Image segmentation method and system based on corrected diffusion model

Abstract

An image segmentation method and system based on a corrected diffusion model is provided that achieves further improvements in image segmentation accuracy. [Solution] The method involves preprocessing an image to be segmented, performing segmentation on the preprocessed image to be segmented, obtaining a primary segmentation result image, performing a difference operation between the correct label of the segmentation result and the primary segmentation result image to obtain a systematic error image, adding Gaussian noise to the systematic error image under Markov chain conditions, calculating the difference between the systematic error image with Gaussian noise added and the systematic error image to which no noise has been added using a loss function, iteratively correcting and reconstructing the systematic error image based on the calculation result of the loss function to obtain a reconstructed systematic error image, and performing an addition operation on the primary segmentation result image and the obtained reconstructed systematic error image to obtain a final segmentation result image.

Description

[Technical field]

[0001] The present invention relates to the field of image segmentation, and more particularly to an image segmentation method and system based on a corrected diffusion model. [Background technology]

[0002] The discussion in this section is merely intended to provide background information related to the present invention and does not necessarily constitute prior art.

[0003] Typically, image segmentation errors can be classified into two types: random errors and systematic errors. Random errors are due to random effects, such as image noise and unpredictable image artifacts, and can be reduced by fusing multiple information labels generated by independent segmentations. However, because random errors are unpredictable, they are very difficult to detect and correct. Systematic errors are errors due to system modes, i.e., the difference between the standard segmentation definition used by system developers to train segmentation models and manual segmentation protocols. Systematic errors are related to a specific model and therefore always occur under specific model conditions. Therefore, machine learning techniques can be used to effectively correct the systematic errors of a specific model.

[0004] In recent years, methods based on machine learning have been widely used in image segmentation. Direct image segmentation using convolutional neural networks or U-Net and its variant models is one of the most widely used techniques. Although traditional techniques can realize image segmentation based on machine learning, the accuracy of the final segmentation is still low due to the following problems:

[0005] Traditional machine learning-based segmentation models do not take into account the problem of segmentation error, that is, when performing image segmentation using a convolutional neural network or U-Net model, the accuracy of the image segmentation results may be affected by factors such as model characteristics, image noise, and artifacts.

[0006] Traditional segmentation models based on machine learning do not consider the correction of systematic errors. That is, although deep learning models such as convolutional neural networks and U-Net can improve the accuracy of image segmentation to a certain extent, most of them cannot automatically handle systematic errors, and they also do not consider how to correct systematic errors. Summary of the Invention

[0007] In order to overcome the above-mentioned shortcomings of the prior art, the present invention provides an image segmentation method and system based on a corrected diffusion model, in which the difference between a systematic error image to which Gaussian noise has been added and a systematic error image to which no noise has been added is calculated using a loss function in the corrected diffusion model, and the systematic error image is corrected and reconstructed to diffuse the systematic error, thereby further improving the image segmentation accuracy.

[0008] In order to achieve the above object, a first aspect of the present invention comprises: performing pre-processing on the segmentation target image to obtain a pre-processed segmentation target image; A step of performing segmentation on the preprocessed segmentation target image using a trained U-Net segmentation network to obtain a primary segmentation result image; A step of performing a difference calculation between the correct answer label of the segmentation result of the segmentation target image and the primary segmentation result image to obtain a systematic error image; adding Gaussian noise to the systematic error image under the condition of a Markov chain, calculating the difference between the systematic error image to which the Gaussian noise has been added and the systematic error image to which no noise has been added using a constructed loss function, and iteratively correcting and reconstructing the systematic error image based on the calculation result of the loss function to obtain a reconstructed systematic error image; performing an addition operation on the primary segmentation result image and the obtained reconstructed systematic error image to obtain a final segmentation result image; The present invention provides an image segmentation method based on a corrected diffusion model, comprising:

[0009] A second aspect of the present invention is an acquisition module for performing pre-processing on the segmentation target image to obtain a pre-processed segmentation target image; A primary segmentation module that performs segmentation on the preprocessed segmentation target image using a trained U-Net segmentation network to obtain a primary segmentation result image; a calculation module for performing a difference calculation between a correct answer label of the segmentation result of the segmentation target image and a primary segmentation result image to obtain a systematic error image; a reconstruction module for adding Gaussian noise to the systematic error image under a Markov chain condition, calculating a difference between the systematic error image to which the Gaussian noise has been added and a systematic error image to which no noise has been added using a constructed loss function, and iteratively correcting and reconstructing the systematic error image based on the calculation result of the loss function to obtain a reconstructed systematic error image; a segmentation module for performing an addition operation on the primary segmentation result image and the obtained reconstructed systematic error image to obtain a final segmentation result image; The present invention provides an image segmentation system based on a corrected diffusion model, comprising:

[0010] A third aspect of the present invention provides a computing device comprising a processor, a memory and a bus, wherein the memory stores machine readable instructions executable by the processor, and when the computing device is operated, the processor and the memory are in communication via the bus, and the machine readable instructions, when executed by the processor, cause the steps of the above-mentioned corrected diffusion model based image segmentation method to be performed.

[0011] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, the computer program being adapted to perform the steps of the above-mentioned corrected diffusion model based image segmentation method when executed by a processor.

[0012] The above-mentioned technical means or means have the following beneficial effects:

[0013] In the present invention, the difference between the systematic error image to which Gaussian noise has been added and the systematic error image to which no noise has been added is calculated using a loss function, and the systematic error image is corrected and reconstructed, thereby diffusing the systematic error and thereby achieving further improvement in image segmentation accuracy.

[0014] In the present invention, by introducing a vector quantization variational autoencoder and applying the compensation diffusion model to the latent space of the powerful autoencoder, the training cost can be reduced and the training of the compensation diffusion model can be realized with limited computational resources.

[0015] In the present invention, by introducing a cross-attention mechanism to improve the flexibility and reliability of the corrected diffusion model, the corrected diffusion model can obtain more flexible and reliable results based on the segmentation target image as a condition input.

[0016] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.

[0017] The accompanying drawings of the specification forming a part of this invention are intended to provide a further understanding of the invention, and the illustrative embodiments of the invention and the description thereof are intended to illustrate the invention and are not to be construed as an undue limitation thereon. [Brief description of the drawings]

[0018] [Figure 1] FIG. 1 is a schematic flow chart of an image segmentation method based on a corrected diffusion model according to a first embodiment of the present invention. [Diagram 2] FIG. 2 is a schematic diagram of the configuration of an inference process of the corrected diffusion model segmentation system in the first embodiment of the present invention. [Diagram 3] FIG. 3 shows the U-Net segmentation result image and the final segmentation result image by the corrected diffusion model segmentation system in the first embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0019] It should be noted that the following detailed description is all exemplary to further explain the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention belongs.

[0020] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments of the present invention.

[0021] Unless contradictory, the embodiments of the present invention and the features of the embodiments can be combined with each other.

[0022] Example 1 In this embodiment, performing pre-processing on the segmentation target image to obtain a pre-processed segmentation target image; A step of performing segmentation on the preprocessed segmentation target image using a trained U-Net segmentation network to obtain a primary segmentation result image; A step of performing a difference calculation between the correct answer label of the segmentation result of the segmentation target image and the primary segmentation result image to obtain a systematic error image; adding Gaussian noise to the systematic error image under the condition of a Markov chain, calculating the difference between the systematic error image to which the Gaussian noise has been added and the systematic error image to which no noise has been added using a constructed loss function, and iteratively correcting and reconstructing the systematic error image based on the calculation result of the loss function to obtain a reconstructed systematic error image; performing an addition operation on the primary segmentation result image and the obtained reconstructed systematic error image to obtain a final segmentation result image; An image segmentation method based on a corrected diffusion model is disclosed, which includes:

[0023] In this embodiment, the technical means will be described in detail by taking an MRI brain tumor image as an example of an image to be segmented.

[0024] In this embodiment, the following is specifically performed on the MRI brain tumor image to be segmented: The acquired MRI brain tumor images to be segmented are processed to have a uniform size, and pre-processing is performed, including standardizing the size of all input images to 128 × 128.

[0025] By performing preprocessing on the original MRI brain tumor images, we can improve computational efficiency and save memory space, which can improve the performance of the network model and make it suitable for segmentation tasks.

[0026] Next, the preprocessed MRI brain tumor image to be segmented is input into a trained U-Net segmentation network to perform segmentation, and the brain tumor segmentation result image is output.

[0027] The training process of the U-Net segmentation network involves taking the labeled MRI brain tumor images as input and the brain tumor segmentation result images as output to train the U-Net segmentation network.

[0028] A difference operation is performed between the real brain tumor segmentation image and the primary segmentation result image obtained by the U-Net segmentation network, and a systematic error image is output.

[0029] The systematic error image is input to the trained compensation diffusion model, the generated systematic error image is compensated by the trained compensation diffusion model, and the trained encoder τ θ The MRI brain tumor image is encoded by the above method, and the systematic error image is mapped to the corrected diffusion model to guide the systematic error image, and the reconstructed systematic error image is output.

[0030] In this embodiment, a corrected diffusion model segmentation network is constructed to perform image segmentation, and the corrected diffusion model segmentation network includes a corrected diffusion model and a U-Net segmentation network. Here, the corrected diffusion model mainly includes a diffusion model, an encoder, a decoder, and a vector quantization variational autoencoder τ θ and cross attention.

[0031] Specifically, in the compensated diffusion model segmentation network, the U-Net segmentation network performs segmentation on the MRI brain tumor image to obtain the primary segmentation result image of the brain tumor. The difference module obtains a systematic error image based on the real brain tumor segmentation image and the primary segmentation result image. The compensated diffusion model module performs forward diffusion and reverse diffusion processes with a parameter θ to obtain a reconstructed systematic error image. A vector quantized variational autoencoder consisting of an encoder E and a decoder D maps the MRI image into a low-dimensional and highly efficient latent space. A cross-attention mechanism improves the flexibility and reliability of the compensated diffusion model. The fusion module obtains the final brain tumor segmentation result image based on the primary segmentation result image and the reconstructed systematic error image.

[0032] In the training stage of the compensated diffusion model segmentation network, the U-Net segmentation network is used to perform segmentation on the MRI brain tumor image to obtain a primary segmentation result image of the brain tumor. The difference module is used to obtain a systematic error image based on the real brain tumor segmentation image and the primary segmentation result image. The compensated diffusion model module is used to obtain a reconstructed systematic error image. Specifically, the systematic error image obtained by the difference module is first subjected to a forward diffusion process and encoded by a vector quantization variational autoencoder. Then, a forward noise addition process is performed T times to obtain a Gaussian noise image. Next, a reverse diffusion process is performed, and the MRI brain tumor image is used as a condition, and a reverse noise removal process is performed T times from the Gaussian noise image, and then the decoder is used to decode the image to obtain a reconstructed systematic error image. The fusion module is used to obtain a final brain tumor segmentation result image based on the primary segmentation result image and the reconstructed systematic error image.

[0033] In the actual application of the corrected diffusion model segmentation network, that is, in the inference step, the U-Net segmentation network performs segmentation on the MRI brain tumor image to obtain the brain tumor segmentation result image. The corrected diffusion model module obtains the reconstructed systematic error image. Specifically, in the inverse diffusion process, the MRI brain tumor image is taken as the input condition, and the Gaussian noise image is subjected to T times of inverse noise removal processes, and then decoded by the decoder to obtain the reconstructed systematic error image. The fusion module obtains the final brain tumor segmentation result image based on the primary segmentation result image and the reconstructed systematic error image.

[0034] In this embodiment, in the training process of the correction diffusion model, in the forward diffusion process, the systematic error image is input, and the Gaussian noise image is output. In the reverse diffusion process, the difference between the predicted output and the actual Gaussian noise output image is calculated by the loss function, and the connection weight update is calculated by the gradient, and a reconstructed systematic error image is gradually generated iteratively from the Gaussian noise image.

[0035] Next, an addition operation is performed on the primary segmentation result image obtained by the U-Net segmentation network and the reconstructed systematic error image obtained by the corrected diffusion model, and the final brain tumor segmentation result image is output.

[0036] In the corrected diffusion model segmentation network, brain tumor segmentation is achieved by the MRI brain tumor images X and the ground truth labels M.

[0037] The process of the correction diffusion model is as follows:

number

[0038] Specifically, the vector quantization variational autoencoder in the above compensated diffusion model is constructed as follows. The U-Net segmentation result image is the primary mask S, and the systematic error image obtained by subtracting the correct label M from the primary mask S is E. For a given systematic error image E, the systematic error image E is encoded by a vector quantization variational autoencoder to obtain a latent space representation Z0 = E(E). Then, the latent space representation Z0 is decoded by a decoder to obtain the reconstructed error image

number

[0039] In this embodiment, the latent variable z is centered at zero by two regularization methods. One is KL regularization, which adds KL divergence to z. The other is VQ regularization, which first learns a codebook of different z's and then regularizes the latent space using a vector quantization layer. By compressing the model, the MRI brain tumor images can enter a latent representation space where high-frequency details and fine details are abstracted. In addition, likelihood-based generative models are more suitable for such low-dimensional and highly efficient latent spaces because they can focus on important semantic bits of the data and can also be trained in a low-dimensional and more computationally efficient space.

[0040] Specifically, the forward diffusion process in the above corrected diffusion model is constructed as follows. In the diffusion process, under the condition of the Markov chain, Gaussian noise is continuously added to the initial data information, that is, Gaussian noise is continuously added to z0, and z0 is changed to z T At each step of noise addition,

number

number

number

[0041] Reparameterization technology allows

number

[0042]

number

[0043] Specifically, the de-diffusion process in the above corrected diffusion model is constructed as follows. The de-spreading process obtains the initial data from the Gaussian noise.

number

number

[0044] Neural Network P θ Using the inverse process q(Z t-1 |Z t , τ θ (X)) is fitted. q(Z t-1 |Z t , τ θ (X),Z0) to P θ Therefore, q(Z t-1 |Z t , τ θ By solving (X, Z0) using Bayes' rule, we finally obtain an exponential function whose exponent is a quadratic function. By using Bayes' rule, we can obtain the mean and variance, and since Z0 cannot be known in advance in the inverse process, we can simply replace Z0 with Z t Therefore,

number

number

[0045] In addition, a cross-attention mechanism is also introduced into the de-diffusion process to enhance the underlying U-Net de-noising network. The de-diffusion process is conditioned on the MRI brain tumor images encoded by the encoder. The MRI brain tumor images are mapped to the hidden layers of the U-Net de-noising network in the diffusion model by the cross-attention mechanism. Then, the conditional noise predictor ε θ We perform T iterative noise removal using T is converted to z0, then input to the decoder and decoded to obtain the reconstruction error diagram

number

[0046] The principle of the diffusion model is T iterative denoising, where each iterative denoising step includes one U-Net denoising network, and in each step, the U-Net denoising network is used to denoise Z t From Z t-1 For example, a sample at time step t is fed into a U-Net denoising network and subjected to implicit denoising to obtain a sample at time step t-1.

[0047] Therefore, the loss function of the corrected diffusion model is

number

[0048] Specifically, the cross-attention mechanism in the above corrected diffusion model is constructed as follows. In principle, the diffusion model can be modeled as a conditional denoising autoencoder, epsilon θ (z t ,t,X)

number

[0049] In this embodiment, we introduce a CLIP image encoder and θ In this way, preprocessing is performed on the MRI brain tumor image X. The CLIP image encoder converts the MRI brain tumor image X into an intermediate representation τ θ (X), and then map it to the hidden layer of the U-Net denoising network in the diffusion model by the cross-attention mechanism. The cross-attention mechanism connects the conditional coding and the hidden layer of the U-Net denoising network, and the input to the cross-attention mechanism is z t and τ θ (X), and the output of the cross attention mechanism is input to the decoder and decoded.

[0050] The back-diffusion process is used to learn the systematic error and guide the model training by adding the information of MRI brain tumor image X at each time step. The information of MRI brain tumor image X is mapped to the intermediate layer of the U-Net denoising network, which improves the performance and ability of the model and contributes to more accurately capturing the features and distribution of the data.

[0051] In the de-diffusion process, we guide the learning of systematic errors by adding information from the MRI brain tumor image X at each time step.

[0052]

number

[0053] To summarize, in this embodiment, the training step of the above corrected diffusion model segmentation network is as follows: Construct a training sample set of tumor-labeled MRI brain tumor images; Constructing a corrected diffusion model segmentation network; Training a corrected diffusion model segmentation network by the training sample set, and stopping the training when the loss function reaches a minimum value or the number of iterations meets a set requirement, and obtaining a trained corrected diffusion model segmentation network; Includes.

[0054] In the training process, based on the sample set, the samples are randomly divided in an 8:2 ratio, of which 80% are training samples and the remaining 20% ​​are test samples. We selected BRATS2019, BRATS2020, and the brain tumor dataset provided by Jun Cheng to evaluate the model. Here, the sample values ​​of BRATS2019 and BRATS2020 are brain tumor images labeled with whole tumor WT, tumor core TC, and enhanced tumor ET, while the brain tumor dataset provided by Jun Cheng contains three types of brain tumor images, and the sample values ​​are all brain tumor images labeled with tumor.

[0055] Here, the compensated diffusion model segmentation network to be trained mainly includes two parts, which correspond to the two modules shown in Figure 1. The first part is a U-Net segmentation network, which takes the MRI brain tumor image as input, and performs segmentation on the MRI brain tumor image through the U-Net network to obtain the primary mask s. The second part is a compensated diffusion model network, which uses the diffusion model principle to compensate for the segmentation systematic error, and obtains the systematic error segmentation result image through the loss function.

[0056] In this embodiment, the loss function of the correction diffusion model system is specifically as follows:

number

[0057] In the training process, the MRI brain tumor image is preprocessed, and the preprocessed MRI brain tumor image is input into the U-Net segmentation network to generate the primary mask s. The obtained error map E is input into the correction diffusion model. The conditional noise predictor ε θ and the vector quantized variational autoencoder τ θ By simultaneously optimizing the loss function, the accuracy of the experiment is further improved, and the reconstruction error diagram

number

number

[0058] In this embodiment, the inference process of the corrected diffusion model network is specifically as follows: As shown in Figure 2, an MRI brain tumor image is taken as input. T Sampled and Vector Quantized Variational Autoencoder τ θ We encode the MRI brain tumor image X by, map it to the U-Net denoising network in the diffusion model by a cross-attention layer, and apply it at each iteration. Then, we use the conditional noise predictor ε θUsing z t , t and τ θ With input z t-1 At t=1, z0 is the prediction of the final systematic error, and the reconstruction error diagram is calculated by decoding z0 by the decoder.

number

number

number

[0059] In this example, brain tumor segmentation experiments were performed on BRATS2019, BRATS2020, and the brain tumor dataset provided by Jun Cheng. The training set of BRATS2019 contained 335 images, which were randomly divided into 268 images for training and 67 images for validation. Similarly, the training set of BRATS2020 contained 369 images, which were randomly divided into 295 images for training and 74 images for validation. In addition, the brain tumor dataset provided by Jun Cheng consisted of 3064 contrast-enhanced images weighted by t1, which were also randomly divided into 2452 images for training and 612 images for validation.

[0060] Comparing the final segmentation results on the BRATS2020 dataset in this embodiment with conventional segmentation techniques, the technique described in this embodiment can achieve better segmentation results, as shown in FIG. 3 and Table 1 below.

[0061] [Table 1]

[0062] Comparing the final segmentation results on the BRATS2019 dataset in this embodiment with conventional segmentation techniques, the technique described in this embodiment can achieve better segmentation effect, as shown in FIG. 3 and Table 2 below.

[0063] [Table 2]

[0064] In this embodiment, the acquired MRI brain tumor image is input to the U-Net network after preprocessing, and segmentation is performed on the MRI brain tumor image to output the brain tumor segmentation result image and generate a primary mask s. Then, a difference operation is performed between the real brain tumor segmentation image and the U-Net segmentation result image s, and the acquired error map E is input to the correction diffusion model network, and ε θ We perform T iterative noise removal and obtain the reconstruction error plot.

number

number

[0065] The accurate MRI brain tumor image segmentation system based on the corrected diffusion model provided by this embodiment uses the diffusion model to diffusion correct the systematic errors generated by the segmentation algorithm, thereby diffusing the systematic errors to further improve the accuracy of MRI brain tumor image segmentation and assist doctors in diagnosing brain tumor diseases.

[0066] The correction diffusion model algorithm provided by this embodiment applies the diffusion model to correct systematic errors and introduces correction learning into the diffusion model to improve the segmentation accuracy of brain tumors.

[0067] In this embodiment, a vector quantization variational autoencoder is introduced, and the compensation diffusion model is applied to the latent space of a powerful autoencoder, thereby reducing the training cost and realizing the training of the compensation diffusion model with limited computational resources.

[0068] In this embodiment, a cross-attention mechanism is introduced to improve the flexibility and reliability of the corrected diffusion model, so that the corrected diffusion model can output more flexible and reliable results based on the condition input of MRI brain tumor images.

[0069] Example 2 In this embodiment, an acquisition module for performing pre-processing on the segmentation target image to obtain a pre-processed segmentation target image; A primary segmentation module that performs segmentation on the preprocessed segmentation target image using a trained U-Net segmentation network to obtain a primary segmentation result image; a calculation module for performing a difference calculation between a correct answer label of the segmentation result of the segmentation target image and a primary segmentation result image to obtain a systematic error image; a reconstruction module for adding Gaussian noise to the systematic error image under a Markov chain condition, calculating a difference between the systematic error image to which the Gaussian noise has been added and a systematic error image to which no noise has been added using a constructed loss function, and iteratively correcting and reconstructing the systematic error image based on the calculation result of the loss function to obtain a reconstructed systematic error image; a segmentation module for performing an addition operation on the primary segmentation result image and the obtained reconstructed systematic error image to obtain a final segmentation result image; The present invention provides an image segmentation system based on a corrected diffusion model, comprising:

[0070] Example 3 The present embodiment aims to provide a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, whereby the method described in embodiment 1 is realized when the program is executed by the processor.

[0071] Example 4 The present embodiment aims to provide a computer-readable storage medium. A computer program is stored in a computer-readable storage medium, and when the program is executed by a processor, the method according to the first embodiment is performed.

[0072] The steps and methods related to the devices in the above Examples 2, 3 and 4 correspond to those in Example 1. For specific embodiments, reference can be made to the relevant description in Example 1. The term "computer-readable storage medium" should be understood to be a single medium or multiple media containing one or more command sets, and should also be understood to include any medium that can store, encode or carry a command set for execution by a processor and can cause a processor to execute any of the methods in the present invention.

[0073] As will be appreciated by those skilled in the art, each module or step of the present invention described above can be implemented by a general-purpose computer device, or alternatively, can be implemented by a program code executable by a computer device, which can be stored in a storage device and executed by the computer device, or can be implemented by creating each of them in an integrated circuit module, or by creating multiple modules or steps in a single integrated circuit module. The present invention is not limited to any specific combination of hardware and software.

[0074] The specific embodiments of the present invention have been described above with reference to the accompanying drawings, but they do not limit the scope of protection of the present invention. As those skilled in the art can understand, various modifications or variations that those skilled in the art can make based on the technical means of the present invention without inventive labor still fall within the scope of protection of the present invention.

Claims

1. performing pre-processing on the segmentation target image to obtain a pre-processed segmentation target image; A step of performing segmentation on the pre-processed segmentation target image using a trained U-Net segmentation network to obtain a primary segmentation result image; A step of performing a difference calculation between the correct answer label of the segmentation result of the segmentation target image and the primary segmentation result image to obtain a systematic error image; adding Gaussian noise to the systematic error image under a Markov chain condition, calculating a difference between the systematic error image to which the Gaussian noise has been added and a systematic error image to which no noise has been added using a constructed loss function, and iteratively correcting and reconstructing the systematic error image based on the calculation result of the loss function to obtain a reconstructed systematic error image; performing an addition operation on the primary segmentation result image and the obtained reconstructed systematic error image to obtain a final segmentation result image; 1. A method for image segmentation based on a corrected diffusion model, comprising:

2. The image segmentation method based on the compensated diffusion model as claimed in claim 1, characterized in that the training of the U-Net segmentation network specifically includes: taking the labeled image as the input of the U-Net segmentation network, taking the image segmentation result as the output of the U-Net segmentation network, and training the U-Net segmentation network.

3. Gaussian noise is added to the systematic error image through a forward diffusion process of a correction diffusion model, specifically, encoding the systematic error image by an encoder; Successively adding Gaussian noise to the encoded systematic error image; 2. The image segmentation method based on the corrected diffusion model according to claim 1, characterized in that the systematic error image after adding Gaussian noise is obtained each time by a reparameterization technique.

4. The image segmentation system based on the compensated diffusion model according to claim 1, characterized in that the segmentation target image is processed by the trained compensated diffusion model to obtain the reconstructed systematic error image, specifically, in the inverse diffusion process of the trained compensated diffusion model, the segmentation target image encoded by the encoder is used as a condition, and the Gaussian noise image is subjected to an inverse noise removal process multiple times, and then decoded by the decoder to obtain the reconstructed systematic error image.

5. The image segmentation system based on the compensated diffusion model according to claim 4, characterized in that the noise removal is performed by a U-Net denoising network, in which the image to be segmented is encoded by a vector quantization variational autoencoder, the encoded image to be segmented is mapped to the intermediate layer of the U-Net denoising network by a cross-attention mechanism, and the iterative noise removal is performed multiple times using a conditional noise predictor, and the latent variables are transformed and then input into the decoder for decoding to obtain a reconstructed systematic error image.

6. The image segmentation system based on the compensated diffusion model according to claim 1, characterized in that the training of the compensated diffusion model is specifically: in the forward diffusion process, the systematic error image is input, and the Gaussian noise image is output; in the reverse diffusion process, the difference between the predicted output and the actual Gaussian noise output image is calculated by the loss function, and the update of the connection weight is calculated by the gradient, so as to iteratively generate a reconstructed systematic error image from the Gaussian noise image in stages.

7. 5. The image segmentation system based on the compensated diffusion model according to claim 4, wherein in the de-diffusion process, an image to be segmented is added at each time step to guide learning of the systematic error.

8. an acquisition module for performing pre-processing on the segmentation target image to obtain a pre-processed segmentation target image; A primary segmentation module that performs segmentation on the preprocessed segmentation target image using a trained U-Net segmentation network to obtain a primary segmentation result image; a calculation module for performing a difference calculation between a correct answer label of the segmentation result of the segmentation target image and a primary segmentation result image to obtain a systematic error image; a reconstruction module for adding Gaussian noise to the systematic error image under a Markov chain condition, calculating a difference between the systematic error image to which the Gaussian noise has been added and a systematic error image to which no noise has been added using a constructed loss function, and iteratively correcting and reconstructing the systematic error image according to the calculation result of the loss function to obtain a reconstructed systematic error image; a segmentation module for performing an addition operation on the primary segmentation result image and the obtained reconstructed systematic error image to obtain a final segmentation result image; 1. An image segmentation system based on a corrected diffusion model, comprising:

9. A computer device comprising a processor, a memory and a bus, the memory storing machine-readable commands executable by the processor, the processor and the memory communicating via the bus when the computer device is operated, and the image segmentation method based on the corrected diffusion model of any one of claims 1 to 7 being performed when the machine-readable commands are executed by the processor.

10. A computer-readable storage medium having a computer program stored thereon, the computer program being operable, when executed by a processor, to perform the image segmentation method based on the corrected diffusion model according to any one of claims 1 to 7.

Images