Joint space angle self-supervision denoising method for diffusion magnetic resonance data
Through the combined spatial angle self-supervised denoising method, paired data are constructed using angle collectors and spatial subsamplers, and combined with denoising networks and regularization losses, the problem of unused spatial angle information in diffuse magnetic resonance imaging denoising is solved. Efficient noise cancellation is achieved.
Patent Information
- Application Number
- CN202410989653.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-23
- Publication Date
- 2024-12-13
AI Technical Summary
The prior art does not fully utilize spatial angle information during the denoising process of diffuse magnetic resonance imaging, resulting in low denoising quality.
Using the joint spatial angle self-supervised denoising method, paired data with the same noise distribution is constructed through the angle collector and the spatial subsampler, combining the denoising network and regularization loss, the denoising model is optimized to realize self-supervised learning.
It significantly improves the denoising effect of diffuse magnetic resonance data, is better than the current advanced methods, effectively eliminates noise and makes up for the internal differences of data.
Smart Images
Figure CN119130837A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical image denoising, and in particular to a joint spatial angle self-supervised denoising method for diffuse magnetic resonance data. Background Art
[0002] In order to improve the denoising quality of diffusion magnetic resonance imaging (dMRI) data, some researchers have made improvements in the acquisition stage, but these methods often rely on hardware upgrades or increased scanning time, and are difficult to popularize in clinical practice. Therefore, post-processing denoising technology has become a research focus, especially extending 2D image denoising technology such as non-local means (NLM) and its derivative algorithms to dMRI data. These methods achieve denoising by weighted averaging of similar structures in the image. There are also algorithms based on the 4D characteristics of dMRI data. For example, some researchers effectively denoise by processing 3D voxels of the spherical surface with different b values and q space (diffusion wave vector space) at the same spatial position. LPCA relies on principal component analysis (PCA) of local blocks of the image, adaptively distributes noise and determines the threshold, achieving good results. However, at high b values, these traditional methods do not always meet the needs. In recent years, data-driven machine learning and deep learning have made significant progress in many fields and have been applied to dMRI denoising. Some researchers use self-supervised learning to perform denoising through full-rank local linear regression, but this will reduce the signal value while denoising. Some researchers use 1D-CNN (Convolutional Neural Networks, CNN) to learn from low-noise data to implement denoising. This method generally requires pairing high-noise and low-noise data, but such data is difficult to obtain clinically. In addition, these methods do not fully utilize spatial angle information. Summary of the invention
[0003] The embodiment of the present invention provides a joint spatial angle self-supervised denoising method for diffusion magnetic resonance data, so as to at least solve the technical problem in the prior art that when denoising diffusion magnetic resonance imaging, spatial angle information is not fully utilized, resulting in low denoising quality.
[0004] According to one aspect of an embodiment of the present invention, a joint spatial angle self-supervised denoising method for diffusion magnetic resonance data is provided. The method may include: acquiring diffusion magnetic resonance data; processing the diffusion magnetic resonance data through an angle collector to obtain angle-similar diffusion magnetic resonance data similar to the diffusion magnetic resonance data; processing the diffusion magnetic resonance data and the angle-similar diffusion magnetic resonance data through a spatial subsampler to obtain first subsampled data, second subsampled data, and angle-similar subsampled data, wherein the original first and second spatial subsampled data correspond to the diffusion magnetic resonance data, and the angle-similar subsampled data correspond to the angle-similar diffusion magnetic resonance data; processing the first subsampled data, the diffusion magnetic resonance data, and the angle-similar diffusion magnetic resonance data through a denoising network to obtain denoised first subsampled data, denoised diffusion magnetic resonance data, and denoised angle-similar diffusion magnetic resonance data, wherein the denoised first subsampled data corresponds to the first subsampled data, the denoised diffusion magnetic resonance data corresponds to the diffusion magnetic resonance data, and the denoised angle-similar diffusion magnetic resonance data corresponds to the angle-similar diffusion magnetic resonance data; The device processes the denoised diffusion magnetic resonance data and the denoised angle-similar diffusion magnetic resonance data to obtain denoised diffusion magnetic resonance data first sub-sampling data, denoised diffusion magnetic resonance data second sub-sampling data and denoised angle-similar diffusion magnetic resonance data sub-sampling data, wherein the denoised diffusion magnetic resonance data first sub-sampling data and the denoised angle-similar diffusion magnetic resonance data second sub-sampling data correspond to the denoised diffusion magnetic resonance data, and the denoised angle-similar diffusion magnetic resonance data sub-sampling data correspond to the denoised angle-similar diffusion magnetic resonance data; based on the denoised first sub-sampling data, the second sub-sampling data, the denoised diffusion magnetic resonance data first sub-sampling data, the denoised diffusion magnetic resonance data second sub-sampling data, the angle-similar sub-sampling data and the denoised angle-similar diffusion magnetic resonance data sub-sampling data, the spatial denoising loss, the spatial regularization loss, the angle denoising loss and the angle regularization loss are calculated to optimize the denoising model; after obtaining the trained denoising model, the diffusion magnetic resonance data is subjected to the denoising model to obtain denoised diffusion magnetic resonance data.
[0005] Optionally, the diffusion magnetic resonance data and the angularly similar diffusion magnetic resonance data are processed by a spatial subsampler to obtain first sub-sampling data, second sub-sampling data and angularly similar sub-sampling data, including: based on spatial information, dividing the diffusion magnetic resonance data into fixed adjacent regions; randomly selecting a first voxel in each adjacent region until all adjacent regions are visited; based on the first voxel, obtaining the first sub-sampling data; randomly selecting a second voxel in the adjacent region of the first voxel until all adjacent regions are visited; based on the second voxel, obtaining the second sub-sampling data; for the angularly similar diffusion magnetic resonance data, only randomly selecting the first voxel in the fixed adjacent region to obtain the angularly similar sub-sampling data.
[0006] Optionally, the denoising network includes: an input layer, 5 Conv3d layers with a convolution kernel of 3, 5 Conv3d layers with a convolution kernel of 1, 10 activation functions and an output layer.
[0007] Optionally, the denoising network processing process is: the input data passes through a Conv3d with a convolution kernel of 1 and then passes through two layers of Conv3d with a convolution kernel of 3, and the result is added, and then undergoes such an operation twice, and finally passes through a Conv3d with a convolution kernel of 1 to obtain the output result, wherein the Rulu activation function needs to be passed after Conv3d.
[0008] Optionally, a loss function is calculated based on the denoised first sub-sampling data, the second sub-sampling data, the denoised diffusion magnetic resonance data first sub-sampling data, the denoised diffusion magnetic resonance data second sub-sampling data, the angular similarity sub-sampling data, and the denoised angular similarity diffusion magnetic resonance data sub-sampling data, including: calculating spatial denoising loss and spatial regularization loss based on the second sub-sampling data, the denoised first sub-sampling data, the denoised diffusion magnetic resonance data first sub-sampling data, and the denoised diffusion magnetic resonance data second sub-sampling data; calculating angular denoising loss and angular regularization loss based on the angular similarity sub-sampling data, the denoised angle similarity diffusion magnetic resonance data sub-sampling data, the denoised first sub-sampling data, and the denoised diffusion magnetic resonance data first sub-sampling data; and optimizing the denoising model by the spatial denoising loss, the spatial regularization loss, the angular denoising loss, and the angular regularization loss.
[0009] Optionally, based on the second sub-sampling data, the denoised first sub-sampling data, the denoised diffusion magnetic resonance data first sub-sampling data, and the denoised diffusion magnetic resonance data second sub-sampling data, spatial denoising loss and spatial regularization loss are calculated, including: calculating the spatial denoising loss based on the second sub-sampling data and the denoised first sub-sampling data; calculating the spatial regularization loss based on the second sub-sampling data, the denoised first sub-sampling data, the denoised diffusion magnetic resonance data first sub-sampling data, and the denoised diffusion magnetic resonance data second sub-sampling data.
[0010] Optionally, angle denoising loss and angle regularization loss are calculated based on angle similarity sub-sampling data, denoised angle similarity diffusion magnetic resonance data sub-sampling data, denoised first sub-sampling data and denoised diffusion magnetic resonance data first sub-sampling data, including: calculating angle denoising loss based on denoised first sub-sampling data and angle similarity sub-sampling data; calculating angle regularization loss based on angle similarity sub-sampling data, denoised angle similarity diffusion magnetic resonance data sub-sampling data, denoised first sub-sampling data and denoised diffusion magnetic resonance data first sub-sampling data.
[0011] Beneficial effects of the present invention:
[0012] The present invention proposes a joint spatial angle self-supervised denoising method for diffusion magnetic resonance data, proposes an effective spatial subsampler and angle sampler, jointly considers spatial and angle information, constructs paired data with the same noise distribution from the diffusion magnetic resonance data, performs self-supervised learning based on the denoising principle, and thus achieves effective denoising. A regularization loss is introduced to compensate for the differences in the data pairs due to different true values, and finally achieves effective denoising. Extensive experimental results show that the denoising effect is significantly better than the current advanced methods. In addition, the present invention solves the problem in the prior art that when denoising diffusion magnetic resonance imaging, spatial angle information is not fully utilized, and the noise of diffusion magnetic resonance data cannot be effectively eliminated. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:
[0014] Figure 1 is a flow chart of a joint spatial angle self-supervised denoising method for diffusion magnetic resonance data according to an embodiment of the present invention;
[0015] Figure 2 (a) is a schematic diagram of the overall structure of a joint spatial angle self-supervised denoising method for diffusion magnetic resonance data according to an embodiment of the present invention;
[0016] Figure 2 (b) is a schematic diagram of a denoising network for performing denoising test on diffusion magnetic resonance data according to an embodiment of the present invention;
[0017] Figure 2 (c) is a schematic diagram of sampling by an angle sampler according to an embodiment of the present invention;
[0018] Figure 2 (d) is a schematic diagram of sampling by a spatial subsampler according to an embodiment of the present invention;
[0019] Figure 3 is a schematic diagram of a denoising network according to an embodiment of the present invention. DETAILED DESCRIPTION
[0020] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only embodiments of a part of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0021] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0022] Example 1
[0023] According to an embodiment of the present invention, a joint spatial angle self-supervised denoising method for diffusion magnetic resonance data is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system comprising at least one set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in an order different from that shown here.
[0024] Figure 1 is a flow chart of a joint spatial angle self-supervised denoising method for diffusion magnetic resonance data according to an embodiment of the present invention. Figure 1 As shown, the method may include the following steps:
[0025] Step S101, acquiring diffusion magnetic resonance data.
[0026] In the technical solution provided in the above step S101 of the present invention, when the target user takes diffusion magnetic resonance data, the diffusion magnetic resonance data is acquired.
[0027] Step S102, processing the diffusion magnetic resonance data by an angle collector to obtain angle-similar diffusion magnetic resonance data similar to the diffusion magnetic resonance data.
[0028] In the technical solution provided in step S102 of the present invention, the diffusion magnetic resonance data is processed by the angle sampler a(·) to obtain angle-similar diffusion magnetic resonance data similar to the diffusion magnetic resonance data. According to an embodiment of the present invention, a schematic diagram of sampling by an angle sampler is provided. Figure 2 (c) is a schematic diagram of an angle sampler sampling according to an embodiment of the present invention, wherein the specific sampling process of the sampler is as follows: Step 1: For each gradient q i , find its angle θ(q i ,qi′ )’s nearest neighbor gradient q i′ Step 2: If θ(q i ,q i′ )>δ, the angle-similar diffusion-weighted data in y is assigned to y', that is, y'(q i )=y(q i ), otherwise the image with adjacent angle is used, i.e. y'(q i )=y(q i′ ). Step 3: Repeat steps 1 and 2 for all gradients to obtain y'. According to an embodiment of the present invention, a schematic diagram of the overall structure of a denoising method for diffusion magnetic resonance data is provided. Figure 2 (a) is a schematic diagram of the overall structure of a denoising method for diffusion magnetic resonance data according to an embodiment of the present invention. Figure 2 (a) The leftmost part is the angle collector sampling the diffusion magnetic resonance data y to obtain the angle-similar diffusion magnetic resonance data y' which is similar to the diffusion magnetic resonance data.
[0029] Step S103, processing the diffusion magnetic resonance data and the angle-similar diffusion magnetic resonance data by a spatial subsampler to obtain first subsampled data, second subsampled data and angle-similar subsampled data, wherein the original first and second spatial subsampled data correspond to the diffusion magnetic resonance data, and the angle-similar subsampled data correspond to the angle-similar diffusion magnetic resonance data.
[0030] In the technical solution provided in step S103 of the present invention, the spatial subsampler samples the diffusion magnetic resonance data to obtain first subsampled data and second subsampled data. According to an embodiment of the present invention, a schematic diagram of sampling by a spatial subsampler is provided. Figure 2 (d) is a schematic diagram of a spatial subsampler sampling according to an embodiment of the present invention, wherein the specific sampling process of the sampler is as follows: Step 1: According to the spatial information, y is divided into K 3 Step 2: Repeat by randomly selecting a voxel in each adjacent region until all adjacent regions have been visited. Step 3: Use the voxels selected in the previous step to construct a voxel of size Sub-sampled data s 1 (y). Step 4: For another sub-sampled data s 2 (y), repeat steps 2 and 3. Figure 2 As shown in (a), the first sub-sampled data is s 1 (y) represents the second sub-sampling data, s 2 (y) represents the angle similarity sub-sampling data with s 1 (y′) represents.
[0031] Step S104, processing the first sub-sampled data, the diffusion magnetic resonance data and the angle-similar diffusion magnetic resonance data through a denoising network to obtain denoised first sub-sampled data, denoised diffusion magnetic resonance data and denoised angle-similar diffusion magnetic resonance data, wherein the denoised first sub-sampled data corresponds to the first sub-sampled data, the denoised diffusion magnetic resonance data corresponds to the diffusion magnetic resonance data, and the denoised angle-similar diffusion magnetic resonance data corresponds to the angle-similar diffusion magnetic resonance data.
[0032] In the technical solution provided in step S104 of the present invention, if Figure 2 (a) shows the denoising network f θ For the first sub-sampled data s 1 (y), the diffusion magnetic resonance data y and the angle-similar diffusion magnetic resonance data y' are processed to obtain the denoised first sub-sampled data f θ (s 1 (y)), denoised diffusion magnetic resonance data f θ (y) and denoised angle-similar diffusion MRI data f θ (y′).
[0033] Step S105, processing the denoised diffusion magnetic resonance data and the denoised angle-similar diffusion magnetic resonance data by a spatial subsampler to obtain denoised diffusion magnetic resonance data first subsampled data, denoised diffusion magnetic resonance data second subsampled data and denoised angle-similar diffusion magnetic resonance data subsampled data, wherein the denoised diffusion magnetic resonance data first subsampled data and the denoised diffusion magnetic resonance data second subsampled data correspond to the denoised diffusion magnetic resonance data, and the denoised angle-similar diffusion magnetic resonance data subsampled data correspond to the denoised angle-similar diffusion magnetic resonance data.
[0034] In the technical solution provided in step S105 of the present invention, if Figure 2 As shown in (a), the de-noised diffusion magnetic resonance data is processed by a spatial sub-sampler to obtain the first sub-sampled data s of the de-noised diffusion magnetic resonance data. 1 (f θ (y)) and the denoised diffusion magnetic resonance data second sub-sampled data s 2 (f θ (y)), the denoised angle similar diffusion magnetic resonance data is processed by a spatial subsampler to obtain the denoised angle similar diffusion magnetic resonance data subsampled data s 1 (f θ (y′)).
[0035] Step S106, based on the denoised first sub-sampling data, the second sub-sampling data, the denoised diffusion magnetic resonance data first sub-sampling data, the denoised diffusion magnetic resonance data second sub-sampling data, the angular similarity sub-sampling data, and the denoised angular similarity diffusion magnetic resonance data sub-sampling data, calculate the spatial denoising loss, the spatial regularization loss, the angular denoising loss and the angular regularization loss to optimize the denoising model.
[0036] In the technical solution provided in the above step S107 of the present invention, the spatial denoising loss, the spatial regularization loss, the angular denoising loss and the angular regularization loss are calculated according to the denoised first sub-sampling data, the second sub-sampling data, the denoised diffusion magnetic resonance data first sub-sampling data, the denoised diffusion magnetic resonance data second sub-sampling data, the angular similarity sub-sampling data, and the denoised angular similarity diffusion magnetic resonance data sub-sampling data, and the denoising model is optimized by the spatial denoising loss, the spatial regularization loss, the angular denoising loss and the angular regularization loss.
[0037] Step S107: After the trained denoising model is obtained, the diffusion magnetic resonance data is processed through the denoising model to obtain denoised diffusion magnetic resonance data.
[0038] In the technical solution provided in step S107 of the present invention, after the denoising model denoises the diffusion magnetic resonance data, denoised diffusion magnetic resonance data is obtained. According to an embodiment of the present invention, a schematic diagram of a denoising network performing a denoising test on diffusion magnetic resonance data is provided. Figure 2 (b) is a schematic diagram of a denoising network performing denoising test on diffusion magnetic resonance data according to an embodiment of the present invention.
[0039] The above method of this embodiment is further introduced below.
[0040] As an optional implementation method, step S103, processing the diffusion magnetic resonance data and the angle-similar diffusion magnetic resonance data by a spatial subsampler to obtain first sub-sampling data, second sub-sampling data and angle-similar sub-sampling data, includes: dividing the diffusion magnetic resonance data into fixed adjacent regions based on spatial information; randomly selecting a first voxel in each adjacent region until all adjacent regions are visited; obtaining first sub-sampling data based on the first voxel; randomly selecting a second voxel in the adjacent region of the first voxel until all adjacent regions are visited; obtaining second sub-sampling data based on the second voxel; for the angle-similar diffusion magnetic resonance data, only randomly selecting the first voxel in the fixed adjacent region to obtain the angle-similar sub-sampling data.
[0041] In this embodiment, step 1: according to the spatial information, the diffusion magnetic resonance data is divided into K 3Step 2: Repeat by randomly selecting a voxel in each adjacent region until all adjacent regions have been visited. Step 3: Use the voxels selected in the previous step to construct a voxel of size The first sub-sampled data s 1 (y). Step 4: For the second sub-sampled data s 2 (y) Repeat steps 2 and 3.
[0042] As an optional implementation manner, in step S104, the denoising network includes: an input layer, 5 Conv3d layers with a convolution kernel of 3, 5 Conv3d layers with a convolution kernel of 1, 10 activation functions and an output layer.
[0043] In this embodiment, according to an embodiment of the present invention, a schematic diagram of a denoising network is provided. Figure 3 It is a schematic diagram of a denoising network according to an embodiment of the present invention, wherein the denoising network includes: an input layer, 5 Conv3d layers with a convolution kernel of 3, 5 Conv3d layers with a convolution kernel of 1, 10 activation functions and an output layer.
[0044] As an optional implementation method, step S104, the denoising network processing process includes: the input data passes through a Conv3d with a convolution kernel of 1 and then passes through two layers of Conv3d with a convolution kernel of 3, and the result is added, and then undergoes such operations twice, and finally passes through a Conv3d with a convolution kernel of 1 to obtain the output result, wherein the Rulu activation function needs to be passed after Conv3d.
[0045] In this embodiment, according to Figure 3 It can be obtained that the input data passes through a Conv3d with a convolution kernel of 1 and then passes through two layers of Conv3d with a convolution kernel of 3, and then the result is added, and then after two such operations, it passes through a Conv3d with a convolution kernel of 1 to get the output result.
[0046] As an optional implementation manner, step S106, based on the denoised first sub-sampling data, the second sub-sampling data, the denoised diffusion magnetic resonance data first sub-sampling data, the denoised diffusion magnetic resonance data second sub-sampling data, the angle similarity sub-sampling data, and the denoised angle similarity diffusion magnetic resonance data sub-sampling data, calculates the loss function, including: calculating the spatial denoising loss and the spatial regularization loss based on the second sub-sampling data, the denoised first sub-sampling data, the denoised diffusion magnetic resonance data first sub-sampling data, and the denoised diffusion magnetic resonance data second sub-sampling data; calculating the angle denoising loss and the angle regularization loss based on the angle similarity sub-sampling data, the denoised angle similarity diffusion magnetic resonance data sub-sampling data, the denoised first sub-sampling data, and the denoised diffusion magnetic resonance data first sub-sampling data; optimizing the denoising model through the spatial denoising loss, the spatial regularization loss, the angle denoising loss, and the angle regularization loss.
[0047] In this embodiment, if Figure 2 (a) shows the denoised first sub-sampled data f θ (s 1 (y)) and the second sub-sampled data s 2 (y) is calculated to obtain the spatial denoising loss L x ; Denoise the first sub-sampled data f θ (s 1 (y)) is similar to the angle sub-sampled data s 1 (y′) is calculated to obtain the angle denoising loss L q . The denoised first sub-sampled data f θ (s 1 (y)), the second sub-sampling data s 2 (y), denoised diffusion magnetic resonance data first sub-sampled data s 1 9f θ (y)) and the denoised diffusion magnetic resonance data second sub-sampled data s 2 (f θ (y)) is calculated to obtain the spatial regularization loss L rx ; The denoised first sub-sampled data (s 1 (y)), angle similarity sub-sampling data s 1 (y′), the first sub-sampled data of denoised diffusion magnetic resonance data s 1 (f θ (y)) and denoised angle-similar diffusion MRI data subsampled data s 1 (f θ (y′)) is calculated to obtain the angle regularization loss L rq The spatial denoising loss, spatial regularization loss, angular denoising loss and angular regularization loss are calculated to optimize the denoising model. The calculation process is shown in formula (1):
[0048] L = L x + γ x L rx +μ(L q + γ q L rq ) (1)
[0049] Among them, L is the target loss, L x is the spatial denoising loss, L rx is the spatial regularization loss, L q is the angle denoising loss, L rq is the angle regularization loss, γ x and γ q are two parameters that control the spatial and angular regularization losses L rx and L rq μ balances the contribution of spatial denoising loss and angular denoising loss. According to a monotonically decreasing function, μ gradually decreases from 1 to 0, which is specifically expressed as where e represents the index of the current training round, and Q is an adjustable parameter. It is worth noting that s 1 (f θ (y))、s 1 (f θ (y′)) and s 2 (f θ (y)) does not participate in gradient back propagation.
[0050] As an optional implementation method, based on the second sub-sampling data, the denoised first sub-sampling data, the denoised diffusion magnetic resonance data first sub-sampling data, and the denoised diffusion magnetic resonance data second sub-sampling data, the spatial denoising loss and the spatial regularization loss are calculated, including: based on the second sub-sampling data and the denoised first sub-sampling data, the spatial denoising loss is calculated; based on the second sub-sampling data, the denoised first sub-sampling data, the denoised diffusion magnetic resonance data first sub-sampling data, and the denoised diffusion magnetic resonance data second sub-sampling data, the spatial regularization loss is calculated.
[0051] In this embodiment, according to formula (2), the spatial denoising loss is obtained, and formula (2) is as follows:
[0052] L x =argmin θ E x,y ||f θ (s 1 (y))-s 2 (y)|| (2)
[0053] Among them, L x is the spatial denoising loss, f θ (s1 (y)) is the denoised first sub-sampled data, s 2 (y) is the second sub-sampling data.
[0054] In this embodiment, according to formula (3), the spatial regularization loss is obtained, and formula (3) is as follows:
[0055]
[0056] Among them, L rx is the spatial regularization loss, f θ (s 1 (y)) is the denoised first sub-sampled data, s 2 (y) is the second sub-sampled data, the denoised diffusion magnetic resonance data is the first sub-sampled data s 1 (f θ (y)) and the denoised diffusion magnetic resonance data second sub-sampled data s 2 (f θ (y)).
[0057] As an optional implementation method, angle denoising loss and angle regularization loss are calculated based on angle similarity sub-sampling data, denoised angle similarity diffusion magnetic resonance data sub-sampling data, denoised first sub-sampling data and denoised diffusion magnetic resonance data first sub-sampling data, including: calculating angle denoising loss based on denoised first sub-sampling data and angle similarity sub-sampling data; calculating angle regularization loss based on angle similarity sub-sampling data, denoised angle similarity diffusion magnetic resonance data sub-sampling data, denoised first sub-sampling data and denoised diffusion magnetic resonance data first sub-sampling data.
[0058] According to formula (4), the angle denoising loss is obtained, and formula (4) is as follows:
[0059] L q =argmin θ E x,y ||f θ (s 1 (y))-s 1 (y′)|| (4)
[0060] Among them, L q is the angle denoising loss, f θ (s 1 (y)) is the denoised first sub-sampled data, s 1 (y′) is the angle similarity sub-sampling data.
[0061] According to formula (5), the angle regularization loss is obtained. Formula (5) is as follows:
[0062]
[0063] L rq is the angle regularization loss, f θ (s 1 (y)) is the denoised first sub-sampled data, s 1 (y′) is the angle similarity sub-sampling data, the first sub-sampling data of the denoised diffusion magnetic resonance data s 1 (f θ (y)) and denoised angle-similar diffusion MRI data subsampled data s 1 (f θ (y′)).
[0064] In an embodiment of the present invention, diffusion magnetic resonance data is acquired; the diffusion magnetic resonance data is processed by an angle collector to obtain angle-similar diffusion magnetic resonance data similar to the diffusion magnetic resonance data; the diffusion magnetic resonance data and the angle-similar diffusion magnetic resonance data are processed by a spatial subsampler to obtain first subsampled data, second subsampled data and angle-similar subsampled data, wherein the original first and second spatial subsampled data correspond to the diffusion magnetic resonance data, and the angle-similar subsampled data correspond to the angle-similar diffusion magnetic resonance data; the first subsampled data, the diffusion magnetic resonance data and the angle-similar diffusion magnetic resonance data are processed by a denoising network to obtain denoised first subsampled data, denoised diffusion magnetic resonance data and denoised angle-similar diffusion magnetic resonance data, wherein the denoised first subsampled data corresponds to the first subsampled data, the denoised diffusion magnetic resonance data corresponds to the diffusion magnetic resonance data, and the denoised angle-similar diffusion magnetic resonance data corresponds to the angle-similar diffusion magnetic resonance data; the denoised diffusion magnetic resonance data and the denoised angle-similar diffusion magnetic resonance data are processed by a spatial subsampler. The denoised diffusion magnetic resonance data is processed to obtain denoised first sub-sampled data of diffusion magnetic resonance data, denoised second sub-sampled data of diffusion magnetic resonance data and denoised sub-sampled data of diffusion magnetic resonance data with similar angles, wherein the denoised first sub-sampled data of diffusion magnetic resonance data and the denoised second sub-sampled data of diffusion magnetic resonance data correspond to the denoised diffusion magnetic resonance data, and the denoised sub-sampled data of diffusion magnetic resonance data with similar angles corresponds to the denoised diffusion magnetic resonance data with similar angles; based on the denoised first sub-sampled data, the second sub-sampled data, the denoised first sub-sampled data of diffusion magnetic resonance data, the denoised second sub-sampled data of diffusion magnetic resonance data and the denoised second sub-sampled data of diffusion magnetic resonance data correspond to the denoised second sub-sampled data of diffusion magnetic resonance data with similar angles. The second sub-sampling data of noisy diffusion magnetic resonance data, the angular similarity sub-sampling data, and the denoised angular similarity diffusion magnetic resonance data sub-sampling data are used to calculate the spatial denoising loss, the spatial regularization loss, the angular denoising loss and the angular regularization loss to optimize the denoising model; after obtaining the trained denoising model, the diffusion magnetic resonance data is passed through the denoising model to obtain denoised diffusion magnetic resonance data. The present invention solves the problem in the prior art that when denoising diffusion magnetic resonance imaging, the spatial angle information is not fully utilized and the noise of the diffusion magnetic resonance data cannot be effectively eliminated.
[0065] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.
[0066] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0067] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.
[0068] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed over multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.
[0069] In addition, each functional unit in each embodiment of the present invention may be integrated into a first processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0070] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.
[0071] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A joint spatial angle self-supervised denoising method for diffusion magnetic resonance data, characterized in that: include: Acquiring diffusion magnetic resonance data; Processing the diffusion magnetic resonance data by an angle collector to obtain angle-similar diffusion magnetic resonance data similar to the diffusion magnetic resonance data; Processing the diffusion magnetic resonance data and the angle-similar diffusion magnetic resonance data by a spatial subsampler to obtain first subsampled data, second subsampled data, and angle-similar subsampled data, wherein the original first and second spatial subsampled data correspond to the diffusion magnetic resonance data, and the angle-similar subsampled data correspond to the angle-similar diffusion magnetic resonance data; Processing the first sub-sampled data, the diffusion magnetic resonance data and the angle-similar diffusion magnetic resonance data through a denoising network to obtain denoised first sub-sampled data, denoised diffusion magnetic resonance data and denoised angle-similar diffusion magnetic resonance data, wherein the denoised first sub-sampled data corresponds to the first sub-sampled data, the denoised diffusion magnetic resonance data corresponds to the diffusion magnetic resonance data, and the denoised angle-similar diffusion magnetic resonance data corresponds to the angle-similar diffusion magnetic resonance data; Processing the denoised diffusion magnetic resonance data and the denoised angle-similar diffusion magnetic resonance data by a spatial subsampler to obtain denoised diffusion magnetic resonance data first subsampled data, denoised diffusion magnetic resonance data second subsampled data, and denoised angle-similar diffusion magnetic resonance data subsampled data, wherein the denoised diffusion magnetic resonance data first subsampled data and the denoised diffusion magnetic resonance data second subsampled data correspond to the denoised diffusion magnetic resonance data, and the denoised angle-similar diffusion magnetic resonance data subsampled data correspond to the denoised angle-similar diffusion magnetic resonance data; Based on the denoised first sub-sampled data, the second sub-sampled data, the denoised diffusion magnetic resonance data first sub-sampled data, the denoised diffusion magnetic resonance data second sub-sampled data, the angular similarity sub-sampled data, and the denoised angular similarity diffusion magnetic resonance data sub-sampled data, a spatial denoising loss, a spatial regularization loss, an angular denoising loss, and an angular regularization loss are calculated to optimize the denoising model; After the trained denoising model is obtained, the diffusion magnetic resonance data is passed through the denoising model to obtain denoised diffusion magnetic resonance data.
2. The method according to claim 1, characterized in that The diffusion magnetic resonance data and the angle-similar diffusion magnetic resonance data are processed by a spatial subsampler to obtain first subsampled data, second subsampled data and angle-similar subsampled data, including: Divide the diffusion magnetic resonance data into fixed adjacent regions based on spatial information; Randomly select a first voxel in each neighboring region until all neighboring regions have been visited; Based on the first voxel, obtaining first sub-sampling data; Randomly select a second voxel in the neighboring area of the first voxel until all neighboring areas have been visited; Based on the second voxel, obtaining second sub-sampling data; For angle-similar diffusion MRI data, only the first voxel is randomly selected in a fixed adjacent region to obtain angle-similar sub-sampling data.
3. The method according to claim 1, characterized in that The denoising network consists of an input layer, 5 Conv3d layers with a convolution kernel of 3, 5 Conv3d layers with a convolution kernel of 1, 10 activation functions and an output layer.
4. The method according to claim 1, characterized in that: The denoising network processing process is as follows: the input data passes through a Conv3d with a convolution kernel of 1 and then passes through two layers of Conv3d with a convolution kernel of 3, and then the result is added, and then this operation is repeated twice, and finally the output result is obtained by passing through a Conv3d with a convolution kernel of 1. Among them, the Rulu activation function needs to be passed after Conv3d.
5. The method according to claim 1, characterized in that Based on the denoised first sub-sampled data, the second sub-sampled data, the denoised diffusion magnetic resonance data first sub-sampled data, the denoised diffusion magnetic resonance data second sub-sampled data, the angle-similar sub-sampled data, and the denoised angle-similar diffusion magnetic resonance data sub-sampled data, a loss function is calculated, including: Calculating a spatial denoising loss and a spatial regularization loss based on the second sub-sampled data, the denoised first sub-sampled data, the denoised diffusion magnetic resonance data first sub-sampled data, and the denoised diffusion magnetic resonance data second sub-sampled data; Calculating an angle denoising loss and an angle regularization loss based on the angle similarity sub-sampled data, the denoised angle similarity diffusion magnetic resonance data sub-sampled data, the denoised first sub-sampled data and the denoised diffusion magnetic resonance data first sub-sampled data; The denoising model is optimized by spatial denoising loss, spatial regularization loss, angular denoising loss and angular regularization loss.
6. The method according to claim 5, characterized in that Based on the second sub-sampled data, the denoised first sub-sampled data, the denoised diffusion magnetic resonance data first sub-sampled data, and the denoised diffusion magnetic resonance data second sub-sampled data, calculating a spatial denoising loss and a spatial regularization loss, including: Calculating a spatial denoising loss based on the second sub-sampled data and the denoised first sub-sampled data; A spatial regularization loss is calculated based on the second sub-sampled data, the denoised first sub-sampled data, the denoised diffusion magnetic resonance data first sub-sampled data, and the denoised diffusion magnetic resonance data second sub-sampled data.
7. The method according to claim 5, characterized in that Based on the angle-similar sub-sampled data, the denoised angle-similar diffusion magnetic resonance data sub-sampled data, the denoised first sub-sampled data and the denoised diffusion magnetic resonance data first sub-sampled data, an angle denoising loss and an angle regularization loss are calculated, including: Calculating the angle denoising loss based on the denoised first sub-sampled data and the angle similarity sub-sampled data; An angle regularization loss is calculated based on the angle similarity sub-sampled data, the denoised angle similarity diffusion magnetic resonance data sub-sampled data, the denoised first sub-sampled data and the denoised diffusion magnetic resonance data first sub-sampled data.
8. A processor, characterized in that: The processor is used to run a program, wherein the program executes any one of the methods in claims 1 to 6 when it is run.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute any one of the methods in claims 1 to 6.