Brain mri super-resolution reconstruction method and system based on implicit neural representation

Abstract

The disclosure provides a brain MRI super-resolution reconstruction method and system based on implicit neural representation, and relates to the technical field of super-resolution image processing. A denoising diffusion probability model adopts a U-Net network framework, a multi-scale low-resolution guide network is introduced in an encoder to extract multi-scale features of a low-resolution MRI image, and the multi-scale features of the low-resolution MRI image are fused to output high-frequency textures and detailed features. The decoder part of the U-Net of the denoising diffusion probability model is replaced by an implicit neural representation. By introducing an implicit image function, a continuous resolution representation of a low-resolution MRI image is obtained, and an arbitrary scale super-resolution MRI image is output by using the obtained continuous resolution representation.

Description

Technical Field

[0001] This disclosure relates to the field of super-resolution image processing technology, specifically to a method and system for super-resolution reconstruction of brain MRI based on implicit neural representations. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Magnetic Resonance Imaging (MRI) is a commonly used medical imaging technique in clinical practice. The resulting three-dimensional MR images provide rich details of tissue anatomy, facilitating accurate early diagnosis of diseases such as tumors. However, in clinical applications, high-quality MRI images are difficult to obtain due to limitations in scanning equipment costs, patient non-cooperation, and scan time. Furthermore, noise interference and insufficient sampling can also lead to image quality degradation during medical image transmission and storage. Therefore, improving the quality and resolution of acquired MRI images has become a research hotspot. Since super-resolution (SR) can improve image quality without changing MRI hardware, this post-processing technique has been widely used to overcome the obstacles to obtaining high-quality MRI images.

[0004] In recent years, deep learning-based systems have shown great potential in MRI SR. Deep generative models, including autoregressive, GAN-based, and procedural models, have become solutions for generating MRI images with rich textures. However, they often suffer from artifacts and are only applicable to predefined scaling factors. To address these issues, many systems have been explored for solving SR tasks at arbitrary scales in natural images, including Meta-SR and LIIF. However, these systems often fail to produce the high-fidelity details required for high magnification.

[0005] Currently, the Denoising Diffusion Probabilistic Model (DDPM) has attracted widespread research interest due to its advantages over GAN-based systems. DDPM excels at generating high-quality images that closely match the distribution of the training data, demonstrating impressive performance in improving the fidelity of MRI images.

[0006] However, the inventors discovered that DDPM-based SR systems are limited to processing specific scaling factors (such as 2×, 3×, and 4×), and changes in the scaling factor lead to a degraded output quality and a significant loss of texture and detail. Therefore, to obtain high-quality SR images at different scaling factors, these systems need to be trained and stored separately for each specific scale, which imposes expensive storage and energy costs on medical professionals who may require SR images with varying scaling factors. Summary of the Invention

[0007] To address the aforementioned issues, this disclosure proposes a method and system for super-resolution reconstruction of brain MRI based on implicit neural representations. By integrating implicit neural representations into a denoising diffusion probability model, high-fidelity brain MRI images at arbitrary scales can be achieved. The image is encoded as a function in a continuous space using implicit neural representations. After integration into the diffusion model, this function is parameterized by a coordinate-based multi-layer perceptron (MLP), thereby enabling more accurate capture of continuous image resolution representations and generating high-quality MRI images with continuous resolution.

[0008] According to some embodiments, the present disclosure adopts the following technical solutions:

[0009] Brain MRI super-resolution reconstruction methods based on implicit neural representations include:

[0010] Acquire low-resolution MRI images, input the low-resolution MRI images into a denoising diffusion probability model, and output super-resolution reconstructed MRI images of arbitrary scale;

[0011] The denoising diffusion probability model employs the U-Net network framework, introducing a multi-scale low-resolution guiding network in the encoder to extract multi-scale features from low-resolution MRI images. These multi-scale features are then fused to output high-frequency texture and detail features. The decoder portion of the U-Net in the denoising diffusion probability model is replaced by implicit neural representations. By introducing implicit image functions, continuous resolution representations of low-resolution MRI images are obtained. Using these continuous resolution representations, super-resolution MRI images of arbitrary scales are output.

[0012] According to some embodiments, the present disclosure adopts the following technical solutions:

[0013] Brain MRI super-resolution reconstruction systems based on implicit neural representations include:

[0014] The data acquisition module is used to acquire low-resolution MRI images;

[0015] The image reconstruction module is used to input the low-resolution MRI image into the denoising diffusion probability model and output a super-resolution reconstructed MRI image of arbitrary scale.

[0016] The denoising diffusion probability model employs the U-Net network framework, introducing a multi-scale low-resolution guiding network in the encoder to extract multi-scale features from low-resolution MRI images. These multi-scale features are then fused to output high-frequency texture and detail features. The decoder portion of the U-Net in the denoising diffusion probability model is replaced by implicit neural representations. By introducing implicit image functions, continuous resolution representations of low-resolution MRI images are obtained. Using these continuous resolution representations, super-resolution MRI images of arbitrary scales are output.

[0017] According to some embodiments, the present disclosure adopts the following technical solutions:

[0018] A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned brain MRI super-resolution reconstruction method based on implicit neural representations.

[0019] According to some embodiments, the present disclosure adopts the following technical solutions:

[0020] An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the brain MRI super-resolution reconstruction method based on implicit neural representation.

[0021] Compared with the prior art, the beneficial effects of this disclosure are as follows:

[0022] This disclosure presents a brain MRI super-resolution reconstruction method based on implicit neural representation, combining implicit neural representation with a denoising diffusion probability model to achieve high-fidelity medical image super-resolution at any scale. Furthermore, a continuous resolution adjustment mechanism is developed, including a multi-scale low-resolution guide network and a scaling factor. The scaling factor can finely adjust the resolution, dynamically regulating the proportion of low-resolution information and synthetic features in the final output, enabling the model to seamlessly meet the needs of continuous resolution adjustment. In addition, the multi-scale low-resolution guide network provides the denoising model with multi-scale low-resolution features to enrich texture information and restore high-frequency details. By combining the advantages of both models, radiologists can more fully visualize lesions, improving the accuracy of medical diagnosis.

[0023] This disclosure achieves high-fidelity brain MRI images at arbitrary scales by integrating implicit neural representations into a denoising diffusion probability model. Leveraging the appealing properties of implicit neural representations, the image is encoded as a function in a continuous space. After integration into the diffusion model, this function is parameterized by a coordinate-based multilayer perceptron (MLP), enabling more accurate capture of continuous image resolution representations. This generates high-quality MRI images with continuous resolution, aiding physicians in diagnosing brain tumors. Attached Figure Description

[0024] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0025] Figure 1 This is a schematic diagram of the reverse diffusion process of the brain MRI super-resolution reconstruction system according to an embodiment of the present disclosure;

[0026] Figure 2 This is a detailed architectural diagram of the t-th iteration of the back diffusion according to an embodiment of the present disclosure;

[0027] Figure 3 The present disclosure provides visualizations of recovery results, error mappings, and magnified local images of different systems on a fastMRI dataset, representing embodiments of this invention.

[0028] Figure 4 This disclosure provides visualizations of recovery results, error mappings, and magnified local images of different systems on the IXI dataset, representing embodiments of the present disclosure.

[0029] Figure 5 Visualization results of different scaling factors s in embodiments of this disclosure;

[0030] The first row displays the reconstructed t2-weighted image, the middle row displays a detailed image of the red-framed area magnified from the first row, and the bottom rows display the error graph between the ground truth and the reconstructed image.

[0031] Figure 6 Visualization results of three systems incorporating low-resolution features according to embodiments of this disclosure;

[0032] The top row shows the complete reference image and the t1-weighted images reconstructed using three different adjustment mechanisms; the bottom row illustrates the differences between the reference image and the reconstructed image. Detailed Implementation

[0033] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0034] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0035] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0036] Example 1

[0037] One embodiment of this disclosure provides a method for super-resolution reconstruction of the brain based on implicit neural representation using MRI, including:

[0038] Step 1: Acquire low-resolution MRI images and input the low-resolution MRI images into the denoising diffusion probability model;

[0039] Step 2: Output super-resolution reconstructed MRI images at any scale;

[0040] The denoising diffusion probability model employs the U-Net network framework, introducing a multi-scale low-resolution guiding network into the encoder to extract multi-scale features from low-resolution MRI images. These multi-scale features are then fused to output high-frequency texture and detail features, specifically referring to minute lesions in the image. The decoder portion of the U-Net in the denoising diffusion probability model is replaced by implicit neural representations. By introducing implicit image functions, continuous resolution representations of the low-resolution MRI images are obtained. Using these continuous resolution representations, super-resolution MRI images of arbitrary scales are output.

[0041] As one embodiment, the specific implementation details of the brain MRI super-resolution reconstruction method based on implicit neural representation disclosed herein are as follows:

[0042] This disclosure integrates implicit neural representations into a denoised diffusion probability model to achieve high-fidelity brain MRI images of arbitrary scales.

[0043] Specifically, using the U-Net network architecture as a denoising block, U-Net denoises noisy low-resolution MRI images. t Encoding as multi-scale feature maps o (i), where i∈{1, ..., W}, and W represents the depth count within the U-Net backbone. Furthermore, an implicit image function is introduced into the U-Net decoder, capable of generating high-quality, resolution-continuous images. This denoising diffusion probability model unifies the iterative diffusion refinement process and the implicit image function within an end-to-end framework.

[0044] This disclosure employs a continuous resolution adjustment mechanism, including a multi-scale low-resolution guiding network and scaling factor modulation.

[0045] The multi-scale low-resolution guided network is introduced into the encoder of the U-Net network, utilizing a convolutional neural network (CNN) as a conditional network to extract multi-resolution conditional features from low-resolution MRI images. It consists of stacked convolutional layers with bilinear filtering downsampling operations and leaky ReLU (Rectified Linear Activation Function) activation.

[0046] Furthermore, a multi-scale low-resolution guided network is introduced into the encoder to extract multi-scale features from low-resolution MRI images. These multi-scale features are then fused to output high-frequency texture and detail features, including:

[0047] First, EDSR is used to extract the original features l from the low-resolution MRI image x. (0) On the one hand, the original features are compared with the MRI image y at step t. t Combined with the input to the U-Net encoder, on the one hand, the original features l (0) The input is fed into a multi-scale low-resolution guided network to obtain multi-scale low-resolution features.

[0048] An initial low-resolution feature f(0) is established using EDSR (Enhanced Deep Residual Networks), and its resolution is made the same as yt using bilinear interpolation. Then, l (0) With y t Connect them together and input the merged result into U-Net to provide initial guidance. Meanwhile, l (0) It is also sent to the CNN, where the feature is progressively downsampled to:

[0049] l (i) =Convo(l (i-1) )

[0050] Among them, l (i-1) It was obtained through downsampling.

[0051] "Convo" indicates a convolutional layer containing a bilinear filtering downsampling operation and a leaky ReLU activation. Notably, the guiding network focuses solely on providing the encoded multi-resolution features. These features are then directly fed into the U-Net without any additional priors, enabling the modeling of the latent representation without relying on additional assumptions or constraints.

[0052] Furthermore, the multi-scale low-resolution features and the output of U-Net are then combined in a certain proportion, and the proportion can be dynamically adjusted by the scaling factor s, which can ensure the fidelity of the final output of the model.

[0053] Scaling factor modulation is responsible for exceeding the limitations of a specific magnification scale. A scaling factor, denoted as 's', is added during the diffusion process. This factor helps to continuously improve resolution and facilitates flexible magnification possibilities.

[0054] First, a value range from 1 to Q is established, where Q represents the upper limit of the scaling factor. During training, a value of s is randomly selected from this range. Then, y is reshaped according to the selected scaling factor s. t This allows control over the resolution of the generated image. The importance of the scaling factor s lies in its ability to finely balance the initial input data l from the guidance network. (i) and the output from the denoising block o (i) The ratio between them. Compared to systems such as cross-attention mechanisms or connection operations, the adaptive MLP (Multilayer Perceptron) maps the scaling factor s to a set of scaling vectors, represented as... Here, i represents the depth index from the different resolution outputs of the guidance network and the denoising block.

[0055] Subsequently, the L2 norm was used to pair and Normalize them, and then use them for channel modulation. (i) and o (i) These components are dynamically merged in an adaptive manner. Essentially, the modulation process is combined with a scaling factor s, as summarized below:

[0056]

[0057]

[0058]

[0059]

[0060] in, and These originate from implicit neural inhibition and noise reduction blocks, respectively.(i) It is the modulation result, l(i) are all multi-scale features obtained by downsampling the original features l(0).

[0061] In this modulation process, a small constant τ = 1e-8 is introduced to prevent zero denominator. (Characteristic diagram) and These originate from implicit neural inhibition and noise reduction blocks, respectively. (i) It is the result of modulation.

[0062] Furthermore, in order to obtain continuous resolution representations of low-resolution MRI images, implicit image functions are introduced, utilizing implicit neural representations to obtain continuous MRI image representations, thus simplifying the entire process.

[0063] Multiple MLPs are integrated into the upsampling stage of the U-Net architecture. These MLPs are used to define parameters of implicit neural representations, facilitating the recovery of low-resolution MRI images with higher fidelity across continuous scales. The continuous coordinates of the hypothetical multi-resolution features are used, denoted as g = {g...} (1) ,...,g (i) ,...g (N)} for reference. These coordinates are obtained from the MLP using a scaling factor s. In the implicit representation, indicator coordinates (modulation result z) are provided. (i) The modulation features around the corresponding coordinates are used as input to determine the corresponding target features.

[0064]

[0065] In the above formula, F i This represents a 2-layer MLP (Multilayer Perceptron) with a hidden dimension of 256.

[0066] Wherein, Fi is a 2-layer MLP with a hidden dimension of 256. and By calculating the distance z respectively (i+1) and g (i+1) The nearest Euclidean distance is used for interpolation.

[0067] Then, the design predicts the super-resolution image y0 through a series of denoising iterations, optimizes the denoising model, and effectively removes noise from the image y0. t The goal is to recover the super-resolution image y0. Furthermore, the objective is to obtain a continuous resolution output; therefore, the denoising model ∈ θ (x, t, s, y) t γ t During training, it is necessary to adapt to any scale while maintaining projection noise ∈ θ The accuracy. In summary, the goal of fine-tuning the denoising network is as follows:

[0068]

[0069] Where ∈ ~ N(0, I), t ~ {1, ..., t}, s ~ U(1, M), and (x, y) are sampled from the training set of low-resolution-high-resolution image pairs. Projected noise ∈ follows a normal distribution, where N represents a normal distribution. t represents the t-th diffusion step, T represents the total diffusion steps, s is a scaling factor with a size ranging from 1 to M, x represents the low-resolution MRI, and y represents the high-resolution MRI. The denoising model ∈ θ This model is equivalent to processing a noisy target image y. t The target image y0 is recovered. Meanwhile, to achieve continuous resolution output, the denoising model ∈ θ It should be applied to arbitrary scales through training, while ensuring that the projected noise ∈ . x refers to the low-resolution image, t refers to the diffusion at step t, and s is the scaling factor.

[0070] As one embodiment, the training process of the denoising diffusion probability model described in this disclosure includes:

[0071] S1. Establishing a Medical Image Super-Resolution Dataset: When using a single classical degradation model to synthesize a training dataset, it differs significantly from real-world medical image degradation, resulting in slightly insufficient LR image reconstruction. Therefore, multiple image degradation models are introduced and combined to perform image degradation processing on the original high-resolution medical images, obtaining corresponding low-resolution images. This simulates the low-resolution images acquired in real-world situations, and such an image pair is used as input.

[0072] S2. Design of a Medical Image Super-Resolution Algorithm: Low-resolution and high-resolution medical images are used as training pairs, with a denoising diffusion model incorporating implicit neural representations as the framework. Improvements were made to the encoder and decoder parts of the U-Net denoising diffusion model. First, a multi-scale low-resolution guiding network was introduced into the encoder part. This network can acquire multi-scale features from low-resolution MRI images and fuse these extracted features into the U-Net encoder, making the model's output contain more high-frequency textures and details. Second, the decoder part of the U-Net was replaced with implicit neural representations. Implicit neural representations can learn the continuous resolution function of MRI images, and using the learned continuous resolution function, MRI images of any scale can be output.

[0073] S3. Training the Medical Image Super-Resolution Model: A model is built based on the algorithm designed in S2 and trained using the PyTorch deep learning framework. The training process consists of two stages. The first stage trains a peak signal-to-noise ratio guided model with L1 loss. The batch size is 8, the epoch is set to 100, and the learning rate is initialized to 1×10⁻⁶.-4 After iteration, the learning rate is reduced by half, and the iteration continues until the loss converges, resulting in the final model.

[0074] S4. Input the low-resolution medical images from the test set into the model to obtain the corresponding super-resolution output results of the medical images.

[0075] In this embodiment, the brain MRI super-resolution reconstruction method based on the diffusion model and implicit neural representation described above is used to output high-fidelity continuous resolution MRI images, specifically including the following steps:

[0076] like Figure 1 The reverse denoising process is shown. T The image represents the noisy image obtained by progressively adding noise to a high-resolution MRI image. t represents the t-th step of the backdiffusion process, s represents the scaling factor, and also indicates the super-resolution factor of the final output. x is the low-resolution image obtained by degrading the high-resolution image, and g represents a set of vector coordinates obtained by passing s through multiple MLP layers. At each step of the backdiffusion denoising process, a multi-scale low-resolution guided network extracts multi-resolution features from the low-resolution MRI image x. These extracted low-resolution features are then used as conditions in the denoising module. This is done to ensure that the final super-resolution result contains richer textures and more high-frequency details, assisting doctors in making more accurate judgments.

[0077] like Figure 2 As shown, the MRI image y from step t is presented. t MRI image y at step t-1 t-1 The detailed architecture used in the process. First, EDSR is used to extract raw features l from the low-resolution MRI image x. (0) On the one hand, the original features are compared with the MRI image y at step t. t Combined with the input to the U-Net encoder, on the one hand, the original features l (0) The input is fed into a multi-scale low-resolution guided network to obtain multi-scale low-resolution features. These features are then combined with the output of U-Net in a specific ratio, which can be dynamically adjusted using a scaling factor s to ensure the fidelity of the final model output.

[0078] In this embodiment, MRI super-resolution experiments were conducted on the IXI and fastMRI datasets. The IXI dataset is a publicly accessible brain image MRI dataset. This dataset consists of real brain MRI images in three modalities: t1-weighted, t2-weighted, and PD-weighted images. Brain imaging regions were extracted from .nii files to obtain relevant and sufficient regions for analysis. Furthermore, to evaluate the effectiveness of the model, 578 healthy brain MRI scans from the IXI dataset were divided into 500 for training, 6 for validation, and 70 for testing. Application center cropping converted each MRI into a new scan image with a resolution of 256×256 pixels per slice. The fastMRI dataset is also a publicly accessible MRI dataset. In the experiments, t2-weighted and PD-weighted brain images were selected from the fastMRI dataset. Similar to the preprocessing of the IXI dataset, a total of 584, 168, and 80 pairs of volumes were prepared. Each volume pair was sliced ​​into 24 pieces, resulting in 14016, 4032, and 1920 pairs of 2D images (256×256) for training, testing, and validation sets, respectively.

[0079] The model disclosed herein was implemented using the PyTorch framework on an NVIDIA Tesla A100 GPU. A 1Q iteration milestone was established for training with a consistent Q× downscaling factor. After the milestone, HR images were stochastically adjusted using a uniformly distributed U(1,Q) for further training with 0.5Q iterations. The Adam optimizer was used, consistent with traditional DDPM. The learning rate for DDPM was set to 1×10⁻⁶. -4 The learning rate of this model is set to 2×10. -5 The batch size is set to 8.

[0080] The proposed system and other state-of-the-art systems were quantitatively evaluated using two widely used metrics: PSNR and SSIM. The model was compared with several recent state-of-the-art systems, including five specific integer-scale super-resolution systems: Bicubic, SwinIR, Guided Diffusion, TTSR, and MASA-SR, and two arbitrary-scale image SR systems: Meta-SR and LIIIF. To evaluate their effectiveness at different scale factors in super-resolution tasks, a two-stage up-and-down strategy was employed: in the upsampling stage, a well-trained model with the highest scale was used to generate an over-magnified SR image. Subsequently, in the downsampling stage, the oversampled super-resolution image was appropriately thinned using bilateral interpolation. Finally, the comparison results and corresponding data of the system described in this embodiment with other systems are as follows: Figure 3 and Figure 4 And as shown in Table 1 below.

[0081] Table 1 shows the PSNR and SSIM values ​​for each system at 2x and 4x super-resolution on the test set.

[0082]

[0083] To demonstrate the superiority of our model over state-of-the-art arbitrary-scale super-resolution systems, we conducted quantitative comparisons with Meta-SR and LIIF on the IXI dataset. Meta-SR, LIIF, and our model were trained within an upgrade range of (1,4) and tested at both in-distribution and out-of-distribution scales. As shown in Table 2, we consistently achieved state-of-the-art results on the IXI dataset. Importantly, for out-of-distribution scales, our model even significantly outperforms existing systems. These results confirm the superiority and generality of our framework over state-of-the-art systems.

[0084] Table 2 shows the quantitative evaluation of continuous super-resolution results on the IXI dataset, obtained by training LR-HR pairs at an 8x scale.

[0085]

[0086] This example demonstrates the importance of the scaling factor. To emphasize the importance of the scaling factor, examples from... Figure 5 Qualitative visual results for 8x super-resolution T2 images from the IXI dataset. Specifically, different s values ​​were assigned to other specific magnification factors. For example, Figure 5 The third column corresponds to a 4.1x T2 image super-resolution output. Clearly, for an 8x T2 image super-resolution output, using a smaller scaling factor with a smaller magnification factor results in blurred textures. As the corresponding magnification factor increases, finer details are synthesized more effectively. This indicates that, in the context of high-magnification super-resolution, the scaling factor tends to assign more weight to the synthesized features. Essentially, the scaling factor cleverly balances scaling features and synthesized details.

[0087] To verify the effectiveness of the multi-scale low-resolution guided network, qualitative experiments were conducted on T1 images from the IXI dataset (8x scale) and on T2-weighted brain images from the fastMRI dataset. Specifically, two comparative models were constructed by replacing the multi-scale low-resolution guided network in the system with two different modulation mechanisms: (1) directly sampling the LR images, or (2) encoding them using an EDSR encoder and then inputting them into the denoising block.

[0088] like Figure 6 As shown in the second column, directly using upsampled low-resolution MRI images as conditioned images often leads to texture blurring. Although pre-incorporating an encoder to extract features mitigates this concern to some extent (e.g., ... Figure 6 (As shown in the third column), but it still exhibits limited ability to generate high-fidelity details, such as edges. In stark contrast, the introduced multi-scale low-resolution guided network employs a parallel architecture, endowing the denoising model with multi-resolution LR features and enhancing the richness of texture information. From Figure 6 As indicated by the middle arrow, it can be observed that, in terms of detail, the multi-scale low-resolution guided network more closely approximates the real ground image compared to other systems, with higher PSNR and SSIM values ​​as shown in Table 3. These observations confirm that the proposed multi-scale LR guided network provides key multi-resolution features for guiding the recovery of details and textures.

[0089] Table 3 Quantitative results of ablation studies on the IXI and fastMRI datasets.

[0090]

[0091] Example 2

[0092] One embodiment of this disclosure provides a brain MRI super-resolution reconstruction system based on implicit neural representations, comprising:

[0093] The data acquisition module is used to acquire low-resolution MRI images;

[0094] The image reconstruction module is used to input the low-resolution MRI image into the denoising diffusion probability model and output a super-resolution reconstructed MRI image of arbitrary scale.

[0095] The denoising diffusion probability model employs the U-Net network framework, introducing a multi-scale low-resolution guiding network in the encoder to extract multi-scale features from low-resolution MRI images. These multi-scale features are then fused to output high-frequency texture and detail features. The decoder portion of the U-Net in the denoising diffusion probability model is replaced by implicit neural representations. By introducing implicit image functions, continuous resolution representations of low-resolution MRI images are obtained. Using these continuous resolution representations, super-resolution MRI images of arbitrary scales are output.

[0096] Example 3

[0097] One embodiment of this disclosure provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the brain MRI super-resolution reconstruction method based on implicit neural representations.

[0098] Example 4

[0099] One embodiment of this disclosure provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the brain MRI super-resolution reconstruction method based on implicit neural representation.

[0100] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0101] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0102] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A brain MRI super-resolution reconstruction method based on implicit neural representation, characterized in that, include: Acquire low-resolution MRI images, input the low-resolution MRI images into a denoising diffusion probability model, and output super-resolution reconstructed MRI images of arbitrary scale; The denoising diffusion probability model employs the U-Net network framework, introducing a multi-scale low-resolution guiding network into the encoder to extract multi-scale features from low-resolution MRI images. These multi-scale features are then fused to output high-frequency texture and detail features. The decoder portion of the U-Net in the denoising diffusion probability model is replaced by implicit neural representations. By introducing implicit image functions, continuous resolution representations of the low-resolution MRI images are obtained. Using these continuous resolution representations, super-resolution MRI images of arbitrary scales are output. Utilizing implicit neural representations to obtain continuous resolution representations of low-resolution MRI images includes: integrating multiple MLPs into the upsampling stage of a U-Net diffusion network framework, where the MLPs are used to define the parameters of the implicit neural representations to facilitate the recovery of low-resolution MRI images with higher fidelity across continuous scales; using the continuous coordinates of hypothetical multi-resolution features as a reference, where the continuous coordinates are obtained from the MLPs using a scaling factor; and providing modulation features around the indicator coordinates as input during the implicit representation process to determine the corresponding target features. The objective is to obtain continuously variable resolution outputs, and to adapt the diffusion network framework to arbitrary scales. The defined objective is: Among them, the denoising model This model is equivalent to processing a noisy target image y. t The target image y0 is recovered in the middle. At the same time, in order to achieve continuous resolution output, the denoising model... It should be trained to be applicable to any scale while ensuring projection noise. x refers to the low-resolution image, t refers to the diffusion at step t, and s is the scaling factor.

2. The brain MRI super-resolution reconstruction method based on implicit neural representation as described in claim 1, characterized in that, The multi-scale low-resolution guided network uses a convolutional neural network as a conditional network to extract multi-scale features from low-resolution MRI images, including convolutional layer stacking, bilinear filtering downsampling, and activation functions.

3. The brain MRI super-resolution reconstruction method based on implicit neural representation as described in claim 1, characterized in that, The decoder portion of the denoising diffusion probability model is replaced by implicit neural representations. By introducing implicit image functions, continuous resolution representations of low-resolution MRI images are obtained, including: In the process of acquiring continuous resolution characterization of low-resolution MRI images, the resolution is continuously improved by scaling factor modulation. A scaling factor is added during the diffusion process, and the low-resolution MRI image is reshaped according to the selected scaling factor, so that the resolution of the generated image can be controlled.

4. The brain MRI super-resolution reconstruction method based on implicit neural representation as described in claim 3, characterized in that, The scaling factor balances the ratio between the initial input low-resolution MRI image and the output multi-scale features. An adaptive MLP is used to map the scaling factor to a set of scaling vectors, the scaling vectors are normalized using the L2 norm, and then channel modulation is performed to dynamically merge them in an adaptive manner.

5. The brain MRI super-resolution reconstruction method based on implicit neural representation as described in claim 1, characterized in that, Using the U-Net network as the architecture of the denoising diffusion probability model, and using the U-Net network as the denoising block, can enable the implementation of a continuous resolution adjustment mechanism, which includes a multi-scale low-resolution guiding network and scaling factor modulation.

6. A brain MRI super-resolution reconstruction system based on implicit neural representations, characterized in that, include: The data acquisition module is used to acquire low-resolution MRI images; The image reconstruction module is used to input the low-resolution MRI image into the denoising diffusion probability model and output a super-resolution reconstructed MRI image of arbitrary scale. The denoising diffusion probability model employs the U-Net network framework, introducing a multi-scale low-resolution guiding network into the encoder to extract multi-scale features from low-resolution MRI images. These multi-scale features are then fused to output high-frequency texture and detail features. The decoder portion of the U-Net in the denoising diffusion probability model is replaced by implicit neural representations. By introducing implicit image functions, continuous resolution representations of the low-resolution MRI images are obtained. Using these continuous resolution representations, super-resolution MRI images of arbitrary scales are output. Utilizing implicit neural representations to obtain continuous resolution representations of low-resolution MRI images includes: integrating multiple MLPs into the upsampling stage of a U-Net diffusion network framework, where the MLPs are used to define the parameters of the implicit neural representations to facilitate the recovery of low-resolution MRI images with higher fidelity across continuous scales; using the continuous coordinates of hypothetical multi-resolution features as a reference, where the continuous coordinates are obtained from the MLPs using a scaling factor; and providing modulation features around the indicator coordinates as input during the implicit representation process to determine the corresponding target features. The objective is to obtain continuously variable resolution outputs, and to adapt the diffusion network framework to arbitrary scales. The defined objective is: Among them, the denoising model This model is equivalent to processing a noisy target image y. t The target image y0 is recovered in the middle. At the same time, in order to achieve continuous resolution output, the denoising model... It should be trained to be applicable to any scale while ensuring projection noise. x refers to the low-resolution image, t refers to the diffusion at step t, and s is the scaling factor.

7. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the brain MRI super-resolution reconstruction method based on implicit neural representation as described in any one of claims 1-5.

8. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the brain MRI super-resolution reconstruction method based on implicit neural representation as described in any one of claims 1-5.

Images