CN109785405A - A method for generating quasi-CT images using multivariate regression of multiple sets of magnetic resonance images - Google Patents

Abstract

A method of generating quasi-CT images using multivariate regression of multiple sets of magnetic resonance images: respectively acquiring a plurality of groups of magnetic resonance images and a corresponding group of CT images from a plurality of trainees, wherein the magnetic resonance images of different sequencing groups of each trainer are acquired by using different magnetic resonance parameters; the magnetic resonance images in the same sequencing group in the multiple groups of magnetic resonance images of different trainees are obtained by using the same magnetic resonance parameters; aligning each image of each set of magnetic resonance images of the same location of each trainer with a corresponding CT image; generating a mapping function corresponding to the same voxel CT value from the intensity values of the multiple groups of magnetic resonance images of all trainees; a quasi-CT image of the subject is generated from the plurality of sets of magnetic resonance images of the subject. The present invention does not require alignment of the trainer image and the target image, thus avoiding errors caused by the alignment process in the atlas method. The invention can be used for simulation calculation and medical imaging of magnetic resonance guide radiotherapy planning.

Description

A method for generating quasi-CT images using multivariate regression of multiple sets of magnetic resonance images

technical field

The present invention relates to the generation of a CT image. In particular, it relates to a method for generating quasi-CT images using multivariate regression of multiple sets of magnetic resonance images.

Background technique

Computed tomography (CT) and magnetic resonance imaging (MRI) are the two main imaging modalities for obtaining three-dimensional images. Computed tomography imaging is an imaging modality traditionally used to create radiation treatment plans. The CT image can accurately describe the geometry of the target, and the CT value corresponding to the CT image can be directly converted into electron density for calculating the radiation dose distribution in the target. However, CT images do not have good contrast for soft tissue, and CT also exposes the target to additional radiation doses. Magnetic resonance imaging modalities are also widely used due to their superior soft tissue contrast compared to CT images. Magnetic resonance imaging does not contain ionizing radiation and can provide functional information such as the target's metabolism.

At present, magnetic resonance images are mainly used to supplement CT images to obtain more accurate anatomical structure contours and tumor targeting, so it is necessary to align the magnetic resonance images of the target with the corresponding CT images. Since MR images and CT images are usually acquired on different machines, MR images and CT images are not perfectly aligned when they overlap; this is a significant source of tumor targeting errors. MR-guided radiation therapy planning utilizes MR images instead of CT images, so tumor targeting will be more accurate. Since magnetic resonance intensity values cannot be directly converted into electron density, methods are needed to accurately convert magnetic resonance images into images corresponding to electron density values, ie, quasi-CT images, also known as pCT images or derived CT images.

The journal article "A Review of Substitute CT Generation for MRI-only Radiation Therapy" by Edmund and Nyholm, RadiatOncol 12:28 (2017), doi:10.1186/s13014-016-0747-y, reviews the various approaches used to create pCT , including the atlas method and the voxel method. For example, the journal article "An Atlas-based Electron Density Mapping Method for Magnetic Resonance Imaging (MRI)-Alone Treatment Planning and Adaptive MRI-Based Prostate Radiation Therapy" by Dowling et al., Int J RadiatOncolBiol Phys 83,5 (2012), doi:10.1016 /j.ijrobp.2011.11.056, the atlas method utilizes pre-existing atlas images as references to help generate quasi-CT images. In the process of generating the quasi-CT image, the atlas MRI image and the atlas CT image are the reference for generating the quasi-CT image from the new subject MRI image. The atlas MR image needs to be aligned with the target's MR image, and the same alignment transformation is applied in the process of fusing the atlas CT image at the same location into the target's quasi-CT image. However, there are some errors in the alignment process of the atlas images and the alignment process from the atlas MR images to the target MR images. The voxel method utilizes a conversion method from magnetic resonance image intensity values to CT values to create quasi-CT images without the need to align the target and trainer magnetic resonance images.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a method for generating quasi-CT images by using multivariate regression of multiple sets of magnetic resonance images.

The technical solution adopted in the present invention is: a method for generating quasi-CT images by using multivariate regression of multiple sets of magnetic resonance images, comprising the following steps:

1) Obtain multiple sets of magnetic resonance images (MRI) and a corresponding set of CT images from several trainers, respectively. The MRI images of different sorted groups of each trainer are obtained using different MRI parameters; The magnetic resonance images in the same sorting group among the multiple groups of magnetic resonance images are acquired using the same magnetic resonance parameters;

2) aligning each image in each group of magnetic resonance images at the same position of each trainer with the corresponding CT image;

3) generating a mapping function corresponding to the CT value of the same voxel from the intensity values of the multiple groups of magnetic resonance images of all trainers;

4) Generating a quasi-CT image of the target person from multiple sets of magnetic resonance images of the target person.

When the magnetic resonance images of different trainers in the same ranking group in step 1) are acquired using different scanning machines and/or different imaging conditions, the magnetic resonance images should be standardized so that all trainers are in the same ranking The mean intensity values of the magnetic resonance images of the groups were the same.

Step 3) includes:

(3.1) Divide the outline of the human body for each magnetic resonance image and CT image respectively;

(3.2) establishing a region segmentation mask for the magnetic resonance images of the first sorting group;

(3.3) using the obtained region segmentation mask to perform region segmentation on magnetic resonance images and CT images of other sorting groups other than the first sorting group at the same position;

(3.4) extracting the corresponding magnetic resonance image intensity value and CT value from the voxels in each area except the exclusion area;

(3.5) Using the multivariate regression method to determine the mapping function of the multivariate high-order polynomial of each region according to the multiple sets of magnetic resonance image intensity values and the average CT value of the voxels with the same multiple sets of magnetic resonance image intensity values:

where _N is the highest degree _{of polynomial} term, CT(S ₁ , . group MRI intensity values, which are independent variables in the mapping function, _i1 is the index _of S1, _im is the index _of Sm, is the fitting coefficient.

The division of the human body contour described in step (3.1) is to separate the human body structure in the image from the surrounding air.

In step (3.2), when there are only bones or soft tissues in the magnetic resonance images of the first sorting group, the magnetic resonance images are regarded as a region; when the magnetic resonance images of the first sorting group have both bones and soft tissues , segment the magnetic resonance image into a bone region, a soft tissue region, and a mixed tissue region, where the mixed tissue region refers to the uncertain part between the bone and the soft tissue, and the boundaries of the regions constitute a region segmentation mask .

In step (3.2), the areas where the anatomical structures are misaligned due to the daily activities of the organs or insufficient image alignment in the multiple sets of magnetic resonance images and CT images at the same location, and the areas outside the body contour are set as Exclusion zone.

Step 4) includes:

(4.1) Use the same magnetic resonance parameters as the trainer to obtain the magnetic resonance images of each group of the target person;

(4.2) Align multiple sets of magnetic resonance images at the same position of the target with each other;

(4.3) When the MR image of the target person is acquired using a different scanning machine and/or different imaging conditions than the trainer, the MRI image of the target person should be standardized so that the target person is in the same position as all trainers. The mean intensity values of the magnetic resonance images of the same sorting group are the same;

(4.4) Use the region segmentation method described in steps (3.2) and (3.3) in step 3) to perform region segmentation on the magnetic resonance images of the first sorting group of the target person and establish a region segmentation mask, and obtain the same position of the target person Regions of magnetic resonance images of other sorting groups other than the first sorting group, wherein the image of the target person does not set an exclusion zone;

(4.5) Extracting multiple sets of magnetic resonance image intensity values from the voxels in the target person's magnetic resonance image;

(4.6) Pass the intensity values of multiple groups of magnetic resonance images of each voxel of the target person through the mapping function of the corresponding area obtained in step 3) step (3.5) to obtain the CT value of the same voxel of the target person;

(4.7) The set of CT values of all voxels of the target person constitutes a quasi-CT image of the target person.

The method for generating quasi-CT images using multivariate regression of multiple sets of magnetic resonance images of the present invention uses a multivariate function to correspond multiple sets of magnetic resonance images of the target person into quasi-CT images. The accuracy of this method is similar to the best existing results of the atlas method, but it is more convenient and faster than the atlas method. The present invention uses the voxel method to directly convert the magnetic resonance image of the target person into a quasi-CT image. Therefore, this method does not need to align the trainer image with the target image, thus avoiding the error caused by the alignment process in the atlas method. The present invention can be used for analog computation and medical imaging of magnetic resonance guided radiation therapy planning.

Description of drawings

1 is an exemplary flowchart of a method for determining a regression model using sets of magnetic resonance images and CT images of a trainer;

FIG. 2 is an exemplary flowchart of generating a quasi-CT image using sets of magnetic resonance images of a subject.

Detailed ways

The method for generating a quasi-CT image by using multivariate regression of multiple sets of magnetic resonance images of the present invention will be described in detail below with reference to the embodiments and the accompanying drawings.

In describing the illustrative examples, the drawings and specific terminology are used for the purpose of clarity only and not for the purpose of limiting this description.

The method for generating quasi-CT images using multivariate regression of multiple sets of magnetic resonance images of the present invention includes the following steps:

1) Obtain multiple sets of magnetic resonance images (MRI) and a corresponding set of CT images from several trainers, respectively. The MRI images of different sorted groups of each trainer are obtained using different MRI parameters; The magnetic resonance images in the same sorting group among the multiple groups of magnetic resonance images are acquired using the same magnetic resonance parameters;

When the MRI images of different trainers in the same ranking group are acquired using different scanning machines and/or different imaging conditions, the MRI images are normalized so that the MR images of all trainers in the same ranking group The average intensity values of the images are the same. Normalization is to ensure that the acquired magnetic resonance image intensity values are self-consistent for different trainers. A simple way to decide whether normalization is needed is to compare histograms of MRI image intensity values for voxels in the same region across different trainers; if the histograms from different trainers show similar shapes but very different peak and mean values value, you need to standardize. In an exemplary normalization process, the average intensity value of each trainer's MRI images in the same rank group is found, and then a correction factor for each trainer is determined to be multiplied by the corresponding trainer's MRI image in the rank group of this rank group. Magnetic resonance image intensity values of voxels so that the mean intensity value of the magnetic resonance images of the sorted group of all trainees is the same.

2) Align each image in each set of magnetic resonance images at the same location of each trainer with the corresponding CT image. The MRI and CT images are usually acquired on different machines, so the MRI and CT images are not perfectly aligned when they overlap. Therefore, it is necessary to align the magnetic resonance image of each trainer and the CT image of the corresponding position through the image alignment technology, so as to make each voxel of the magnetic resonance image of the trainer and the CT image correspond to each other.

3) Generate a mapping function corresponding to the CT value of the same voxel from the intensity values of multiple sets of magnetic resonance images of all trainers; including:

(3.1) Divide the outline of the human body for each magnetic resonance image and CT image respectively; the dividing the outline of the human body is to separate the human body structure in the image from the surrounding air.

(3.2) Establish a region segmentation mask for the magnetic resonance images of the first sorting group; wherein,

When there are only bones or soft tissues in the magnetic resonance images of the first sorting group, the magnetic resonance images are regarded as a region; when the magnetic resonance images of the first sorting group contain both bones and soft tissues, the magnetic resonance images The image is divided into bone area, soft tissue area, and mixed tissue area. The mixed tissue area refers to the uncertain part between bone and soft tissue, and the boundary of each area constitutes an area segmentation mask; Regions of misalignment of anatomical structures due to the daily activities of organs or inadequate image alignment in the MRI and CT images and regions outside the contours of the human body were set as exclusion zones.

(3.3) using the obtained region segmentation mask to perform region segmentation on magnetic resonance images and CT images of other sorting groups other than the first sorting group at the same position;

(3.4) extracting the corresponding magnetic resonance image intensity value and CT value from the voxels in each area except the exclusion area;

(3.5) Using the multivariate regression method to determine the mapping function of the multivariate high-order polynomial of each region according to the multiple sets of magnetic resonance image intensity values and the average CT value of the voxels with the same multiple sets of magnetic resonance image intensity values:

where _N is the highest degree _{of polynomial} term, CT(S ₁ , . group MRI intensity values, which are independent variables in the mapping function, _i1 is the index _of S1, _im is the index _of Sm, is the fitting coefficient.

4) Generating a quasi-CT image of the target person from multiple sets of magnetic resonance images of the target person. include:

(4.1) Obtain each group of magnetic resonance images of the target person using the same magnetic resonance parameters as those of the trainer in the same ranking group. In a preferred example, the magnetic resonance images of the subject and the trainer are generated by the same magnetic resonance scanner; in other examples, the magnetic resonance images of the trainer and the target may be generated by different magnetic resonance scanners.

(4.2) Align multiple sets of magnetic resonance images at the same position of the target with each other;

(4.3) When the MR image of the target person is acquired using a different scanning machine and/or different imaging conditions than the trainer, the MRI image of the target person should be standardized so that the target person is in the same position as all trainers. The mean intensity values of the magnetic resonance images of the same sorting group are the same;

(4.4) Use the region segmentation method described in steps (3.2) and (3.3) in step 3) to perform region segmentation on the magnetic resonance images of the first sorting group of the target person and establish a region segmentation mask, and obtain the same position of the target person Regions of magnetic resonance images of other sorting groups other than the first sorting group, wherein the image of the target person does not set an exclusion zone;

(4.5) Extracting multiple sets of magnetic resonance image intensity values from the voxels in the target person's magnetic resonance image;

(4.6) Pass the intensity values of multiple groups of magnetic resonance images of each voxel of the target person through the mapping function of the corresponding area obtained in step 3) step (3.5) to obtain the CT value of the same voxel of the target person;

(4.7) The set of CT values of all voxels of the target person constitutes a quasi-CT image of the target person.

A specific example is given below:

1 is an exemplary flowchart of a method of determining a regression model using sets of magnetic resonance images and CT images of a trainer.

At step 110, the training data are two sets of magnetic resonance (referred to as MRI1 and MRI2 sets) images obtained from multiple trainers using the same MRI scanner and corresponding parts obtained from these trainers using the same CT scanner CT images. The MR images in the same ranking group among the multiple sets of MR images from different trainers were acquired using the same MR parameters and imaging conditions. The magnetic resonance images of the different ranking groups of each trainer are acquired using different magnetic resonance parameters, such as magnetic resonance image sequence parameters (T1-weighted, T2-weighted) with different contrast properties. The data for each trainer consisted of two sets of magnetic resonance images and a set of corresponding CT images to serve as ground truth. Since the magnetic resonance images of different trainers in the same ranking group were acquired using the same magnetic resonance parameters and imaging conditions, there was no need to normalize the magnetic resonance images of the trainers.

At step 120, the magnetic resonance image and the CT image are usually acquired on different machines, so the magnetic resonance image and the CT image will not be perfectly aligned when overlapping. Therefore, it is necessary to align the magnetic resonance image of each trainer and the CT image of the corresponding position through the image alignment technology, so as to make each voxel of the magnetic resonance image of the trainer and the CT image correspond to each other.

In step 130, firstly, the outline of the human body is divided for each magnetic resonance image and the CT image, the human body structure in the image is separated from the surrounding air, and then the region is segmented for each image. Region segmentation is first performed for the magnetic resonance images of the first sorted group. When there are only bones or soft tissues in the magnetic resonance images of the first sorting group, the magnetic resonance images are regarded as a region; when the magnetic resonance images of the first sorting group contain both bones and soft tissues, the magnetic resonance images The image is segmented into three regions: bone region, soft tissue region, and mixed tissue region. The skeletal region contains only defined skeletal anatomy, including cortical and spongy (cancellous) skeletal structures. Soft tissue areas only include all defined non-skeletal anatomical structures. The remaining area remaining between the area where the bone area and the soft tissue area are in contact (ie, the indeterminate area) is the mixed tissue area. In addition, in the multiple sets of magnetic resonance images and CT images at the same location, the areas where the anatomical structures are misaligned due to the daily activities of organs or insufficient image alignment, and the areas outside the outline of the human body are set as exclusion zones. Region segmentation can be done manually or using computer-generated trainer profiles. The boundary of each region in the magnetic resonance image of the first sorting group constitutes a region segmentation mask; then the obtained region segmentation mask is used for the magnetic resonance images and CT images of other sorting groups other than the first sorting group at the same position Region segmentation is performed.

At step 140, magnetic resonance image intensity values and CT values of the voxels are extracted from each image of the training data. The data for each voxel is a triplet consisting of the magnetic resonance image intensity value from the MRI1 group, the magnetic resonance image intensity value from the MRI2 group, and the corresponding CT value. Data extracted from voxels in a given region of each trainer are grouped together; however, voxels in exclusion zones are excluded from the numerical extraction in order to ensure a true correlation between MRI image intensity values and CT values .

Step 150 performs multivariate regression on each region, that is, uses the function of the two groups of magnetic resonance image intensity values to fit the corresponding average CT value (that is, the voxels in the same region of the training data are in the two groups of magnetic resonance image intensity value intervals. average value within). The multivariate regression in this example uses the following mapping function for a bivariate high-degree polynomial:

Among them, the highest polynomial term is N=30, CT(S ₁ , S ₂ ) is the mapping function and is the dependent variable, and the magnetic resonance image intensity value S ₁ of the MRI1 group and the magnetic resonance image intensity value S ₂ of the MRI2 group are the independent variables. is the fitting coefficient.

FIG. 2 is an exemplary flowchart of generating a quasi-CT image using sets of magnetic resonance images of a subject.

At step 210, two sets of magnetic resonance images (MRI1 group and MRI2 group) of the target are acquired using the same type of magnetic resonance scanner as the trainer and using the same magnetic resonance parameters and imaging conditions as the trainer's same ranking group. Since the MR images of the target were acquired using the same model of MR scanner and imaging conditions as the trainer, normalization of the MR images of the target was not required.

In step 220, the two sets of magnetic resonance images of the same position of the target are aligned with each other.

In step 230, the region segmentation method described in step 130 is used to perform region segmentation on the magnetic resonance images of the first sorted group of the target person and establish a region segmentation mask, and acquire the region of the magnetic resonance image of the second sorted group of the target person at the same position ; The image of the target does not set an exclusion zone.

In step 240, the magnetic resonance image intensity value of each voxel of the target person is extracted, that is, the magnetic resonance image intensity value from the MRI1 group and the magnetic resonance image intensity value from the MRI2 group.

In step 250, the multivariate function for each region determined in step 150 is applied to the corresponding region of the target person. First use the two sets of magnetic resonance image intensity values of each voxel of the target as an independent variable, and then use the binary high-order polynomial of the region where the voxel is determined in step 150 to obtain the dependent variable as the CT value of this voxel .

In step 260, all the quasi-CT values corresponding to each voxel of the target constitute a quasi-CT image. This image can be used for simulation calculations and medical imaging for MRI-guided radiation therapy planning of the target.

Claims (7)

1. A method for generating quasi-CT images using multivariate regression of multiple sets of magnetic resonance images, comprising the steps of:

1) acquiring a plurality of sets of Magnetic Resonance Images (MRI) and a corresponding set of CT images from a plurality of trainees respectively, wherein the MRI images of different sequencing sets of each trainer are acquired by using different MRI parameters; the magnetic resonance images in the same sequencing group in the multiple groups of magnetic resonance images of different trainees are obtained by using the same magnetic resonance parameters;

2) aligning each image of each set of magnetic resonance images of the same location of each trainer with a corresponding CT image;

3) generating a mapping function corresponding to the same voxel CT value from the intensity values of the multiple groups of magnetic resonance images of all trainees;

4) a quasi-CT image of the subject is generated from the plurality of sets of magnetic resonance images of the subject.

2. The method of claim 1, wherein when the magnetic resonance images of different trainees in the same ordered group are acquired by different scanning machines and/or different imaging conditions in step 1), the magnetic resonance images are normalized so that the mean intensity values of the magnetic resonance images of the same ordered group are the same among all trainees.

3. The method of generating quasi-CT images using multivariate regression of multiple sets of magnetic resonance images as set forth in claim 1, wherein step 3) comprises:

(3.1) dividing the human body outline for each magnetic resonance image and each CT image respectively;

(3.2) creating a region segmentation mask for the magnetic resonance images of the first ordered set;

(3.3) using the obtained region segmentation mask for carrying out region segmentation on the magnetic resonance image and the CT image of other sequencing groups except the first sequencing group at the same position;

(3.4) extracting respective magnetic resonance image intensity values and CT values from voxels in each region outside the exclusion zone;

(3.5) determining a mapping function of a multiple high degree polynomial for each region using a multivariate regression method based on the multiple sets of magnetic resonance image intensity values and the average CT values of voxels having the same multiple sets of magnetic resonance image intensity values:

where N is the polynomial maximum degree, CT (S)₁，…，S_m) For the mapping function being a dependent variable, S₁Is a first set of magnetic resonance image intensity values, being an independent variable in a mapping function, S_mIs the m-th set of magnetic resonance image intensity values, being the independent variable, i, in the mapping function₁Is S₁Index of (1), i_mIs S_mThe index of (a) is,are fitting coefficients.

4. The method of claim 2, wherein the step of dividing the contour of the human body in step (3.1) is to separate the structure of the human body from the surrounding air in the image.

5. The method of claim 3, wherein in the step (3.2), when only bone or soft tissue exists in the first ordered set of magnetic resonance images, the magnetic resonance images are used as a region; when the magnetic resonance image of the first ordering group has bone and soft tissue at the same time, the magnetic resonance image is divided into a bone region, a soft tissue region and a mixed tissue region, the mixed tissue region refers to an uncertain part between the bone and the soft tissue, and the boundary of each region forms a region division mask.

6. The method of claim 3, wherein in the step (3.2), the regions of the multi-set magnetic resonance image and the CT image at the same position where the anatomical structure is not aligned due to the daily movement of the organ or insufficient image alignment, and the regions outside the contour of the human body are set as exclusion regions.

7. The method of generating quasi-CT images using multivariate regression of multiple sets of magnetic resonance images as set forth in claim 1, wherein step 4) comprises:

(4.1) acquiring each group of magnetic resonance images of the target person by using the same magnetic resonance parameters of the same sequencing group of the trainer;

(4.2) aligning a plurality of groups of magnetic resonance images of the same position of the target person with each other;

(4.3) when the magnetic resonance image of the target person is obtained by using a different scanning machine and/or different imaging conditions from those of the trainer, standardizing the magnetic resonance image of the target person to make the average intensity value of the magnetic resonance images of the same sequencing group at the same position of the target person and all the trainees be same;

(4.4) carrying out region segmentation on the magnetic resonance image of the first sequencing group of the target person and establishing a region segmentation mask by adopting the region segmentation mode in the steps (3.2) and (3.3) in the step 3), and acquiring the regions of the magnetic resonance images of other sequencing groups except the first sequencing group at the same position of the target person, wherein the image of the target person is not provided with an exclusion region;

(4.5) extracting a plurality of groups of magnetic resonance image intensity values of voxels in the magnetic resonance image of the target;

(4.6) passing the intensity values of the multiple groups of magnetic resonance images of each voxel of the target through the mapping function of the corresponding region obtained in the step (3) and the step (3.5) of the step 3) to obtain the CT value of the same voxel of the target;

and (4.7) forming the quasi-CT image of the target by the set of CT values of all voxels of the target.