A training method for MRI reconstruction network based on speedup segmentation

Abstract

A training method for an MRI reconstruction network based on speedup ratio segmentation, firstly, by deeply segmenting the target speedup ratio, a series of MRI data with different speedup ratios are constructed; secondly, based on a sequential training method, from low speedup ratio to high speedup ratio, the mapping relationship between each undersampled MRI image and the fully sampled MRI image is sequentially learned, and only a small number of iterative trainings are performed for each speedup ratio; finally, multiple iterative trainings are performed for the target speedup ratio. The method provided by the present invention can effectively utilize small-scale training samples to reconstruct high-multiple undersampled MRI images, and has strong practicality.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, in particular to the field of medical image technology of magnetic resonance imaging, and is mainly used for realizing accelerated reconstruction of medical magnetic resonance imaging. Background Art

[0002] Magnetic Resonance Imaging (MRI) is a medical imaging technology that uses a strong magnetic field and a specific radio frequency pulse sequence to obtain detailed images of human tissues or organs. It has the advantages of no radiation, high image quality, and rich contrast information, and is widely used in clinical diagnosis. However, the biggest limitation of MRI is that it takes a long time to acquire images. On the one hand, it is easy to cause motion artifacts in the image. On the other hand, patients need to endure the discomfort caused by long-term static acquisition in a closed device, and it cannot be used for patients in critical conditions. Therefore, since the birth of MRI, fast MRI technology has been a hot research issue.

[0003] Recent studies have shown that deep learning techniques can effectively improve the quality of MRI reconstruction images and achieve reconstruction with higher acceleration ratios. The training of deep neural networks often requires large-scale training data. For MRI reconstruction, large-scale MRI training data is often obtained from different scanners. The main problems are: (1) Although it comes from different MRI scanners, the cost of data acquisition is still high; (2) MRI training data from different scanners have large differences in resolution and contrast, which makes network training difficult; (3) MRI data from the target scanner also often have huge differences in resolution and contrast. Therefore, in practical applications, the network model obtained by training based on training data from multiple MRI scanners is difficult to achieve good generalization performance on a specific target scanner. In order to obtain larger-scale training data for the target scanner, data augmentation is usually required. Multi-scale augmentation of MRI data based on different sampling ratios is a common method. Existing methods make full use of the correlation between multi-scale data and adopt a progressive sequential training method: first use low-multiple under-sampled data to train the network, and then gradually use higher-multiple under-sampled data to train the network until the reconstruction of under-sampled data at the target acceleration ratio can be completed well. A major shortcoming of this multi-scale sequential training method is that the division of sampling ratios is still relatively rough, and the span of sampling ratios from low-multiple undersampling data to the next high-multiple undersampling data is still large. Therefore, if the target sampling ratio is high, it is difficult for the network trained at the previous low-multiple undersampling ratio to learn enough prior knowledge to effectively assist in the reconstruction of MRI data at the next sampling ratio, and it is also difficult to obtain high reconstruction performance. Summary of the invention

[0004] In order to overcome the problems existing in the existing multi-scale MRI sequential training methods, the present invention proposes a training method for an MRI reconstruction network based on speedup ratio subdivision, namely: the speedup ratio of the under-sampled data used in the training process is divided more finely, and only a small number of iterations are performed for the training data of each speedup ratio (except the target speedup ratio); while sufficient iterations and optimization are performed for the training data of the target speedup ratio.

[0005] The technical solution provided by the present invention is as follows:

[0006] A training method for an MRI reconstruction network based on speedup segmentation comprises the following steps:

[0007] Step 1: Select the MRI k-space data of n subjects collected by the target scanner to form a set Y = {y 1 ,y 2 ,…,y n},in, represents k-space;

[0008] Step 2: Build a deep neural network model f CNN (·; Θ), where Θ represents f CNN Parameters;

[0009] Step 3: Randomly initialize f CNN The parameter set Θ = Θ (0) , so that the best reconstructed image quality

[0010] Step 4: Set the total number of iterations s=0 and the index of the current speedup ratio t=1;

[0011] Step 5: Calculate the current speedup u t :

[0012]

[0013] Where u0=u T is the target speedup ratio, T is the preset speedup ratio subdivision number;

[0014] Step 6: Subdivision speedup u t , build a training dataset:

[0015]

[0016] The superscript i represents the index of the training sample, and x i =F -1 y i , F -1 represents the two-dimensional inverse Fourier transform, here, Denote the mask matrix for undersampling the magnetic resonance image in k-space by u t times;

[0017] Step 7 further randomly divides the training set into a training set and a validation set Let the training cycle e of the current acceleration ratio t = 1;

[0018] Step 8 randomly divides the training set into B subsets Let the current iteration batch b = 1, and E t be the maximum training cycle preset for the current acceleration ratio u t ;

[0019] Step 9 For all calculate of the reconstructed image

[0020]

[0021] Step 10 For a given loss function Optimize the following objective formula by the gradient descent method:

[0022]

[0023] Step 11 If t < T, jump to the above Step 13, otherwise, execute the next step;

[0024] Step 12 Evaluate the reconstruction image quality of the deep neural network f on the validation set CNN (·; Θ (s) ):

[0025]

[0026] where Q(·,·) is the quality evaluation function;

[0027] Step 13 If Let Θ * = Θ (s) ;

[0028] Step 14 Let s = s + 1, b = b + 1. If b ≤ B, jump to the above Step 9, otherwise, execute the next step;

[0029] Step 15 Let e t = e t + 1. If e t ≤ E t , jump to the above Step 8, otherwise, execute the next step;

[0030] Step 16: Let t=t+1. If t≤T, jump to the above step 5. Otherwise, stop the iteration.

[0031] The beneficial effects of the present invention are that it can effectively utilize small-scale training samples to reconstruct highly undersampled MRI images, and has strong practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flow chart of the training method of the MRI reconstruction network based on acceleration ratio subdivision proposed by the present invention.

[0033] Figure 2 Schematic diagram of the training process of the MRI reconstruction network based on the acceleration ratio segmentation proposed in the present invention. In the figure, y is the fully sampled k-space data, is a series of undersampled k-space data at different acceleration ratios, where u T =u0 is the target acceleration ratio; the arrow → indicates the direction of data transmission, and the arrow → indicates the model evaluation function, whose source input is the network model to be evaluated, and its output points to the optimal network model obtained by the evaluation.

[0034] Figure 3 is the network structure of the deep cascade network used in the present invention, wherein, Figure 3 (a) is the overall structure of the network: n c The subnets are cascaded, each of which contains a convolutional neural network module and a DC layer; Figure 3 (b) shows the specific structure of the i-th subnet: Convolutional neural network module CNN i In d Convolutional layers Cascaded, in CNN i The output end of the DC layer is residually connected to the input end; the DC layer first performs Fourier transform on the input image, and then performs data consistency operation on the Fourier transform result and the undersampled k-space data. DC , and finally converted into image output through inverse Fourier transform.

[0035] Figure 4 There are two ways to undersample k-space data. The left picture is Gaussian undersampling, and the right picture is central undersampling.

[0036] Figure 5This is a visual comparison of the reconstruction effects of 8-fold undersampling MRI images. From top to bottom and from left to right, they are the real image, the zero-filled image after 8-fold undersampling, the reconstructed image of the network model obtained based on the single-scale training method, the reconstructed image of the network model obtained based on multi-scale hybrid training, and the reconstructed image of the network model based on the acceleration ratio segmentation proposed by the present invention. One region of interest is selected in each image and displayed in the lower right corner of the image after being enlarged. DETAILED DESCRIPTION

[0037] The present invention will be further described below in conjunction with the accompanying drawings.

[0038] Reference Figure 1 to Figure 5 , a training method for MRI reconstruction network based on acceleration ratio segmentation, the main process and the training process based on acceleration ratio segmentation are as follows Figure 1 and Figure 2 As shown, the tool includes the following steps:

[0039] Step 1: Select the MRI k-space data of n subjects collected by the target scanner to form a set Y = {y 1 ,y 2 ,…,y n},in, represents k-space;

[0040] Step 2: Build a deep neural network model f CNN (·; Θ), where Θ represents f CNN The parameter, f CNN (·; Θ) can be any neural network that can be used for image reconstruction. Commonly used network structures include deep cascade networks, such as Figure 3 As shown;

[0041] Step 3: Randomly initialize f CNN The parameter set Θ = Θ (0) The initialization method can be to sample from a Gaussian distribution with a mean of 0 and a variance of 1 as the initial weight, and set the initial value of the bias term to 0, so that the best reconstructed image quality

[0042] Step 4: Set the total number of iterations s=0 and the current subdivision speedup ratio index t=1;

[0043] Step 5: Calculate the current segmentation speedup u according to the index value t t , calculate the current acceleration ratio u t :

[0044]

[0045] Where u0=u Tis the target speedup ratio, T is the preset speedup ratio subdivision number, and in specific implementation, T can be set to 100;

[0046] Step 6: Subdivision speedup u t , build a training dataset:

[0047]

[0048] The superscript i represents the index of the training sample, and x i =F -1 y i , F -1 represents the two-dimensional inverse Fourier transform, here, Indicates the u-scaling of the magnetic resonance image in k-space t There are two common ways to undersample k-space data: Gaussian undersampling and center undersampling, as shown in Figure 4 As shown in the left and right figures of , in order to facilitate the segmentation of k-space data and progressive training, the central undersampling method can be used;

[0049] Step 7: The training set Further randomly split into training sets and validation set Let the training cycle e of the current acceleration ratio be t =1;

[0050] Step 8: The training set Randomly divide into B subsets The batch size (BatchSize) can be taken as 5, and the current iteration batch b = 1, E t For the current acceleration ratio u t The preset maximum training cycle, when the number of speedup subdivisions T = 100, the speedup of each subdivision is E t Can be set to 2, the target speedup ratio E T Can be set to 70;

[0051] Step 9 For all calculate The reconstructed image

[0052]

[0053] Step 10: For a given loss function By using the gradient descent method, the following objective formula is optimized:

[0054]

[0055] Here, the loss function The norm, the norm or the Charbonnier penalty function can be adopted, where the Charbonnier penalty function is defined as: where x1 and x2 are the reconstructed image and the original image respectively, and ∈ is a constant, and ∈ can be set to 0.001;

[0056] In step 11, if t < T, jump to step 13 above, otherwise, execute the next step;

[0057] In step 12, evaluate the reconstructed image quality of the deep neural network f (·; Θ CNN (·; Θ (s) ) on the validation set:

[0058]

[0059] where Q(·, ·) is a quality evaluation function, and in specific implementation, the peak signal-to-noise ratio (PSNR) or the structural similarity (SSIM) can be adopted as the quality evaluation function;

[0060] In step 13, if Let Θ * = Θ (s) ;

[0061] In step 14, let s = s + 1, b = b + 1. If b ≤ B, return to step 9 above; otherwise, execute the next step;

[0062] In step 15, let e t = e t + 1. If e t ≤ E t , jump to step 8 above; otherwise, execute the next step;

[0063] In step 16, let t = t + 1. If t ≤ T, jump to step 5 above; otherwise, stop the iteration.

[0064] Next, the beneficial effects of the training method of the MRI reconstruction network based on acceleration ratio subdivision proposed by the present invention are verified through experiments.

[0065]

[0066] Table 1

[0067]

[0068] Table 2

[0069] First, verify the influence of the subdivision degree T of the acceleration ratio on the performance of the MRI reconstruction network. Select as​ Figure 3 The deep cascade network shown is an MRI reconstruction network, and T is set to 10, 50, and 100 respectively; to ensure the consistency of the total number of training periods, corresponding to T = 10, T = 50, and T = 100, the number of training periods under each subdivision speedup ratio is set to 20, 4, and 2 respectively, that is, a total of 200 cycles of training, and at the target speedup ratio, continue training for 70 cycles; use the Adam optimization algorithm for training, and the initial learning rate is set to 0.001. The learning rate remains unchanged during the first 200 cycles of training, and in the next 70 cycles (i.e., the training stage for the target speedup ratio), the learning rate decays to 1 / 10 of the original every 40 cycles. For the three cases where the target speedup ratio is 4, 8, and 12, Table 1 reports the reconstruction performance of the deep cascade network under the three speedup ratio subdivision methods of T=10, T=50, and T=100. It can be seen that: with the increase of the subdivision degree, the network performance gradually improves in the reconstruction models of the three speedup ratios, and with the increase of the speedup ratio, the increase in the subdivision degree has a more obvious improvement on the network effect.

[0070] Secondly, the beneficial effects of the method proposed in the present invention are illustrated by comparing the reconstruction effects of five training methods, namely, the collar filling method (no training), single-scale training, multi-scale mixed training, multi-scale sequential training, and sequential training based on speedup ratio subdivision proposed in the present invention. Three different speedup ratios are selected for multi-scale mixed training and multi-scale sequential fine-tuning training. Specifically, for 4 times, 8 times, and 12 times the target speedup ratios, the speedup ratio combinations used in actual training are {2,3,4}, {2,4,8}, and {4,8,12}, respectively. The training method based on speedup ratio subdivision proposed in the present invention still selects T=100. The experimental results are shown in Table 2. It can be found that the multi-scale data mixed training model has a performance degradation compared to the single-scale training, which may be because the mixed data interferes with the network reconstruction target; multi-scale sequential fine-tuning training is better than direct training with the target single-scale data; the proposed multi-scale sequential training method based on speedup ratio subdivision achieves the best reconstruction effect. Compared with the single-scale training results, the PSNR / SSIM in the undersampled image reconstruction with 4x, 8x and 12x speedup ratios are improved by 0.20 / 0.0003, 0.31 / 0.0019 and 0.38 / 0.0029 respectively, and a greater improvement is achieved in the reconstruction with higher speedup ratios. Figure 5 Further visual comparison of the reconstructed images of various reconstruction methods was performed, and it can be seen that visually, the method proposed in the present invention is still the best.

[0071] The contents described in the embodiments of this specification are merely enumerations of implementation forms of the inventive concept and are for illustrative purposes only. The protection scope of the present invention should not be considered to be limited to the specific forms described in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that can be thought of by ordinary technicians in this field based on the inventive concept.

Claims

1. A training method for an MRI reconstruction network based on speedup segmentation, characterized in that: The method comprises the following steps: Step 1: Select the MRI k-space data of n subjects collected by the target scanner to form a set Y = {y 1 ,y 2 , ..., y n },in, represents k-space; Step 2: Build a deep neural network model f CNN (·; Θ), where Θ represents f CNN Parameters; Step 3: Randomly initialize f CNN The parameter set Θ = Θ (0) , so that the best reconstructed image quality Step 4: Set the total number of iterations s=0 and the index of the current speedup ratio t=1; Step 5: Calculate the current speedup ratio u t : Where u0=u T is the target speedup ratio, T is the preset speedup ratio subdivision number; Step 6: Subdivision speedup u t , build a training dataset: The superscript i represents the index of the training sample, and x i =F -1 y i , F -1 represents the two-dimensional inverse Fourier transform, here, Indicates the u-scaling of the magnetic resonance image in k-space t times undersampling mask matrix; Step 7: The training set Further randomly split into training sets and validation set Let the training cycle e of the current acceleration ratio be t =1; Step 8: The training set Randomly divide into B subsets Let the current iteration batch b = 1, E i For the current acceleration ratio u t Preset maximum training period; Step 9 For all calculate The reconstructed image Step 10: For a given loss function l(·,·), optimize the following objective formula by gradient descent: Step 11: If t<T, jump to step 13, otherwise, execute the next step; Step 12: On the validation set Evaluate the deep neural network f on CNN (·; Θ (s) )’s reconstructed image quality: Among them, Q(·,·) is the quality assessment function; Step 13 If make Θ * =Θ (s) ; Step 14: Let s = s + 1, b = b + 1. If b ≤ B, jump to the above step 9. Otherwise, go to the next step. Step 15: Let e t =e t +1, if e t ≤E t , skip to step 8 above, otherwise, proceed to the next step; Step 16: Let t=t+1. If t≤T, jump to the above step 5. Otherwise, stop the iteration.

Images