Image reconstruction method, system and terminal of diffusion bridge model based on knowledge distillation

Abstract

According to the image reconstruction method and system of the knowledge distillation-based diffusion bridge model and the terminal provided by the invention, the diffusion bridge model is utilized to directly generate the high-resolution image from the low-resolution image, the traditional Gaussian noise initial state is abandoned, and the credibility and anatomical consistency of the reconstruction result are improved. According to the method, the training strategy of knowledge distillation is utilized, an image reconstruction teacher network model needing multi-step reasoning is compressed into an ultimate image reconstruction student network model needing only one-step reasoning, the image fidelity is guaranteed, meanwhile, the calculation time is shortened, and the clinical real-time requirement is met. Moreover, some extra data without pairing are introduced in the training process of the image reconstruction student network model, so that the model learns more data samples, and the generalization of the model is improved through the semi-supervised training strategy.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence and medical imaging technology, and in particular to an image reconstruction method, system, and terminal based on a diffusion bridge model of knowledge distillation. Background Art

[0002] In the field of medical imaging, especially super-resolution (SR) reconstruction of neonatal brain magnetic resonance imaging (MRI), generating high-resolution (HR) thin-slice images from clinically acquired low-resolution (LR) thick-slice images is a clinically valuable yet challenging task. Due to the developmental characteristics of neonatal brain tissue, such as low contrast between gray and white matter and blurred tissue boundaries, there is a significant difference in interslice resolution between clinical thick-slice MRI and high-resolution thin-slice MRI (typically one thick 2D slice corresponds to approximately six thin slices). Traditional interpolation models or deep learning-based super-resolution techniques struggle to generate images with high fidelity and anatomical consistency.

[0003] In recent years, generative adversarial networks (GANs) and diffusion models have shown potential in the field of medical image synthesis. Existing methods establish a mapping relationship between low-resolution input and high-resolution output through end-to-end supervised learning. For example, the U-Net-based SRGAN extracts multi-scale features through residual blocks and combines adversarial loss to improve image perception quality. As a generative model based on probability distribution learning, the diffusion model generates high-quality data by gradually reversing the noise addition process. In the field of medical imaging, the diffusion model is used for image generation, denoising, and super-resolution reconstruction, and its generation capability has been proven to be superior to traditional generative adversarial networks (GANs). However, existing technologies still have the following core defects:

[0004] (1) Insufficient fidelity: Traditional end-to-end supervised learning methods rely on limited paired LR-HR data for training. When the gray matter-white matter contrast of neonatal MRI is naturally low, the model is prone to produce false textures (such as artifacts and abnormal gray matter clumps), and it is difficult to model the complex nonlinear mapping relationship between thick-slice MRI and thin-slice MRI, resulting in poor biological rationality and anatomical consistency of the generated images, which cannot meet clinical diagnostic needs.

[0005] (2) Difficulty in integrating whole-brain information: Existing technologies are mostly based on independent processing of 2D slices, ignoring the continuity of the three-dimensional structure of the brain and failing to fully utilize the three-dimensional spatial context information. This results in poor continuity between layers in the reconstruction results and makes it difficult to ensure the consistency of the whole-brain anatomical structure.

[0006] (3) Excessive inference time: Although the diffusion model can generate high-fidelity images through probabilistic sampling, it requires multi-step iterative denoising, which makes single inference time-consuming and the generation efficiency low, making it difficult to meet clinical real-time requirements.

[0007] (4) Unstable generation quality: Although GAN can generate visually realistic images, its training process is unstable and prone to mode collapse or artifacts. This is especially evident in data-scarce neonatal MRI scenarios, where it is difficult to cover the diverse developmental characteristics of the neonatal brain.

[0008] (5) Scarcity of paired data: It is expensive to obtain paired thick-thin MRI data in clinical scenarios, and the amount of neonatal data is limited, which restricts the performance of supervised learning methods. Summary of the Invention

[0009] In view of the shortcomings of the prior art described above, the purpose of this application is to provide an image reconstruction method, system and terminal based on a diffusion bridge model of knowledge distillation, which is used to solve the technical problems of long inference time and insufficient fidelity in medical image reconstruction in the prior art.

[0010] To achieve the above-mentioned purpose and other related purposes, the first aspect of the present application provides an image reconstruction method of a diffusion bridge model based on knowledge distillation, comprising: collecting multiple brain thick layer image data and brain thin layer image data corresponding to each of the brain thick layer image data; performing data preprocessing on each of the brain thick layer image data and the brain thin layer image data corresponding to each of the brain thick layer image data to obtain a training data set; using the training data set to train a diffusion bridge model, and based on a preset loss function, adjusting the parameters of the denoising network structure in the diffusion bridge model to construct an image reconstruction teacher network model; constructing an initial image reconstruction student network model of single-step mapping based on the image reconstruction teacher network model; optimizing the initial image reconstruction student network model based on the knowledge distillation optimization objective function to obtain an ultimate image reconstruction student network model; reconstructing the acquired real-time low-resolution brain image data based on the ultimate image reconstruction student network model to generate target high-resolution image data.

[0011] In some embodiments of the first aspect of the present application, the training data set includes multiple data pairs, each data pair including original low-resolution brain image data and actual high-resolution brain image data corresponding to the original low-resolution brain image data; wherein, the method of constructing the image reconstruction teacher network model includes: inputting the original low-resolution brain image data and the actual high-resolution brain image data corresponding to the original low-resolution brain image data into the diffusion bridge model for training, and generating noise sample image data based on the diffusion forward algorithm; based on the noise sample image data and the original low-resolution brain image data, calling the denoising network structure in the diffusion bridge model to generate the first predicted high-resolution image data corresponding to the original low-resolution brain image data; based on the actual high-resolution brain image data corresponding to the original low-resolution brain image data and the first predicted high-resolution image data, and based on the preset loss function, adjusting the parameters of the denoising network structure in the diffusion bridge model to construct the image reconstruction teacher network model.

[0012] In some embodiments of the first aspect of the present application, the original low-resolution brain image data and actual high-resolution brain image data corresponding to the original low-resolution brain image data are input into the diffusion bridge model for training, and a method for generating noisy sample image data based on a diffusion forward algorithm includes:

[0013] q t (x t |x0,x T )=N(x t ;μ t (x0,x T ),∑t);

[0014] Among them, q t (x t |x0,x T ) represents conditional probability; N represents normal distribution; x t represents the noise sample image data of the current time step; x0 represents the actual high-resolution brain image data corresponding to the original low-resolution brain image data; x T represents the original low-resolution brain image data; μ t represents the mean of the noise sample image data; ∑t represents the variance of the noise sample image data; t represents the current time step; T represents the preset total time step.

[0015] In some embodiments of the first aspect of the present application, the method of optimizing the initial image reconstruction student network model based on the knowledge distillation optimization objective function to obtain the ultimate image reconstruction student network model includes: generating second predicted high-resolution image data corresponding to the original low-resolution brain image data according to the image reconstruction teacher network model and the diffusion inverse algorithm; determining the knowledge distillation optimization objective function according to the knowledge distillation loss function, the degradation loss function and the cycle consistency loss function; iteratively training the initial image reconstruction student network model using the original low-resolution brain image data and the second predicted high-resolution image data corresponding to the original low-resolution brain image data, and optimizing the initial image reconstruction student network model based on the knowledge distillation optimization objective function to obtain the ultimate image reconstruction student network model.

[0016] In some embodiments of the first aspect of the present application, the method of generating the second predicted high-resolution image data corresponding to the original low-resolution brain image data includes: calculating the intermediate image data of the previous time step based on the noise sample image data of the current time step and based on the image reconstruction teacher network model and the diffusion inverse algorithm; performing multiple inferences on the intermediate image data of the previous time step to generate the second predicted high-resolution image data corresponding to the original low-resolution brain image data.

[0017] In some embodiments of the first aspect of the present application, the method of determining the knowledge distillation optimization objective function based on the knowledge distillation loss function, the degradation loss function, and the cycle consistency loss function includes:

[0018] L=L distill +L degrade +L cycle ;

[0019]

[0020] Among them, L represents the knowledge distillation optimization objective function; L distill represents the knowledge distillation loss function; L degrade represents the degradation loss function; L cycle represents the cycle consistency loss function; MSE represents the mean square error function; represents the initial image reconstruction student network model; F θ represents the denoising network structure in the image reconstruction teacher network model; x T represents the original low-resolution brain image data; x0 represents the actual high-resolution brain image data corresponding to the original low-resolution brain image data; T represents the preset total time step; represents the predicted low-resolution image data generated by the initial image reconstruction student network model; θ represents the learnable parameters.

[0021] In some embodiments of the first aspect of the present application, a method for calculating intermediate image data at a previous time step based on the noise sample image data at the current time step and the image reconstruction teacher network model and the diffusion inverse algorithm includes:

[0022]

[0023] Among them, x p Represents the intermediate image data of the previous time step; x t Represents the noise sample image data of the current time step; σ t represents the first previous time step factor; σ p represents the second previous time step factor; F θ represents the denoising network structure in the image reconstruction teacher network model; θ represents the learnable parameters; t represents the current time step; I represents Gaussian noise that obeys the normal distribution.

[0024] In some embodiments of the first aspect of the present application, the preset loss function includes:

[0025] L diffusion =MSE(x0-F θ (x t , x T , t;θ)));

[0026] Among them, L diffusion represents the preset loss function; MSE represents the mean square error function; x0 represents the actual high-resolution brain image data corresponding to the original low-resolution brain image data; F θ represents the denoising network structure in the image reconstruction teacher network model; x t Represents the noise sample image data of the current time step; x T Represents the original low-resolution brain image data; t represents the current time step; T represents the preset total time step; θ represents the learnable parameter.

[0027] To achieve the above-mentioned purpose and other related purposes, the second aspect of the present application provides an image reconstruction system of a diffusion bridge model based on knowledge distillation, comprising: a data acquisition module for acquiring multiple brain thick layer image data and brain thin layer image data corresponding to each of the brain thick layer image data; a data preprocessing module for performing data preprocessing on each of the brain thick layer image data and the brain thin layer image data corresponding to each of the brain thick layer image data to obtain a training data set; a teacher network construction module for training a diffusion bridge model using the training data set, and adjusting the parameters of the denoising network structure in the diffusion bridge model based on a preset loss function to construct an image reconstruction teacher network model; a student network construction module for constructing an initial image reconstruction student network model of single-step mapping based on the image reconstruction teacher network model; optimizing the initial image reconstruction student network model based on the knowledge distillation optimization objective function to obtain an ultimate image reconstruction student network model; and an image reconstruction module for reconstructing the acquired real-time low-resolution brain image data based on the ultimate image reconstruction student network model to generate target high-resolution image data.

[0028] To achieve the above-mentioned purpose and other related purposes, the third aspect of the present application provides an electronic terminal, including a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to implement the image reconstruction method of the diffusion bridge model based on knowledge distillation as described above.

[0029] As described above, the image reconstruction method, system, and terminal of the diffusion bridge model based on knowledge distillation of the present application have the following beneficial effects:

[0030] (1) Improved generation credibility and accuracy: The diffusion bridge model used in the forward process fuses the information of low-resolution images with that of high-resolution images, enabling the model to learn the complex mapping relationship from low resolution to high resolution and enhancing the model's robustness to noise. In the reverse process, the low-resolution image is used as the starting point, and denoising is continuously performed through the denoising network structure, finally obtaining a high-resolution image. This abandons the traditional Gaussian noise initial state and introduces structural consistency information, thereby improving the structural fidelity, credibility, and anatomical consistency of the reconstruction results.

[0031] (2) Reduce model inference time: Using the knowledge distillation training strategy, the image reconstruction teacher network model that requires multiple steps of inference is compressed into the ultimate image reconstruction student network model that only requires one step of inference. While ensuring image fidelity, it reduces the computation time and meets clinical real-time requirements. In addition, the image reconstruction student network model introduces some unpaired additional data during the training process. The image reconstruction teacher network model generates high-resolution prediction results through reverse inference, which serve as soft labels for the initial image reconstruction student network model, thereby guiding the learning of the initial image reconstruction student network model and enabling the model to learn more data samples. This semi-supervised training strategy increases the generalization of the model.

[0032] (3) Alleviating data scarcity: The diffusion bridge model can effectively utilize limited paired data (LR-HR) for training and enhance data expression capabilities through probability distribution learning. It solves the problem of low contrast between gray matter and white matter in neonatal brain tissue, which easily generates false textures. It enables the model to adapt to low-contrast scenes and generate high-resolution images that conform to the anatomical structure, alleviating the problem of scarce paired data in the field of medical imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Shown is a flowchart of an image reconstruction method based on a diffusion bridge model of knowledge distillation in one embodiment of the present application.

[0034] Figure 2 Shown is a flowchart of data preprocessing in one embodiment of the present application.

[0035] Figure 3 Shown is a flowchart of constructing an image reconstruction teacher network model in one embodiment of the present application.

[0036] Figure 4 Shown is a schematic diagram of the architecture of the diffusion bridge model training in one embodiment of the present application.

[0037] Figure 5 Shown is an architectural diagram of a denoising network structure in one embodiment of the present application.

[0038] Figure 6 Shown is a schematic diagram of the architecture of the initial image reconstruction student network model optimization in one embodiment of the present application.

[0039] Figure 7 Shown is a flowchart of constructing the ultimate image reconstruction student network model in one embodiment of the present application.

[0040] Figure 8 Shown is a flowchart of image reconstruction using the ultimate image reconstruction student network model in one embodiment of the present application.

[0041] Figure 9Shown is a schematic block diagram of an image reconstruction system based on a diffusion bridge model of knowledge distillation in one embodiment of the present application.

[0042] Figure 10 Shown is a structural schematic diagram of an electronic terminal in one embodiment of the present application. DETAILED DESCRIPTION

[0043] The following describes the embodiments of the present application through specific examples. Those skilled in the art can easily understand the other advantages and effects of the present application from the content disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0044] Before further explaining the present invention in detail, the nouns and terms involved in the embodiments of the present invention are explained. The nouns and terms involved in the embodiments of the present invention are subject to the following interpretations:

[0045] <1> MRI (Magnetic Resonance Imaging): Magnetic resonance imaging is a medical imaging technology that uses strong magnetic fields and radio frequency pulses to generate detailed images of the internal structures of the human body.

[0046] <2> Super-Resolution (SR): Super-resolution is a technique that reconstructs high-resolution (HR) images from low-resolution (LR) images using algorithms or models. In medical imaging, SR is primarily used to improve the spatial resolution of images and enhance detail clarity, helping doctors more accurately observe tiny anatomical structures or lesions.

[0047] <3> LR (Low Resolution): Low resolution refers to the lack of information on spatial details in images or data, which appears as blur, unclear edges or missing textures.

[0048] <4> HR (High Resolution): High resolution refers to images or data with higher detail clarity and fineness, capable of presenting richer structures and information.

[0049] <5> GAN (Generative Adversarial Network): Generative adversarial network is a deep learning framework based on game theory. It consists of two parts: the generator and the discriminator. Data generation is achieved through adversarial training between the two.

[0050] <6> U-Net: U-Net is a convolutional neural network (CNN) architecture used for medical image segmentation.

[0051] <7> SRGAN (Super-Resolution Generative Adversarial Network): Super-resolution generative adversarial network is a super-resolution method based on generative adversarial network (GAN).

[0052] <8> Diffusion Model: Diffusion model learns the data distribution by gradually adding noise (forward process) and gradually denoising (reverse process), and finally generates target data (such as images) from random noise.

[0053] <9> Diffusion Bridge Model: The diffusion bridge model introduces conditional constraints in the diffusion process, so that the generation process gradually transitions from a specific initial state (such as a low-resolution image) to a target state (such as a high-resolution image) rather than starting from pure noise.

[0054] <10> Unet: It is an encoder-decoder neural network used for medical image segmentation. It has a symmetrical U-shaped architecture and fuses shallow details with deep semantic information through skip connections, improving segmentation performance while maintaining positioning accuracy.

[0055] In the field of medical imaging, generating high-resolution (HR) thin-slice images from low-resolution (LR) thick-slice images faces severe challenges. Due to the low contrast and blurred boundaries of neonatal brain tissue, and the significant resolution difference between thick and thin slices (e.g., 1:6 inter-slice correspondence), traditional interpolation or deep learning-based super-resolution methods (e.g., SRGAN) have difficulty in ensuring anatomical consistency and high fidelity. Although generative adversarial networks (GANs) and diffusion models have shown potential in medical image synthesis, existing technologies still have the following shortcomings: (1) Insufficient fidelity: supervised learning relies on limited paired data, easily generates false textures, and has difficulty modeling the mapping relationship between thick and thin slices; (2) Difficulty in integrating whole-brain information: 2D slice processing destroys the spatial continuity of the brain and has poor inter-slice consistency; (3) Long inference time: The multi-step iteration of the diffusion model leads to a long inference time; (4) Unstable generation quality: GAN is prone to artifacts or mode collapse, especially when data is scarce; (5) Scarce paired data: Paired thick and thin MRI of neonates is difficult to obtain, which restricts the performance of supervised learning. These deficiencies limit the reliability and practicality of existing methods in clinical diagnosis.

[0056] In order to solve the problems in the above-mentioned background technology, the present application provides an image reconstruction method, system and terminal based on the diffusion bridge model of knowledge distillation, aiming to solve the technical problems of long inference time and insufficient fidelity in medical image reconstruction in the existing technology.

[0057] To facilitate understanding of the embodiments of this application, first Figure 1 Detailed description. Figure 1 The following is a flow chart of an image reconstruction method based on a diffusion bridge model using knowledge distillation in an embodiment of the present invention. The image reconstruction method based on a diffusion bridge model using knowledge distillation in this embodiment mainly includes the following steps:

[0058] S101: Acquire a plurality of thick-slice brain image data and thin-slice brain image data corresponding to each of the thick-slice brain image data.

[0059] In this embodiment, thick-slice brain image data (low-resolution images) of multiple subjects scanned at the same time and thin-slice brain image data (high-resolution images) corresponding to the thick-slice brain image data are collected to obtain multiple sets of paired low-resolution images and high-resolution images.

[0060] S102: performing data preprocessing on each of the brain thick slice image data and the brain thin slice image data corresponding to each of the brain thick slice image data to obtain a training data set.

[0061] In this embodiment, data preprocessing includes but is not limited to: linear interpolation resampling preprocessing, cropping preprocessing, expansion preprocessing, scalp segmentation preprocessing, neck cutting preprocessing, linear registration preprocessing or normalization preprocessing.

[0062] In this embodiment, if Figure 2 The figure shows a flow chart of data preprocessing in an embodiment of the present invention. The brain thick slice image data is first preprocessed by linear interpolation resampling to adjust the resolution of the brain thick slice image data to 0.8*0.8*0.8mm. 3 After that, the image is cropped and expanded to a uniform size (such as 192*224*192), and finally the scalp segmentation and normalization preprocessing are performed to obtain the original low-resolution brain image data x T .

[0063] In this embodiment, the brain thin-layer image data is first pre-processed by neck segmentation, and then linear registration pre-processing is performed to ensure that the thick and thin layer paired data are aligned and have the same size and resolution. Finally, scalp segmentation pre-processing and normalization pre-processing are performed to obtain the original low-resolution brain image data x T The corresponding actual high-resolution brain image data x0.

[0064] In this embodiment, linear interpolation resampling preprocessing is a mathematical method based on weighted averaging of adjacent data points to estimate the value of a new position. In medical imaging, linear interpolation resampling preprocessing is used to adjust the resolution of the image.

[0065] In this embodiment, the cropping preprocessing is to remove the invalid background area in the image (such as the black border of the MRI scan) and retain only the area containing the target (such as brain tissue) to avoid background noise interfering with model training.

[0066] In this embodiment, the expansion preprocessing is to expand the image to a uniform size by padding (such as zero value, edge replication or reflection) when the image size is smaller than the target size, so as to meet the input size requirement of the deep learning model.

[0067] In this embodiment, the scalp segmentation preprocessing is to remove non-brain tissues (such as the skull) in the MRI, retain the brain parenchyma (gray matter, white matter, cerebrospinal fluid), and reduce the interference of non-brain tissues.

[0068] In this embodiment, the neck cutting preprocessing is to remove the irrelevant structures in and around the neck during brain MRI analysis, leaving only the head region, thereby eliminating the interference of the neck on registration.

[0069] In this embodiment, the linear registration pre-processing is to spatially align the thin-slice image and the thick-slice image after neck resection so that the two have the same size, resolution, and anatomical position, thereby ensuring the consistency of the paired data.

[0070] In this embodiment, the normalization preprocessing is to adjust the image intensity distribution to eliminate the influence of differences in scanning equipment and protocols, so as to make the data distribution of different samples consistent.

[0071] S103: Using the training data set to train a diffusion bridge model, and based on a preset loss function, adjusting the parameters of the denoising network structure in the diffusion bridge model to construct an image reconstruction teacher network model.

[0072] In this embodiment, if Figure 3 FIG. 1 shows a flow chart of constructing an image reconstruction teacher network model according to an embodiment of the present invention. The training dataset includes multiple data pairs, each data pair including original low-resolution brain image data and actual high-resolution brain image data corresponding to the original low-resolution brain image data. The method of constructing the image reconstruction teacher network model includes:

[0073] S1031: Inputting the original low-resolution brain image data and the actual high-resolution brain image data corresponding to the original low-resolution brain image data into the diffusion bridge model for training, and generating noise sample image data based on a diffusion forward algorithm.

[0074] In this embodiment, if Figure 4 FIG. 1 shows a schematic diagram of the architecture of the diffusion bridge model training in an embodiment of the present invention. The original low-resolution brain image data and the actual high-resolution brain image data corresponding to the original low-resolution brain image data are input into the diffusion bridge model for training, and the method of generating noise sample image data based on the diffusion forward algorithm includes:

[0075] q t (x t |x0,x T )=N(x t ;μ t (x0,x T ),∑t); Formula (1)

[0076] Among them, q t (x t |x0,x T ) represents conditional probability; N represents normal distribution; x t represents the noise sample image data of the current time step; x0 represents the actual high-resolution brain image data corresponding to the original low-resolution brain image data; x T represents the original low-resolution brain image data; μ t represents the mean of the noise sample image data; ∑t represents the variance of the noise sample image data; t represents the current time step; T represents the preset total time step.

[0077] In this embodiment, the original low-resolution brain image data x T Represents the least clear image, original low-resolution brain image data x T The corresponding actual high-resolution brain image data x0 represents the clearest image, and the noise sample image data x t The clarity is x T Between x and x0 is a noisy, poor-quality image. The original low-resolution brain image data and its corresponding actual high-resolution brain image data are input into the diffusion bridge model. According to formula (1), a forward process is used to sample noisy sample image data for time steps 0-T, where the preset total time step T is generally set to 1000.

[0078] In this embodiment, the mean value μ of the noise sample image data is t is calculated as follows:

[0079]

[0080] Among them, μ trepresents the mean value of the noise sample image data; x0 represents the actual high-resolution brain image data corresponding to the original low-resolution brain image data; x T represents the original low-resolution brain image data; σ t represents the first previous time step factor; represents the late time step factor; β t represents the noise variance scheduling parameter; t represents the current time step; T represents the preset total time step.

[0081] In this embodiment, β t Represents the noise variance scheduling parameter, which is a linear increasing function related to the time step. When the early time step is close to 0, The smaller it is, the closer the mean is to x0, preserving high-resolution structural information. When the time step approaches T, The smaller it is, the closer the mean is to x T , ensuring consistency with the input low-resolution data, so that high- and low-resolution image data are gradually mixed to achieve a smooth transition, avoid information mutation, and enable the model to gradually learn the mapping relationship between the two.

[0082] In this embodiment, the variance of the noise sample image data is calculated as follows:

[0083]

[0084] Where ∑t represents the variance of the noise sample image data; σ t represents the first previous time step factor; Represents the later time step factor; I represents Gaussian noise that obeys the normal distribution; t represents the current time step; T represents the preset total time step.

[0085] In this embodiment, the variance ∑t of the noise sample image data is given by and The noise contribution of high-resolution and low-resolution data is balanced by introducing Gaussian noise I that obeys N(0, 1) distribution, which can avoid training divergence caused by too large or too small variance.

[0086] In this embodiment, in the forward process of the diffusion bridge model, the original low-resolution brain image data x is converted into T It is dynamically weighted fused with the corresponding actual high-resolution brain image data x0, and superimposed with the noise variance scheduling parameter β t Controlled Gaussian noise is used to generate noisy sample image data, simulating the real image degradation process, enabling the model to efficiently learn complex mapping from low resolution to high resolution and enhancing the model's robustness to noise.

[0087] S1032: Based on the noise sample image data and the original low-resolution brain image data, calling a denoising network structure in a diffusion bridge model to generate first predicted high-resolution image data corresponding to the original low-resolution brain image data.

[0088] In this embodiment, if Figure 5 As shown, the architecture diagram of the denoising network structure in the embodiment of the present invention is shown. According to the noise sample image data and the original low-resolution brain image data, the denoising network structure in the diffusion bridge model is called. The denoising network structure is a Unet structure, and the input is the original low-resolution brain image data and the noise sample image data of the current time step. First, the image passes through the convolution input layer to encode the channel into the feature dimension, and then undergoes four residual modules to extract features and fuse them with the time information features. Accompanied by four downsampling, the self-attention mechanism operation is performed in the self-attention module of the high-level semantic features. The upsampling part structure is similar to the downsampling. Finally, the convolution output layer outputs the single-channel predicted high-resolution image, that is, the first predicted high-resolution image data corresponding to the original low-resolution brain image data.

[0089] S1033: According to the actual high-resolution brain image data corresponding to the original low-resolution brain image data and the first predicted high-resolution image data, and based on the preset loss function, adjust the parameters of the denoising network structure in the diffusion bridge model to construct an image reconstruction teacher network model.

[0090] In this embodiment, the preset loss function includes:

[0091] L diffusion =MSE(x0-F θ (x t , x T , t; θ)); Formula (VI)

[0092] Among them, L diffusion represents the preset loss function; MSE represents the mean square error function; x0 represents the actual high-resolution brain image data corresponding to the original low-resolution brain image data; F θ represents the denoising network structure in the image reconstruction teacher network model; x t Represents the noise sample image data of the current time step; x T Represents the original low-resolution brain image data; t represents the current time step; T represents the preset total time step; θ represents the learnable parameter.

[0093] In this embodiment, during the training process of the diffusion bridge model, according to the current time step t, the noise sample image data x at the current time step t and the original low-resolution brain image data x T, through the denoising network structure F θ , predict the first predicted high-resolution image data corresponding to F θ (x t , x T , t; θ). Subsequently, according to the preset loss function, namely formula (six), the difference between the first predicted high-resolution image data and the actual high-resolution brain image data is calculated, and the parameters of the denoising network structure are adjusted through backpropagation to minimize the global pixel value difference between the predicted output and the true high-resolution image. Specifically, during the training phase, the model optimizes the loss function through gradient descent and gradually updates the network weights, so that the denoising network structure can accurately recover high-resolution details from the noisy sample image data and the original low-resolution brain image data input, thereby improving the accuracy and quality of super-resolution reconstruction. This process is iteratively optimized, and ultimately enables the model to generate high-resolution brain images that both conform to the anatomical structure and retain key pathological features.

[0094] It's worth noting that traditional diffusion models gradually add noise to the target distribution, ultimately approximating a Gaussian distribution, and then use a learnable neural network to inversely denoise the generated image. However, in image reconstruction tasks, these models typically begin the generation process with pure Gaussian noise as the starting point for reconstruction, then repeatedly use a denoising network to remove some of the noise. This lacks structural information, and the low-resolution image is fed only as a conditional input to the denoising network. Gradually conditionally denoising the image, ultimately yielding a high-resolution image, can lead to poor consistency in the reconstruction results.

[0095] The present application adopts a diffusion bridge model, which has paired original low-resolution brain image data and actual high-resolution brain image data. In the forward process, the original low-resolution brain image data and its corresponding actual high-resolution brain image data are dynamically weighted and fused, and a diffusion bridge is directly modeled between them. The original low-resolution brain image data is directly integrated into the intermediate variable, and Gaussian noise controlled by the noise variance scheduling parameter is superimposed to generate noise sample image data, simulating the real image degradation process, so that the model can learn the complex mapping from low resolution to high resolution and enhance the model's robustness to noise. In the reverse process, the low-resolution image is used as the starting point, and denoising is continuously performed through the denoising network structure to finally obtain a high-resolution image. Compared with using pure noise as the starting point, the low-resolution image retains rich structural details, thus providing a more informative starting point for subsequent reconstruction, which helps to improve the quality and consistency of image reconstruction.

[0096] S104: Constructing an initial image reconstruction student network model of single-step mapping based on the image reconstruction teacher network model.

[0097] In this embodiment, after the diffusion bridge model is trained, an image reconstruction teacher network model is obtained. The image reconstruction teacher network model can achieve super-resolution conversion from low-resolution images to high-resolution images through multi-step iterations, but it usually takes 15-30 steps, which is time-consuming. Therefore, the knowledge distillation framework is used to optimize the denoising network structure in the image reconstruction teacher network model into a single-step super-resolution network to construct an initial image reconstruction student network model of single-step mapping. The initial image reconstruction student network model copies the denoising network structure in the image reconstruction teacher network model and inherits its weights as initial parameters.

[0098] S105: Based on the knowledge distillation optimization objective function, the initial image reconstruction student network model is optimized to obtain the ultimate image reconstruction student network model.

[0099] In this embodiment, if Figure 6 As shown in FIG, a schematic diagram of the architecture of the student network model optimization for initial image reconstruction according to an embodiment of the present invention is shown. Figure 7 FIG. 1 shows a flow chart of constructing the ultimate image reconstruction student network model according to an embodiment of the present invention. Based on the knowledge distillation optimization objective function, the initial image reconstruction student network model is optimized to obtain the ultimate image reconstruction student network model. The method includes:

[0100] S1051: Generating second predicted high-resolution image data corresponding to the original low-resolution brain image data according to the image reconstruction teacher network model and the diffusion inverse algorithm, wherein the method includes:

[0101] (1) According to the noise sample image data of the current time step, and based on the image reconstruction teacher network model and the diffusion inverse algorithm, the intermediate image data of the previous time step is calculated.

[0102] (2) Perform multiple inferences on the intermediate image data of the previous time step to generate second predicted high-resolution image data corresponding to the original low-resolution brain image data.

[0103] In this embodiment, according to the noise sample image data of the current time step, and based on the image reconstruction teacher network model and the diffusion inverse algorithm, the method of calculating the intermediate image data of the previous time step includes:

[0104]

[0105] Among them, x p Represents the intermediate image data of the previous time step; x t Represents the noise sample image data of the current time step; σ t represents the first previous time step factor; σ p represents the second previous time step factor; Fθ It represents the denoising network structure in the image reconstruction teacher network model; θ represents the learnable parameter; t represents the current time step; I represents Gaussian noise subject to a normal distribution.

[0106] In this embodiment, after training the diffusion bridge model, an image reconstruction teacher network model is obtained, and the reverse inference process from time step T to 0 is performed using the image reconstruction teacher network model. The preliminary predicted high-resolution image is directly calculated using the output at the current time step t. The preliminarily predicted high-resolution image is not the final image to be obtained, but is only estimated through the denoising network structure. Then, using Equation (VII), the intermediate image data at the previous time step p (p < t) is calculated, and based on the intermediate image data at the previous time step, multi-step inference (iteration) is performed until the current time step t = 0, when the second predicted high-resolution image data corresponding to the original low-resolution brain image data is obtained, and the inference ends.

[0107] In this embodiment, during the reverse inference process, constraints are combined with the low-resolution input, and denoising is continuously performed through the denoising network structure to ensure that the generated second predicted high-resolution image data is consistent with the original low-resolution brain image data in anatomical structure.

[0108] S1052: Determine the knowledge distillation optimization objective function according to the knowledge distillation loss function, the degradation loss function, and the cycle consistency loss function.

[0109] In this embodiment, the method for determining the knowledge distillation optimization objective function according to the knowledge distillation loss function, the degradation loss function, and the cycle consistency loss function includes:

[0110] L = L distill + L degrade + L cycle ; Equation (VIII)

[0111]

[0112] where L represents the knowledge distillation optimization objective function; L distill represents the knowledge distillation loss function; L degrade represents the degradation loss function; L cycle represents the cycle consistency loss function; MSE represents the mean square error function; represents the initial image reconstruction student network model; F θ represents the denoising network structure in the image reconstruction teacher network model; x T represents the original low-resolution brain image data; x0 represents the actual high-resolution brain image data corresponding to the original low-resolution brain image data; T represents the preset total time step; represents the predicted low-resolution image data generated by the initial image reconstruction student network model; θ represents the learnable parameters.

[0113] In this embodiment, a degradation loss function and a cycle consistency loss function are introduced to comprehensively improve the generalization ability and reconstruction quality of the student network model for initial image reconstruction. Specifically, the degradation loss function encourages the student network model for initial image reconstruction to not only focus on the reconstruction of high-resolution images, but also to have the ability to restore the original low-resolution images from the generated high-resolution images, thereby enhancing the network's understanding and modeling capabilities of structural features, reducing the risk of overfitting, avoiding focusing only on high-resolution reconstruction, and enhancing feature learning. The cycle consistency loss function further constrains the network, that is, when the degraded image is input again into the initial image reconstruction student network model for super-resolution reconstruction, it is required to restore the original high-resolution image as much as possible to enhance the robustness of the model. This mechanism effectively improves the fidelity of the image reconstruction results, while strengthening the mapping consistency between low-resolution images and high-resolution images, which helps to build a more stable and robust super-resolution model.

[0114] In this embodiment, the knowledge distillation optimization objective function is the weighted sum of the knowledge distillation loss function, the degradation loss function, and the cycle consistency loss function, which jointly optimize the initial image reconstruction student network model.

[0115] S1053: Using the original low-resolution brain image data and the second predicted high-resolution image data corresponding to the original low-resolution brain image data, the initial image reconstruction student network model is iteratively trained, and based on the knowledge distillation optimization objective function, the initial image reconstruction student network model is optimized to obtain the ultimate image reconstruction student network model.

[0116] In this embodiment, because the image reconstruction teacher network model has time perception capabilities, the total time step T is preset and fixed in the initial image reconstruction student network model. The image reconstruction teacher network model generates high-resolution images through multi-step iterations, while the initial image reconstruction student network model reconstructs through a single-step mapping. This constructs a knowledge distillation loss function to guide the initial image reconstruction student network model in learning the generative capabilities of the image reconstruction teacher network model. Furthermore, distillation learning can introduce some additional unpaired data. The image reconstruction teacher network model generates high-resolution prediction results through reverse reasoning, which serve as soft labels for the initial image reconstruction student network model, thereby guiding the learning of the initial image reconstruction student network model and enabling the model to learn more data samples. This semi-supervised training strategy increases the generalization of the model.

[0117] In this embodiment, the prediction results of the image reconstruction teacher network model are used to guide the training of the initial image reconstruction student network model. Combined with the knowledge distillation optimization objective function, the parameters of the initial image reconstruction student network model are iteratively optimized through gradient descent back propagation, and the performance of the image reconstruction teacher network model is gradually approached. The ultimate image reconstruction student network model generated inherits the reconstruction ability of the image reconstruction teacher network model, compresses the teacher network that requires multiple steps of reasoning into a student network that only requires one step of reasoning, and significantly reduces the computational complexity, making it suitable for real-time clinical applications.

[0118] S106: Reconstructing the acquired real-time low-resolution brain image data based on the ultimate image reconstruction student network model to generate target high-resolution image data.

[0119] In this embodiment, if Figure 8 As shown in FIG, a flow chart of image reconstruction using the ultimate image reconstruction student network model in an embodiment of the present invention is shown. θ Represents the ultimate image reconstruction student network model. Data preprocessing is performed on the acquired real-time low-resolution brain image data to ensure that it aligns with the distribution of the training data used to construct the ultimate image reconstruction student network model. Based on the preprocessed real-time low-resolution brain image data, the ultimate image reconstruction student network model can generate the target high-resolution image data in a single step, reducing model inference time.

[0120] It is worth noting that the image reconstruction method of the diffusion bridge model based on knowledge distillation in this application has the following advantages:

[0121] (1) Improved generation credibility and accuracy: The diffusion bridge model used in the forward process fuses the information of low-resolution images with that of high-resolution images, enabling the model to learn the complex mapping relationship from low resolution to high resolution and enhancing the model's robustness to noise. In the reverse process, the low-resolution image is used as the starting point, and denoising is continuously performed through the denoising network structure, finally obtaining a high-resolution image. This abandons the traditional Gaussian noise initial state and introduces structural consistency information, thereby improving the structural fidelity, credibility, and anatomical consistency of the reconstruction results.

[0122] (2) Reduce model inference time: Using the knowledge distillation training strategy, the image reconstruction teacher network model that requires multiple steps of inference is compressed into the ultimate image reconstruction student network model that only requires one step of inference. While ensuring image fidelity, it reduces the computation time and meets clinical real-time requirements. In addition, the image reconstruction student network model introduces some unpaired additional data during the training process. The image reconstruction teacher network model generates high-resolution prediction results through reverse inference, which serve as soft labels for the initial image reconstruction student network model, thereby guiding the learning of the initial image reconstruction student network model and enabling the model to learn more data samples. This semi-supervised training strategy increases the generalization of the model.

[0123] (3) Alleviating data scarcity: The diffusion bridge model can effectively utilize limited paired data (LR-HR) for training and enhance data expression capabilities through probability distribution learning. It solves the problem of low contrast between gray matter and white matter in neonatal brain tissue, which easily generates false textures. It enables the model to adapt to low-contrast scenes and generate high-resolution images that conform to the anatomical structure, alleviating the problem of scarce paired data in the field of medical imaging.

[0124] Furthermore, the image reconstruction method of the diffusion bridge model based on knowledge distillation in this application improves the detail expression of the image, making the anatomical structures such as the gray-white matter boundary of the brain, sulci and gyri clearer, which is helpful for the identification and clinical evaluation of early abnormalities. Higher-resolution images also enhance the accuracy of subsequent image registration and automatic segmentation, providing reliable support for developmental research and atlas construction. This technology can standardize thick-layer data collected from different centers into a unified high-resolution format, promote the integration and sharing of cross-center data, and enhance the comparability and wide applicability of multi-center research.

[0125] In the embodiments of this application, terms such as "first" and "second" are used to distinguish between identical or similar items with substantially the same function or effect. For example, the first predicted high-resolution image data and the second predicted high-resolution image data are used solely to distinguish between different predicted high-resolution image data and do not limit their order. Those skilled in the art will understand that terms such as "first" and "second" do not limit the quantity or execution order, and do not necessarily imply that they are different.

[0126] It should be noted that in the embodiments of this application, words such as "exemplary" or "for example" represent examples, illustrations, or descriptions. Any embodiment or design described in this application as "exemplary" or "for example" should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0127] In the embodiments of the present application, "at least one" refers to one or more, and "more" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: the existence of A alone, the existence of A and B at the same time, and the existence of B alone, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can represent: a, b, c, ab, ac, bc or abc, where a, b, c can be single or multiple.

[0128] Figure 9 It is a schematic block diagram of an image reconstruction system of a diffusion bridge model based on knowledge distillation provided in an embodiment of the present application.

[0129] like Figure 9 As shown, the image reconstruction system 900 of the diffusion bridge model based on knowledge distillation includes:

[0130] The data acquisition module 901 is used to acquire a plurality of brain thick slice image data and brain thin slice image data corresponding to each of the brain thick slice image data.

[0131] The data preprocessing module 902 is configured to perform data preprocessing on each of the brain thick slice image data and the brain thin slice image data corresponding to each of the brain thick slice image data to obtain a training data set.

[0132] The teacher network construction module 903 is used to train the diffusion bridge model using the training data set and adjust the parameters of the denoising network structure in the diffusion bridge model based on a preset loss function to construct an image reconstruction teacher network model.

[0133] The student network construction module 904 is used to construct an initial image reconstruction student network model of single-step mapping based on the image reconstruction teacher network model; based on the knowledge distillation optimization objective function, the initial image reconstruction student network model is optimized to obtain the ultimate image reconstruction student network model.

[0134] The image reconstruction module 905 is used to reconstruct the acquired real-time low-resolution brain image data based on the ultimate image reconstruction student network model to generate target high-resolution image data.

[0135] It should be understood that the specific process of each module executing the above corresponding steps has been described in detail in the above method embodiment, and for the sake of brevity, it will not be repeated here.

[0136] It should also be understood that the division of modules in the embodiments of the present application is illustrative and is merely a logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in the various embodiments of the present application may be integrated into a single processor, or may exist physically separately, or two or more modules may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or software functional modules.

[0137] Figure 10 : is a schematic block diagram of an electronic terminal provided in an embodiment of the present application. The electronic terminal includes a memory, a processor, and a computer program stored in the memory, and the processor executes the computer program to implement the image reconstruction method of the diffusion bridge model based on knowledge distillation as described above. Figure 10 As shown, the electronic terminal 1000 includes: at least one processor 1001, a memory 1002, at least one network interface 1003 and a user interface 1005. The various components in the device are coupled together via a bus system 1004. It is understood that the bus system 1004 is used to achieve connection and communication between these components. In addition to including a data bus, the bus system 1004 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 1004 is not described in detail. Figure 10 Various buses are labeled as bus systems.

[0138] The user interface 1005 may include a display, a keyboard, a mouse, a trackball, a click gun, keys, buttons, a touch pad or a touch screen.

[0139] It will be appreciated that the memory 1002 may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM) or a programmable read-only memory (PROM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static random access memory (SRAM) and synchronous static random access memory (SSRAM). The memory described in the embodiments of the present invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0140] The memory 1002 in the embodiment of the present invention is used to store various categories of data to support the operation of the electronic terminal 1000. Examples of such data include: any executable program for operating on the electronic terminal 1000, such as an operating system 10021 and an application 10022; the operating system 10021 includes various system programs, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks. The application 10022 can include various applications, such as a media player (Media Player), a browser (Browser), etc., for implementing various application services. The image reconstruction method based on the diffusion bridge model based on knowledge distillation provided in the embodiment of the present invention can be included in the application 10022.

[0141] The methods disclosed in the above embodiments of the present invention can be applied to or implemented by processor 1001. Processor 1001 may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by hardware integrated logic circuits in processor 1001 or by software instructions. The above processor 1001 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 1001 can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. The general-purpose processor 1001 may be a microprocessor or any conventional processor. The steps of the accessory optimization method provided in the embodiments of the present invention can be directly implemented and executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium located in a memory. The processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above method.

[0142] In an exemplary embodiment, the electronic terminal 1000 may be used by one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), and complex programmable logic devices (CPLDs) to execute the aforementioned method.

[0143] As used in this specification, the terms "component," "module," "system," and the like are used to represent computer-related entities, hardware, firmware, a combination of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, and / or a computer. By way of illustration, both an application running on a computing device and a computing device can be a component. One or more components can reside in a process and / or an execution thread, and a component can be located on a computer and / or distributed between two or more computers. In addition, these components can be executed from various computer-readable media having various data structures stored thereon. Components can communicate, for example, via local and / or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component on a local system, a distributed system, and / or a network, such as the Internet interacting with other systems via signals).

[0144] Those skilled in the art will appreciate that the various illustrative logical blocks and steps described in conjunction with the embodiments disclosed herein can be implemented using electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0145] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0146] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0147] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0148] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0149] In the above embodiments, the functions of each functional unit can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions (programs). When the computer program instructions (program) are loaded and executed on a computer, the process or function according to the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media integrated therein. Available media may be magnetic media (e.g., floppy disks, hard disks, tapes), optical media (e.g., high-density digital video discs (DVDs), or semiconductor media (e.g., solid state disks (SSDs)).

[0150] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and other media that can store program codes.

[0151] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

[0152] In summary, the present application provides an image reconstruction method, system and terminal based on a diffusion bridge model of knowledge distillation, which uses the diffusion bridge model to directly generate high-resolution images from low-resolution images, abandons the traditional Gaussian noise initial state, and improves the credibility and anatomical consistency of the reconstruction results. By utilizing the training strategy of knowledge distillation, the image reconstruction teacher network model that requires multiple steps of reasoning is compressed into an ultimate image reconstruction student network model that only requires one step of reasoning. While ensuring image fidelity, it reduces the computing time and meets clinical real-time requirements. Moreover, the image reconstruction student network model introduces some unpaired additional data during the training process, so that the model learns more data samples. This semi-supervised training strategy increases the generalization of the model. Therefore, the present application effectively overcomes the various shortcomings in the prior art and has a high industrial utilization value.

[0153] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Anyone skilled in the art may modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by one of ordinary skill in the art without departing from the spirit and technical concepts disclosed in this application shall be covered by the claims of this application.

Claims

1. An image reconstruction method based on a diffusion bridge model of knowledge distillation, characterized in that: include: Acquiring a plurality of brain thick-slice image data and brain thin-slice image data corresponding to each of the brain thick-slice image data; performing data preprocessing on each of the brain thick slice image data and the brain thin slice image data corresponding to each of the brain thick slice image data to obtain a training data set; Using the training data set to train a diffusion bridge model, and based on a preset loss function, adjusting the parameters of the denoising network structure in the diffusion bridge model to construct an image reconstruction teacher network model; According to the image reconstruction teacher network model, an initial image reconstruction student network model of single-step mapping is constructed; Based on the knowledge distillation optimization objective function, the initial image reconstruction student network model is optimized to obtain the ultimate image reconstruction student network model; Based on the ultimate image reconstruction student network model, the acquired real-time low-resolution brain image data is reconstructed to generate target high-resolution image data.

2. The image reconstruction method of the diffusion bridge model based on knowledge distillation according to claim 1 is characterized in that: The training data set includes multiple data pairs, each data pair includes original low-resolution brain image data and actual high-resolution brain image data corresponding to the original low-resolution brain image data; wherein the method of constructing the image reconstruction teacher network model includes: Inputting the original low-resolution brain image data and actual high-resolution brain image data corresponding to the original low-resolution brain image data into the diffusion bridge model for training, and generating noisy sample image data based on a diffusion forward algorithm; Invoking a denoising network structure in a diffusion bridge model according to the noise sample image data and the original low-resolution brain image data to generate first predicted high-resolution image data corresponding to the original low-resolution brain image data; According to the actual high-resolution brain image data corresponding to the original low-resolution brain image data and the first predicted high-resolution image data, and based on the preset loss function, the parameters of the denoising network structure in the diffusion bridge model are adjusted to construct an image reconstruction teacher network model.

3. The image reconstruction method of the diffusion bridge model based on knowledge distillation according to claim 2 is characterized in that: Inputting the original low-resolution brain image data and actual high-resolution brain image data corresponding to the original low-resolution brain image data into the diffusion bridge model for training, and generating noise sample image data based on a diffusion forward algorithm includes: q t (x) t |x0,x T )=N(x t ;μ t (x0, x T ),∑t); Among them, q t (x t |x0,x T ) represents conditional probability; N represents normal distribution; x t represents the noise sample image data of the current time step; x0 represents the actual high-resolution brain image data corresponding to the original low-resolution brain image data; x T represents the original low-resolution brain image data; μ t represents the mean of the noise sample image data; ∑t represents the variance of the noise sample image data; t represents the current time step; T represents the preset total time step.

4. The image reconstruction method of the diffusion bridge model based on knowledge distillation according to claim 2 is characterized in that: Based on the knowledge distillation optimization objective function, the initial image reconstruction student network model is optimized to obtain the ultimate image reconstruction student network model. The method includes: generating second predicted high-resolution image data corresponding to the original low-resolution brain image data based on the image reconstruction teacher network model and the diffusion inverse algorithm; Determine the knowledge distillation optimization objective function based on the knowledge distillation loss function, degradation loss function, and cycle consistency loss function; The initial image reconstruction student network model is iteratively trained using the original low-resolution brain image data and the second predicted high-resolution image data corresponding to the original low-resolution brain image data, and the initial image reconstruction student network model is optimized based on the knowledge distillation optimization objective function to obtain the ultimate image reconstruction student network model.

5. The image reconstruction method of the diffusion bridge model based on knowledge distillation according to claim 4 is characterized in that: The method of generating the second predicted high-resolution image data corresponding to the original low-resolution brain image data includes: According to the noise sample image data of the current time step, and based on the image reconstruction teacher network model and the diffusion inverse algorithm, the intermediate image data of the previous time step is calculated; Multiple inferences are performed on the intermediate image data of the previous time step to generate second predicted high-resolution image data corresponding to the original low-resolution brain image data.

6. The image reconstruction method of the diffusion bridge model based on knowledge distillation according to claim 4 is characterized in that: The method of determining the knowledge distillation optimization objective function according to the knowledge distillation loss function, the degradation loss function, and the cycle consistency loss function includes: L=L distill +L degrade +L cycle ; Among them, L represents the knowledge distillation optimization objective function; L distill represents the knowledge distillation loss function; L degrade represents the degradation loss function; L cycle represents the cycle consistency loss function; MSE represents the mean square error function; represents the initial image reconstruction student network model; F θ represents the denoising network structure in the image reconstruction teacher network model; x T represents the original low-resolution brain image data; x0 represents the actual high-resolution brain image data corresponding to the original low-resolution brain image data; T represents the preset total time step; represents the predicted low-resolution image data generated by the initial image reconstruction student network model; θ represents the learnable parameters.

7. The image reconstruction method of the diffusion bridge model based on knowledge distillation according to claim 5, characterized in that: According to the noise sample image data of the current time step, and based on the image reconstruction teacher network model and the diffusion inverse algorithm, the method of calculating the intermediate image data of the previous time step includes: Among them, x p Represents the intermediate image data of the previous time step; x t Represents the noise sample image data of the current time step; σ t represents the first previous time step factor; σ p represents the second previous time step factor; F θ represents the denoising network structure in the image reconstruction teacher network model; θ represents the learnable parameters; t represents the current time step; I represents Gaussian noise that obeys the normal distribution.

8. The image reconstruction method of the diffusion bridge model based on knowledge distillation according to claim 2, characterized in that: The preset loss function includes: L diffusion =MSE(x0-F θ (x t ,x T ,t;θ))); Among them, L diffusion represents the preset loss function; MSE represents the mean square error function; x0 represents the actual high-resolution brain image data corresponding to the original low-resolution brain image data; F θ represents the denoising network structure in the image reconstruction teacher network model; x t Represents the noise sample image data of the current time step; x T represents the original low-resolution brain image data; t represents the current time step; T represents the preset total time step; θ represents the learnable parameter.

9. An image reconstruction system based on a diffusion bridge model of knowledge distillation, characterized in that: include: a data acquisition module, configured to acquire a plurality of thick-slice brain image data and thin-slice brain image data corresponding to each of the thick-slice brain image data; a data preprocessing module, configured to perform data preprocessing on each of the brain thick slice image data and the brain thin slice image data corresponding to each of the brain thick slice image data, to obtain a training data set; a teacher network construction module, configured to train a diffusion bridge model using the training data set and, based on a preset loss function, adjust parameters of a denoising network structure in the diffusion bridge model to construct an image reconstruction teacher network model; A student network construction module is used to construct an initial image reconstruction student network model of single-step mapping based on the image reconstruction teacher network model; based on the knowledge distillation optimization objective function, the initial image reconstruction student network model is optimized to obtain the ultimate image reconstruction student network model; The image reconstruction module is used to reconstruct the acquired real-time low-resolution brain image data based on the ultimate image reconstruction student network model to generate target high-resolution image data.

10. An electronic terminal comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the image reconstruction method based on the diffusion bridge model of any one of claims 1 to 8.

Images