A bias correction self-supervised deep learning magnetic resonance image denoising method

Abstract

This invention discloses a bias-correcting self-supervised deep learning method for magnetic resonance imaging (MRI) denoising. The method involves acquiring noisy MRI images and their corresponding squared data to construct training and testing sets; building a self-supervised denoising network; training the self-supervised denoising network based on the training set; squaring the noisy MRI image to be processed and inputting it into the trained self-supervised denoising network to further obtain a background mask and a denoised MRI image. This invention utilizes only noisy MRI images as the training set, reducing the manpower required to collect noise-free MRI images, and can correct biases caused by non-zero mean noise, providing more accurate denoising results, thus having practical application value. No additional manual optimization is required during application, and the denoising effect is excellent.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a self-supervised deep learning magnetic resonance image denoising method for bias correction, applicable to magnetic resonance image denoising. Background Technology

[0002] Magnetic Resonance Imaging (MRI) is a commonly used imaging technique in the medical field, providing crucial information about the structure and function of biological tissues and organs. However, noise is unavoidable during MRI scans, significantly impacting data accuracy and image quality. Therefore, denoising of acquired MRI data is necessary to improve image quality. Denoising methods can be categorized into two types: one involves improving the performance of the imaging equipment, such as increasing magnetic field strength or extending scan time, but this increases cost and examination time while causing discomfort to the patient; the other method utilizes image post-processing algorithms to denoise low-quality MRI images. This method does not alter the original imaging process, is easily integrated into existing equipment, and has practical application value.

[0003] In recent years, deep learning-based magnetic resonance imaging (MRI) denoising methods have demonstrated superior performance, showing significant advantages in denoising effect, speed, and model application. Deep learning-based MRI denoising algorithms can be trained on external datasets to autonomously establish a nonlinear mapping model from noisy MRI images to noise-free MRI images. However, most existing deep learning MRI denoising methods rely on paired noisy and noise-free images for effective training, which is often difficult to achieve in practice. To enable training of denoising models even without paired images, Krull et al. proposed a blind spot network for denoising single noisy images. This method trains the network in a self-supervised manner, requiring only the noisy image. However, this self-supervised denoising method requires zero mean noise in the image, while MRI noise follows a Ricean distribution, which does not meet the zero-mean noise requirement. Therefore, directly applying self-supervised denoising to MRI images will lead to denoising bias.

[0004] In summary, the existing technology has the following problems:

[0005] 1. Some solutions require both noisy and noiseless MRI images, which is difficult to achieve in real-world scenarios.

[0006] 2. Other solutions do not require noise-free images when applied, but may cause denoising deviations due to non-zero mean noise. Summary of the Invention

[0007] The purpose of this invention is to address the aforementioned problems in existing technologies by proposing a bias-correcting self-supervised deep learning method for denoising magnetic resonance images. This method trains the denoising network without the presence of noise-free magnetic resonance images, while simultaneously correcting biases caused by non-zero mean noise. After model training, it enables fast and accurate denoising of magnetic resonance images.

[0008] The above-mentioned objective of this invention is achieved through the following technical solution:

[0009] A bias-correction self-supervised deep learning method for magnetic resonance imaging denoising includes the following steps:

[0010] Step 1: Obtain noisy magnetic resonance images and corresponding squared noisy magnetic resonance images, and construct training and testing sets based on the squared noisy magnetic resonance images.

[0011] Step 2: Construct a self-supervised denoising network, which includes an encoder and a decoder;

[0012] Step 3: Input the squared data of noisy magnetic resonance images from the training set obtained in Step 1 into the self-supervised denoising network;

[0013] Step 4: Define the self-supervised denoising loss function L;

[0014] Step 5: Based on the self-supervised denoising loss function set in Step 4, use the training set generated in Step 1 to train the self-supervised denoising network constructed in Step 2 end-to-end, and save the parameters of the self-supervised denoising network.

[0015] Step 6: Perform the squaring process in Step 1 on the noisy magnetic resonance image to be processed to obtain the squared data of the magnetic resonance image to be processed, and then input it into the self-supervised denoising network trained in Step 5 to obtain the denoised magnetic resonance image squared data.

[0016] Step 7: Segment the predicted denoised magnetic resonance image squared data obtained in Step 6 to obtain the background mask, and calculate the noise variance based on the background mask;

[0017] Step 8: Subtract twice the noise variance obtained in Step 7 from the squared data of the predicted denoised magnetic resonance image obtained in Step 6, and then take the square root to obtain the denoised magnetic resonance image.

[0018] As described above, the squared data of the noisy magnetic resonance imaging in step 1 are calculated based on the following formula:

[0019]

[0020] Where x represents a noisy magnetic resonance imaging image, and ⊙ represents a dot product. This represents the squared data of the noisy magnetic resonance imaging obtained.

[0021] As described above in step 2, the encoder includes one input layer and multiple downsampling layers in sequence, and the decoder includes multiple upsampling layers and one output layer in sequence. The number of upsampling layers is the same as the number of downsampling layers. The input layer includes a convolutional layer, a group normalization layer, and an activation layer, and the output layer includes an adaptive fusion module and a convolutional layer.

[0022] As described above, the downsampling layer includes a convolutional layer, a group normalization layer, and an activation layer.

[0023] The feature map output from the last downsampling layer is used as the input to the first upsampling layer.

[0024] The first upsampling layer includes an upsampling module, a convolutional layer, a group normalization layer, and an activation layer.

[0025] The i-th upsampling layer includes an adaptive fusion module, an upsampling module, a convolutional layer, a group normalization layer, and an activation layer, where i is the upsampling layer index other than 1.

[0026] The feature map output by the (S+1-i)th downsampling layer and the feature map output by the (i-1)th upsampling layer are used as the input to the adaptive fusion module in the ith upsampling layer, where S is the number of upsampling and downsampling layers.

[0027] The adaptive fusion module in the i-th upsampling layer includes the first convolutional layer, the first activation layer, the second convolutional layer, and the second activation layer.

[0028] As described above, in the i-th upsampling layer, the adaptive fusion module stacks the feature maps output from the (S+1-i)-th downsampling layer and the (i-1)-th upsampling layer according to their channels, and uses them as the input to the first convolutional layer of the adaptive fusion module in the i-th upsampling layer. The second activation layer of the adaptive fusion module in the i-th upsampling layer outputs a weight map m. The feature map output from the (i-1)-th upsampling layer is then multiplied by the weight map m calculated in the adaptive fusion module in the i-th upsampling layer to obtain the feature map output from the (i-1)-th upsampling layer. The dot product feature map corresponding to the image is obtained by multiplying the feature map output by the (S+1-i)th downsampling layer with (1 - the weight map m calculated by the adaptive fusion module in the ith upsampling layer). The dot product feature map corresponding to the feature map output by the (i-1)th upsampling layer is then added pixel by pixel to obtain the fused feature map output by the adaptive fusion module in the ith upsampling layer.

[0029] As described above, the fused feature map output by the adaptive fusion module in the i-th upsampling layer is input to the upsampling module in the i-th upsampling layer. The feature map output by the upsampling module in the i-th upsampling layer is then input to the convolutional layer of the i-th upsampling layer. The feature map output by the convolutional layer of the i-th upsampling layer is then passed through the group normalization layer and activation layer in the i-th upsampling layer to obtain the feature map output by the i-th upsampling layer.

[0030] As described above, the adaptive fusion module in the output layer uses the feature map output from the input layer and the feature map output from the last upsampling layer, stacked by channel, as the input to the first convolutional layer of the adaptive fusion module in the output layer. The second activation layer of the adaptive fusion module in the output layer outputs a weight map m. The feature map output from the input layer is then multiplied by the weight map m calculated by the adaptive fusion module in the output layer to obtain the multiplied feature map corresponding to the feature map output from the output layer. The feature map output from the last upsampling layer is then multiplied by (1 - the weight map m calculated by the adaptive fusion module in the output layer) to obtain the multiplied feature map corresponding to the feature map output from the last upsampling layer, thus obtaining the fused feature map output by the adaptive fusion module in the output layer.

[0031] As described above, the fused feature map output by the adaptive fusion module in the output layer is used as the input to the convolutional layer in the output layer, and the convolutional layer in the output layer outputs the predicted denoised magnetic resonance image squared data.

[0032] As mentioned above, the self-supervised denoising loss function L is based on the following formula:

[0033]

[0034] Where ||·||2 represents the F-norm operation, is the squared data of noisy magnetic resonance imaging, n is a random mask following a Bernoulli distribution, and f θ Represents a self-supervised denoising network, f θ (·) represents the output of the self-supervised denoising network.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] 1. This invention constructs a self-supervised denoising network and trains it using a pre-defined self-supervised denoising loss function even without noise-free MRI images. It then uses the estimated noise variance to remove denoising bias from the squared data of the obtained denoised MRI images, effectively ensuring the denoising performance of the MRI images. Compared to existing deep learning-based MRI image denoising methods, this invention only requires noisy MRI images as the training set, reducing the manpower required to collect noise-free MRI images. Furthermore, it can correct the bias caused by non-zero mean noise, providing more accurate denoising results and possessing practical application value.

[0037] 2. Compared with traditional magnetic resonance imaging denoising methods, this invention does not require additional parameter adjustment after the self-supervised denoising network is trained. It has a fast denoising speed, can better preserve image details, and has excellent results. Attached Figure Description

[0038] Figure 1 This is a flowchart of the present invention;

[0039] Figure 2 This is a schematic diagram of the structure of a self-supervised denoising network;

[0040] Figure 3 This is a schematic diagram of the adaptive fusion module.

[0041] Figure 4 The diagram illustrates the denoising effect; where (a) is a low signal-to-noise ratio 2D magnetic resonance brain image, and (b) is a denoised 2D magnetic resonance brain image.

[0042] Figure 5 This is a schematic diagram comparing the performance of the present invention with other methods on the constructed dataset;

[0043] Table 1 is a quantitative comparison table of the performance of the present invention and other methods. Detailed Implementation

[0044] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0045] A bias-correction self-supervised deep learning method for magnetic resonance imaging denoising includes the following steps:

[0046] Step 1: Acquire noisy MRI images. Square these noisy MRI images to obtain noisy MRI image squared data. In this embodiment, the noisy MRI image squared data is divided into a training set and a test set. In this embodiment, 2000 high signal-to-noise ratio (SNR) 2D MRI brain images are acquired from the publicly available MRI brain dataset IXI, and Ricean noise with a noise level of 9% is added to obtain 2000 low SNR 2D MRI brain images as noisy MRI images. Square these 2000 low SNR 2D MRI brain images to obtain 2000 noisy MRI image squared data (squared data of low SNR 2D MRI brain images).

[0047]

[0048] Where x represents a noisy magnetic resonance imaging image, and ⊙ represents a dot product. The noisy MRI image squared data is represented by the obtained data. 1200 noisy MRI image squared data (squared data of low signal-to-noise ratio 2D MRI brain images) were used as the training set, and 800 noisy MRI image squared data (squared data of low signal-to-noise ratio 2D MRI brain images) were used as the test set. The size of the noisy MRI image squared data (squared data of low signal-to-noise ratio 2D MRI brain images) was 256×256.

[0049] Step 2: Construct a self-supervised denoising network. The self-supervised denoising network includes an encoder and a decoder, such as... Figure 2 As shown.

[0050] The encoder includes one input layer and multiple downsampling layers in sequence. In this embodiment, the encoder includes one input layer and four downsampling layers in sequence.

[0051] The input layer includes convolutional layers, group normalization layers, and activation layers. The convolutional kernel size in the input layer is 3×3, the number of input channels is 1, the number of output channels is 64, the input feature map size is 256×256, and the output feature map size is 256×256.

[0052] All four downsampling layers include convolutional layers, group normalization layers, and activation layers; the convolutional layer in the k-th downsampling layer has a kernel size of 3×3 and 64×2 input channels. k-1 The number of output channels is 64×2 k The input feature map size is The output feature map size is In the encoder, the activation function for each activation layer is the LeakyReLU function.

[0053] The decoder consists of multiple upsampling layers and one output layer, with the number of upsampling layers matching the number of downsampling layers. In this embodiment, the number of upsampling layers and downsampling layers, S, is 4 each, and the decoder consists of 4 upsampling layers and 1 output layer.

[0054] The first upsampling layer includes an upsampling module, a convolutional layer, a group normalization layer, and an activation layer. The feature map output from the last downsampling layer serves as the input to the first upsampling layer. The upsampling module uses bilinear interpolation. The output feature map size of the upsampling module is four times the input feature map size. The input feature map size is 16×16, and the output feature map size is 32×32. The convolutional layer in the first upsampling layer has a 3×3 kernel size, 1024 input channels, and 512 output channels. The input feature map size of the first upsampling layer is 16×16, and the output feature map size is 32×32.

[0055] The i-th upsampling layer includes an adaptive fusion module, an upsampling module, a convolutional layer, a group normalization layer, and an activation layer (i is the upsampling layer number other than 1). The feature maps output by the (5-i)-th downsampling layers and the feature maps output by the (i-1)-th upsampling layers serve as inputs to the adaptive fusion module in the i-th upsampling layer. The adaptive fusion module in the i-th upsampling layer includes, in sequence, the first convolutional layer, the first activation layer, the second convolutional layer, and the second activation layer, where the first convolutional layer has a 3×3 kernel size and 64×2 input channels. 6-i Number of output channels: 64×2 5-i The input feature map size is The output feature map size is The first activation layer uses the LeakyReLU function; the second convolutional layer has a 3×3 kernel size and 64×2 input channels. 5-i The number of output channels is 64×2 5-i The input feature map size is The output feature map size is The activation function for the second activation layer is the Sigmoid function.

[0056] In the i-th upsampling layer, the adaptive fusion module stacks the feature maps output from the (5-i)-th downsampling layer and the (i-1)-th upsampling layer by channel, and uses them as the input to the first convolutional layer of the adaptive fusion module. The second activation layer of the adaptive fusion module outputs a weight map m. The feature map output from the (i-1)-th upsampling layer is then multiplied by the weight map m output from the second activation layer of the adaptive fusion module to obtain the feature map output from the (i-1)-th upsampling layer. The corresponding dot product feature map is then multiplied with (1 - the weight map m output by the second activation layer of the adaptive fusion module in the i-th upsampling layer) to obtain the dot product feature map corresponding to the feature map output by the i-th downsampling layer. Then, the dot product feature map corresponding to the feature map output by the i-1 upsampling layer is added pixel by pixel to the dot product feature map corresponding to the feature map output by the i-th downsampling layer to obtain the fused feature map output by the adaptive fusion module in the i-th upsampling layer.

[0057] The number of input channels for the adaptive fusion module in the i-th upsampling layer is 64×2. 6-i The number of output channels is 64×2 5-iThe fused feature map output from the adaptive fusion module in the i-th upsampling layer is input to the upsampling module in the i-th upsampling layer. The feature map output from the upsampling module in the i-th upsampling layer is then input to the convolutional layer of the i-th upsampling layer. The feature map output from the convolutional layer of the i-th upsampling layer is then processed by the group normalization layer and activation layer in the i-th upsampling layer to obtain the feature map output by the i-th upsampling layer. The upsampling method of the upsampling module in the i-th upsampling layer is bilinear interpolation. The size of the output feature map of the upsampling module is 4 times the size of the input feature map. The size of the input feature map is... The output feature map size is The kernel size of the convolutional layer in the i-th upsampling layer is 3×3, and the number of input channels is 64×2. 5-i The number of output channels is 64×2 4-i The size of the input feature map of the i-th upsampling layer is The output feature map size is Where i = 2, 3, 4.

[0058] The output layer includes an adaptive fusion module and convolutional layers. The adaptive fusion module in the output layer consists of a first convolutional layer, a first activation layer, a second convolutional layer, and a second activation layer, in that order. The input to the adaptive fusion module includes the feature map output from the last upsampling layer and the feature map output from the input layer. The adaptive fusion module uses the feature map output from the input layer and the feature map output from the last upsampling layer, stacked by channel, as the input to the first convolutional layer of the adaptive fusion module. The second activation layer of the adaptive fusion module outputs a weight map m. The feature map output from the input layer is then multiplied by the weight map m output from the second activation layer of the adaptive fusion module to obtain the multiplied feature map corresponding to the output feature map. The feature map output from the last upsampling layer is then multiplied by (1 - the weight map m output from the second activation layer of the adaptive fusion module) to obtain the multiplied feature map corresponding to the last upsampling layer. This results in the fused feature map output by the adaptive fusion module of the output layer.

[0059] In this embodiment, the feature maps output from the input layer and the fourth upsampling layer serve as inputs to the adaptive fusion module in the output layer. The adaptive fusion module in the output layer includes a first convolutional layer, a first activation layer, a second convolutional layer, and a second activation layer. The first convolutional layer has a 3×3 kernel size, 128 input channels, 64 output channels, an input feature map size of 256, and an output feature map size of 256. The activation function of the first activation layer is LeakyReLU. The second convolutional layer has a 3×3 kernel size, 64 input channels, 64 output channels, an input feature map size of 256, and an output feature map size of 256. The activation function of the second activation layer is the Sigmoid function. In the output layer, the adaptive fusion module uses the feature map output from the input layer and the feature map output from the fourth upsampling layer, stacked by channel, as the input to the first convolutional layer of the adaptive fusion module. The second activation layer of the adaptive fusion module outputs a weight map m. The feature map output from the input layer is then multiplied by the weight map m output from the second activation layer of the adaptive fusion module to obtain the multiplied feature map corresponding to the input layer. The feature map output from the fourth upsampling layer is then multiplied by (1 - the weight map m output from the second activation layer of the adaptive fusion module) to obtain the multiplied feature map corresponding to the fourth upsampling layer. The multiplied feature maps corresponding to the input layer and the fourth upsampling layer are added pixel by pixel to obtain the fused feature map output by the adaptive fusion module of the output layer.

[0060] The fused feature map output by the adaptive fusion module in the output layer serves as the input to the convolutional layer in the output layer. The convolutional layer in the output layer outputs the predicted squared data of the denoised magnetic resonance image. The adaptive fusion module in the output layer has 128 input channels and 64 output channels. The convolutional layer has a 3×3 kernel size, 64 input channels, and 1 output channel. The input feature map size and the output feature map size are both 256×256.

[0061] The adaptive fusion module generates a corresponding weight map based on the feature map, and then performs feature fusion based on the weight map. This can promote the fusion of effective features and reduce the impact of unfavorable features.

[0062] Each upsampling layer is followed by a group normalization layer, and then an activation layer is added after the group normalization layer, with the activation function being the LeakReLU function.

[0063] Step 3: Input the squared data of noisy magnetic resonance images from the training set obtained in Step 1 into the self-supervised denoising network.

[0064] Step 4: Set the self-supervised denoising loss function L. In this embodiment, the self-supervised denoising loss function L is set as follows:

[0065]

[0066] Where ||·||2 represents the F-norm operation, is the squared data of noisy magnetic resonance imaging, n is a random mask following a Bernoulli distribution, and f θ This represents a self-supervised denoising network, where θ is the parameter of the self-supervised denoising network, and f... θ (·) represents the output of the self-supervised denoising network.

[0067] Step 5: Based on the total loss function set in Step 4, use the training set generated in Step 1 to train the self-supervised denoising network constructed in Step 2 end-to-end. The learning rate of the network is initialized to 0.0001, the batch size is set to 8, and the network is trained on the PyTorch platform using the ADAMW optimizer. Training is stopped after the total number of iterations reaches 100, and the parameters of the self-supervised denoising network are saved.

[0068] Step 6: Perform the squaring process in Step 1 on the noisy magnetic resonance image to be processed to obtain the squared data of the magnetic resonance image to be processed, and then input it into the self-supervised denoising network trained in Step 5 to obtain the squared data of the denoised magnetic resonance image.

[0069] In this embodiment, the squared data of the magnetic resonance images to be processed (low signal-to-noise ratio 2D magnetic resonance brain image squared data) corresponding to the noisy magnetic resonance images to be processed in the test set generated in step 1 is input into the self-supervised denoising network trained in step 5 to obtain the corresponding predicted denoised magnetic resonance image squared data (denoised 2D magnetic resonance brain image squared data).

[0070] Step 7: Segment the predicted denoised MRI image squared data (denoised 2D MRI brain image squared data) obtained in Step 6 to obtain the background mask, and calculate the noise variance based on the background mask.

[0071] This example uses a threshold to segment the predicted squared data of the denoised MRI image obtained in step 6, obtaining a background mask. Then, the background portion of the noisy MRI image is extracted using the background mask, and the noise variance is calculated based on maximum likelihood estimation. The noise distribution of the background portion of the noisy MRI image satisfies...

[0072]

[0073] Where M represents the pixel value of the background portion of the noisy magnetic resonance image to be processed, and P... M (M) is the probability density function of the background portion M of the noisy magnetic resonance image to be processed, σ 2Let I be the noise variance, I0 be the first-order modified Bessel function, and I0(0) be the value of the first-order modified Bessel function corresponding to the background with a signal value of 0.

[0074] Step 8: Subtract twice the noise variance obtained in Step 7 from the squared data of the predicted denoised magnetic resonance image obtained in Step 6, and then take its square root to obtain the denoised magnetic resonance image.

[0075] In this example, the predicted denoised MRI image square data (denoised 2D MRI brain image square data) obtained in step 6 is subtracted by twice the noise variance obtained in step 7, and then its square root is calculated to obtain the denoised MRI image (final denoised 2D MRI brain image).

[0076]

[0077] Where y represents the squared data of the magnetic resonance image to be processed. For denoised magnetic resonance imaging.

[0078] The denoised MRI images obtained in step 8 (final denoised 2D MRI brain images) are compared with the noisy MRI images (low signal-to-noise ratio 2D MRI brain images) in the test set obtained in step 1. The peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) indices are calculated for both the denoised MRI images (denoised 2D MRI brain images) and the corresponding noisy MRI images to be processed (noisy 2D MRI brain images). Figure 4 As shown, (a) is the noisy MRI image to be processed (noisy 2D MRI brain image), and (b) is the denoised MRI image obtained by the method of the present invention (denoised 2D MRI brain image). The results show that the present invention can effectively remove noise, increasing the peak signal-to-noise ratio of the noisy MRI image to be processed (noisy 2D MRI brain image) from 19.60 to 29.04, thus obtaining a high-quality denoised 2D MRI brain image.

[0079] The proposed method is compared with other methods, and the results are as follows: Figure 5As shown in Table 1, GT represents the high signal-to-noise ratio (SNR) 2D MRI brain image obtained in step 1; Noisy represents the noisy MRI image to be processed, with a PSNR of 16.60 and an SSIM of 0.378; MF represents median filtering, with a PSNR of 22.08 and an SSIM of 0.424; NLM represents nonlocal mean filtering, with a PSNR of 21.95 and an SSIM of 0.441; UNLM represents unbiased nonlocal mean filtering, with a PSNR of 26.33 and an SSIM of 0.581; U-Net is a deep learning method trained in a self-supervised manner, with a PSNR of 22.17 and an SSIM of 0.568; the method of this invention has a PSNR of 29.04 and an SSIM of 0.886. Compared with other methods, this invention better preserves image details while achieving denoising.

[0080] Table 1. Quantitative comparison of the performance of the present invention with other methods.

[0081] Noisy MF NLM UNLM U-Net Ours PSNR 19.60 22.08 21.95 26.33 22.17 29.04 SSIM 0.378 0.424 0.441 0.581 0.568 0.886

[0082] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0083] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0084] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0085] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0086] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A self-supervised deep learning method for bias correction in magnetic resonance imaging denoising, characterized in that, Includes the following steps: Step 1: Obtain noisy magnetic resonance images and corresponding squared noisy magnetic resonance images, and construct training and testing sets based on the squared noisy magnetic resonance images. Step 2: Construct a self-supervised denoising network, which includes an encoder and a decoder; Step 3: Input the squared data of noisy magnetic resonance images from the training set obtained in Step 1 into the self-supervised denoising network; Step 4: Define the self-supervised denoising loss function L; Step 5: Based on the self-supervised denoising loss function set in Step 4, use the training set generated in Step 1 to train the self-supervised denoising network constructed in Step 2 end-to-end, and save the parameters of the self-supervised denoising network. Step 6: Perform the squaring process in Step 1 on the noisy magnetic resonance image to be processed to obtain the squared data of the magnetic resonance image to be processed, and then input it into the self-supervised denoising network trained in Step 5 to obtain the denoised magnetic resonance image squared data. Step 7: Segment the predicted denoised magnetic resonance image squared data obtained in Step 6 to obtain the background mask, and calculate the noise variance based on the background mask; Step 8: Subtract twice the noise variance obtained in Step 7 from the squared data of the predicted denoised magnetic resonance image obtained in Step 6, and then take the square root to obtain the denoised magnetic resonance image.

2. The self-supervised deep learning magnetic resonance image denoising method for bias correction according to claim 1, characterized in that, The squared data of the noisy magnetic resonance imaging in step 1 is calculated based on the following formula: Where x represents a noisy magnetic resonance imaging image, and ⊙ represents a dot product. This represents the squared data of the noisy magnetic resonance imaging obtained.

3. The self-supervised deep learning magnetic resonance image denoising method for bias correction according to claim 1, characterized in that, In step 2, the encoder includes one input layer and multiple downsampling layers in sequence, and the decoder includes multiple upsampling layers and one output layer in sequence. The number of upsampling layers is the same as the number of downsampling layers. The input layer includes a convolutional layer, a group normalization layer and an activation layer, and the output layer includes an adaptive fusion module and a convolutional layer.

4. The self-supervised deep learning magnetic resonance image denoising method for bias correction according to claim 3, characterized in that, The downsampling layer includes a convolutional layer, a group normalization layer, and an activation layer. The feature map output from the last downsampling layer is used as the input to the first upsampling layer. The first upsampling layer includes an upsampling module, a convolutional layer, a group normalization layer, and an activation layer. The i-th upsampling layer includes an adaptive fusion module, an upsampling module, a convolutional layer, a group normalization layer, and an activation layer, where i is the upsampling layer index other than 1. The feature map output by the (S+1-i)th downsampling layer and the feature map output by the (i-1)th upsampling layer are used as the input to the adaptive fusion module in the ith upsampling layer, where S is the number of upsampling and downsampling layers. The adaptive fusion module in the i-th upsampling layer includes the first convolutional layer, the first activation layer, the second convolutional layer, and the second activation layer.

5. The self-supervised deep learning magnetic resonance image denoising method for bias correction according to claim 4, characterized in that, In the i-th upsampling layer, the adaptive fusion module stacks the feature maps output from the (S+1-i)-th downsampling layer and the (i-1)-th upsampling layer by channel, and uses them as the input to the first convolutional layer of the adaptive fusion module. The second activation layer of the adaptive fusion module outputs a weight map m. The feature map output from the (i-1)-th upsampling layer is then multiplied by the weight map m calculated in the adaptive fusion module to obtain the feature map output from the (i-1)-th upsampling layer. The corresponding dot product feature map, the feature map output by the (S+1-i)th downsampling layer is then multiplied by (1 - the weight map m calculated by the adaptive fusion module in the ith upsampling layer) to obtain the dot product feature map corresponding to the feature map output by the (S+1-i)th downsampling layer. The dot product feature map corresponding to the feature map output by the (i-1)th upsampling layer is then added pixel by pixel to the dot product feature map corresponding to the feature map output by the (S+1-i)th downsampling layer to obtain the fused feature map output by the adaptive fusion module in the ith upsampling layer.

6. The self-supervised deep learning magnetic resonance image denoising method for bias correction according to claim 5, characterized in that, The fused feature map output by the adaptive fusion module in the i-th upsampling layer is input to the upsampling module in the i-th upsampling layer. The feature map output by the upsampling module in the i-th upsampling layer is then input to the convolutional layer of the i-th upsampling layer. The feature map output by the convolutional layer of the i-th upsampling layer is then passed through the group normalization layer and activation layer in the i-th upsampling layer to obtain the feature map output by the i-th upsampling layer.

7. The self-supervised deep learning magnetic resonance image denoising method for bias correction according to claim 6, characterized in that, In the output layer, the adaptive fusion module uses the feature maps output from the input layer and the feature maps output from the last upsampling layer, stacked by channels, as input to the first convolutional layer of the adaptive fusion module. The second activation layer of the adaptive fusion module outputs a weight map m. The feature map output from the input layer is then multiplied by the weight map m calculated by the adaptive fusion module in the output layer to obtain the multiplied feature map corresponding to the feature map output from the output layer. The feature map output from the last upsampling layer is then multiplied by (1 - the weight map m calculated by the adaptive fusion module in the output layer) to obtain the multiplied feature map corresponding to the feature map output from the last upsampling layer, thus obtaining the fused feature map output by the adaptive fusion module of the output layer.

8. The self-supervised deep learning magnetic resonance image denoising method for bias correction according to claim 7, characterized in that, The fused feature map output by the adaptive fusion module in the output layer is used as the input of the convolutional layer in the output layer, and the convolutional layer in the output layer outputs the predicted denoised magnetic resonance image squared data.

9. The self-supervised deep learning magnetic resonance image denoising method for bias correction according to claim 1, characterized in that, The self-supervised denoising loss function L is based on the following formula: Where ||·||2 represents the F-norm operation, is the squared data of noisy magnetic resonance imaging, n is a random mask following a Bernoulli distribution, and f θ Represents a self-supervised denoising network, f θ (·) represents the output of the self-supervised denoising network.

Images