Multi-contrast MRI super-resolution reconstruction method based on conditional guided diffusion model

Abstract

This invention relates to the field of image processing technology and discloses a multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model. The method includes designing a dual-branch conditional information encoder, combining a feature correction module and a feature fusion module to correct and fuse cross-modal conditional features, generating semantically consistent conditional fusion features; and designing a denoising U-Net architecture, introducing a feature injection module into the decoder to modulate conditional features, achieving adaptive conditional information guidance. This multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model achieves refined processing and efficient utilization of cross-modal conditional information by decoupling the fusion and guidance processes of conditional information. It reduces interference noise introduced by modal differences and achieves accurate conditional information guidance while maintaining the stability of the denoising process, mitigating artifact generation or structural distortion during denoising and reconstruction.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model. Background Technology

[0002] Magnetic Resonance Imaging (MRI) is one of the cornerstone technologies in the field of modern medical imaging. Its imaging is based on the abundant hydrogen nuclei in the human body. By applying radio frequency pulses in a strong magnetic field to excite proton resonance and detecting the radio frequency signals released during relaxation, high-contrast anatomical images are finally formed through spatial coding and mathematical reconstruction. Compared with computed tomography (CT) and ultrasound imaging (US), MRI has the inherent advantages of being non-invasive and without radiation damage. In addition, MRI can provide good soft tissue contrast for the diagnosis of brain diseases and can finely distinguish the gray matter, white matter and other structures of the brain, making it play an irreplaceable role in the diagnosis of brain tumors, screening for cerebrovascular malformations and early detection of neurodegenerative diseases. While MRI holds irreplaceable value in clinical diagnosis, its application faces two inherent limitations. First, due to hardware performance and signal-to-noise ratio (SNR) bottlenecks, conventional clinical scans acquire images with limited spatial resolution, resulting in insufficient ability to depict fine anatomical structures. Second, image quality and scanning efficiency are difficult to balance. Patients need to remain absolutely still for extended periods during MRI scans. Attempting to improve SNR and resolution by extending scan time significantly increases the risk of patient movement, introducing motion artifacts. This not only fails to improve diagnostic value but may even lead to examination failure. Therefore, in clinical practice, compromises often need to be made between resolution, scan time, and SNR. In many cases, resolution must be sacrificed for acceptable scan time. This contradiction has spurred an urgent need for MRI super-resolution technology. This technology aims to use advanced computer algorithms to post-process acquired low-resolution MRI images, reconstructing images with higher spatial resolution and clearer details. This breaks the limitations of hardware and physical laws, achieving the goal of "improving image quality from a software perspective," and providing a highly promising solution to the aforementioned clinical dilemmas. In recent years, MRI super-resolution technology has achieved great success, especially the multi-contrast method that extracts texture information from reference modal images to guide the super-resolution reconstruction of the target modality. Due to the imaging characteristics of MRI, multi-contrast MRI images can be obtained. These images provide supplementary information about anatomical structures. For example, T1W-weighted images focus on observing anatomical structures, while T2W-weighted images focus on observing pathological edema structures. T1-weighted (T1W) images and PD-weighted (PDW) images usually have shorter scan times than T2-weighted (T2W) images. Clinicians usually fully sample high-resolution T1W and PDW images and undersample low-resolution T2W images. Therefore, multi-contrast MRI super-resolution reconstruction methods usually use high-resolution T1W or PDW images as reference modalities and assist low-resolution T2W images to complete the super-resolution reconstruction task. In recent years, the Diffusion Model (DDPM), as a popular generative model, has been widely used in multi-contrast MRI super-resolution reconstruction methods because it can generate high-fidelity images while effectively avoiding problems such as mode collapse. Current methods often use low-resolution T2W images of the target modality and high-resolution T1W images of the reference modality as conditional guidance information, inputting them into the noise prediction network of the diffusion model for processing. However, there are significant semantic differences between cross-modal images. A clear edge or bright area in the T1W image of the reference modality may correspond to completely different grayscale features in the T2W image of the target modality. T1W and T2W images are similar in that... The signal responses of a single tissue can be diametrically opposed (e.g., cerebrospinal fluid). Furthermore, T2W images, which serve as an important constraint, are downsampled, resulting in significant loss of high-frequency details (e.g., tissue boundaries, small lesions), structural blurring, and partial volume effects. This further exacerbates the semantic inconsistencies between cross-modal images. Unfortunately, existing diffusion-model-based multi-contrast MRI super-resolution methods merely fuse cross-modal images, failing to effectively mitigate the differences between them and failing to fully utilize their complementary features. This introduces semantic noise, leading to reconstruction biases and unexpected artifacts in the reconstructed images, which greatly affects the accuracy of medical diagnosis. Summary of the Invention

[0003] The purpose of this invention is to provide a multi-contrast MRI super-resolution reconstruction method based on a condition-guided diffusion model, in order to solve the problems mentioned in the background art, such as insufficient utilization of cross-modal conditional information, difficulty in effectively coordinating semantic differences, and the introduction of noise and redundancy in the reconstruction process of multi-contrast MRI super-resolution model based on diffusion model.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model, comprising the following steps: S1: Construct a multi-contrast MRI super-resolution reconstruction network model based on a conditionally guided diffusion model; S2: Obtain multi-contrast MRI image data from publicly available datasets; S3: Perform data preprocessing to obtain low-resolution target contrast MRI image data; S4: Diffuse model T-step forward noise addition process, gradually adding noise from the high-resolution target contrast image to generate a Gaussian noise image; S5: Cross-modal condition information is processed by a dual-branch condition information encoder; S6: The noisy image is processed by the denoising U-Net network; S7: The U-Net network outputs the noise added to the prediction at the current time step and then denoises it to obtain the denoised image at the current step; S8: Diffusion model T-step reverse denoising process, gradually denoising the Gaussian noise image to generate a super-resolution target contrast image; S9: Set the loss function and iteratively train the model until the loss converges; S10: Obtain the model weights after training, perform the diffusion model inference process, and combine cross-modal conditional information to gradually denoise from Gaussian noise to obtain a super-resolution target contrast image.

[0005] Preferably, in step S1, the multi-contrast MRI super-resolution reconstruction method based on the conditionally guided diffusion model includes a dual-branch conditional information encoder, a denoising U-Net network, and a diffusion model workflow. The dual-branch conditional information encoder includes: a SwingTransformer backbone network, a feature correction module (FRM), and a feature fusion module (FFM); The denoising U-Net network includes: a denoising U-Net encoder, a denoising U-Net bottleneck layer, a denoising U-Net decoder, and a feature injection module (FIM); The diffusion model workflow includes: a forward noise addition process, a backward noise reduction process, and an inference reconstruction process.

[0006] Preferably, in step S2: multi-contrast MRI high-resolution image data are obtained using the publicly available IXI dataset and BraTS2020 dataset, respectively; Specifically, the PDW modality was obtained from the IXI dataset as the MRI reference contrast image data, and the T2W modality was obtained as the target contrast image data. The T1W modality was obtained from the BraTS2020 dataset as the MRI reference contrast image data, and the T2W modality was obtained as the target contrast image data.

[0007] Preferably, in step S3: data preprocessing is performed on the original dataset images; Each MRI image was cropped into multiple slices, and further k-space truncation was used to obtain T2W low-resolution target contrast images with 2x and 4x downsampling to simulate the data degradation process. The low-resolution image obtained after K-space truncation loses high-frequency information, resulting in visual blurring. It can be used to simulate low-resolution MRI image data obtained from rapid clinical scans.

[0008] Preferably, in step S4: the diffusion model uses a T-step forward noise addition process to improve the contrast of the high-resolution target image. Gaussian noise image generated by progressively adding noise At any time step t, the current noisy image Can be derived from the original image It is obtained directly through sampling using the reparameterization formula, and is expressed as: ; in To follow a standard normal distribution random noise, For the cumulative noise dispatch coefficient, as Increase Approaching zero, making It approximately follows a standard Gaussian distribution.

[0009] Preferably, in step S5: cross-modal condition information and The process will involve a dual-branch conditional information encoder, and the specific steps are as follows: S5.1 First, use the SwinTransformer backbone network to extract low-resolution target modal images. With reference modal image The encoder features multi-scale characteristics and consists of four levels, with each level undergoing a 2x downsampling and the number of channels increasing accordingly. The processing flow of the dual-branch condition information encoder is as follows: ; ; ; in: Indicates the current network layer index. Indicates the first The SwinTransformer block of the layer, and They represent the first The layer's feature correction module and feature fusion module, and For the first time after STB processing Cross-modal condition characteristics of the layer and For the first after FRM correction Layer-cross modal condition characteristics, and At the same time, it serves as the input feature for the next layer STB. The first one obtained after FFM The layer conditional fusion features are used as the input features of the current layer denoising U-Net decoder feature injection module FIM; S5.2: The Feature Correction Module (FRM) utilizes the spatial and channel correlations of cross-modal conditional features to adaptively correct and filter differences and noise between different modalities, gradually mitigating semantic differences between modalities. It first performs spatial correction, given cross-modal conditional features. and The features of the two branches are concatenated along the channel dimension and a spatial weight map is generated through a multilayer perceptron, as shown in the following formula: ; in This is a spatial weighting graph of the two conditional feature branches. For feature segmentation, For feature splicing, Using the Sigmoid activation function and MLP (Multilayer Perceptron), the number of concatenated feature channels is mapped from 2C to 2 to generate two spatial weight maps. The spatially corrected features can be represented as: ; in This represents element-wise multiplication. These are learnable parameters that control the injection intensity of spatial correction. Based on spatial correction, FRM further introduces a channel correction mechanism to achieve global semantic correction, concatenating the spatially corrected conditional features to obtain... Then, the global feature vector is extracted through three pooling operations and concatenated to obtain the pooled feature vector. : ; in, , and Average pooling, max pooling, and standard deviation pooling operations are performed respectively, followed by channel weights generated using MLP and the sigmoid activation function: ; in The channel weight map for the two conditional feature branches results in the final channel correction feature: ; in and This represents the cross-modal conditional characteristics after two-stage FRM correction, and the learnable parameters. With the spatial correction stage Complementary, controlling the injection intensity of the channel correction; S5.3: The Feature Fusion Module (FFM) generates semantically consistent and detail-rich fused features through cross-attention. Taking the LR branch query Ref branch as an example, it first generates the query matrix through linear projection. ,at this time To reduce computational overhead, spatial downsampling will be performed on the Ref branch: ; in This represents a depthwise convolution operation with stride R, which generates a key-value matrix. and ,at this time In this setting, the attention computation scale is... Reduced to This reduces the computational complexity from Reduce to Then, cross-attention is calculated: ; in, This indicates cross-attention calculation. This represents the enhanced features obtained by fusing information from the LR branch with information from the Ref branch. The Ref branch also queries the LR branch to obtain the corresponding enhanced features. Then, the enhanced features are integrated into the original feature representation to obtain the final interactive enhanced features. and : ; in Presentation layer normalization operation, As a linear projection matrix, the two interactively enhanced features are concatenated along the channel dimension and then passed through a gated feedforward network (FFN) to generate the final conditional fusion features. .

[0010] Preferably, in step S6: noisy image The data will be processed using a denoising U-Net network. The specific steps are as follows: S6.1 Denoising U-Net encoder extracts noisy images Multi-scale features: The encoder contains 4 levels, each with a 2x downsampling, and the number of channels increases accordingly. The U-net encoder denoises each layer of the noisy image. Extracting multi-scale features yields the processed features: ; in, For the first The features of the layer after being processed by the encoder, Indicates encoder processing; S6.2, Denoising U-Net bottleneck layer for processing noisy images The deepest feature layer: The bottleneck layer acts as a bridge connecting contextual information and spatial detail recovery, processing the features after passing through the fourth-layer U-Net encoder. Output features Entering the fourth layer U-Net decoder: ; in, For the input features of the fourth layer decoder, For the output features of the fourth layer encoder, Indicates bottleneck layer processing; S6.3, Denoising U-Net Decoder for processing noisy images Multi-scale features: The decoder corresponds to the encoder and contains 4 levels. Each level performs a 2x upsampling, and the number of channels decreases accordingly. Before entering each layer of the decoder, the noisy features to be processed are... First, the processed features of the corresponding encoder layer are compared. Perform a skip connection, and then proceed to the current decoder layer for processing: ; in, Indicates the first Features of the layer after being processed by the decoder Indicates decoder processing; S6.4 Feature Injection Module (FIM): Adaptive modulation conditional fusion features are injected into the decoder to achieve adaptive conditional information guidance. This module is located after each residual block of the decoder. Due to the conditional features... With noisy features Because of differences in distribution characteristics, direct fusion can lead to feature mismatch. Therefore, FIM first performs instance normalization on the conditional features to remove their original distribution information. ; in Represents the normalized features. and These are the mean and standard deviation, respectively. Then, a learnable convolutional network is used to learn the bias term of the modulation parameters from the noisy features: ; in , These are the learned mean and standard deviation modulation parameters, respectively. and For convolutional network processing, a residual learning strategy is then employed to adjust its parameters and combine them with the statistical properties of the noisy features: ; in and This represents the modulated mean and standard deviation. and for The mean and standard deviation, after normalization, are used to impart modulated distribution characteristics through affine transformation: ; in This represents the modulated conditional features, where the content of the conditional features is preserved, while their distribution has been modulated to match the noisy features in the current denoising stage. Finally, the modulated conditional features and the original noisy features are first concatenated along the channel dimension, then dimensionality-reduced and activated by a multilayer perceptron, and finally injected into the denoising U-Net main process. .

[0011] Preferably, in step S7: the denoising U-Net network outputs the added noise predicted at the current time step, and calculates the mean of the denoised image based on the predicted noise, using the following formula: ; in: This represents the mean value of the denoised image obtained in the current step. , , , All of these require inputting network information. , All of these are preset noise control coefficients for the diffusion process. This indicates the added noise in the predictions of the denoising U-Net network; During training, the model optimizes its parameters by minimizing the difference between the predicted noise and the actual noise. .

[0012] Preferably, in step S8: the diffusion model will undergo T-step reverse denoising processing, utilizing cross-modal conditional information. and From standard Gaussian noise Initially, based on the mean and variance predicted by the denoising network, images from the previous time step by step are sampled and generated until a super-resolution target contrast image is generated. Its single-step sampling formula is as follows: ; in, This represents the image from the previous time step generated after denoising. This represents the posterior mean of the predictions made by the denoised U-Net network. Indicates preset Time-of-flight noise standard deviation This represents the standard Gaussian noise of the sample.

[0013] Preferably, in step S9: a loss function is set for model training, and the model is iteratively trained until the loss converges. The loss function is the Charbonnier loss. ; in The noise predicted by the model at time step t. This is the actual noise added during the forward diffusion process. is a smoothing constant used to control the smoothness of the loss function near zero. By minimizing this loss function, the diffusion model learns to accurately predict the noise distribution in each denoising step. In step S10: After the model training is completed and the model weights are obtained, the diffusion model inference process can be carried out. Combined with cross-modal conditional information, the super-resolution target contrast image can be obtained by gradually denoising from pure Gaussian noise.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model: 1. By designing a dual-branch conditional information encoder comprising a SwingTransformer backbone network, a Feature Correction Module (FRM), and a Feature Fusion Module (FFM), the semantic differences of cross-modal conditional information are effectively mitigated by utilizing the spatial and channel dual correction mechanism in the FRM. Complementary high-frequency details are efficiently extracted through the cross-attention mechanism in the FFM, generating semantically consistent high-quality conditional fusion features. Simultaneously, a Feature Injection Module (FIM) is introduced into the denoising U-Net decoder to perform adaptive normalization transformation on the conditional features, achieving precise conditional guidance and ensuring the consistency and stability of the denoising process. Ultimately, while significantly suppressing artifacts and structural distortion, high-resolution MRI images with rich anatomical details are reconstructed, effectively solving the reconstruction bias problem caused by modal differences in current diffusion-based MRI multi-contrast super-resolution reconstruction methods. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the learning method of the present invention; Figure 2 A schematic diagram of the network model architecture constructed by the method of this invention; Figure 3 This figure shows the qualitative analysis results of the method of this invention compared with other advanced MRI super-resolution methods. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figures 1-3 This invention provides a technical solution: a multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model.

[0018] First, the flowchart of the multi-contrast MRI super-resolution reconstruction method based on the conditionally guided diffusion model is shown in the attached instruction manual. Figure 1 The network architecture diagram is shown in the attached manual. Figure 2 As shown in the appendix, the qualitative analysis results of the method described in this invention compared with other advanced MRI super-resolution methods are as follows. Figure 3 As shown, this invention is implemented on two NVIDIA GeForce RTX 3090 GPUs, trained for 200,000 iterations using the AdamW optimizer, with an initial learning rate of 1e-4, a batch size of 8, a diffusion step size T of 1000, an inference step size of 50, and reconstruction scales of ×2 and ×4.

[0019] The construction of the multi-contrast MRI super-resolution reconstruction method based on the conditionally guided diffusion model includes the following steps: S1: Construct a multi-contrast MRI super-resolution reconstruction network model based on a conditionally guided diffusion model; Specifically, the multi-contrast MRI super-resolution reconstruction method based on the conditional guided diffusion model includes a dual-branch conditional information encoder, a denoising U-Net network, and a diffusion model workflow; The dual-branch conditional information encoder includes: a SwingTransformer backbone network, a feature correction module (FRM), and a feature fusion module (FFM); The denoising U-Net network includes: a denoising U-Net encoder, a denoising U-Net bottleneck layer, a denoising U-Net decoder, and a feature injection module (FIM); The diffusion model workflow includes: forward denoising, backward denoising, and inference reconstruction. S2: Obtain multi-contrast MRI image data from publicly available datasets; Specifically, high-resolution multi-contrast MRI image data were obtained using the publicly available IXI dataset and BraTS2020 dataset, respectively. Specifically, the PDW modality was obtained from the IXI dataset as the MRI reference contrast image data, and the T2W modality was obtained as the target contrast image data; the T1W modality was obtained from the BraTS2020 dataset as the MRI reference contrast image data, and the T2W modality was obtained as the target contrast image data. S3: Perform data preprocessing to obtain low-resolution target contrast MRI image data; Specifically, each MRI image was cropped into 20 slices with a resolution of 224×224, and further truncated using k-space to obtain T2W low-resolution target contrast images with 2x and 4x downsampling, simulating the data degradation process. S4: Diffuse model T-step forward noise addition process, gradually adding noise from the high-resolution target contrast image to generate a Gaussian noise image; Specifically, the diffusion model uses a T-step forward noise addition process to improve the contrast of high-resolution target images. Gaussian noise image generated by progressively adding noise At any time step t, the current noisy image Can be derived from the original image It is obtained directly through sampling using the reparameterization formula, and is expressed as: ; in To follow a standard normal distribution random noise, For the cumulative noise dispatch coefficient, as Increase Approaching zero, making It approximately follows a standard Gaussian distribution; S5: Cross-modal condition information is processed by a dual-branch condition information encoder; Specifically, the target modality image is first extracted using the SwingTransformer backbone network. With reference modal image Multi-scale features at each level The processing flow of the dual-branch condition information encoder is as follows: ; ; ; in: Indicates the current network layer index. Indicates the first The SwinTransformer block of the layer, and They represent the first The layer's feature correction module and feature fusion module, and For the first time after STB processing Cross-modal condition characteristics of the layer and For the first after FRM correction Layer-cross modal conditional features (which also serve as input features for the next layer STB). The first one obtained after FFM The layer conditional fusion features are used as the input features of the current layer denoising U-Net decoder feature injection module FIM; Then, given cross-modal conditional features and The features of the two branches are concatenated along the channel dimension and a spatial weight map is generated using an MLP, as shown in the following formula: ; in This is a spatial weighting graph of the two conditional feature branches. For feature segmentation, For feature splicing, Using the Sigmoid activation function and MLP (Multilayer Perceptron), the number of concatenated feature channels is mapped from 2C to 2 to generate two spatial weight maps. The spatially corrected features can be represented as: ; in This represents element-wise multiplication. These are learnable parameters that control the injection intensity for spatial correction, and are obtained by stitching together the spatially corrected conditional features. Then, the global feature vector is extracted through three pooling operations and concatenated to obtain the pooled feature vector. : ; in, , and Average pooling, max pooling, and standard deviation pooling operations are performed respectively, followed by channel weights generated using MLP and the sigmoid activation function: ; in The channel weight map for the two conditional feature branches results in the final channel correction feature: ; in and This represents the cross-modal conditional characteristics after two-stage FRM correction, and the learnable parameters. With the spatial correction stage Complementary, controlling the injection intensity of the channel correction; Finally, conditional feature fusion will be performed. Taking the LR branch query Ref branch as an example, the query matrix will first be generated through linear projection. ,at this time To reduce computational overhead, spatial downsampling will be performed on the Ref branch: ; in This represents a depthwise convolution operation with stride R, which generates a key-value matrix. and ,at this time Then, cross-attention is calculated: ; in, This indicates cross-attention calculation. This represents the enhanced features obtained by fusing information from the LR branch with information from the Ref branch. The Ref branch also queries the LR branch to obtain the corresponding enhanced features. Then, the enhanced features are integrated into the original feature representation to obtain the final interactive enhanced features. and : ; in Presentation layer normalization operation, As a linear projection matrix, the two interactively enhanced features are concatenated along the channel dimension and then passed through a gated feedforward network (FFN) to generate the final conditional fusion features. ; S5: The noisy image is processed by the denoising U-Net network; Specifically, firstly, the denoising U-Net encoder extracts the noisy image. Multi-scale features at each level The U-net encoder denoises each layer of the noisy image. Extracting multi-scale features yields the processed features: ; in, For the first The features of the layer after being processed by the encoder, Indicates encoder processing; Next, the denoising U-Net bottleneck layer processes the noisy image. The deepest features are output and fed into the fourth layer of the U-Net decoder: ; in, For the input features of the fourth layer decoder, For the output features of the fourth layer encoder, Indicates bottleneck layer processing; Next, the denoising U-Net decoder processes the noisy image. The multi-scale features, the noisy features to be processed before entering each layer of the decoder. First, the processed features of the corresponding encoder layer are compared. Perform a skip connection, and then proceed to the current decoder layer for processing: ; in, Indicates the first Features of the layer after being processed by the decoder Indicates decoder processing; Finally, the feature injection module first performs instance normalization on the conditional features to remove their original distribution information: ; in and These are the mean and standard deviation, respectively. Then, a learnable convolutional network is used to learn the bias term of the modulation parameters from the noisy features: ; in , These are the learned mean and standard deviation modulation parameters, respectively. and For convolutional network processing, a residual learning strategy is then employed to adjust its parameters and combine them with the statistical properties of the noisy features: ; in and This represents the modulated mean and standard deviation. and for The mean and standard deviation, after normalization, are used to impart modulated distribution characteristics through affine transformation: ; in This represents the modulated conditional features. The modulated conditional features and the original noisy features are first concatenated along the channel dimension, then dimensionality reduced and activated by a multilayer perceptron, and finally injected into the denoising U-Net main process. ; S7: The U-Net network outputs the noise added to the prediction at the current time step and then denoises it to obtain the denoised image at the current step; Specifically, the denoising U-Net network outputs the added noise predicted at the current time step, and calculates the mean of the denoised image based on the predicted noise. The relevant formula is as follows: ; in: This represents the mean value of the denoised image obtained in the current step. , , , All of these require inputting network information. , All of these are preset noise control coefficients for the diffusion process. This indicates the added noise in the predictions of the denoising U-Net network; During training, the model optimizes its parameters by minimizing the difference between the predicted noise and the actual noise. ; S8: Diffusion model T-step reverse denoising process, gradually denoising the Gaussian noise image to generate a super-resolution target contrast image; The diffusion model will undergo T-step inverse denoising, utilizing cross-modal conditional information. and From standard Gaussian noise Initially, based on the mean and variance predicted by the denoising network, images from the previous time step by step are sampled and generated until a super-resolution target contrast image is generated. Its single-step sampling formula is as follows: ; in, This represents the image from the previous time step generated after denoising. This represents the posterior mean of the predictions made by the denoised U-Net network. Indicates preset Time-of-flight noise standard deviation The sampled standard Gaussian noise; S9: Set the loss function and iteratively train the model until the loss converges; Specifically, a loss function will be set for model training, and the model will be iteratively trained until the loss converges. The loss function is the Charbonnier loss. ; in The noise predicted by the model at time step t. This is the actual noise added during the forward diffusion process. is a smoothing constant used to control the smoothness of the loss function near zero. By minimizing this loss function, the diffusion model learns to accurately predict the noise distribution in each denoising step. S10: Obtain the model weights after training, perform the diffusion model inference process, and combine cross-modal conditional information to gradually denoise from Gaussian noise to obtain a super-resolution target contrast image. Specifically, after the model training is completed and the model weights are obtained, the diffusion model inference process can be carried out. By combining cross-modal conditional information, the super-resolution target contrast image can be obtained by gradually denoising pure Gaussian noise.

[0020] This invention presents a quantitative comparison of the reconstruction performance of a conditionally guided diffusion model-based multi-contrast MRI super-resolution reconstruction method with other multi-contrast MRI super-resolution reconstruction methods. The evaluation metrics used were Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and the Level-Limited Image Similarity Scale (LPIPS). Super-resolution reconstructions at ×2 and ×4 scales were performed on the IXI and BraTS2020 datasets. The upper arrow indicates higher performance, and the lower arrow indicates lower performance. The best results are marked in bold (see Table 1). A qualitative comparison of the error maps of this invention's conditionally guided diffusion model-based multi-contrast MRI super-resolution reconstruction method with other multi-contrast MRI super-resolution reconstruction methods is shown below. Figure 3 As shown, in Figure 3 The error qualitative graphs shown list the reconstructed MRI images and corresponding error graphs of the method of the present invention and other comparative methods. The bright outlines of the error graphs represent reconstruction deviations. The brighter the color of the error graph, the worse the effect. It can be observed that the method of the present invention achieved the minimum error at each reconstruction magnification in the two datasets.

[0021] Table 1

[0022] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model, characterized in that: Includes the following steps: S1: Construct a multi-contrast MRI super-resolution reconstruction network model based on a conditionally guided diffusion model; S2: Obtain multi-contrast MRI image data from publicly available datasets; S3: Perform data preprocessing to obtain low-resolution target contrast MRI image data; S4: Diffuse model T-step forward noise addition process, gradually adding noise from the high-resolution target contrast image to generate a Gaussian noise image; S5: Cross-modal condition information is processed by a dual-branch condition information encoder; S6: The noisy image is processed by the denoising U-Net network; S7: The U-Net network outputs the noise added to the prediction at the current time step and then denoises it to obtain the denoised image at the current step; S8: Diffusion model T-step reverse denoising process, gradually denoising the Gaussian noise image to generate a super-resolution target contrast image; S9: Set the loss function and iteratively train the model until the loss converges; S10: Obtain the model weights after training, perform the diffusion model inference process, and combine cross-modal conditional information to gradually denoise from Gaussian noise to obtain a super-resolution target contrast image.

2. The multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model according to claim 1, characterized in that: In step S1, the multi-contrast MRI super-resolution reconstruction method based on the conditionally guided diffusion model includes a dual-branch conditional information encoder, a denoising U-Net network, and a diffusion model workflow. The dual-branch conditional information encoder includes: a SwinTransformer backbone network, a feature correction module (FRM), and a feature fusion module (FFM); The denoising U-Net network includes: a denoising U-Net encoder, a denoising U-Net bottleneck layer, a denoising U-Net decoder, and a feature injection module (FIM). The diffusion model workflow includes: a forward noise addition process, a backward noise reduction process, and an inference reconstruction process.

3. The multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model according to claim 1, characterized in that: In step S2: multi-contrast MRI high-resolution image data are obtained using the publicly available IXI dataset and BraTS2020 dataset, respectively; Specifically, the PDW modality was obtained from the IXI dataset as the MRI reference contrast image data, and the T2W modality was obtained as the target contrast image data. The T1W modality was obtained from the BraTS2020 dataset as the MRI reference contrast image data, and the T2W modality was obtained as the target contrast image data.

4. The multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model according to claim 1, characterized in that: In step S3: data preprocessing is performed on the original dataset images; Each MRI image was cropped into multiple slices, and then k-space truncation was used to obtain T2W low-resolution target contrast images with 2x and 4x downsampling to simulate the data degradation process.

5. The multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model according to claim 1, characterized in that: In step S4: the diffusion model uses a T-step forward noise addition process to improve the contrast of the high-resolution target image. Gaussian noise image generated by progressively adding noise At any time step t, the current noisy image Can be derived from the original image It is obtained directly through sampling using the reparameterization formula, and is expressed as: ； in To follow a standard normal distribution random noise, For the cumulative noise dispatch coefficient, as Increase Approaching zero, making It approximately follows a standard Gaussian distribution.

6. The multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model according to claim 1, characterized in that: In step S5: cross-modal condition information and The process will involve a dual-branch conditional information encoder, and the specific steps are as follows: S5.1 First, use the SwinTransformer backbone network to extract low-resolution target modal images. With reference modal image The encoder features multi-scale characteristics and consists of four levels, with each level undergoing a 2x downsampling and the number of channels increasing accordingly. The processing flow of the dual-branch condition information encoder is as follows: ； ； ； in: Indicates the current network layer index. Indicates the first The SwinTransformer block of the layer, and They represent the first The layer's feature correction module and feature fusion module, and For the first time after STB processing Cross-modal condition characteristics of the layer and For the first after FRM correction Layer-cross modal condition characteristics, and At the same time, it serves as the input feature for the next layer STB. The first one obtained after FFM The layer conditional fusion features are used as the input features of the current layer denoising U-Net decoder feature injection module FIM; S5.2: The feature correction module utilizes the spatial and channel correlation of cross-modal conditional features to adaptively correct and filter differences and noise between different modalities, gradually mitigating semantic differences between modalities. First, spatial correction is performed, given the cross-modal conditional features. and The features of the two branches are concatenated along the channel dimension and a spatial weight map is generated through a multilayer perceptron, as shown in the following formula: ； in This is a spatial weighting graph of the two conditional feature branches. For feature segmentation, For feature splicing, Using the Sigmoid activation function and MLP (Multilayer Perceptron), the number of concatenated feature channels is mapped from 2C to 2 to generate two spatial weight maps. The spatially corrected features can be represented as: ； in This represents element-wise multiplication. These are learnable parameters that control the injection intensity of spatial correction. Based on spatial correction, FRM further introduces a channel correction mechanism to achieve global semantic correction, concatenating the spatially corrected conditional features to obtain... Then, the global feature vector is extracted through three pooling operations and concatenated to obtain the pooled feature vector. : ； in, , and Average pooling, max pooling, and standard deviation pooling operations are performed respectively, followed by channel weights generated using MLP and the sigmoid activation function: ； in The channel weight map for the two conditional feature branches results in the final channel correction feature: ； in and This represents the cross-modal conditional characteristics after two-stage FRM correction, and the learnable parameters. With the spatial correction stage Complementary, controlling the injection intensity of the channel correction; S5.3: The feature fusion module generates semantically consistent and detail-rich fused features through cross-attention. Taking the LR branch query Ref branch as an example, it first generates the query matrix through linear projection. ,at this time To reduce computational overhead, spatial downsampling will be performed on the Ref branch: ； in This represents a depthwise convolution operation with stride R, which generates a key-value matrix. and ,at this time In this setting, the attention computation scale is... Reduced to This reduces the computational complexity from Reduce to Then, cross-attention is calculated: ； in, This indicates cross-attention calculation. This represents the enhanced features obtained by fusing information from the LR branch with information from the Ref branch. The Ref branch also queries the LR branch to obtain the corresponding enhanced features. Then, the enhanced features are integrated into the original feature representation to obtain the final interactive enhanced features. and : ； in Presentation layer normalization operation, As a linear projection matrix, the two interactively enhanced features are concatenated along the channel dimension and then passed through a gated feedforward network to generate the final conditional fusion feature: 。 7. The multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model according to claim 1, characterized in that: In step S6: noisy image The data will be processed using a denoising U-Net network. The specific steps are as follows: S6.1 Denoising U-Net encoder extracts noisy images Multi-scale features: The encoder contains 4 levels, each with a 2x downsampling, and the number of channels increases accordingly. The U-net encoder denoises each layer of the noisy image. Extracting multi-scale features yields the processed features: ； in, For the first The features of the layer after being processed by the encoder, Indicates encoder processing; S6.2, Denoising U-Net bottleneck layer for processing noisy images The deepest feature layer: The bottleneck layer acts as a bridge connecting contextual information and spatial detail recovery, processing the features after passing through the fourth-layer U-Net encoder. Output features Entering the fourth layer U-Net decoder: ； in, For the input features of the fourth layer decoder, For the output features of the fourth layer encoder, Indicates bottleneck layer processing; S6.3, Denoising U-Net Decoder for processing noisy images Multi-scale features: The decoder corresponds to the encoder and contains 4 levels. Each level performs a 2x upsampling, and the number of channels decreases accordingly. Before entering each layer of the decoder, the noisy features to be processed are... First, the processed features of the corresponding encoder layer are compared. Perform a skip connection, and then proceed to the current decoder layer for processing: ； in, Indicates the first Features of the layer after being processed by the decoder Indicates decoder processing; S6.4 Feature Injection Module: Adaptive modulation conditional fusion features are injected into the decoder to achieve adaptive conditional information guidance. This module is located after each residual block of the decoder. Due to the conditional features... With noisy features Due to differences in distribution characteristics, direct fusion can lead to feature mismatch. Therefore, FIM first performs instance normalization on the conditional features to remove their original distribution information. ； in Represents the normalized features. and These are the mean and standard deviation, respectively. Then, a learnable convolutional network is used to learn the bias term of the modulation parameters from the noisy features: ； in , These are the learned mean and standard deviation modulation parameters, respectively. and For convolutional network processing, a residual learning strategy is then employed to adjust its parameters and combine them with the statistical properties of the noisy features: ； in and This represents the modulated mean and standard deviation. and for The mean and standard deviation, after normalization, are used to impart modulated distribution characteristics through affine transformation: ； in This represents the modulated conditional features, where the content of the conditional features is preserved, while their distribution has been modulated to match the noisy features in the current denoising stage. Finally, the modulated conditional features and the original noisy features are first concatenated along the channel dimension, then dimensionality-reduced and activated by a multilayer perceptron, and finally injected into the denoising U-Net main process. 。 8. The multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model according to claim 1, characterized in that: In step S7: the denoising U-Net network outputs the added noise predicted at the current time step, and calculates the mean of the denoised image based on the predicted noise. The relevant formula is as follows: ； in: This represents the mean value of the denoised image obtained in the current step. , , , All of these require inputting network information. , All of these are preset noise control coefficients for the diffusion process. This indicates the added noise in the predictions of the denoising U-Net network; During training, the model optimizes its parameters by minimizing the difference between the predicted noise and the actual noise. .

9. The multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model according to claim 1, characterized in that: In step S8: the diffusion model will undergo a T-step reverse denoising process, utilizing cross-modal conditional information. and From standard Gaussian noise Initially, based on the mean and variance predicted by the denoising network, images from the previous time step by step are sampled and generated until a super-resolution target contrast image is generated. Its single-step sampling formula is as follows: ； in, This represents the image from the previous time step generated after denoising. This represents the posterior mean of the predictions made by the denoised U-Net network. Indicates preset Time-of-flight noise standard deviation This represents the standard Gaussian noise of the sample.

10. The multi-contrast MRI super-resolution reconstruction method based on a conditionally guided diffusion model according to claim 1, characterized in that: In step S9: a loss function is set for model training, and the model is iteratively trained until the loss converges. The loss function is the Charbonnier loss. ； in The noise predicted by the model at time step t. This is the actual noise added during the forward diffusion process. is a smoothing constant used to control the smoothness of the loss function near zero. By minimizing this loss function, the diffusion model learns to accurately predict the noise distribution in each denoising step. In step S10: After the model training is completed and the model weights are obtained, the diffusion model inference process can be carried out. Combined with cross-modal conditional information, the super-resolution target contrast image can be obtained by gradually denoising from pure Gaussian noise.

Images