A low-field magnetic resonance stroke diffusion weighted imaging image quality enhancement method

Abstract

The present application relates to the technical field of image processing and deep learning, more particularly to a low-field magnetic resonance stroke diffusion weighted imaging image quality enhancement method, comprising the following steps: constructing an original input image pair; performing a preprocessing operation on the original input image pair; segmenting the low-field DWI using a pre-trained UNet segmentation network to obtain a corresponding stroke lesion mask; dividing the paired low-field DWI, high-field DWI and corresponding lesion mask into a training set and a test set; constructing a difference-guided conditional diffusion model; training the difference-guided conditional diffusion model using the training set; performing generation testing and effect evaluation on the trained model using the test set, and selecting the model weight with the optimal performance for saving; using the difference-guided conditional diffusion model with the saved optimal weight to perform end-to-end clarification processing on the low-field DWI to be processed, and outputting the quality-enhanced DWI image.

Description

Technical Field

[0001] This invention relates to the fields of image processing and deep learning technology, and more specifically to a method for enhancing the image quality of low-field magnetic resonance imaging (MRI) diffusion-weighted imaging in stroke patients. Background Technology

[0002] Traditional interpolation techniques essentially utilize existing pixel information from the original image. Noise interference in low-field images renders interpolation methods ineffective, and the noise cannot be eliminated. Furthermore, deep learning-based methods such as convolutional neural networks (CNNs), generative adversarial networks (GANs), and diffusion models also have limitations. CNNs, due to their insufficient modeling capabilities, struggle to fully recover complex textures and fine-grained structures. GANs, due to their architectural limitations, are prone to gradient collapse. Existing diffusion model-based methods primarily employ a globally uniform noise scheduling strategy.

[0003] Therefore, these methods fail to effectively utilize the global distribution differences between low and high fields in the image, as well as prior information about lesion masks, to construct adaptive, region-aware noise guidance mechanisms. This limitation hinders the enhancement of the signal-to-noise ratio in healthy tissue regions while preserving fine structural details within stroke lesions. Secondly, most methods rely on single-stream or shallow conditional integration strategies, which fail to adequately capture the inherent multi-scale and hierarchical features of low-field DWI and neglect explicit lesion modeling, often leading to loss of detail and structural distortion. Thirdly, existing optimization strategies ignore task heterogeneity within the diffusion time step and employ static loss reweighting strategies, thus limiting the enhancement of perceptual supervision of clinically critical structural and textural features at low noise levels.

[0004] Therefore, how to provide a method for enhancing the image quality of low-field magnetic resonance imaging (MRI) diffusion-weighted imaging in stroke is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a method for enhancing the image quality of low-field magnetic resonance imaging (MRI) diffusion-weighted imaging in stroke, aiming to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for enhancing the image quality of low-field magnetic resonance imaging (MRI) for stroke diffusion-weighted imaging includes the following steps: S1. Construct original input image pairs, the original input image pairs including paired low field strength DWI and high field strength DWI; S2. Perform preprocessing operations on the original input image pairs, including skull dissection, image registration, and data standardization. S3. Use the pre-trained UNet segmentation network to segment the low-field DWI and obtain the corresponding stroke lesion mask. S4. Divide the paired low-field DWI, high-field DWI and corresponding lesion masks into training set and test set; S5. Construct a difference-guided conditional diffusion model, which includes a difference-guided adaptive forward noise module, a two-stream multi-scale conditional coding module, and a time-adaptive loss optimization module. S6. Train the difference-guided conditional diffusion model using the training set; S7. Use the test set to perform generation testing and performance evaluation on the trained model, and save the weights of the model with the best performance. S8. Using a difference-guided conditional diffusion model that preserves the optimal weights, end-to-end sharpening of the low-field DWI to be processed is performed, and the quality-enhanced DWI image is output.

[0007] Furthermore, the specific process of the preprocessing operation in S2 is as follows: The skull was removed by removing the skull in low-field DWI and high-field DWI respectively by setting a preset threshold, thus completing the skull dissection. A nonlinear registration method was used, with the high-field DWI as the template, and the low-field DWI was registered into the high-field DWI template. The registered low-field DWI and high-field DWI have a size of 256×256×13, which correspond to the height, width and number of slices of the image, respectively. The registered paired data are standardized using the maximum-minimum normalization method, mapping the data to the numerical range of [0,1].

[0008] Furthermore, the pre-trained UNet segmentation network used in S3 has a segmentation accuracy of 95%, and the segmented stroke lesion mask is used to guide the training and generation processes of the differentially guided conditional diffusion model.

[0009] Furthermore, in S4, the paired low-field DWI, high-field DWI, and corresponding lesion masks are divided into training and test sets in a 9:1 ratio.

[0010] Furthermore, the first stage of the difference-guided adaptive forward noise module is: the difference between the low-field DWI and the high-field DWI is introduced into the forward diffusion process as a directional drift term, so that the diffusion trajectory is consistent with the intrinsic distribution between the low-field and high-field DWI. The diffusion process after introducing the directional drift term satisfies the following equation: (1), in, , , Indicates high-field DWI, Indicates low-field DWI, This represents the time-controlled noise scheduling factor. The hyperparameter representing the overall noise intensity is 2. For time t The noise addition result is given by the following formula: (2).

[0011] Furthermore, the second stage of the difference-guided adaptive forward noise module is as follows: calculate the difference map based on low-field DWI and the corresponding lesion mask, and adaptively adjust the noise injection intensity of each region according to the value in the difference map. Stronger noise is injected into healthy tissue regions with higher difference values, and weaker noise is injected into stroke lesion regions with lower difference values. The difference map is defined according to the following formula: (3), in, Represents a difference plot. This represents the mask for stroke lesions on low-field DWI. Represents the Hadamard product; The forward diffusion process after introducing the difference map satisfies the following equation: (4), in, Represents the identity matrix; The corresponding posterior distribution of the reverse process satisfies the following equation: (5).

[0012] Furthermore, the dual-stream multi-scale conditional coding module includes two parallel coding branches: a low-field DWI coding branch and a lesion mask coding branch. The low-field DWI coding branch is used to extract multi-scale hierarchical intrinsic features from low-field DWI. During the decoding process, the feature map extracted by this branch is spliced ​​layer by layer with the corresponding feature map of the upsampling path of the generator model decoder. The lesion mask encoding branch is used to extract multi-scale spatial structure and morphological features from the stroke lesion mask. During the decoding process, the structural features extracted by this branch are input to the shallow decoding layer of the generation model. The features extracted from the two branches provide context and spatial constraints for the generative model in a dense multi-scale conditional fusion manner, guiding the generative model to perceive and reconstruct the stroke lesion area.

[0013] Furthermore, the time-adaptive loss optimization module constructs a loss function based on parameterized v, specifically as follows: Define target vector Satisfy the following formula: (6), in: (7), The predicted output of the generative model is: (8), in The parameterized function for generating the model. The conditional information consists of multi-scale feature maps from low-field DWI and feature maps from lesion masks. For diffusion time steps; Predicting losses for and The mean square error between them satisfies: (9), Furthermore, regarding Perform inverse parameterization to obtain the generated DWI ,as follows: (10).

[0014] Furthermore, based on the generated DWI Constructing pixel-level MSE loss and perceptual loss: The pixel-level MSE loss satisfy: (11), The perceived loss satisfy: (12), in For the first pre-trained feature extraction network Layer output, This represents the number of network layers.

[0015] Furthermore, a time-adaptive weighting function is introduced. Construct the total loss, the time-adaptive weighted function satisfy: (13), Where T is a hyperparameter representing the total time step; Total loss Satisfy the following formula: (14), Among them, the weighting coefficient , .

[0016] Compared to existing technologies, this invention utilizes paired low-field and high-field DWI data. The task is formulated as an image translation task. First, a difference-guided adaptive forward noise addition process is constructed to improve the model by preserving lesions while significantly enhancing the texture information of surrounding brain tissue. Furthermore, a proposed dual-stream multi-scale conditional coding is used to extract multi-scale intrinsic low-field DWI features and prior spatial information of stroke lesions. Finally, a temporally adaptive multi-loss optimization strategy is proposed, which adaptively adjusts the contributions of pixel-level loss and perceptual loss at different diffusion time steps, thereby emphasizing structural and texture reconstruction under different noise levels. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a data preprocessing method according to the present invention.

[0019] Figure 2 This is a schematic diagram of the structure of a difference-guided adaptive conditional diffusion model according to the present invention.

[0020] Figure 3 This is a schematic diagram of the process of a dual-stream multi-scale conditional coding according to the present invention.

[0021] Figure 4 This is a schematic flowchart of a weighted imaging image quality enhancement method according to the present invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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.

[0023] This invention provides a method for enhancing the quality of low-field magnetic resonance imaging (MRI) diffusion-weighted imaging (DWI) images of stroke. Based on a difference-guided conditional diffusion model, the method enhances the quality of low-field DWI images, improves the signal-to-noise ratio, suppresses noise, and preserves the structural information of stroke lesions to the maximum extent. This makes the reconstructed images approach the overall quality of high-field DWI, thus aiding in clinical diagnosis.

[0024] like Figure 4 As shown, the specific steps of the method are as follows: S1: Construct the original input image pair The original input images included paired low-field-strength (0.23T) DWI and high-field-strength (3.0T) DWI.

[0025] S2: Perform preprocessing operations on the original input image pairs. Preprocessing operations include skull dissection, data registration, and data standardization, as detailed below: Skull dissection: The skull of low-field and high-field DWI is removed separately by setting a preset threshold.

[0026] Image registration: Using a nonlinear registration method, the high-field DWI data is used as a template, and the low-field DWI is registered into the high-field DWI template using nonlinear registration. The registered low-field and high-field DWI have a size of 256×256×13, where height, width and number of slices are represented respectively.

[0027] Data standardization: Using the maximum and minimum value normalization method, paired data are mapped to the range [0,1].

[0028] The data preprocessing process is as follows: Figure 1 As shown.

[0029] S3: Segmenting low-field DWI using a pre-trained UNet network The pre-trained segmentation network UNet is used to segment the low-field data in the training set to obtain the corresponding lesion segmentation mask; the segmentation accuracy of the pre-trained UNet network reaches 95%, and the segmented lesion mask is used as input to participate in the training of the differential-guided conditional diffusion model.

[0030] S4: Building the dataset The low-field DWI, high-field DWI, and corresponding low-field DWI lesion masks of the registration pair were divided into training and test sets in a 9:1 ratio.

[0031] S5: Construct a difference-guided conditional diffusion model A difference-guided conditional diffusion model is constructed by first building a difference-guided adaptive forward noise addition module, then a two-stream multi-scale conditional coding module, and finally an adaptive temporal loss optimization strategy. The overall model structure is as follows: Figure 2 As shown.

[0032] The following is a detailed explanation of each module: (1) Difference-guided adaptive forward noise addition process To enhance the signal-to-noise ratio of low-field DWI while preserving the fine structural integrity of stroke lesions, a difference-guided adaptive forward noise injection process is proposed. Traditional diffusion models rely on globally uniform Gaussian noise injection, neglecting the heterogeneous sensitivity of anatomical structures to random perturbations, often leading to excessive damage to stroke lesion details. This method directly embeds structural differences into the forward diffusion process, achieving controlled noise injection that can perceive both overall pattern differences and lesion regions. The difference-adaptive guidance mechanism comprises two stages: the first stage captures the difference between low-field and high-field DWI, and the second stage captures the difference between low-field DWI and the corresponding lesion mask.

[0033] In the first stage, the forward diffusion process explicitly includes the difference between low-field and high-field DWI. As a directional drift term, the design guides the diffusion trajectory to maintain consistency with the intrinsic distribution between low-field and high-field DWIs, rather than a purely random evolution.

[0034] The diffusion process after introducing the first stage becomes: (1), in , , Indicates high-field DWI, Indicates low-field DWI, This represents the time-controlled noise scheduling factor. The hyperparameter representing the overall noise intensity is set to 2. For time t The noise addition result is given by the following formula: (2).

[0035] In the second stage, the forward diffusion process utilizes a difference map calculated from low-field DWI and a lesion mask. During diffusion, the noise perturbation is adaptively adjusted based on the values ​​in this difference map: regions with higher difference values ​​(e.g., healthy tissue) are injected with stronger noise, while regions with lower difference values ​​(e.g., stroke lesions) receive weaker noise. This adaptive mechanism prompts the model to preferentially preserve lesion features in low-field DWI.

[0036] The difference diagram is defined as follows: (3), in, Represents a difference plot. This represents the mask for stroke lesions on low-field DWI. This represents the Hadamard product.

[0037] The diffusion process after introducing the second stage is represented as follows: (4), in, Represents the identity matrix; The forward diffusion process is based on formula (3). The posterior distribution of the reverse process can be derived through analysis as follows: (5), (2) Dual-stream multi-scale conditional coding The model uses parameterized functions Modeling the reverse generation process, it predicts The definition is as follows: (6), in It is the model's predicted output. The conditional information consists of multi-scale feature maps of low-field DWI and feature maps of the corresponding lesion masks.

[0038] To effectively utilize the intrinsic features of low-field DWI derived from low-field imaging characteristics and the spatial prior information embedded in the stroke lesion mask, a two-stream multi-scale conditional coding framework is introduced, with the structure as follows: Figure 3 As shown, the architecture contains two parallel branches: one for extracting multi-scale feature information from low-field DWI, and the other for extracting multi-scale feature information from stroke lesion masks.

[0039] The DWI branch focuses on extracting multi-scale, hierarchical intrinsic features from low-field DWI, while the stroke lesion masking branch captures the spatially defined structural and morphological features of the stroke lesion region. During decoding, the feature maps extracted by the DWI branch are concatenated layer by layer with the corresponding feature maps on the upsampling path of the decoder in the generative model; simultaneously, the stroke lesion masking branch provides its extracted structural features to the shallow decoding layer in the generative model.

[0040] This dense multi-scale conditional fusion mechanism provides the generative model with rich context and spatial constraints, explicitly guiding and enhancing its ability to perceive and reconstruct stroke lesion areas, thereby improving target generation. The accuracy is improved. Simultaneously, this mechanism effectively suppresses over-smoothing during training, thus preserving stroke lesion features on low-field DWI in the enhanced DWI images.

[0041] (3) Time-adaptive loss optimization strategy To further stabilize the training process and enhance the perceptual quality of the generated DWI, a time-adaptive loss optimization strategy based on parameterized v is proposed.

[0042] definition The target vector for loss calculation is used using the following formula: (7), in: (8), The main training objective is defined as follows: and Mean square error between: (9), Furthermore, regarding Perform inverse parameterization to obtain the generated DWI ,as follows: (10), To improve the perception quality of fine details in DWI generated under low noise levels, a time-adaptive loss supervision mechanism is introduced, which employs a weighting function that decays over time. As shown in the following formula: (11), Where T is a hyperparameter representing the total time step.

[0043] Pixel-level loss DWI for supervision generation Perceived loss The generated DWI is further optimized based on perceived quality, as shown in the following formula: (12), (13), in This represents a pre-trained feature extraction network that computes perceptual loss. Indicates the number of floors.

[0044] The complete loss is represented as: (14), in, , .

[0045] S6: Training the Model The difference-guided conditional diffusion model is trained using the training set constructed in S4. The training process is optimized by backpropagation based on the total loss function mentioned above, and the model parameters are updated.

[0046] S7: Evaluate the model and find the optimal weights The model is tested using data from the test set, the generated results are evaluated, and the weights of the model with the best performance are selected and saved.

[0047] S8: Image Quality Enhancement Inference Using a trained generative model, end-to-end clear MRI images of the low-field DWI to be processed are generated, outputting enhanced DWI images. The inference process is as follows: Figure 2 As shown in section C, the reverse diffusion process is completed in T iterations, ultimately yielding an enhanced high-resolution DWI image.

[0048] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0049] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for enhancing the image quality of low-field magnetic resonance imaging (MRI) for stroke diffusion-weighted imaging, characterized in that, Includes the following steps: S1. Construct original input image pairs, the original input image pairs including paired low field strength DWI and high field strength DWI; S2. Perform preprocessing operations on the original input image pairs, including skull dissection, image registration, and data standardization. S3. Use the pre-trained UNet segmentation network to segment the low-field DWI and obtain the corresponding stroke lesion mask. S4. Divide the paired low-field DWI, high-field DWI and corresponding lesion masks into training set and test set; S5. Construct a difference-guided conditional diffusion model, which includes a difference-guided adaptive forward noise module, a two-stream multi-scale conditional coding module, and a time-adaptive loss optimization module. S6. Train the difference-guided conditional diffusion model using the training set; S7. Use the test set to perform generation testing and performance evaluation on the trained model, and save the weights of the model with the best performance. S8. Using a difference-guided conditional diffusion model that preserves the optimal weights, end-to-end sharpening processing is performed on the low-field DWI to be processed, and the quality-enhanced DWI image is output.

2. The method for enhancing the image quality of low-field magnetic resonance imaging (MRI) for stroke according to claim 1, characterized in that, The specific process of the preprocessing operation in S2 is as follows: The skull was removed by removing the skull in low-field DWI and high-field DWI respectively by setting a preset threshold, thus completing the skull dissection. A nonlinear registration method was used, with the high-field DWI as the template, and the low-field DWI was registered into the high-field DWI template. The registered low-field DWI and high-field DWI have a size of 256×256×13, which correspond to the height, width and number of slices of the image, respectively. The registered paired data are standardized using the maximum-minimum normalization method, mapping the data to the numerical range of [0,1].

3. The method for enhancing the image quality of low-field magnetic resonance imaging (MRI) for stroke according to claim 1, characterized in that, The pre-trained UNet segmentation network used in S3 has a segmentation accuracy of 95%, and the resulting stroke lesion mask is used to guide the training and generation processes of the differentially guided conditional diffusion model.

4. The method for enhancing the image quality of low-field magnetic resonance imaging (MRI) for stroke according to claim 1, characterized in that, In S4, the paired low-field DWI, high-field DWI, and corresponding lesion masks are divided into training and test sets in a 9:1 ratio.

5. The method for enhancing the image quality of low-field magnetic resonance imaging (MRI) for stroke according to claim 1, characterized in that, The first stage of the difference-guided adaptive forward noise module is: the difference between the low-field DWI and the high-field DWI is introduced into the forward diffusion process as a directional drift term, so that the diffusion trajectory is consistent with the intrinsic distribution between the low-field and high-field DWI. The diffusion process after introducing the directional drift term satisfies the following equation: （1）， in, , , Indicates high-field DWI, Indicates low-field DWI, Indicates the time-controlled noise scheduling factor. The hyperparameter representing the overall noise intensity is 2. For time t The noise addition result is given by the following formula: （2）。 6. The method for enhancing the image quality of low-field magnetic resonance imaging (MRI) for stroke according to claim 1, characterized in that, The second stage of the difference-guided adaptive forward noise module is as follows: calculate the difference map based on low-field DWI and the corresponding lesion mask, and adaptively adjust the noise injection intensity of each region according to the value in the difference map. Stronger noise is injected into healthy tissue regions with higher difference values, and weaker noise is injected into stroke lesion regions with lower difference values. The difference map is defined according to the following formula: (3)， in, Represents a difference plot. This represents the mask for stroke lesions on low-field DWI. Represents the Hadamard product; The forward diffusion process after introducing the difference map satisfies the following equation: （4）， in, Represents the identity matrix; The corresponding posterior distribution of the reverse process satisfies the following equation: (5）。 7. The method for enhancing the image quality of low-field magnetic resonance imaging (MRI) for stroke according to claim 1, characterized in that, The dual-stream multi-scale conditional coding module includes two parallel coding branches: a low-field DWI coding branch and a lesion mask coding branch. The low-field DWI coding branch is used to extract multi-scale hierarchical intrinsic features from low-field DWI. During the decoding process, the feature map extracted by this branch is spliced ​​layer by layer with the corresponding feature map of the upsampling path of the generator model decoder. The lesion mask encoding branch is used to extract multi-scale spatial structure and morphological features from the stroke lesion mask. During the decoding process, the structural features extracted by this branch are input to the shallow decoding layer of the generation model. The features extracted from the two branches provide context and spatial constraints for the generative model in a dense multi-scale conditional fusion manner, guiding the generative model to perceive and reconstruct the stroke lesion area.

8. The method for enhancing the image quality of low-field magnetic resonance imaging (MRI) for stroke according to claim 1, characterized in that, The time-adaptive loss optimization module constructs a loss function based on parameterized v, and the specific process is as follows: Define target vector Satisfy the following formula: (6)， in: (7)， The predicted output of the generative model is: (8)， in The parameterized function for generating the model. The conditional information consists of multi-scale feature maps from low-field DWI and feature maps from lesion masks. For diffusion time steps; Predicting losses for and The mean square error between them satisfies: (9)， Furthermore, regarding Perform inverse parameterization to obtain the generated DWI ,as follows: (10）。 9. The method for enhancing the image quality of low-field magnetic resonance imaging (MRI) for stroke according to claim 8, characterized in that, Based on the generated DWI Constructing pixel-level MSE loss and perceptual loss: The pixel-level MSE loss satisfy: (11)， The perceived loss satisfy: (12), in For the first pre-trained feature extraction network Layer output, This represents the number of network layers.

10. The method for enhancing the image quality of low-field magnetic resonance imaging (MRI) for stroke according to claim 9, characterized in that, Introducing a time-adaptive weighting function Construct the total loss, the time-adaptive weighted function satisfy: (13)， Where T is a hyperparameter representing the total time step; Total loss Satisfy the following formula: (14)， Among them, the weighting coefficient , .

Images