Reconstruction method for breaking through low-field magnetic resonance imaging quality limit

Abstract

The invention discloses a reconstruction method for breaking through a low-field magnetic resonance imaging quality limit. The method comprises the following steps: acquiring and constructing a low-field data set; acquiring and constructing a high field data set; respectively training a basic super-resolution model from a low-field low-resolution image excited by 1-N times to a low-field high-resolution image excited by N times; comparing the influence of different excitation times on the performance of the basic super-resolution model; training a cross-field super-resolution model from a low-field and N-time excited high-resolution image to a high-field and ultrahigh-resolution image; cascading the basic super-resolution model and the cross-field super-resolution model to construct an enhanced two-stage cross-field super-resolution model; a high-field and ultrahigh-resolution image is reconstructed from a low-field, low-excitation-frequency and low-resolution image through an enhanced two-stage cross-field super-resolution model. According to the invention, the low-field magnetic resonance scanner can realize high-quality imaging breaking through the magnetic field intensity constraint within a short scanning time.

Description

Technical Field

[0001] The invention belongs to the field of medical imaging technology, and in particular relates to a reconstruction method that breaks through the quality limit of low-field magnetic resonance imaging. Background Art

[0002] Low-field (< 0.3 T) MRI scanners have the advantages of low cost, high flexibility, and the ability to provide point-of-care imaging, and have attracted widespread attention in the field of medical imaging. However, due to physical limitations, low-field MRI scanners cannot obtain high-quality images within an acceptable scanning time, which seriously affects the realization of advanced functions and accurate diagnosis.

[0003] The existing basic super-resolution methods use low-resolution images to reconstruct high-resolution images, thereby improving the acquisition speed of low-field imaging. However, the images reconstructed by this method do not exceed the constraints of magnetic field strength, that is, images with high-field effects cannot be obtained. In addition, some cross-field super-resolution methods use high-field images to build data sets and train super-resolution models to reconstruct high-resolution images with high signal-to-noise ratio from low-resolution low-field images. The signal-to-noise ratio of the images reconstructed by this method is significantly improved, but the improvement in image resolution is not obvious, and the spatial resolution cannot reach the high-field level. There are two reasons for this phenomenon. First, cross-field super-resolution requires the model to handle the tasks of denoising and super-resolution at the same time. Due to the limited fitting ability of the model, although the signal-to-noise ratio of the reconstructed image is significantly improved, the improvement in image resolution is not obvious; second, in these methods, the resolution of the high-field image is degraded to the highest resolution level that can be obtained in the low field, which leads to the limit of cross-field super-resolution cannot break through the resolution of low-field imaging.

[0004] Therefore, it is necessary to propose a reconstruction method that breaks through the quality limit of low-field MRI, and reconstructs images with a level close to high-field and ultra-high resolution using low-field and low-resolution images obtained with a shorter scanning time. Summary of the invention

[0005] In view of the above-mentioned defects or shortcomings in the existing methods, the present invention provides a reconstruction method that breaks through the quality limit of low-field magnetic resonance imaging. A basic super-resolution model and a cross-field super-resolution model are trained using a low-field data set and a high-field ultra-high resolution data set respectively. The two models are then cascaded to construct an enhanced two-stage cross-field super-resolution model. Finally, a high-quality image with a magnetic field strength limit that exceeds the low-field is reconstructed from the low-field, low-resolution image through the enhanced two-stage cross-field super-resolution model.

[0006] In order to achieve the above object, the present invention adopts the following technical scheme:

[0007] A reconstruction method that breaks through the quality limit of low-field magnetic resonance imaging includes the following steps:

[0008] Step 1: Collection and construction of low-field datasets, including,

[0009] Step 1.1: Under low field conditions, collect low-resolution images of the target object being excited 1 to N times, N ≥ 4, and N is a natural number;

[0010] Step 1.2: Under low field conditions, collect high-resolution images of the target object being excited N times;

[0011] Step 1.3: construct data pairs of low-resolution images of low-field 1 to N excitations and high-resolution images of low-field N excitations respectively;

[0012] Step 2: Collection and construction of high-field data sets, including,

[0013] Step 2.1: Under high field conditions, collect an ultra-high resolution image of the target object being excited once;

[0014] Step 2.2: performing degradation processing on the ultra-high resolution image to obtain a high-field degraded image;

[0015] Step 2.3: constructing a data pair of a high-field degraded image and the super-high resolution image;

[0016] Step 3: Train the basic super-resolution model from low-field, 1~N-excitation low-resolution images to low-field, N-excitation high-resolution images respectively;

[0017] Step 4: Compare the effects of different excitation times on the performance of the basic super-resolution model, including:

[0018] Step 4.1: Compare the image quality difference between the basic super-resolution result of the low-resolution image of the target object being stimulated 1 to N times under the low-field condition and the high-resolution image of the target object being stimulated N times under the low-field condition;

[0019] Step 4.2: Obtaining the minimum number of times the low-resolution image with a quality close to that of the high-resolution image can be obtained through a basic super-resolution model through an evaluation of visual effects;

[0020] Step 5: Train a cross-field super-resolution model from low-field, N-shot high-resolution images to high-field, ultra-high-resolution images;

[0021] Step 6: Cascade the basic super-resolution model and the cross-field super-resolution model to construct an enhanced two-stage cross-field super-resolution model;

[0022] Step 7: Reconstructing a high-field, super-high-resolution image from a low-field, low-excitation-number, low-resolution image by an enhanced two-stage cross-field super-resolution model, wherein the low-excitation-number is ≤ N-1;

[0023] Step 8: End.

[0024] Furthermore, the low-resolution image and the high-resolution image in step 1 are obtained from multiple scans of the same target object at the same position; and the data of each data pair are obtained from different target objects.

[0025] Furthermore, in step 2, ultra-high resolution refers to a resolution that cannot be obtained by a low-field magnetic resonance scanner within an acceptable scanning time, including acquisition resolution and spatial resolution; the high-field degraded image is obtained by downsampling the ultra-high resolution image, adding Gaussian noise, and adding Rice noise.

[0026] Furthermore, the basic super-resolution model and the cross-field super-resolution model are both 2x super-resolution models.

[0027] Furthermore, the minimum number of excitations in step 4 refers to the number of excitations that can obtain the low-resolution image with the same quality as the high-resolution image through the basic super-resolution model.

[0028] Furthermore, the first level of the enhanced two-level cross-field super-resolution model in step 6 is a basic super-resolution model, and the second level is a cross-field super-resolution model; the input of the basic super-resolution model is a low-field, low-resolution image with the least number of excitations, and the output is used as an intermediate feature of the enhanced two-level cross-field super-resolution model; the input of the cross-field super-resolution model is the intermediate feature, and the output is the output of the enhanced two-level cross-field super-resolution model.

[0029] Furthermore, the image reconstructed by the enhanced two-stage cross-field super-resolution model in step 7 has a quality that exceeds the low-field magnetic field strength constraint and has an effect close to that of the high-field, ultra-high resolution image, including contrast, signal-to-noise ratio, and spatial resolution.

[0030] The beneficial effects of the present invention are:

[0031] First, when collecting low-field data, each data pair of the present invention is a multiple scan of the same subject in the same position, so that the basic super-resolution model can learn the mapping relationship between low-resolution images with low field and low excitation times and high-resolution images with low field and multiple excitation times in real scanning conditions.

[0032] Secondly, while ensuring the low-field basic super-resolution performance, the present invention reduces the number of excitations of low-field, low-resolution images, and further improves the imaging speed for obtaining high-quality low-field, high-resolution images.

[0033] Third, the present invention uses high-field ultra-high resolution data to construct a cross-field super-resolution dataset, so that the cross-field super-resolution model can reconstruct images with high-field ultra-high resolution effect from low-field high-resolution images.

[0034] Fourth, the present invention cascades the basic super-resolution model and the cross-field super-resolution model and constructs an enhanced two-stage cross-field super-resolution model, so that the reconstructed image can first reach the limit level of low-field imaging quality and then further reach the level of high-field high-quality imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of a reconstruction method for breaking through the quality limit of low-field magnetic resonance imaging according to the present invention;

[0036] Figure 2 A schematic diagram of training a basic super-resolution model and a cross-field super-resolution model of the present invention;

[0037] Figure 3 It is a comparison diagram of the basic super-resolution image of the present invention with low field and 1 to 4 excitations and the high-resolution image with low field and 4 excitations;

[0038] Figure 4 Schematic diagram of an enhanced two-stage cross-field super-resolution model of the present invention;

[0039] Figure 5 This is a comparison diagram of the effects of the enhanced two-stage cross-field super-resolution model of the present invention at each stage;

[0040] Figure 6 A comparison diagram of the application effect of the embodiment of the present invention on a 0.2 T low-field scanner and a high-field ultra-high resolution image. DETAILED DESCRIPTION

[0041] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0042] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0043] like Figure 1The flowchart shown is a reconstruction method that breaks through the quality limit of low-field magnetic resonance imaging. In the following, high excitation times refers to the number of excitations that can obtain better image quality, such as 4 times. In the basic super-resolution of the present invention, the image with a high excitation number is set as the reference image. As mentioned above, the high excitation number N is not limited to a specific value. The following is a specific description using N=4 as an example. Specifically, the method includes the following steps:

[0044] Step 1: Collection and construction of low-field datasets, including,

[0045] Step 1.1: Collect low-resolution images with low field and 1 to 4 excitations;

[0046] Step 1.2: Acquire high-resolution images at low field and 4 excitations;

[0047] Step 1.3: Construct data pairs of low-resolution images of low-field 1 to 4 excitations and high-resolution images of low-field and 4 excitations. The data pairs are low-resolution images of low-field 1 excitation and high-resolution images of low-field and 4 excitations, low-resolution images of low-field 2 excitations and high-resolution images of low-field and 4 excitations, low-resolution images of low-field 3 excitations and high-resolution images of low-field and 4 excitations, and low-resolution images of low-field 4 excitations and high-resolution images of low-field and 4 excitations.

[0048] The low-field data set of multiple subjects was collected. For each subject, low-resolution images of 1 to 4 excitations and high-resolution images of 4 excitations were collected continuously at the same imaging position. The low-field low-resolution image was 128×128, and the 1.87 mm 2 The spatial resolution of low field high resolution is 256×256, and the acquisition resolution is 0.93 mm 2 spatial resolution.

[0049] Step 2: Collection and construction of high-field data sets, including,

[0050] Step 2.1: Under high field conditions, collect an ultra-high resolution image of the target object being excited once;

[0051] The high-field data set of multiple subjects was collected, and the ultra-high-resolution image with the same imaging orientation as the low-field data set was collected for each subject, and the number of excitations was 1. Among them, ultra-high resolution refers to the resolution that cannot be obtained by low-field magnetic resonance scanners within an acceptable scanning time, including acquisition resolution and spatial resolution; high-field ultra-high resolution is an acquisition resolution of 512×512, 0.45 mm 2 spatial resolution.

[0052] The data pairs in all low-field data sets are divided into training set, validation set and test set in a ratio of 8:1:1.

[0053] Step 2.2: performing degradation processing on the ultra-high resolution image to obtain a high-field degraded image;

[0054] Firstly, the resolution of the high-field super-resolution image is adjusted multiple times from 512×512→170×170→512×512→256×256 by using the bicubic interpolation downsampling method, so that the image is visually close to the resolution of the low-field high-resolution image. Then, the adjusted image is normalized. Finally, Gaussian noise with a standard deviation of 0.005 and Rice noise with a standard deviation of 0.002 are added to the normalized image and then normalized again to obtain the final high-field degraded image.

[0055] Step 2.3: Construct a data pair of the high-field degraded image and the super-high resolution image.

[0056] The data pairs in all high-field data sets are divided into training set, validation set and test set in a ratio of 8:1:1.

[0057] Step 3: Train the basic super-resolution model from low-field, 1~4-shot low-resolution images to low-field, 4-shot high-resolution images respectively.

[0058] As shown in Figure 2, the figure above is a schematic diagram of training a basic super-resolution model from a low-field, 1 to 4-times-excitation low-resolution image to a low-field, 4-times-excitation high-resolution image. The basic super-resolution model is trained from a low-field, 1 to 4-times-excitation low-resolution image to a high-field, 4-times-excitation image, and then the basic super-resolution model with the best performance corresponding to 4 different excitation times is obtained.

[0059] Step 4: Compare the effects of different excitation times on the performance of the basic super-resolution model, including:

[0060] Step 4.1: Compare the image quality difference between the basic super-resolution results of the low-resolution images of the target object being stimulated 1 to 4 times under low field conditions and the high-resolution images of the target object being stimulated 4 times under low field conditions;

[0061] Step 4.2: Obtain the minimum number of times the low-resolution image can be stimulated to obtain a quality close to that of the high-resolution image through basic super-resolution through visual effect evaluation.

[0062] Through actual observation, the basic super-resolution results of low-field, 1-4-shot low-resolution images are evaluated. Figure 3As shown in the figure, from left to right, they correspond to the basic super-resolution results of low-resolution images with low field and 1 to 4 excitations, and the high-resolution images with low field and 4 excitations. The white arrows in the figure point to the difference in details. It can be seen that the image quality of the third and fourth images is close to that of the fifth image, and there is no loss of detail information. There is a blurring effect in the first and second images, and there is a loss of detail information to varying degrees. Therefore, the minimum number of excitations required to obtain a low-field, low-resolution image with a quality close to that of a high-resolution image with low field and 4 excitations through basic super-resolution is 3.

[0063] Step 5: Train a cross-field super-resolution model from the low-field, 4-shot high-resolution images to the high-field, ultra-resolution images.

[0064] like Figure 2 The figure shows a schematic diagram of a cross-field super-resolution model for training high-field degraded images to high-field, ultra-high resolution images. In this case, a high-field degraded image is used to simulate a low-field, 4-times-excited high-resolution image. By training a cross-field super-resolution model from a high-field degraded image to a high-field, ultra-high resolution image, a cross-field super-resolution model with the best performance is obtained.

[0065] Step 6: Cascade the basic super-resolution model and the cross-field super-resolution model to construct an enhanced two-stage cross-field super-resolution model.

[0066] As shown in Figure 4, the basic super-resolution model and the cross-field super-resolution model are cascaded by merging the input and output to construct an enhanced two-stage cross-field super-resolution model. Among them, the first stage of the enhanced two-stage cross-field super-resolution model is the basic super-resolution model, and the second stage is the cross-field super-resolution model. The input of the basic super-resolution model is a low-field, low-resolution image with the least number of excitations, and the output is the intermediate feature of the enhanced two-stage cross-field super-resolution model; the input of the cross-field super-resolution model is the intermediate feature, and the output is the output of the enhanced two-stage cross-field super-resolution model.

[0067] Step 7: Reconstruct a high-field, super-resolution image from the low-field, low-excitation-number, low-resolution image through an enhanced two-stage cross-field super-resolution model.

[0068] The enhanced two-stage cross-field super-resolution model is fed with a low-field, three-shot low-resolution image to obtain intermediate features that are close to the low-field, four-shot high-resolution image effect. The enhanced two-stage cross-field super-resolution model further processes the intermediate features to output a final reconstructed image that is close to the high-field, ultra-high-resolution image effect.

[0069] As shown in Figure 5, the first row shows the low-resolution image and its detail map of the low-field, three-shot, the second row shows the intermediate features and its detail map output by the enhanced two-stage cross-field super-resolution model, and the third row shows the final image and its detail map output by the enhanced two-stage cross-field super-resolution model. It can be observed that from the first row to the third row, the image quality is significantly improved step by step in terms of resolution and signal-to-noise ratio.

[0070] Step 8: Finalize the design.

[0071] In one embodiment, based on the present invention, low-field data are collected on a 0.2 T low-field mobile scanner and a data set is constructed, and high-field data are collected on a 1.5 T high-field scanner and a data set is constructed. First, an enhanced two-stage cross-field super-resolution model is obtained. Then, a 0.2 T low field, 3 excitations, an acquisition resolution of 128×128, and a spatial resolution of 1.87 mm are performed. 2 The imaging experiment took 3.3 minutes for a single scan. Finally, the image was input into the enhanced two-stage cross-field super-resolution model to obtain a reconstructed image close to the high-field and ultra-high resolution effect. As shown in Figure 6, from left to right, the first image is a low-field, 3-shot image acquired on a 0.2 T low-field scanner with an acquisition resolution of 128×128 and a spatial resolution of 1.87 mm. 2 The first image is the intermediate feature and its detail map output by the enhanced two-stage cross-field super-resolution model, the third image is the final image and its detail map output by the enhanced two-stage cross-field super-resolution model, and the fourth image is a high-field image acquired by a 1.5T high-field scanner, with a single excitation, an acquisition resolution of 512×512, and a spatial resolution of 0.45 mm. 2 Images and their details. In the first image, the image quality is poor and some tissue structure information is missing. In the second image, the image quality is improved to a certain extent, the tissue structure information is complete, and the quality of this image reaches the limit level that a 0.2 T scanner can obtain within an acceptable scanning time. In the third image, the image quality is significantly improved, the tissue structure information is rich, the texture features are obvious, the spatial resolution is high, and the signal-to-noise ratio is high. Compared with the fourth image, the image quality of the third image reaches the level of 1.5 T high-field ultra-high resolution images.

[0072] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

[0073] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A reconstruction method that breaks through the quality limit of low-field magnetic resonance imaging, characterized in that: The following steps are involved: Step 1: Collection and construction of low-field datasets, including, Step 1.1: Under low field conditions, collect low-resolution images of the target object being excited 1 to N times, N ≥ 4, and N is a natural number; Step 1.2: Under low field conditions, collect high-resolution images of the target object being excited N times; Step 1.3: construct data pairs of low-resolution images of low-field 1 to N excitations and high-resolution images of low-field N excitations respectively; Step 2: Collection and construction of high-field data sets, including, Step 2.1: Under high field conditions, collect an ultra-high resolution image of the target object being excited once; Step 2.2: performing degradation processing on the ultra-high resolution image to obtain a high-field degraded image; Step 2.3: constructing a data pair of a high-field degraded image and the super-high resolution image; Step 3: Train the basic super-resolution model from low-field, 1~N-excitation low-resolution images to low-field, N-excitation high-resolution images respectively; Step 4: Compare the effects of different excitation times on the performance of the basic super-resolution model, including: Step 4.1: Compare the image quality difference between the basic super-resolution result of the low-resolution image of the target object being stimulated 1 to N times under the low-field condition and the high-resolution image of the target object being stimulated N times under the low-field condition; Step 4.2: Obtaining the minimum number of times the low-resolution image with a quality close to that of the high-resolution image can be obtained through a basic super-resolution model through an evaluation of visual effects; Step 5: Train a cross-field super-resolution model from low-field, N-shot high-resolution images to high-field, ultra-high-resolution images; Step 6: Cascade the basic super-resolution model and the cross-field super-resolution model to construct an enhanced two-stage cross-field super-resolution model; Step 7: Reconstructing a high-field, super-high-resolution image from a low-field, low-excitation-number, low-resolution image by an enhanced two-stage cross-field super-resolution model, wherein the low-excitation-number is ≤ N-1; Step 8: End.

2. The reconstruction method for breaking through the quality limit of low-field magnetic resonance imaging according to claim 1, characterized in that: The low-resolution image and the high-resolution image in step 1 are obtained from multiple scans of the same target object at the same position; the data of each data pair are obtained from different target objects.

3. The reconstruction method for breaking through the quality limit of low-field magnetic resonance imaging according to claim 1, characterized in that: In step 2, ultra-high resolution refers to a resolution that cannot be obtained by a low-field magnetic resonance scanner within an acceptable scanning time, including acquisition resolution and spatial resolution; the high-field degraded image is obtained from the ultra-high resolution image by downsampling, adding Gaussian noise, and adding Rice noise.

4. The reconstruction method for breaking through the quality limit of low-field magnetic resonance imaging according to claim 1, characterized in that: The basic super-resolution model and the cross-field super-resolution model are both 2x super-resolution models.

5. The reconstruction method for breaking through the quality limit of low-field magnetic resonance imaging according to claim 1, characterized in that: The minimum number of excitations in step 4 refers to the number of excitations that can be used to obtain the low-resolution image with the same quality as the high-resolution image through the basic super-resolution model.

6. The reconstruction method for breaking through the quality limit of low-field magnetic resonance imaging according to claim 1, characterized in that: In step 6, the first level of the enhanced two-level cross-field super-resolution model is a basic super-resolution model, and the second level is a cross-field super-resolution model; the input of the basic super-resolution model is a low-field, low-resolution image with the least number of excitations, and the output is used as an intermediate feature of the enhanced two-level cross-field super-resolution model; the input of the cross-field super-resolution model is an intermediate feature, and the output is the output of the enhanced two-level cross-field super-resolution model.

7. The reconstruction method for breaking through the quality limit of low-field magnetic resonance imaging according to claim 1, characterized in that: The image reconstructed by the enhanced two-stage cross-field super-resolution model in step 7 has a quality that exceeds the low-field magnetic field strength constraint and has an effect close to that of the high-field, ultra-high resolution image, including contrast, signal-to-noise ratio, and spatial resolution.

Images