阅读说明：本技术 一种基于解剖先验数据的磁共振成像图像重建方法 (Magnetic resonance imaging image reconstruction method based on anatomical prior data ) 是由 刘懿龙 朱瑞星 于 2021-08-17 设计创作，主要内容包括：本发明涉及磁共振成像技术领域,尤其涉及一种基于解剖先验数据的磁共振成像图像重建方法,包括：获取训练数据,并将训练数据按解剖部位进行分类,针对每个解剖部位重建一个子成像模型；在实际扫描时,通过定位扫描或导航信号识别待成像的解剖部位,根据所识别的解剖部位选择对应的子成像模型,并使用所选的子成像模型来重建图像。本发明的有益效果在于：针对各个解剖部位进行图像重建,获得的磁共振图像具有更好的图像质量,比如噪声水平更低、伪影更少等；此外,对于相同的图像质量需要更少的磁共振成像数据,因此所需的扫描时间更少。(The invention relates to the technical field of magnetic resonance imaging, in particular to a magnetic resonance imaging image reconstruction method based on anatomical prior data, which comprises the following steps: acquiring training data, classifying the training data according to anatomical parts, and reconstructing a sub-imaging model for each anatomical part; during actual scanning, an anatomical part to be imaged is identified through a positioning scanning or navigation signal, a corresponding sub-imaging model is selected according to the identified anatomical part, and an image is reconstructed by using the selected sub-imaging model. The invention has the beneficial effects that: performing image reconstruction on each anatomical part, wherein the obtained magnetic resonance image has better image quality, such as lower noise level, fewer artifacts and the like; furthermore, less magnetic resonance imaging data is required for the same image quality, and thus less scanning time is required.)

1. A magnetic resonance imaging image reconstruction method based on anatomical prior data is characterized in that a plurality of corresponding sub-imaging models are formed aiming at each anatomical part of a target object in advance before the magnetic resonance image reconstruction is carried out, and the specific training process comprises the following steps:

step A1, collecting training data of the target object, and dividing the training data into a plurality of training samples according to each anatomical part of the target object;

step A2, respectively modeling a plurality of anatomical parts by using a plurality of training samples, so as to form a corresponding sub-imaging model for each anatomical part and a corresponding reference image for each anatomical part;

the method further comprises a magnetic resonance image reconstruction process by using the plurality of sub-imaging models, and specifically comprises the following steps:

step B1, performing a main scan on the target object to acquire magnetic resonance data, performing at least one acquisition of positioning information, and comparing the acquired positioning information with the reference image to identify the currently scanned anatomical region;

step B2, according to the identified anatomical part, calling the corresponding sub-imaging model, and reconstructing the magnetic resonance data into an anatomical image corresponding to the anatomical part by using the sub-imaging model;

and step B3, combining the anatomical images corresponding to the anatomical parts to form a whole image of the target object.

2. The method of claim 1, wherein each of the training samples comprises a pair of a training image obtained by acquisition from a magnetic resonance imaging system or a computer simulation system of magnetic resonance imaging and k-space training data, wherein the k-space training data is used as an input of the sub-imaging model and the training image is used as an output of the sub-imaging model.

3. The method of reconstructing a magnetic resonance image according to claim 2, wherein in the step a1, the method of acquiring the training data includes:

step A11, acquiring original k-space training data in a full sampling mode;

step A12, forming a first training image from the raw k-space training data;

step A13, performing affine transformation on the first training image to obtain a second training image;

step A14, acquiring sparse sampling k-space data from the second training image in a down-sampling mode to serve as the k-space training data finally input into the sub-imaging model, and taking the second training image as the output of the sub-imaging model.

4. The magnetic resonance imaging image reconstruction method according to claim 3, characterized in that the step A2 specifically includes:

step A21, performing image registration and averaging on the second training image to obtain the reference image;

step a22, using the sparsely sampled k-space data obtained in step a14 as the k-space training data, using the second training image as the training image, and forming one training sample for each pair of the k-space training data and the training image, so as to train the corresponding sub-imaging model.

5. A method of magnetic resonance imaging image reconstruction of claim 2, characterized in that the k-space training data in the training samples are either fully sampled or under-sampled by post-processing.

6. The magnetic resonance imaging image reconstruction method according to claim 5, wherein the k-space training data is two-dimensional data or three-dimensional data;

when the k-space training data is the two-dimensional data, each of the anatomical sites includes at least one imaging slice;

when the k-space training data is the three-dimensional data, each of the anatomical regions includes at least one imaging patch.

7. The method of claim 1, wherein the sampling trajectory of the training data comprises a cartesian sampling trajectory or a non-cartesian sampling trajectory.

8. The magnetic resonance imaging image reconstruction method according to claim 1, characterized in that in step B1, the localization information is acquired by performing at least one localization scan before performing the main scan, the procedure of the localization scan specifically including:

executing at least one positioning scan to obtain a positioning image, and carrying out image registration on the positioning image and the reference image to obtain a first transformation matrix;

acquiring a scanning position of the positioning scanning from a magnetic resonance imaging system to obtain a second transformation matrix of the positioning scanning relative to the scanning position of the main scanning;

and multiplying the first transformation matrix and the second transformation matrix to obtain a third transformation matrix, and obtaining the anatomical part corresponding to the main scanning through the third transformation matrix.

9. The method of reconstructing an mri image according to claim 1, wherein in step B1, a navigation signal is acquired during the execution of the main scan, a navigation image is obtained by performing an inverse fourier transform on the navigation signal, and the navigation image and the reference image are image-matched to obtain a third transformation matrix, and the anatomical region corresponding to the main scan is obtained by the third transformation matrix.

10. The magnetic resonance image reconstruction method according to claim 1, characterized in that the network architecture adopted by the sub-imaging model comprises a convolutional neural network and/or a generative countermeasure network and/or a self-encoder.

11. The method of claim 1, wherein the training data includes magnetic resonance data of various lesion situations during the training of the sub-imaging model.

Technical Field

The invention relates to the technical field of Magnetic Resonance Imaging (MRI), in particular to a Magnetic Resonance image reconstruction method based on anatomical prior data.

Background

With the rapid development of the medical industry, the magnetic resonance imaging technology is widely applied to clinical diagnosis of various diseases. The raw data acquired by the magnetic resonance imaging system is data in a frequency domain space (i.e., k-space data), and the raw data is converted into a magnetic resonance image by a series of signal processing methods (i.e., image reconstruction methods). The image reconstruction method determines the quality of the magnetic resonance image to a certain extent, and is particularly important for obtaining high-quality images when only partial k-space data are acquired in order to save scanning time.

Generally, image reconstruction can be performed in conjunction with a priori knowledge. Existing image reconstruction methods typically use a priori knowledge from such sources as receive coil sensitivities (e.g., SENSE, GRAPPA, etc. parallel imaging image reconstruction algorithms) and transform domain data sparsity (e.g., compressed sensing methods), but prior art methods do not exist for separately reconstructing images based on the anatomical location of the scan.

Disclosure of Invention

In view of the above problems in the prior art, a magnetic resonance imaging image reconstruction method is provided.

The specific technical scheme is as follows:

the invention comprises a magnetic resonance imaging image reconstruction method based on anatomical prior data, wherein a plurality of corresponding sub-imaging models are formed aiming at each anatomical part of a target object in advance before the magnetic resonance image reconstruction is carried out, and the specific training process comprises the following steps:

step A1, collecting training data of the target object, and dividing the training data into a plurality of training samples according to each anatomical part of the target object;

step A2, respectively modeling a plurality of anatomical parts by using a plurality of training samples, so as to form a corresponding sub-imaging model for each anatomical part and a corresponding reference image for each anatomical part;

the method further comprises a magnetic resonance image reconstruction process by using the plurality of sub-imaging models, and specifically comprises the following steps:

step B1, performing a main scan on the target object to acquire magnetic resonance data, performing at least one acquisition of positioning information, and comparing the acquired positioning information with the reference image to identify the currently scanned anatomical region;

step B2, according to the identified anatomical part, calling the corresponding sub-imaging model, and reconstructing the magnetic resonance data into an anatomical image corresponding to the anatomical part by using the sub-imaging model;

and step B3, combining the anatomical images corresponding to the anatomical parts to form a whole image of the target object.

Optionally, each of the training samples includes a pair of training images acquired by acquisition from a magnetic resonance imaging system or acquired from a computer simulation system of magnetic resonance imaging and k-space training data, wherein the k-space training data is used as an input of the sub-imaging model, and the training images are used as an output of the sub-imaging model.

Optionally, in the step a1, the method for acquiring the training data includes:

step A11, acquiring original k-space training data in a full sampling mode;

step A12, forming a first training image from the raw k-space training data;

step A13, performing affine transformation on the first training image to obtain a second training image;

step A14, acquiring sparse sampling k-space data from the second training image in a down-sampling mode to serve as the k-space training data finally input into the sub-imaging model, and taking the second training image as the output of the sub-imaging model.

Optionally, the step a2 specifically includes:

step A21, performing image registration and averaging on the second training image to obtain the reference image;

step a22, using the sparsely sampled k-space data obtained in step a14 as the k-space training data, using the second training image as the training image, and forming one training sample for each pair of the k-space training data and the training image, so as to train the corresponding sub-imaging model.

Optionally, the k-space training data in the training samples are fully sampled or under-sampled by post-processing.

Optionally, the k-space training data is two-dimensional data or three-dimensional data;

when the k-space training data is the two-dimensional data, each of the anatomical sites includes at least one imaging slice;

when the k-space training data is the three-dimensional data, each of the anatomical regions includes at least one imaging patch.

Optionally, the sampling trajectory of the training data includes a cartesian sampling trajectory or a non-cartesian sampling trajectory.

Optionally, in step B1, the positioning information is acquired by performing at least one positioning scan before the main scan is performed, where the positioning scan specifically includes:

executing at least one positioning scan to obtain a positioning image, and carrying out image registration on the positioning image and the reference image to obtain a first transformation matrix;

acquiring a scanning position of the positioning scanning from a magnetic resonance imaging system to obtain a second transformation matrix of the positioning scanning relative to the scanning position of the main scanning;

and multiplying the first transformation matrix and the second transformation matrix to obtain a third transformation matrix, and obtaining the anatomical part corresponding to the main scanning through the third transformation matrix.

Optionally, in step B1, a navigation signal is acquired during the execution of the main scan, a navigation image is obtained from the navigation signal through inverse fourier transform, an image matching is performed between the navigation image and the reference image to obtain a third transformation matrix, and the anatomical region corresponding to the main scan is obtained through the third transformation matrix.

Optionally, the network architecture adopted by the sub-imaging model includes a convolutional neural network and/or a generative countermeasure network and/or a self-encoder.

Optionally, in the training process of the sub-imaging model, the training data includes magnetic resonance data of various lesion situations.

The technical scheme of the invention has the following advantages or beneficial effects: the method comprises the steps of training sub-imaging models aiming at all anatomical parts by using the magnetic resonance data of all anatomical positions, and respectively reconstructing images of all the anatomical parts by using the sub-imaging models in the subsequent scanning process so as to obtain better magnetic resonance image quality; meanwhile, in the subsequent scanning process, the original magnetic resonance imaging data with reduced sampling or low signal-to-noise ratio can be used for reconstructing a magnetic resonance image with better quality, so that the scanning time is effectively shortened.

Drawings

Embodiments of the present invention will now be described more fully hereinafter with reference to the accompanying drawings. The drawings are, however, to be regarded as illustrative and explanatory only and are not restrictive of the scope of the invention.

FIG. 1 is a schematic diagram of training a sub-imaging model for an anatomical region in an embodiment of the present invention;

FIG. 2 is a schematic flow chart of image reconstruction using sparsely sampled k-space data according to an embodiment of the present invention;

FIG. 3 is an overall flowchart of a model training process and an image reconstruction process in an embodiment of the present invention;

FIG. 4 is a flow chart illustrating the training of sub-imaging models for various anatomical regions in an embodiment of the present invention;

FIG. 5 is a schematic diagram illustrating the identification of an anatomical region using scout scan according to an embodiment of the present invention;

fig. 6 is a schematic diagram illustrating the principle of identifying an anatomical region by using a navigation signal according to an embodiment of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

It should be noted that the embodiments and features of the embodiments may be combined with each other without conflict.

The invention is further described with reference to the following drawings and specific examples, which are not intended to be limiting.

The invention includes a magnetic resonance imaging image reconstruction method based on anatomical prior data, before magnetic resonance image reconstruction, a plurality of corresponding sub-imaging models are formed in advance for each anatomical part of a target object, as shown in fig. 1 and 4, and a specific training process includes:

step A1, collecting training data of a target object, and dividing the training data into a plurality of training samples according to each anatomical part of the target object;

step a2, using a plurality of training samples to respectively model a plurality of anatomical regions to form a corresponding sub-imaging model for each anatomical region and a corresponding reference image for each anatomical region.

In particular, human anatomy structures still have great similarity among different individuals, and most clinical magnetic resonance scans image a to-be-imaged part in a preset standard anatomical orientation, so that images obtained by magnetic resonance imaging also have great similarity. For example, for magnetic resonance imaging of the brain, one common preset anatomical orientation is to take a cross-sectional view and make the imaging plane parallel to the Anterior Commissure (AC) -Posterior Commissure (PC) line (AC-PC) to acquire a multi-slice image covering the entire brain region from the cranial apex to the cranial base. Thus, with anatomical prior knowledge and the scanned anatomical region, the image content at a particular region to be imaged is largely predictable. The anatomical prior data can be extracted from a huge magnetic resonance image data set and used for training sub-imaging models for all anatomical parts, the sub-imaging models are used for reconstructing images of all anatomical parts in the subsequent scanning process so as to obtain better magnetic resonance image quality, original magnetic resonance imaging data with reduced sampling or low signal-to-noise ratio can be used for reconstructing a magnetic resonance image with better quality in the subsequent scanning process, and meanwhile, the scanning time can be shortened.

The invention aims to improve the quality of image reconstruction of magnetic resonance imaging by combining anatomical prior knowledge, refine magnetic resonance data of various anatomical orientations through a deep learning method (such as a convolutional neural network, a generation countermeasure network and the like), and embody the refined magnetic resonance data in a reconstruction model (namely a sub-imaging model) obtained by training. More specifically, a training data set comprising magnetic resonance images and corresponding k-space data may be acquired directly using a magnetic resonance imaging system or using computer simulation and divided into a plurality of training samples according to anatomical regions of a magnetic resonance imaging scan, for each anatomical region at least one sub-imaging model being trained to predict images from the k-space data. As shown in fig. 1, images are first extracted from specific imaging slices, and an image reconstruction model is trained from sparse or low signal-to-noise ratio k-space data obtained from these imaging slices. Different models (models 1-N) will be trained for different sub-imaging sites (1-N). This concept can be implemented by Convolutional Neural Network (CNN), a challenge generation network (GAN), a Collaborative challenge generation network (CollaGAN), or an auto encoder (AutoEncoder) method.

The method further includes performing a magnetic resonance image reconstruction process using the plurality of sub-imaging models, as shown in fig. 2, specifically including:

step B1, performing main scanning to the target object to obtain magnetic resonance data, performing at least one acquisition of positioning information, and comparing the acquired positioning information with a reference image to identify the currently scanned anatomical part;

b2, calling a corresponding sub-imaging model according to the identified anatomical part, and reconstructing the magnetic resonance data into an anatomical image of the corresponding anatomical part by using the sub-imaging model;

step B3, merging the anatomical images corresponding to the respective anatomical regions to form an overall image of the target object.

Specifically, when scanning the target object, at least one positioning information acquisition is performed except for the main scanning, and in the image reconstruction stage, the image imaging layer/sub-imaging part is firstly identified from the positioning scanning, and then a corresponding reconstruction model is selected. This strategy is expected to enable more efficient image reconstruction from potentially less or/and noisier k-space data. The positioning information is matched with the reference image, the anatomical part corresponding to the main scanning can be identified, the sub-imaging model corresponding to the anatomical part is called for image reconstruction, or the main scanning data is further divided into sub-imaging parts, a reconstruction model is obtained by training each sub-imaging part or the adjacent layer of the sub-imaging part, the main scanning data (including multi-layer two-dimensional data, three-dimensional volume data, two-dimensional dynamic data, three-dimensional dynamic volume data and the like) is reconstructed on each sub-imaging part, the anatomical images generated by each sub-imaging part are merged, stored, transmitted or displayed, and finally an integral magnetic resonance image aiming at the target object is formed; furthermore, less magnetic resonance imaging data is required for the same image quality, and thus less scanning time is required.

As an alternative embodiment, each training sample comprises a pair of training images acquired by acquisition from the magnetic resonance imaging system or acquired from the magnetic resonance image simulation system and k-space training data, wherein the k-space training data is used as an input of the sub-imaging model, and the training images are used as an output of the sub-imaging model.

In the model training phase, training data for a specific anatomical region, such as a cross-section of the brain, is first prepared. The training data consists of k space training data and training images of multiple pairs of magnetic resonance imaging, then the training data is divided into a plurality of training samples according to anatomical parts, and a corresponding training sample is used for training and reconstructing a sub-imaging model aiming at each anatomical part; these sub-imaging models can each predict a magnetic resonance image from k-space data of a given magnetic resonance imaging. As shown in fig. 4, at least one set of reference images is generated during the division of the training data, so that the actual scanning process can be located by comparing the coordinates of the reference images. And storing the trained sub-imaging model and the reference image of the anatomical part for subsequent scanning. When a target object to be scanned is placed in a magnetic resonance imaging system for imaging, a scout scan or navigator signal needs to be acquired in addition to the main scan to identify the anatomical region of the main scan data.

As an alternative embodiment, as shown in fig. 4, in step a1, the method for acquiring training data includes:

step A11, acquiring original k-space training data in a full sampling mode;

step A12, forming a first training image from the raw k-space training data;

step A13, performing affine transformation on the first training image to obtain a second training image;

and A14, acquiring sparse sampling k-space data from the second training image in a down-sampling mode to serve as k-space training data of the final input sub-imaging model, and taking the second training image as the output of the sub-imaging model.

FIG. 4 illustrates a training process for reconstructing sub-imaging models in some embodiments. Firstly, original magnetic resonance imaging k-space data or a first training image needs to be converted into a second training image, namely, the orientation, the geometric position, the size and the like of the training images are consistent as much as possible through affine transformation. This is because, for the massive amount of raw k-space data or images that we can obtain, the imaged orientation, geometric position, size, etc. may be different, and we need to make them as consistent as possible through affine transformation (affine transformation), so that the effect of training the model is better. Specifically, the first training image may be registered by an image registration algorithm such as mutual information maximization (mutual information maximization), so as to obtain a second training image with the same orientation, geometric position, and size as much as possible. In some embodiments, the second training images may be averaged to form the reference image.

Specifically, the second training image is divided into N parts according to the anatomical region to be imaged. For example, one anatomical site is formed at intervals from front to back, or top to bottom, or left to right. And undersampling the magnetic resonance imaging k-space data according to a preset k-space sampling track to form sparsely sampled magnetic resonance imaging k-space data. For each anatomical region, the sparsely sampled k-space data and the corresponding magnetic resonance image are combined into a pair of training samples, which are respectively used as input and output of the sub-imaging model, and used for training the sub-imaging model to predict an image by using the k-space data. During the training process, different objective functions may be used, for example, cross-entropy minimization, mean square error minimization, or hinge loss minimization may be used.

Further, as shown in fig. 4, step a2 specifically includes:

step A21, carrying out image registration and averaging on the second training image to obtain a reference image;

step a22, using the sparse sampling k-space data obtained in step a14 as k-space training data, using the second training image as a training image, and forming a training sample for each pair of k-space training data and training image, so as to train a corresponding sub-imaging model.

As an alternative embodiment, the k-space training data in the training samples are fully sampled or undersampled by post-processing. The k-space training data are two-dimensional data or three-dimensional data; when the k-space training data is two-dimensional data, each anatomical part comprises at least one imaging slice; when the k-space training data is three-dimensional data, each anatomical region includes at least one imaging block.

As an alternative embodiment, the sampling trajectory of the training data comprises a cartesian sampling trajectory or a non-cartesian sampling trajectory.

As an alternative embodiment, as shown in fig. 5, in step B1, the positioning information is acquired by performing at least one positioning scan before the main scan is performed, and the positioning scan specifically includes:

executing at least one positioning scanning to obtain a positioning image, and carrying out image registration on the positioning image and a reference image to obtain a first transformation matrix;

acquiring a scanning position of positioning scanning from a magnetic resonance imaging system to obtain a second transformation matrix of the positioning scanning relative to the scanning position of main scanning;

and multiplying the first transformation matrix and the second transformation matrix to obtain a third transformation matrix, and obtaining the anatomical part corresponding to the main scanning through the third transformation matrix.

Fig. 5 illustrates a method of identifying anatomical regions using scout scans in some embodiments. Through an image registration algorithm such as mutual information maximization, a first transformation matrix (affine transformation) from the positioning image to the reference image is obtained. A second transformation matrix (corresponding to a rigid body transformation) of the system default scan orientation (i.e., the scan orientation of the scout scan image) relative to the scan orientation of the main scan is then obtained from the magnetic resonance imaging system, which is typically set manually by an operator or automatically adjusted by some algorithm. The third transformation matrix is obtained by multiplying the second transformation matrix and the first transformation matrix; using the third transformation matrix described above, the coordinates of the main scan may be mapped to the first training image coordinates. Then, the anatomical region corresponding to the main scan can be confirmed by coordinate transformation.

As another alternative, in step B1, a navigation signal is acquired during the execution of the main scan, a navigation image is obtained from the navigation signal through an inverse fourier transform, an image matching is performed between the navigation image and the reference image to obtain a third transformation matrix, and an anatomical region corresponding to the main scan is obtained through the third transformation matrix.

Fig. 6 illustrates a method of identifying an anatomical region using a navigation signal in some embodiments. The navigation image may be obtained by inverse fourier transform using the navigation signal, and the image between the navigation image and the training reference image is matched, similarly to the process using scout scan in the above-described embodiment, to obtain the third transformation matrix.

As an alternative embodiment, in the training process of the sub-imaging model, the training data contains magnetic resonance data of various lesion situations. For example, brain image data with different diseases can be acquired to train the sub-imaging model, so that the sub-imaging model can reconstruct a brain image with better quality aiming at the magnetic resonance data of the brain of a patient with diseases in the actual scanning process.

The embodiment of the invention has the beneficial effects that: the method comprises the steps of training sub-imaging models aiming at all anatomical parts by using the magnetic resonance data of all anatomical positions, and respectively reconstructing images of all the anatomical parts by using the sub-imaging models in the subsequent scanning process so as to obtain better magnetic resonance image quality; meanwhile, in the subsequent scanning process, the original magnetic resonance imaging data with reduced sampling or low signal-to-noise ratio can be used for reconstructing a magnetic resonance image with better quality, so that the scanning time is effectively shortened.

While the invention has been described with reference to a preferred embodiment, it will be understood by those skilled in the art that various changes in form and detail may be made therein without departing from the spirit and scope of the invention.