Self-supervised deep learning magnetic resonance image denoising method for deviation correction

Abstract

The invention discloses a self-supervised deep learning magnetic resonance image denoising method for deviation correction, and the method comprises the steps: obtaining a noisy magnetic resonance image and corresponding noisy magnetic resonance image square data, and constructing a training set and a test set; constructing a self-supervised denoising network; training the self-supervised denoising network based on the training set; and performing square processing on a noisy magnetic resonance image to be processed, inputting the processed noisy magnetic resonance image to the trained self-supervised denoising network, and further obtaining a background mask and a denoised magnetic resonance image. According to the method, only noisy magnetic resonance images are used as a training set, manpower resources consumed for collecting noise-free magnetic resonance images are reduced, deviation caused by non-zero mean noise can be corrected, a more accurate denoising result is provided, and the method has practical application value. During application, extra manual optimization is not needed, and the denoising effect is excellent.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically relates to a bias-corrected self-supervised deep learning magnetic resonance image denoising method, which is suitable for magnetic resonance image denoising. Background Art

[0002] Magnetic resonance imaging (MRI) is a commonly used imaging technology in the medical field, which can provide important information about the structure and function of biological tissues and organs. However, noise is difficult to avoid during MRI scanning, which will significantly affect the accuracy of the data and the image quality. Therefore, it is necessary to denoise the acquired MRI data to improve the image quality. There are two denoising methods: one is to improve the performance of the imaging equipment, such as increasing the magnetic field strength or extending the scanning time, but this method will increase the cost and examination time, and cause discomfort to the subjects; the other method is to use image post-processing algorithms to denoise the acquired low-quality MRI images. This method does not change the original imaging process, is easy to integrate into existing equipment, and has practical application value.

[0003] In recent years, deep learning-based MRI denoising methods have shown superior performance, with significant advantages in denoising effect, denoising speed and model application. MRI denoising algorithms based on deep learning can autonomously establish a nonlinear mapping model from noisy MRI images to noise-free MRI images through external data set training. However, most of the existing deep learning denoising methods for MRI images rely on paired noisy and noise-free images for effective training, but in practical applications, it is often difficult to collect these paired images at the same time. In order to achieve the training of denoising models in the absence of paired images, Krull et al. proposed a blind spot network for denoising a single noisy image. This method trains the network in a self-supervised manner and only requires noisy images. However, this self-supervised denoising method requires that the image noise mean is zero, while the noise of MRI images follows the Rice distribution and does not meet the requirement of zero noise mean. Therefore, if self-supervised denoising is directly applied to MRI images, it will lead to denoising deviation.

[0004] In summary, the prior art has the following problems:

[0005] 1. Some solutions require both noisy MRI images and noise-free MRI images, which is difficult to achieve in actual scenarios.

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

[0007] The purpose of the present invention is to address the above-mentioned problems existing in the prior art and propose a bias-corrected self-supervised deep learning magnetic resonance image denoising method, which trains the denoising network without noise-free magnetic resonance images and corrects the bias caused by non-zero mean noise. After the model training is completed, the magnetic resonance images can be quickly and accurately denoised.

[0008] The above-mentioned object of the present invention is achieved by the following technical solutions:

[0009] A bias-corrected self-supervised deep learning magnetic resonance image denoising method comprises the following steps:

[0010] Step 1, obtaining noisy magnetic resonance images and corresponding noisy magnetic resonance image square data, and constructing a training set and a test set based on the noisy magnetic resonance image square data;

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

[0012] Step 3, inputting the noisy magnetic resonance image square data of the training set obtained in step 1 into the self-supervised denoising network;

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

[0014] Step 5: According to the self-supervised denoising loss function set in step 4, the self-supervised denoising network constructed in step 2 is trained end-to-end using the training set generated in step 1, and the parameters of the self-supervised denoising network are saved;

[0015] Step 6, performing the square processing in step 1 on the noisy magnetic resonance image to be processed to obtain the square data of the magnetic resonance image to be processed, and then inputting it into the self-supervised denoising network trained in step 5 to obtain the denoised square data of the magnetic resonance image;

[0016] Step 7, segmenting the squared denoised magnetic resonance image data predicted in step 6 to obtain a background mask, and calculating the noise variance based on the background mask;

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

[0018] The square data of the noisy MRI image in step 1 above is calculated based on the following formula:

[0019]

[0020] Where x represents the noisy MRI image, ⊙ represents the dot product, represents the obtained noisy magnetic resonance image square data.

[0021] As described above, in step 2, the encoder includes 1 input layer and multiple downsampling layers in order, the decoder includes multiple upsampling layers and 1 output layer in order, the number of upsampling layers is consistent with the number of downsampling layers, the input layer includes a convolution layer, a group normalization layer and an activation layer, and the output layer includes an adaptive fusion module and a convolution layer.

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

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

[0024] The first upsampling layer includes upsampling module, convolution layer, group normalization layer and activation layer.

[0025] The i-th upsampling layer includes an adaptive fusion module, an upsampling module, a convolution layer, a group normalization layer, and an activation layer, where i is the sequence number of the upsampling layer except 1.

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

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

[0028] As mentioned above, the adaptive fusion module in the i-th upsampling layer stacks the feature map output by the S+1-i-th downsampling layer and the feature map output by the i-1-th upsampling layer according to the channel as the input of the first convolution layer of the adaptive fusion module in the i-th upsampling layer, and the second activation layer of the adaptive fusion module in the i-th upsampling layer outputs the weight map m; the feature map output by 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 by the i-1-th upsampling layer 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 i-th upsampling layer) to obtain the dot multiplication feature map corresponding to the feature map output by the S+1-i-th downsampling layer. The dot multiplication feature map corresponding to the feature map output by the i-1-th upsampling layer and the dot multiplication feature map corresponding to the feature map output by the S+1-i-th downsampling layer are added pixel by pixel to obtain the fused feature map output by the adaptive fusion module in the i-th upsampling layer.

[0029] As mentioned 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, and the feature map output by the upsampling module in the i-th upsampling layer is then input to the convolution layer of the i-th upsampling layer. The feature map output by the convolution layer of the i-th upsampling layer passes through the group normalization layer and the 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 stacks the feature map output by the input layer and the feature map output by the last upsampling layer as the input of the first convolutional layer of the adaptive fusion module in the output layer according to the channel stacking, and the second activation layer of the adaptive fusion module in the output layer outputs the weight map m; the feature map output by the input layer is then point-multiplied with the weight map m calculated by the adaptive fusion module in the output layer to obtain the point-multiplied feature map corresponding to the feature map output by the output layer, and the feature map output by the last upsampling layer is then point-multiplied with (1-the weight map m calculated by the adaptive fusion module in the output layer) to obtain the point-multiplied feature map corresponding to the feature map output by the last upsampling layer, and the fused feature map output by the adaptive fusion module of the output layer is obtained.

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

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

[0033]

[0034] Among them, ∥·∥2 is the F norm operation, is the square data of noisy magnetic resonance images, n is a random mask following Bernoulli distribution, f θ represents the 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 beneficial effects:

[0036] 1. The present invention builds a self-supervised denoising network, uses the set self-supervised denoising loss function to train the self-supervised denoising network in the absence of noise-free magnetic resonance images, and uses the estimated noise variance to remove the denoising deviation from the obtained denoised magnetic resonance image square data, effectively ensuring the denoising performance of the magnetic resonance image. Compared with the existing magnetic resonance image denoising method based on deep learning, the present invention only needs noisy magnetic resonance images as training sets, reduces the human resources consumed in collecting noise-free magnetic resonance images, and can correct the deviation caused by non-zero mean noise, providing more accurate denoising results, and has practical application value.

[0037] 2. Compared with the traditional magnetic resonance image denoising method, after the self-supervised denoising network training is completed, the present invention does not require additional parameter adjustment, has a fast denoising speed, can better preserve image details, and has excellent results. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flow chart of the present invention;

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

[0040] Figure 3 It is a structural diagram of the adaptive fusion module;

[0041] Figure 4 Schematic diagram of denoising effect; (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 It is a schematic diagram of the performance comparison between the present invention and other methods on the constructed data set;

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

[0044] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below in conjunction with embodiments. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0045] A bias-corrected self-supervised deep learning magnetic resonance image denoising method specifically comprises the following steps:

[0046] Step 1, obtain noisy magnetic resonance images, perform square operations on these noisy magnetic resonance images, and obtain noisy magnetic resonance image square data. In this embodiment, the noisy magnetic resonance image square data is divided into a training set and a test set. In this embodiment, 2000 high signal-to-noise ratio 2D magnetic resonance brain images are obtained from the magnetic resonance brain public data set IXI, and Rice noise with a noise level of 9% is added to obtain 2000 low signal-to-noise ratio 2D magnetic resonance brain images as noisy magnetic resonance images. The 2000 low signal-to-noise ratio 2D magnetic resonance brain images are squared to obtain 2000 noisy magnetic resonance image square data (square data of low signal-to-noise ratio 2D magnetic resonance brain images).

[0047]

[0048] Where x represents the noisy MRI image, ⊙ represents the dot product, Represents the obtained noisy magnetic resonance image square data. 1200 noisy magnetic resonance image square data (square data of low signal-to-noise ratio 2D magnetic resonance brain images) are used as training sets, and 800 noisy magnetic resonance image square data (square data of low signal-to-noise ratio 2D magnetic resonance brain images) are used as test sets. The size of the noisy magnetic resonance image square data (square data of low signal-to-noise ratio 2D magnetic resonance brain images) is 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 shown.

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

[0051] The input layer includes a convolution layer, a group normalization layer, and an activation layer. The convolution kernel size in the convolution layer of 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] The four downsampling layers all include convolutional layers, group normalization layers, and activation layers; the convolutional kernel size in the kth downsampling layer is 3×3, and the number of input channels is 64×2. 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 of each activation layer is the LeakyReLU function.

[0053] The decoder includes multiple upsampling layers and one output layer in order, wherein the number of upsampling layers is consistent with the number of downsampling layers. In this embodiment, the number of layers S of the upsampling layers and the downsampling layers is 4, and the decoder includes 4 upsampling layers and one output layer in order.

[0054] The first upsampling layer includes an upsampling module, a convolution layer, a group normalization layer, and an activation layer. The feature map output by the last downsampling layer is used as the input of the first upsampling layer. The upsampling method of the upsampling module 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 input feature map size is 16×16, and the output feature map size is 32×32. The convolution kernel size in the convolution layer of the first upsampling layer is 3×3, the number of input channels is 1024, and the number of output channels is 512. 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 convolution layer, a group normalization layer, and an activation layer (i is the sequence number of the upsampling layer except 1). The feature maps output by the 5-i downsampling layers and the feature maps output by the i-1 upsampling layers are used as inputs to the adaptive fusion module in the i-th upsampling layer. The adaptive fusion module in the i-th upsampling layer includes the first convolution layer, the first activation layer, the second convolution layer, and the second activation layer in order. The convolution kernel size of the first convolution layer is 3×3, and the number of input channels is 64×2. 6-i , output channel number 64×2 5-i , the input feature map size is The output feature map size is The activation function of the first activation layer is the LeakyReLU function; the convolution kernel size of the second convolution layer is 3×3, and the number of input channels is 64×2 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 of the second activation layer is the Sigmoid function.

[0056] The adaptive fusion module in the i-th upsampling layer stacks the feature map output by the 5-i downsampling layer and the feature map output by the i-1 upsampling layer according to the channel as the input of the first convolution layer of the adaptive fusion module in the i-th upsampling layer, and the second activation layer of the adaptive fusion module in the i-th upsampling layer outputs the weight map m; the feature map output by the i-1 upsampling layer is then multiplied by the weight map m output by the second activation layer of the adaptive fusion module in the i-th upsampling layer to obtain the feature map output by the i-1 upsampling layer. The corresponding dot product operation feature map, the feature map output by the 5-i th downsampling layer is then dot-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 operation feature map corresponding to the feature map output by the 5-i th downsampling layer. Then, the dot product operation feature map corresponding to the feature map output by the i-1 th upsampling layer and the dot product operation feature map corresponding to the feature map output by the 5-i downsampling layer are added pixel by pixel to obtain the fused feature map output by the adaptive fusion module in the i-th upsampling layer.

[0057] The number of input channels of 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 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 convolution layer of the i-th upsampling layer. The feature map output by the convolution layer of the i-th upsampling layer passes through the group normalization layer and the 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 feature map output by 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 convolution kernel size of the convolution 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 input feature map size 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 a convolution layer. The adaptive fusion module in the output layer includes the first convolution layer, the first activation layer, the second convolution layer and the second activation layer in order. The input of the adaptive fusion module of the output layer includes the feature map output by the last upsampling layer and the feature map output by the input layer. The adaptive fusion module in the output layer stacks the feature map output by the input layer and the feature map output by the last upsampling layer as the input of the first convolution layer of the adaptive fusion module in the output layer according to channel stacking. The second activation layer of the adaptive fusion module in the output layer outputs a weight map m. The feature map output by the input layer is then point-multiplied with the weight map m output by the second activation layer of the adaptive fusion module in the output layer to obtain a point-multiplied feature map corresponding to the feature map output by the output layer. The feature map output by the last upsampling layer is then point-multiplied with (1-the weight map m output by the second activation layer of the adaptive fusion module in the output layer) to obtain a point-multiplied feature map corresponding to the feature map output by the last upsampling layer, and a fused feature map output by the adaptive fusion module of the output layer is obtained.

[0059] In this embodiment, the feature map output by the input layer and the feature map output by the fourth upsampling layer are used as inputs of the adaptive fusion module in the output layer. The adaptive fusion module in the output layer includes the first convolution layer, the first activation layer, the second convolution layer and the second activation layer, wherein the convolution kernel size of the first convolution layer is 3×3, the number of input channels is 128, the number of output channels is 64, the input feature map size is 256, and the output feature map size is 256; the activation function of the first activation layer is LeakyReLU; the convolution kernel size of the second convolution layer is 3×3, the number of input channels is 64, the number of output channels is 64, the input feature map size is 256, and the output feature map size is 256; the activation function of the second activation layer is Sigmoid function. The adaptive fusion module in the output layer uses the feature map output by the input layer and the feature map output by the 4th upsampling layer as the input of the first convolutional layer of the adaptive fusion module in the output layer according to channel stacking, and the second activation layer of the adaptive fusion module in the output layer outputs the weight map m; the feature map output by the input layer is then point-multiplied with the weight map m output by the second activation layer of the adaptive fusion module in the output layer to obtain the point-multiplication feature map corresponding to the feature map output by the input layer, the feature map output by the 4th upsampling layer is then point-multiplied with (1-the weight map m output by the second activation layer of the adaptive fusion module in the output layer) to obtain the point-multiplication feature map corresponding to the feature map output by the 4th upsampling layer, the point-multiplication feature map corresponding to the feature map output by the input layer and the point-multiplication feature map corresponding to the feature map output by the 4th 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 is used as the input of the convolution layer in the output layer, and the convolution layer in the output layer outputs the predicted denoised magnetic resonance image square data. The number of input channels of the adaptive fusion module in the output layer is 128, the number of output channels is 64, the convolution kernel size in the convolution layer is 3×3, the number of input channels is 64, the number of output channels is 1, the feature map size of the output layer input is 256×256, and the output feature map size is 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, which can promote the fusion of effective features and reduce the impact of unfavorable features.

[0062] There is a group normalization layer after the convolution layer of each upsampling layer, and an activation layer is added after the group normalization layer, and the activation function is the LeakReLU function.

[0063] Step 3: Input the noisy magnetic resonance image square data of 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:

[0065]

[0066] Among them, ∥·∥2 is the F norm operation, is the square data of noisy magnetic resonance images, n is a random mask following Bernoulli distribution, f θ represents the self-supervised denoising network, 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. According to the total loss function set in step 4, the self-supervised denoising network constructed in step 2 is trained end-to-end using the training set generated in step 1. 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. The 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 square processing in step 1 on the noisy magnetic resonance image to be processed to obtain the square 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 square data of the magnetic resonance image.

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

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

[0071] This example uses the threshold to segment the predicted denoised magnetic resonance image square data obtained in step 6 to obtain a background mask, and then uses the background mask to extract the background part of the noisy magnetic resonance image, and calculates the noise variance based on the maximum likelihood estimation. The noise distribution of the background part of the noisy magnetic resonance image satisfies

[0072]

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

[0074] Step 8, subtracting twice the noise variance obtained in step 7 from the square data of the predicted denoised magnetic resonance image obtained in step 6, and then calculating the square root thereof to obtain the denoised magnetic resonance image;

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

[0076]

[0077] Where y is the square data of the magnetic resonance image to be processed, To denoise magnetic resonance images.

[0078] The denoised MRI image (final denoised 2D MRI brain image) obtained in step 8 is compared with the noisy MRI image (low signal-to-noise ratio 2D MRI brain image) in the test set obtained in step 1, and the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) index of the denoised MRI image (denoised 2D MRI brain image) and the corresponding noisy MRI image to be processed (noisy 2D MRI brain image) are calculated respectively, as follows: Figure 4 As shown, (a) is a noisy magnetic resonance image to be processed (noisy 2D magnetic resonance brain image), and (b) is a denoised magnetic resonance image (denoised 2D magnetic resonance brain image) obtained by the method of the present invention. It can be seen from the result figure that the present invention can effectively remove noise, and the peak signal-to-noise ratio of the noisy magnetic resonance image to be processed (noisy 2D magnetic resonance brain image) is increased from 19.60 to 29.04, thereby obtaining a high-quality denoised 2D magnetic resonance brain image.

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

[0080] Table 1 Quantitative comparison of the performance of the present invention and 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 provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0083] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

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

[0085] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through a computer program. 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-mentioned methods.

[0086] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A bias-corrected self-supervised deep learning magnetic resonance image denoising method, characterized in that: The following steps are involved: Step 1, obtaining noisy magnetic resonance images and corresponding noisy magnetic resonance image square data, and constructing a training set and a test set based on the noisy magnetic resonance image square data; Step 2: Build a self-supervised denoising network, which includes an encoder and a decoder; Step 3, inputting the noisy magnetic resonance image square data of the training set obtained in step 1 into the self-supervised denoising network; Step 4: Set the self-supervised denoising loss function L; Step 5: According to the self-supervised denoising loss function set in step 4, the self-supervised denoising network constructed in step 2 is trained end-to-end using the training set generated in step 1, and the parameters of the self-supervised denoising network are saved; Step 6, performing the square processing in step 1 on the noisy magnetic resonance image to be processed to obtain the square data of the magnetic resonance image to be processed, and then inputting it into the self-supervised denoising network trained in step 5 to obtain the denoised square data of the magnetic resonance image; Step 7, segmenting the squared denoised magnetic resonance image data predicted in step 6 to obtain a background mask, and calculating the noise variance based on the background mask; Step 8: Subtract twice the noise variance obtained in step 7 from the square data of the predicted denoised magnetic resonance image obtained in step 6, and then calculate the square root to obtain the denoised magnetic resonance image.

2. The method for denoising magnetic resonance images by self-supervised deep learning with bias correction according to claim 1, characterized in that: The noisy magnetic resonance image square data in step 1 is calculated based on the following formula: Where x represents the noisy MRI image, ⊙ represents the dot product, represents the obtained noisy magnetic resonance image square data.

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

4. The method for denoising magnetic resonance images by self-supervised deep learning with bias correction according to claim 3, characterized in that: The downsampling layer includes a convolution layer, a group normalization layer and an activation layer, The feature map output by the last downsampling layer is used as the input of the first upsampling layer. The first upsampling layer includes upsampling module, convolution layer, group normalization layer and activation layer. The i-th upsampling layer includes an adaptive fusion module, an upsampling module, a convolution layer, a group normalization layer, and an activation layer, where i is the sequence number of the upsampling layer except 1. The feature map output by the S+1-ith downsampling layer and the feature map output by the i-1th upsampling layer are used as the input of the adaptive fusion module in the i-th upsampling layer, where S is the number of upsampling layers and downsampling layers; The adaptive fusion module in the i-th upsampling layer includes the first convolution layer, the first activation layer, the second convolution layer and the second activation layer.

5. The method for denoising magnetic resonance images by self-supervised deep learning with bias correction according to claim 4, characterized in that: The adaptive fusion module in the i-th upsampling layer stacks the feature map output by the S+1-i-th downsampling layer and the feature map output by the i-1-th upsampling layer according to the channel as the input of the first convolution layer of the adaptive fusion module in the i-th upsampling layer, and the second activation layer of the adaptive fusion module in the i-th upsampling layer outputs the weight map m; the feature map output by 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 by the i-1-th upsampling layer The corresponding dot multiplication feature map, the feature map output by the S+1-i-th downsampling layer is then dot-multiplied with (1-the weight map m calculated by the adaptive fusion module in the i-th upsampling layer) to obtain the dot multiplication feature map corresponding to the feature map output by the S+1-i-th downsampling layer, the dot multiplication feature map corresponding to the feature map output by the i-1-th upsampling layer and the dot multiplication feature map corresponding to the feature map output by the S+1-i-th downsampling layer are added pixel by pixel to obtain the fused feature map output by the adaptive fusion module in the i-th upsampling layer.

6. The method for denoising magnetic resonance images by self-supervised deep learning with 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 convolution layer of the i-th upsampling layer, the feature map output by the convolution layer of the i-th upsampling layer passes through the group normalization layer and the activation layer in the i-th upsampling layer, and the feature map output by the i-th upsampling layer is obtained.

7. The method for denoising magnetic resonance images by self-supervised deep learning with bias correction according to claim 6, characterized in that: The adaptive fusion module in the output layer stacks the feature map output by the input layer and the feature map output by the last upsampling layer as the input of the first convolution layer of the adaptive fusion module in the output layer according to the channel, and the second activation layer of the adaptive fusion module in the output layer outputs the weight map m; the feature map output by the input layer is then point-multiplied with the weight map m calculated by the adaptive fusion module in the output layer to obtain the point-multiplied feature map corresponding to the feature map output by the output layer, the feature map output by the last upsampling layer is then point-multiplied with (1-the weight map m calculated by the adaptive fusion module in the output layer) to obtain the point-multiplied feature map corresponding to the feature map output by the last upsampling layer, and the fused feature map output by the adaptive fusion module of the output layer is obtained.

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

9. The method for denoising magnetic resonance images by self-supervised deep learning with bias correction according to claim 1, characterized in that: The self-supervised denoising loss function L is based on the following formula: Among them, ∥·∥2 is the F norm operation, is the square data of noisy magnetic resonance images, n is a random mask following Bernoulli distribution, f θ represents the self-supervised denoising network, f θ (·) represents the output of the self-supervised denoising network.

Images