model training method is obtained by training.
10. The OCT image segmentation device according to claim 9, wherein the OCT image segmentation device further comprises an area area calculation module and a lesion area acquisition module;

    The area area calculation module is used to calculate the area of each lesion image to obtain the area of the lesion area;

    The focal area acquisition module is used to calculate the weighted sum of the area of each focal area to obtain the focal area
11. A computer device, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, characterized in that, when the processor executes the computer-readable instructions, it is implemented as follows step:

    Acquiring a training sample image set, the training sample image set including the original OCT image and the gold standard image;

    Input the original OCT image into a preset generator model for segmentation processing to obtain a first segmented image;

    Comparing the first segmented image with the gold standard image through a preset discriminator model to obtain a comparison result, calculating a loss function of the generator model according to the comparison result, and according to the loss The function updates the generator model;

    Use the updated generator model to convert the first segmented image into a second segmented image, input the second segmented image and the gold standard image into a preset discriminator model, and update according to the binary cross entropy The preset discriminator model to obtain an updated discriminator model;

    Iteratively training the updated generator model and the updated discriminator model until the loss function of the updated discriminator model converges, then iterative training is stopped, and the updated after the iterative training is stopped The generator model is determined as the image lesion segmentation model.
12. The computer device according to claim 11, wherein the inputting the original OCT image into a preset generator model for segmentation processing to obtain a first segmented image includes:

    Input the original OCT image into a set of down-sampled blocks of the preset generator model to obtain a feature map corresponding to the original OCT image, wherein the set of down-sampled blocks is composed of N down-sampled blocks connected in sequence , N is a positive integer;

    The feature map is input into an abstract arrangement block to obtain the first segmented image, wherein the abstract arrangement block is composed of M arrangement combination units connected in sequence, and M is a positive integer.
13. The computer device according to claim 11, wherein the iterative training of the updated generator model and the updated discriminator model includes:

    Use the updated discriminator model to reverse adjust the updated generator model.
14. A computer device, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, characterized in that, when the processor executes the computer-readable instructions, it is implemented as follows step:

    Obtain the OCT image to be processed;

    Input the to-be-processed OCT image into an image lesion segmentation model for segmentation to obtain a lesion image, wherein the image lesion segmentation model is obtained by training using the segmentation model training method according to any one of claims 1 to 3. of.
15. The computer device of claim 14, wherein after the OCT image to be processed is input into an image lesion segmentation model for segmentation to obtain a lesion image, the processor executes the computer-readable instructions The following steps are also implemented:

    Calculating the area of each lesion image to obtain the area of the lesion area;

    The weighted sum is used to calculate the area of each lesion area to obtain the lesion area.
16. One or more readable storage media storing computer-readable instructions, which when executed by one or more processors, cause the one or more processors to perform the following steps:

    Acquiring a training sample image set, the training sample image set including the original OCT image and the gold standard image;

    Input the original OCT image into a preset generator model for segmentation processing to obtain a first segmented image;

    Comparing the first segmented image with the gold standard image through a preset discriminator model to obtain a comparison result, calculating a loss function of the generator model according to the comparison result, and according to the loss The function updates the generator model;

    Use the updated generator model to convert the first segmented image into a second segmented image, input the second segmented image and the gold standard image into a preset discriminator model, and update according to the binary cross entropy The preset discriminator model to obtain an updated discriminator model;

    Iteratively training the updated generator model and the updated discriminator model until the loss function of the updated discriminator model converges, then iterative training is stopped, and the updated after the iterative training is stopped The generator model is determined as the image lesion segmentation model.
17. The readable storage medium according to claim 16, wherein the inputting the original OCT image into a preset generator model for segmentation processing to obtain a first segmented image includes:

    Input the original OCT image into a set of down-sampled blocks of the preset generator model to obtain a feature map corresponding to the original OCT image, wherein the set of down-sampled blocks is composed of N down-sampled blocks connected in sequence , N is a positive integer;

    The feature map is input into an abstract arrangement block to obtain the first segmented image, wherein the abstract arrangement block is composed of M arrangement combination units connected in sequence, and M is a positive integer.
18. The readable storage medium of claim 16, wherein the iterative training of the updated generator model and the updated discriminator model includes:

    Use the updated discriminator model to reverse adjust the updated generator model.
19. One or more readable storage media storing computer-readable instructions, which when executed by one or more processors, cause the one or more processors to perform the following steps:

    Obtain the OCT image to be processed;

    Input the to-be-processed OCT image into an image lesion segmentation model for segmentation to obtain a lesion image, wherein the image lesion segmentation model is obtained by training using the segmentation model training method according to any one of claims 1 to 3. of.
20. The readable storage medium of claim 19, wherein after the OCT image to be processed is input into an image lesion segmentation model for segmentation to obtain a lesion image, the computer-readable instruction is When multiple processors execute, the one or more processors also execute the following steps:

    Calculating the area of each lesion image to obtain the area of the lesion area;

    The weighted sum is used to calculate the area of each lesion area to obtain the lesion area.

Images