A deep learning oriented low-field MRI image optimization enhancement method

Abstract

The application discloses a kind of deep learning-oriented low-field MRI image optimization enhancement method.Due to the limitation of different MRI imaging instruments, it is difficult to obtain a large number of paired high-low field data for training of deep learning model in practice, therefore, a deep learning-oriented low-field MRI image simulation method is proposed in this paper, combining the characteristics of low signal-to-noise ratio and low resolution of low-field images, a large number of one-to-one corresponding high-low field paired data are obtained for training by undersampling and adding K-space noise to high-field images. Further, a low-field MRI image enhancement method based on noise adaptive perception is proposed, which uses the method of adding noise and scaling in the k-space of high-field MRI to simulate low-field data, and uses a noise adaptive perception branch to predict the noise in the low-field image, so as to optimize the image more finely. The experimental results show that whether using high-field simulation data or real data collected on a low-field instrument, the method has good characteristics in reducing the noise of MRI images and retaining image details, and the model trained with simulation data can be directly applied to unseen real data, and has great potential in solving the actual low-field image quality problem.

Description

Technical Field

[0001] This invention relates to an optimization and enhancement method for low-field MRI images based on deep learning, belonging to the field of computer image processing technology. Background Technology

[0002] Low-field magnetic resonance imaging (MRI) systems offer a more cost-effective way to acquire diagnostic images in medical imaging. This method boasts advantages such as low cost, minimal shielding requirements, and small device size, making it more practical in resource-constrained environments. However, the clinical application of low-field MRI systems has significant drawbacks. Due to the weak signal at low field strengths, noise and low resolution issues in the acquired data ultimately manifest as significant noise interference, low resolution, and blurred details in the reconstructed images, greatly affecting image quality.

[0003] Noise and resolution issues in low-field MRI typically manifest in two main forms: random noise and low spatial resolution. Random noise is primarily caused by system electronic noise, thermal noise, and random interference during signal acquisition, appearing as point-like or granular interference in the image. Furthermore, because the signal intensity is lower at low fields while the noise signal remains relatively constant, this significantly reduces the image's signal-to-noise ratio (SNR). Low resolution issues manifest as blurred image details and indistinct edges, due to the weaker signal intensity at low field strengths, leading to a significant decrease in spatial resolution.

[0004] Traditional methods for removing noise and artifacts in low-field MRI fall into two categories: one starts from the raw data domain (k-space), optimizing signal acquisition strategies or using filters to preprocess k-space data to reduce the impact of noise and artifacts; the other starts from the reconstructed image domain, utilizing various post-processing techniques (such as nonlocal mean filtering and wavelet transform) to extract noise and artifact structures, achieving denoising and artifact removal. However, these methods have limited effectiveness in handling low-resolution problems, require extensive experience in image reconstruction and signal processing, and are not particularly effective under low-field conditions. Furthermore, due to inconsistencies in the hardware performance and noise characteristics of different devices, traditional methods typically require parameter adjustments for specific equipment, making them difficult to promote as a universal solution.

[0005] In recent years, with the development of deep learning technology, data-driven low-field MRI denoising and resolution enhancement methods have gradually become a research hotspot. These methods, through training neural network models, can automatically learn the features of noise, artifacts, and low-resolution images, achieving efficient noise suppression, artifact removal, and resolution enhancement during reconstruction. For example, super-resolution reconstruction methods based on convolutional neural networks can recover high-resolution details from low-resolution MRI images, while effectively suppressing random noise and structured artifacts when combined with denoising modules. Furthermore, generative adversarial networks (GANs) have been widely applied to low-field MRI image enhancement tasks, generating high-quality images through adversarial learning, significantly improving the clinical application value of low-field MRI.

[0006] Despite significant advancements in low-field MRI denoising and resolution enhancement using deep learning techniques, several challenges remain. For instance, training deep learning models requires a large volume of high-quality low-field and high-field MRI paired data, which is difficult to obtain in practical applications. Therefore, this invention proposes a low-field data simulation method for deep learning. Based on obtaining sufficient high-field-low-field training data pairs, it proposes an optimized enhancement method for low-field MRI images that combines sliding self-attention and channel attention mechanisms. Summary of the Invention

[0007] This invention addresses the poor performance of traditional enhancement methods in low-field MRI images by proposing an optimized enhancement method based on deep learning. The entire process of this method is as follows: Figure 2 As shown, a low-field MRI noise simulation method was used to obtain one-to-one paired high-field-low-field training data pairs for network training; a dual-path denoising and enhancement network was designed, which takes low-field, low-quality images as input and outputs enhanced, high-quality images, thereby improving the signal-to-noise ratio of the images while ensuring the preservation of image details.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] A method for simulating low-field MRI noise images based on the principle of MRI effect noise is proposed.

[0010] Step 1. Scale the high-field MRI image. To match the low-resolution characteristics of actual low-field MRI images, the 1.5T high-field MRI image is first scaled. Scale reduces the spatial resolution of the image, making it closer to the characteristics of a real low-field image.

[0011] (1)

[0012] in For the original high-field image, To reduce the resolution of the image.

[0013] Step 2. Add noise to the image obtained in Step 1. In actual low-field MRI systems, the sources of noise are very complex, with thermal noise, electrical noise, and coil noise all coupling into the final MRI image. Since the original MRI image data is K-space data, this invention adds mixed random Gaussian noise to the K-space of the image obtained in Step 1, and the standard deviation of the added random Gaussian noise varies according to the range of the real and imaginary parts of the K-space.

[0014] (2) (3)

[0015] Where F is the Fourier transform. for K-space data, and They are The real and imaginary parts, and These represent random Gaussian noise added to the real and imaginary parts, respectively. This is the K-space data for the final simulated low-field MRI image.

[0016] Step 3. Perform inverse Fourier transform on the K-space data obtained in Step 2 to obtain the final simulated low-field data. Such low-field images are closer to real low-field MRI images in terms of resolution and noise distribution, and can form one-to-one training data pairs with 1.5T high-field images, which is convenient for training using supervised deep learning.

[0017] (4)

[0018] in This is the inverse Fourier transform. It is a simulated low-field K-space. This is the final simulated low-field image. In the test, LPIPS and FID were used as indicators to measure the enhancement effect of low-field MRI images.

[0019] We propose a deep learning-based low-field image optimization method based on a hybrid attention mechanism. This method overcomes the problem of obtaining a one-to-one correspondence between high-field and low-field images, which is difficult to achieve in practice, compared to the aforementioned low-field MRI noise simulation method based on the principle of MRI effect noise. With sufficient training data, we designed a dual-path denoising and enhancement network for low-field image optimization, and also designed the training and testing process for the entire network.

[0020] Step 1: Design a dual-path noise reduction and enhancement network.

[0021] Our network employs a dual-path structure, consisting of a noise-adaptive sensing branch for predicting noise in the input low-field MRI image and a denoising backbone for recovering image details.

[0022] The denoising backbone employs a hybrid attention network structure, consisting of multiple hybrid attention modules and a cascaded 3x3 convolutional layer. First, the image passes through a convolutional layer to extract shallow features. These shallow features are then progressively processed to extract deeper features, which are finally aggregated by a convolutional layer. Each hybrid attention module comprises multiple hybrid attention mechanism blocks, an overlapping cross-attention block, a convolutional layer, and a residual connection. Specifically, the hybrid attention block uses a Swin transformer-like structure and introduces a channel attention mechanism. This design utilizes more global image information, activating more pixels for feature learning, thus enhancing the network's reconstruction capabilities. The overlapping cross-attention block consists of an overlapping cross-attention layer and an MLP layer, similar to a standard Swin transformer block. By directly establishing cross-window connections, the overlapping cross-attention block enhances the representativeness of window self-attention. This design effectively preserves image details while removing noise during feature extraction. Furthermore, the introduction of overlapping cross-attention reduces the number of parameters required by the network, lowering training costs.

[0023] In the noise adaptive perception branch, we designed a UNET-like structure, consisting of a symmetric encoder-decoder architecture. The encoder comprises two downsampling stages forming a progressive feature abstraction, each consisting of two convolutional ReLU combinations. After the two downsampling stages, a max-pooling layer halves the feature space resolution. After the two-stage encoder, a bottleneck module constructs a high-level semantic feature representation while maintaining the same dimensionality as the input features. The decoder, through a symmetric structure with the encoder module, achieves feature refinement and resolution restoration. In the first upsampling stage, transposed convolutions (2×2 kernels, stride 2) double the feature map resolution, reducing the number of channels to 64. The second-stage encoder features are fused through lateral connections and processed by 64-channel dual convolutions (3×3 kernels) for feature fusion. The second upsampling stage uses transposed convolutions with the same parameters to restore the original resolution, reducing the number of channels to 32. By concatenating with the first-stage encoder features, a 32-channel dual convolution completes the detail reconstruction. Finally, the output layer uses a 1×1 convolution kernel to map the 32-channel features to the target dimension, directly generating a noise prediction map. This branch can perceive detailed information about noise from the input image.

[0024] Step 2: Design the network training and testing process.

[0025] (1) First, we use a low-field MRI noise image simulation method based on the principle of MRI effect noise to obtain a large number of one-to-one corresponding training dataset pairs from high-field images, and use a dual-path denoising and enhancement network for supervised training. Furthermore, as shown in the figure, we use image restoration loss and noise prediction loss to jointly supervise the network training, thereby obtaining better image noise removal and detail restoration effects. During the testing phase, since there are no high-quality images that correspond one-to-one with the real low-field images, the metrics SSIM and PSNR, which are generally used for image optimization and enhancement tasks, are not suitable for low-field image optimization. We introduce FID and LPIPS metrics to evaluate the results. Specifically, we calculate FID and LPIPS between the high-field dataset and the enhanced low-field dataset, and measure the similarity between the optimized low-field image and the high-field image by the distance between the two datasets in the feature space. This measurement method is more scientific.

[0026] Compared with the prior art, the advantages of the present invention are as follows: (1) It proposes a low-field MRI image simulation method, which provides data support for the application of supervised learning in the field of low-field MRI image enhancement. This method fully considers the MRI system and low-field MRI noise characteristics, and the simulated low-field MRI data is comprehensive, rich in features, and closer to the actual artifacts. (2) It has a significant effect on removing noise from low-field MRI images and has practical potential. This method fully integrates attention mechanisms, fully considers various information of the image at both levels of network design, and designs a dual-path denoising enhancement network. The experimental results of simulated data and real data show that the method has good loop removal performance. (3) It restores image details and ensures that the image after denoising enhancement does not produce distortion or information loss. The dual-path denoising enhancement network proposed by this method not only has a good removal effect, but also shows good performance in restoring image details and removing noise, which meets the actual diagnostic needs of doctors. (4) It runs fast. The optimized enhancement method runs on GPU, with low computational requirements and short actual test time. (5) It has universality and parameter-free adjustment. This method requires no parameter adjustment and has good optimization and enhancement effects for any low-field MRI device. (6) This invention is an automated, end-to-end method that can be integrated into the reconstruction software of MRI devices. Only the corresponding buttons need to be designed to obtain high-quality images with ring artifacts removed. Attached Figure Description

[0027] Figure 1 This is a flowchart of the present invention.

[0028] Figure 2 This is a diagram of the dual-path denoising enhancement network structure.

[0029] Figure 3 The graph shows the test results of the method.

[0030] Figure 4 A schematic diagram of adding noise to the K-space. Detailed Implementation

[0031] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0032] Example: Figure 1 As shown, this invention provides a data simulation and optimization enhancement method for low-field MRI images based on deep learning. The enhancement test results are as follows: Figure 4 As shown.

[0033] I. Prerequisites for the Implementation Examples

[0034] The training equipment used an NVIDIA 3090 GPU (24GB VRAM), an Intel Xeon Platinum 8375C CPU (32 cores, 64 threads), and 128GB of RAM. The testing equipment supported NVIDIA RTX 3090 or higher GPUs (24GB VRAM) or equivalent computing power. The operating system was Ubuntu 20.04 LTS, and the deep learning framework used was PyTorch 1.12.1+cu116, along with CUDA 11.6 and CuDNN 8.4.1. Auxiliary libraries included NumPy 1.23.5, SciPy 1.9.3, SimpleITK 2.2.1, and PIL 9.3.0. High-field MRI data used the publicly available datasets IXI and fastMRI, while the real low-field MRI test data consisted of 10 brain images acquired using a 0.1T device.

[0035] The specific steps are as follows:

[0036] Step 1: High-field MRI Image Resolution Reduction (Simulating Low-field, Low-resolution Characteristics) The original 1.5T high-field MRI image is read using the SimpleITK library. 2D brain tissue region slices are extracted, and background blank slices are removed to obtain a single-channel grayscale image. The PyTorch.nn.functional.interpolate function (mode='bilinear', align_corners=True) is used to perform bilinear interpolation, scaling the image size to the resolution of the target field intensity (e.g., if the voxel size of the target low-field intensity image is twice that of the high-field image, then the original high-field image is resized to half its original size). Then, the formula is used...

[0037]

[0038] Pixel values ​​are normalized to the range [0, 1] to avoid feature distortion caused by subsequent noise addition, where I is the input 2D brain slice. and These are the minimum and maximum grayscale values ​​of the image, respectively.

[0039] Step 2: K-space noise addition operation (simulating low-field noise characteristics) The normalized image obtained in Step 1 is subjected to a 2D Fast Fourier Transform using the SciPy.fftpack.fft2 function, converting the image domain data into K-space frequency domain data containing real and imaginary parts; the real part of the K-space is then calculated. and the virtual part maximum value and Determine the standard deviation of the real noise. (Using 0.1T as the simulated target), standard deviation of imaginary noise (Using 0.1T as the simulation target). Gaussian noise with a mean of 0 is generated using the NumPy.random.normal function and superimposed onto the real and imaginary parts of the K-space to obtain noisy complex data in the K-space. Finally, the SciPy.fftpack.ifft2 function is used to perform an inverse fast Fourier transform, converting the frequency domain data back to the image domain to obtain a simulated low-field MRI image. .

[0040] Step 3: Construction of high-field-simulated low-field data pairs. Pair the simulated low-field images obtained in Step 2 with the corresponding original high-field images to form "high-field clear image - low-field simulated image" data pairs. A total of 20,000 training data pairs, 800 validation data pairs, and 200 test data pairs are constructed. The data is randomly divided according to the case number and there are no overlapping cases to avoid data leakage. Finally, all data pairs are converted into PyTorch tensor format (dimensions [B, C, H, W]) to improve the speed of reading training data.

[0041] Step 4: Design and Construction of Dual-Path Denoising Enhancement Network. A dual-path parallel network structure is built based on the PyTorch framework. The denoising backbone consists of an input layer → shallow convolution (3×3 kernels, 64 channels, ReLU activation) → 6 hybrid attention modules (including SW-MSA, channel attention CAB, and overlapping cross attention OCA) → deep feature aggregation convolution (3×3 kernels, 1 channel) → output layer. The noise adaptive perception branch consists of an encoder (2 downsampling stages, 3×3 kernels convolutional layers, 64→128 channels, ReLU activation, max pooling 2×2 stride 2) → bottleneck module (128-channel convolution + BN layer) → decoder (2 upsampling stages, transposed convolution 2×2 kernel stride 2, 128→64→32 channels) → noise prediction output layer (1×1 kernel, 1 channel). Dual-path feature fusion is achieved through concatenation and 1×1 convolution dimensionality reduction to enhance feature consistency.

[0042] Step 5: The network training process uses a joint loss function. . To enhance the MSE loss of the image and the high-field sharpness image, To predict the MSE loss between the noise map and the real noise map, the batch size was set to 8, the number of training epochs to 100, and the initial learning rate to 1e-4 (cosine annealing scheduling strategy, T_max=100, eventually reduced to 1e-6). The optimizer used was AdamW (betas=(0.9, 0.999), weight decay 1e-5). An early stopping strategy was adopted (monitoring the validation set metrics, stopping training if there is no optimization for 10 consecutive epochs, and saving the optimal model weights). The training set / validation set loss values ​​were output every 5 epochs, and the validation set enhancement effect was visualized every 20 epochs to ensure stable convergence of training.

[0043] Step 6: Real Low-Field MRI Image Enhancement Test. Ten real low-field MRI images were read and normalized to the range [0, 1] (without adding extra noise). The trained optimal model weights were loaded, gradient calculation was turned off (torch.no_grad()), and forward propagation was performed. The noise prediction map was output through the noise adaptive sensing branch, and the enhanced image was output through the denoising backbone. The final result was obtained after feature fusion. The enhanced image was denormalized from [0, 1] and saved in DICOM format for clinical observation and quantitative evaluation.

[0044] Effectiveness evaluation:

[0045] This invention discloses a data simulation and optimization enhancement method for low-field MRI images based on deep learning. The test results of this method are as follows: Figure 3As shown in the diagram, the testing procedure is as follows: low-field MRI image data is input into the model, and the deep learning model is activated to enhance the low-field images. After the model's inference, a high-quality denoised and enhanced MRI image is output. Comparison with the original image shows that the signal-to-noise ratio and detail recovery are significantly improved after denoising and enhancement. This algorithm has low memory usage, fast running speed, and can process images in a short time, removing noise from low-field MRI images, optimizing details, enhancing image quality, and promoting clinical applications.

[0046] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any transformations or substitutions that can be conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A data simulation and optimization enhancement method for low-field MRI images oriented to deep learning, characterized in that, The specific steps for low-field data simulation using MRI data are as follows: Step 1: Interpolate the 1.5T high-field MRI image to reduce its size to half and decrease its resolution. Step 2: Divide the image processed in Step 2 into real and imaginary parts in K-space. Add random Gaussian noise with zero mean and standard deviation within a certain range to the real and imaginary parts respectively to obtain noisy frequency domain data. Then perform inverse Fourier transform on this frequency domain data to obtain simulated low-field MRI images. Step 3: Pair the image obtained after processing in Step 2 with the original 1.5T high-field MRI image to obtain a high-field-simulated low-field image data pair; Step 4: Design a dual-path denoising and enhancement network, which consists of a noise adaptive sensing branch and a denoising backbone; Step 5: Use the high-field-simulated low-field image data obtained in Step 3 to train the dual-path denoising enhancement network; Step 6: Input the real low-field MRI image data into the trained dual-channel denoising and enhancement network for optimization and enhancement, and obtain optimized high-quality low-field MRI images, which significantly improves the resolution and signal-to-noise ratio of the low-field images.

2. The data simulation and optimization enhancement method for low-field MRI images based on deep learning according to claim 1, characterized in that, The image interpolation operation mentioned in step 1 simulates the low resolution problem of low-field images through interpolation.

3. The method of claim 1, wherein the method further comprises: determining a plurality of parameters of the low-field MRI system based on the plurality of simulated low-field MRI images. The K-space noise addition operation mentioned in step 2 divides the low-resolution MR image into real and imaginary parts in K-space, and adds random Gaussian noise to both parts. Finally, the noisy K-space data is transformed into the image domain using inverse Fourier transform to obtain a simulated low-field image. Since the original MRI image data is a frequency domain image, and the complex coupling noise in the low-field image data is treated as Gaussian noise, the standard deviation parameter of the noise is set to a random value related to the K-space value range of the image during simulation to better simulate the noise of a real MRI system.

4. The method of claim 1, wherein, The image pair proposed in step 3 is a dataset. Through the degradation in steps 1 and 2, the resolution and signal-to-noise ratio of the high-field 1.5T image are reduced, resulting in a simulated low-field image. This simulated low-field image is then combined with the original high-field 1.5T image to obtain a high-field-noise image data pair.

5. The method of claim 1, wherein, Step 4: Design a dual-path denoising and enhancement network. The network uses a dual-path structure, where the noise adaptive perception branch perceives the noise features of the input low-field MRI image, while the denoising backbone extracts the image details at different scales and predicts the noise. It also uses a combination of channel attention and sliding window attention mechanisms to fuse the deep features and channel information of the noisy image. By combining the features of these two branches, the dual-path denoising and enhancement network significantly improves the denoising effect and information preservation ability of the image.

6. The low-field MRI image data simulation and optimization enhancement method based on deep learning according to claim 1, wherein, Step 5 uses the high-field-noise image data pair to train the dual-path denoising enhancement network. The training loss function is the sum of the MSE losses between the enhanced image and the high-field 1.5T image, the predicted noise data, and the actual noise data. The gradient is calculated through the loss, and the network parameters are updated using the stochastic gradient descent optimization method (AdamW).

7. The data simulation and optimization enhancement method for low-field MRI images based on deep learning according to claim 1, characterized in that, The dual-path denoising enhancement network testing process proposed in step 6 involves inputting the low-field MRI image into a trained dual-path denoising network with fixed weight parameters to obtain a high-quality denoised MRI image.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements a data simulation and optimization enhancement method for low-field MRI images based on deep learning, as described in any one of claims 1 to 7.

9. A computer readable storage medium having stored thereon computer instructions, wherein, When executed by a processor, the computer instructions implement a data simulation and optimization enhancement method for low-field MRI images based on deep learning, as described in any one of claims 1-7.

Images