A fully automatic reconstruction method for the human pyramidal tract based on multi-dimensional cross-modal image fusion technology

Abstract

The present application relates to a fully automatic reconstruction method and device for the pyramidal tract of the human brain using multimodal image fusion technology, which can perform multimodal fusion registration on DTI images and MRI image sequences of the human brain, and perform fully automatic three-dimensional reconstruction of the pyramidal tract of the human brain based on the registration.

Description

[0001] The present invention relates to the field of medical image processing, and particularly to a method and device for fully automatic reconstruction of the pyramidal tract of the human brain. Background Art

[0002] Since magnetic resonance diffusion tensor imaging (DTI) can non-invasively display the morphology of nerve fibers, it has been widely applied in medical clinics. It has played a crucial role in the diagnosis and treatment of diseases in the brain, spinal cord, kidneys, muscles, peripheral nerves, etc., especially in the application of brain nerve diseases. As another common magnetic resonance image, MRI is also widely used in the diagnosis of human brain diseases. T1, T1C, and DTI in MRI can only provide two-dimensional image pictures and cannot provide three-dimensional intracranial anatomical structures and lesion models. Therefore, a three-dimensional model of the intracranial pyramidal tract enables clinicians to clearly understand the situation of nerve fiber bundles visually from the three-dimensional model, especially the information on the abnormal position and orientation of the pyramidal tract caused by tumor mass effect and lesions clinically, as well as the spatial relationship between the pyramidal tract, tumor, blood vessels, and surrounding tissues at the lesion location, providing a preoperative solid disease model for clinicians during surgery and further designing and planning surgical plans and evaluating surgical risks on the solid model, which has become a widespread research hotspot.

[0003] Currently, the three-dimensional reconstruction techniques of the human brain pyramidal tract are generally divided into three categories: One solution is to reconstruct the pyramidal tract by manually placing seed points on the image obtained after tensor calculation of a single DTI image; one solution requires the use of commercial software such as mimics and is highly dependent on the manual settings and operations of medical staff; another solution requires a large number of device parameters to convert the gradient table to achieve the registration effect. Summary of the Invention

[0004] The present invention relates to a method for fully automatic reconstruction of the human brain pyramidal tract based on multi-dimensional cross-modal image fusion technology, including the following steps:

[0005] Step 1, data quality inspection, including inspection of the quality parameters of the reference image sequence and the DTI sequence;

[0006] Step 2, data preprocessing, including data format conversion, data correction, and fitting a tensor model;

[0007] Step 3, segmenting and extracting the human brain fiber bundle based on the tensor model fitted in Step 2;

[0008] Step 4, saving the human brain fiber bundle data obtained by segmentation and extraction as a model file;

[0009] Step 5, performing fusion registration with the reference sequence to obtain a rotation matrix;

[0010] Step 6: Based on the rotation matrix obtained in Step 5, rotate the model file in Step 4 to obtain a human brain pyramidal tract model that more conforms to the actual situation.

[0011] Beneficial effects: The above-mentioned fully automatic reconstruction method for the human brain pyramidal tract proposed by the present invention can avoid the problems brought by the three commonly used schemes in the prior art. There is no need to manually place seed points during the reconstruction process after selecting the image sequence; it avoids excessive dependence on specific commercial software; it does not require a large number of device parameters, and adopts the method of first reconstructing and then registering the model to the reference sequence to ensure the registration accuracy. Description of the Drawings

[0012] Figure 1 is the human brain MRI image obtained by the present invention;

[0013] Figure 2 is the human brain DTI image obtained by the present invention;

[0014] Figure 3 is the image of the reconstructed human brain pyramidal tract of the present inventor. Detailed Embodiments

[0015] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

[0016] Diffusion tensor imaging technology is a new nuclear magnetic resonance technology developed on the basis of diffusion weighted imaging technology, and it realizes imaging according to the diffusion movement of water molecules in the brain. In clinical medical applications, anisotropic information in biological tissues can be extracted according to the movement of water molecules, and then the nerve fibers in the tissues can be reconstructed. Nerve fibers mainly originate from pyramidal cells in the cerebral cortex, so they are called pyramidal tracts. The reconstruction results of the pyramidal tracts are useful for obtaining information on the structural and functional connections in the brain, and at the same time providing effective information for doctors' diagnosis.

[0017] The present application proposes a fully automatic reconstruction method for the human brain pyramidal tract based on multi-dimensional cross-modal image fusion technology, which mainly includes six steps: data quality inspection, data preprocessing, fiber bundle segmentation and extraction, model generation, fusion registration, and model rotation. The reconstruction of the pyramidal tract requires the use of a reference sequence (T1 or T1C) and a DTI sequence.

[0018] I. Data Quality Inspection

[0019] Check the imaging scan parameters, including parameters such as TR, voxel size, and scan duration, to ensure the suitability and uniformity of the parameters. The reference sequence (TI) should meet the requirement of isotropic voxel with a resolution less than or equal to 1 mm, and the diffusion magnetic resonance (DTI) should meet the requirement of isotropic voxel with a resolution of 2 mm, a b-value of 800 - 1000, and no less than 30 diffusion gradient directions, with more than 64 recommended.

[0020] II. Data preprocessing

[0021] 1. Data format conversion:

[0022] Convert the reference sequence and DTI raw sequence images from DICOM format to NIFIT format and the corresponding file organization form.

[0023] 2. Generate b-value and gradient table:

[0024] When converting DTI to NIFIT format, bvec (gradient table) and bval (b-value) files will be generated simultaneously.

[0025] 3. Head motion and eddy current correction:

[0026] During the data scanning process, head movement will introduce noise to the data, and the eddy current effect during the imaging process will also cause image errors. Therefore, head motion and eddy current correction of the data are required to reduce errors. That is, remove the image artifacts caused by the eddy current effect from the magnetic resonance coil, and at the same time remove the motion artifacts through the registration between DWI images.

[0027] 4. b0 image extraction:

[0028] Extract the b0 image of DTI.

[0029] 5. Skull stripping:

[0030] Perform scalp and bone stripping operations on the b0 image to generate a mask.

[0031] 6. Fitting the tensor model:

[0032] Before performing tensor fitting, first use the mask generated from the b0 image in the previous step to perform scalp and skull stripping on the corrected DTI image. Then, based on the DTI image after removing the scalp and skull, the gradient table (bevc) and b-value (bval) files, perform tensor fitting to calculate the tensor model based on each voxel.

[0033] 7. Generate intermediate files:

[0034] Generate intermediate files based on the tensor model calculated in step 6 for subsequent tractography and segmentation of the pyramidal tract.

[0035] III. Fiber bundle segmentation and extraction:

[0036] 1. Whole-brain fiber tracking:

[0037] In the prior art, many tracking algorithms for the corticospinal tract of the human brain have been proposed, including the fiber tract continuous tracking algorithm, the tensor line technique, the vector normalization selection tracking algorithm, the probabilistic tracking method, the high-order tensor model, and so on. The fiber tract continuous tracking algorithm is easily affected by noise, the accuracy of the tensor line technique needs to be improved, and the vector normalization selection tracking algorithm and the probabilistic tracking method cannot solve the problems of fiber crossing and fiber bifurcation of the human brain corticospinal tract. In the present invention, the high-order tensor model technology is adopted to perform whole-brain fiber tract tracking. By tracking the fibers of multiple medical samples, multiple fibers are obtained, and after clustering them, a corticospinal tract template is obtained. Subsequently, based on the corticospinal tract template obtained by clustering, the corticospinal tract is automatically segmented to ensure the segmentation efficiency and segmentation accuracy.

[0038] Specifically, 40,000 fibers are respectively tracked and screened from n = 20 medical sample images for the generation of the corticospinal tract template. Since the fibers tracked from the original medical sample images are different in length, morphology, etc., and there are a small number of abnormal fibers that will affect the clustering effect of the fiber bundle, it is necessary to calculate the similarity between each pair of fibers and eliminate the abnormal fibers.

[0039] For example, the set of tracked fibers is {a 1 , a 2 , …, a n}, where n is a positive integer. Calculate the similarity between each pair of fibers to obtain a similarity histogram, and determine whether the similarity histogram conforms to the normal distribution. If it conforms to the normal distribution, determine the similarity threshold t according to the significance level of the similarity histogram. For the fiber a 1 , first calculate the similarity f 12 between this fiber and all other fibers, f 13 , …, f 1n . This similarity can be calculated using the average nearest point distance between two fibers. For any fiber a ij for which there exists f i greater than the similarity threshold, it is considered that a i is an abnormal fiber and does not participate in the clustering.

[0040] After eliminating the abnormal fibers, the k-means algorithm or other common clustering algorithms can be used for fiber clustering to generate a corticospinal tract template.

[0041] 2. Fiber segmentation:

[0042] Based on the corticospinal tract template, the current image is automatically segmented for the human brain corticospinal tract, and the actually required CST corticospinal tract is extracted.

[0043] The above-mentioned brain fiber tracking and segmentation processes are based on intermediate files.

[0044] IV. Model Generation:

[0045] 1. Model Saving

[0046] Export the segmented pyramidal tract values as a tck model.

[0047] V. Fusion Registration:

[0048] To further restore the authenticity of the reconstructed human brain pyramidal tract model, a fusion registration operation is performed here, which needs to be fused and registered with the reference sequence (T1).

[0049] The rotation matrix obtained from the fusion registration:

[0050] Fuse and register the extracted b0 image with the reference sequence, and at the same time obtain the matrix rotation matrix. During the registration process, the specific steps are as follows:

[0051] Perform scalp and skull removal on the reference sequence image T1, register it with the b0 image after scalp and skull removal, and the two registered images are used as two images to be registered A(x,y) and B(x,y) respectively. Fuse the registered images to obtain the fused image Ffusion(x,y).

[0052] Furthermore, the fusion process includes:

[0053] (1) The reference image after scalp and skull removal after registration is A(x,y), and the b0 image after scalp and skull removal after registration is B(x,y);

[0054] (2) Calculate the common feature C of A(x,y) and B(x,y), calculate the correlation degrees of the common feature C with A(x,y) and B(x,y) respectively, denoted as Re(C,A(x,y)) and Re(C,B(x,y)), and the value range is (0,1);

[0055] (3) Generate the fused image Ffusion(x,y) of the DTI image reference image based on the common feature C, B(x,y) and the reference sequence image T1. The formula is as follows:

[0056] Ffusion(x,y)=aT1+bC+zB(x,y)

[0057] Where a, b, and z are weight values respectively, b = Re(C,A(x,y)), z = Re(C,B(x,y)).

[0058] Obtain the matrix rotation matrix based on the above fusion registration results.

[0059] VI. Model Rotation:

[0060] By using the model file generated in the fourth step and substituting the matrix rotation matrix obtained in the previous step, the rotated human brain pyramidal tract model can be obtained.

[0061] Through a fully automatic reconstruction method for the human brain pyramidal tract based on multi-dimensional cross-modal image fusion technology provided by the present invention, a three-dimensional model of the intracranial pyramidal tract is obtained, enabling clinicians to clearly understand the situation of nerve fiber bundles visually from the three-dimensional model, especially the information on the abnormal position and orientation of the pyramidal tract caused by tumor mass effect and lesions clinically, as well as the spatial relationship between the pyramidal tract, tumor, blood vessels and surrounding tissues at the lesion location. This provides preoperative information for clinicians and further enables the design and planning of surgical plans and assessment of surgical risks on a physical model. Three-dimensional reconstruction of the pyramidal tract in the human brain fiber bundle and the display of the relationship between the pyramidal tract, tumor and blood vessels through a three-dimensional solid model are of great significance for guiding surgeons in surgical planning.

Claims

1. A method for fully automatic reconstruction of human brain pyramidal beams based on multi-dimensional cross-modal image fusion technology. It is characterized in that The steps include: Step 1: Data quality check, including quality parameter check of reference image sequence and DTI sequence; Step 2: Data preprocessing, including data format conversion, data correction, and fitting tensor models, further includes: Step 2.1, data format conversion, converting the reference image sequence and DTI original sequence images from DICOM format to NIFIT format and the corresponding file organization form; Step 2.2, generating a gradient table and b-value file while converting the data format; Step 2.3: Perform head motion and eddy current correction to remove image artifacts caused by the eddy current effect of the magnetic resonance coil, and remove motion artifacts by DWI image registration; Step 2.4, extract the b0 image of the DTI image, perform scalp and skull removal on the b0 image to obtain a mask; Step 2.5, performing scalp and skull removal operation on the DTI processed in step 2.3 based on the mask, and performing tensor fitting calculation in combination with the gradient table and the b-value file; Step 2.6, save the fitted tensor model as an intermediate file for subsequent pyramidal beam segmentation extraction; Step 3: Segmenting and extracting human brain fiber bundles based on the fitted tensor model in step 2, further comprising: Step 3.1: whole-brain fiber tracking, fiber removal, screening of representative fibers, and clustering to generate fiber bundle templates; Step 3.2, based on the pyramidal beam template, automatically segment the pyramidal beam of the human brain in the current image and extract the CST pyramidal beam actually needed; Step 4, saving the human brain fiber bundle data obtained by segmentation and extraction as a model file; Step 5: Fuse and register the b0 image with the reference sequence to obtain a rotation matrix; Step 6: Based on the rotation matrix obtained in step 5, the model file in step 4 is rotated to obtain a human brain pyramidal tract model that is more in line with the actual situation.

2. The method according to claim 1, wherein the data quality check further comprises: include: The reference image sequence must have a resolution less than or equal to 1 mm isovoxel, and the DTI sequence must have a resolution of 2 mm isovoxel.

3. The method according to claim 1, wherein step 3.1 further comprises: include: Step 3.1.1, using the high-order tensor model technology, perform the tracking of the whole-brain fiber bundles to obtain a fiber set {a 1 , a 2 , …, a n}, where n is a positive integer; Step 3.1.2, for fiber a i Calculate the similarity f between this fiber and all other fibers ij , where both i and j are positive integers less than or equal to n and are different from each other. For any fiber a where f ij is greater than the similarity threshold i , consider ai as an abnormal fiber and eliminate it; Step 3.1.3, clustering the fiber set after removing abnormal fibers.

4. The method according to claim 1, wherein the step 5 further comprises: include: Step 5.1, perform scalp and skull removal operation on the reference sequence image, and align it with the b0 image after the scalp and skull are removed. The two images after the alignment are used as two images to be fused, A(x, y) and B(x, y), respectively. Among them, the reference image after the scalp and skull are removed after the alignment is A(x, y), and the b0 image after the scalp and skull are removed after the alignment is B(x, y); Step 5.2, calculate the common feature C of A(x,y) and B(x,y), and calculate the correlation between the common feature C and A(x,y) and B(x,y), respectively, denoted as Re(C,A(x,y)) and Re(C,B(x,y)), with a value range of (0,1); Step 5.3, generate the fused image Ffusion(x,y) of the DTI image reference image according to the common feature C, B(x,y), and the reference sequence image T1. The formula is as follows: Ffusion(x,y)=aT1+bC+zB(x,y) where a, b, and z are weight values respectively, b = Re(C,A(x,y)), z = Re(C,B(x,y)); Step 5.5, based on the above fusion result, obtain the rotation matrix registered to the coordinate system of the reference image sequence.

Images