WO2024011647A1

WO2024011647A1 - Deep learning-based thalamus individualized map drawing method guided by group prior knowledge

Info

Publication number: WO2024011647A1
Application number: PCT/CN2022/106161
Authority: WO
Inventors: 樊令仲; 高超宏; 吴霞
Original assignee: 中国科学院自动化研究所
Priority date: 2022-07-12
Filing date: 2022-07-18
Publication date: 2024-01-18
Also published as: CN115082586A; CN115082586B

Abstract

The present invention belongs to the field of brain map drawing using magnetic resonance imaging, and particularly relates to a deep learning-based thalamus individualized map drawing method and system guided by group prior knowledge, and a device, which are aimed at solving the problems in the prior art of the accuracy, robustness and repeatability of a drawn individualized thalamus map being relatively poor. The present method comprises: acquiring thalamus ROIs of individual brains of subjects; acquiring a group prior map of the individual brains of the subjects; eliminating the group prior map from the thalamus ROIs; acquiring an individual feature, inputting same into an individualized classification model, so as to obtain a prediction probability value vector of a voxel in each undefined region, and then generating a map of the undefined region; and combining the group prior map and maps of undefined regions, so as to generate individualized thalamus maps of the subjects. In the present invention, individualized thalamus parcellation and map drawing are guided by using high-credibility group prior knowledge, such that the accuracy, robustness and repeatability of drawn individualized thalamus maps are improved.

Description

Group prior-guided deep learning-based personalized map drawing method of thalamus

Technical field

The invention belongs to the field of brain mapping of magnetic resonance imaging, and specifically relates to a method, system and equipment for individualized mapping of the thalamus based on deep learning guided by group priori.

Background technique

The thalamus is a relay nucleus in the brain that participates in brain functional circuits such as hearing, vision, movement, somatosensory, emotion, memory and learning. In clinical practice, the thalamus, as the target of deep brain stimulation (DBS), is involved in the regulation and treatment of neuropsychiatric diseases such as Parkinson's disease, epilepsy, multiple sclerosis, vegetative awakening, schizophrenia, and essential tremor. Therefore, the drawing of fine and accurate thalamic maps is the key to thalamic research. In existing studies, researchers use methods such as histological section staining and magnetic resonance imaging to characterize the structural or functional characteristics of the thalamus, and use this as the basis for thalamic division. Among these methods for thalamic partitioning, histological section staining is regarded as the gold standard for thalamic partitioning, which can only be performed on isolated brain specimens, is not reproducible and relies on manual marking by anatomists. Rapidly developing magnetic resonance imaging can non-invasively characterize features within the thalamus, including local microstructure, anatomical connections, and functional connections. Based on this, the data-driven thalamic partitioning process has gradually become a research hotspot in thalamic partitioning. According to the modality of magnetic resonance imaging, existing thalamic partitioning can be divided into different methods based on three modalities of structural, diffusion, and functional imaging data. Among the thalamic partitioning methods based on these three magnetic resonance modality data, the thalamic partitioning based on diffusion magnetic resonance is the closest to the anatomical structure of the thalamus. Diffusion magnetic resonance imaging can provide two types of diffusion information, namely fiber bundle connections and local diffusion characteristics. In early studies, researchers found that thalamic divisions based on fiber tract connections did not correspond well to the actual thalamic anatomy, while thalamic divisions based on local diffusion characteristics were basically consistent with the anatomical structure. Therefore, thalamic partitioning based on local diffusion characteristics is the most direct method to characterize the local microstructure of the thalamus.

Most of the above-mentioned thalamic partitioning studies use a group of subjects. Based on the thalamic partitioning results in their individual spaces, the thalamic partitions of different subjects are mapped to the same space through manual marking or automatic registration methods, thereby constructing a group-level thalamus. Map. This method can objectively and unbiasedly reflect the intrinsic partitioning pattern of the thalamus, such as the number of subregions, and the consistent attributes at the group level, such as the structural and functional connectivity patterns of thalamic subnuclei and other brain regions. However, with the deepening of research, researchers have found that there are obvious differences in brain partition patterns between individuals, whether in the cerebral cortex or subcortical nuclei, and that individual-specific brain partitions are better reflective than group-level brain partitions. Individual characteristics, such as cognitive, developmental, aging, and disease characteristics. In addition, in clinical practice, especially in fields such as precision medicine, subject-specific brain maps play an important role, such as preoperative diagnosis, efficacy prediction, target positioning, etc. Therefore, the importance of personalized brain mapping has gradually attracted the attention of researchers. Some individualized map drawing methods have also been gradually developed, which are roughly divided into three types: direct partitioning of single subjects, individual registration of group maps, and individual partitioning guided by group priors. The first method relies on high-quality magnetic resonance imaging data and robust partitioning algorithms, including fiber projection method, spectral clustering method, edge detection method, region growing method, etc.; the second method directly registers group maps to the individual space, thus treating the group map as an individual map; the third method first constructs the group map and treats the group map as prior knowledge of individual partitions to assist subsequent individual partitions. The first partitioning method is suitable for nuclei with specific fiber tract projections, such as the subthalamic nucleus, medial globus pallidus, etc. The second method is suitable for subjects with severe brain damage or in whom diffusion magnetic resonance imaging is not possible, such as patients with brain tumors and patients with metal implants in the head. The third method is suitable for the cerebral cortex and nuclei with larger volumes and richer individual characteristics. Therefore, in order to accurately construct an individualized thalamic map, an individualized partitioning strategy guided by group prior can be used.

With the development of magnetic resonance imaging technology, multi-b-value High Angular Resolution Diffusion Imaging (HARDI) has greatly improved the efficiency and accuracy of modeling and estimation of local diffusion directions. Whether in neuroscience research or clinical scanning, multi-b-value high-angle resolution diffusion magnetic resonance imaging has gradually become a trend. At the same time, in the methodological research on local dispersion direction estimation, different from the classic dispersion tensor model, various existing q-space sampling methods support the construction of higher-order dispersion models. For example, diffusion kurtosis imaging, diffusion spectrum imaging, Q-ball imaging and its derived multi-spherical shell imaging, etc. Among such methods, multi-spherical shell multi-tissue restricted spherical deconvolution (MSMT-CSD) can directly estimate the diffusion direction distribution function (ODF) of brain tissue. This method relies on multi-b-value HARDI and is currently the most suitable for extracting thalamus. method of local diffusion characteristics. In addition, deep learning has developed rapidly in the past decade, and its powerful data fitting and classification capabilities have made it brilliant in various fields including neuroscience. In this context, it is gradually becoming possible to build an automated individualized thalamic map drawing method based on deep learning models by combining local diffusion characteristics and group prior-guided individualized partitioning strategies.

Contents of the invention

In order to solve the above-mentioned problems in the existing technology, that is, to solve the problem that the existing individualized thalamus map drawing method cannot construct a high-confidence group prior and cannot use the group prior to guide the drawing of the individualized thalamus map, resulting in The problem of poor accuracy, robustness and repeatability of the drawn individualized thalamus map. In the first aspect of the present invention, a group prior guided method for drawing an individualized thalamus map based on deep learning is proposed. The method includes:

S100, obtain the group thalamic probability map under the first threshold and perform binarization to obtain the mask of the group thalamic probability map as the first mask; register the first mask to the individual subject The diffusion magnetic resonance space of the brain is used to obtain the thalamus ROI of the subject's individual brain;

S200, obtain the group thalamic probability map under the second threshold, and register it to the diffusion magnetic resonance space of the subject's individual brain to obtain the group prior map of the subject's individual brain;

S300, in the diffusion magnetic resonance space of the subject's individual brain, eliminate the group prior map in the thalamus ROI and obtain a mask of the undefined thalamus area as the second mask;

S400, combined with the second mask, calculate the coefficients of the 45-dimensional spherical harmonic function and the position coordinates of the 3-dimensional diffusion magnetic resonance space of each voxel in the group prior map, and merge the two into a 48-dimensional Feature vector, as an individual feature; input the individual features into the pre-built individualized classification model to obtain the predicted probability value vector of the voxels in each undefined area, and assign the largest sub-region label corresponding to the predicted probability value vector As the final sub-region label of the voxel, a map of the undefined area is generated; the individualized classification model is built based on a deep learning neural network;

S500, merge the group prior map and the map of the undefined area to generate a subject's individualized thalamus map;

Wherein, the construction method of the group thalamic probability map is:

A100, obtain the structural MRI images and diffusion tensor MRI images of the individual brains of N subjects as input images; N is a positive integer;

A200, sequentially perform HCP minimum preprocessing, ROI registration, and ROI postprocessing on the input image to obtain the final thalamus ROI of the subject's individual brain, and obtain the final thalamus ROI of the subject's individual brain through the ODF estimation algorithm. Local diffusion characteristics of each voxel within the ROI;

A300, combined with the local diffusion characteristics obtained by A200, calculates the similarity between voxels and performs clustering to obtain the clustering results of voxels within the final thalamic ROI of the subject's individual brain;

A400, register the clustering results obtained by A300 to the standard space, and perform label remapping to obtain the sub-region label corresponding to each voxel in the final thalamic ROI of the individual subject's brain, that is, N individual subjects are obtained thalamic partitioning of the brain in standard space;

A500, calculate the sub-region probability value corresponding to each voxel in the final thalamus ROI of the individual brain of the subject, remove the voxel points with the largest sub-region probability value lower than the first threshold, and then construct the remaining voxel points The thalamic probability map at the group level is the group thalamic probability map; the sub-region probability value is the ratio of the number of subjects in each sub-region to the total number of subjects.

In some preferred embodiments, the input image is sequentially subjected to HCP minimum pre-processing, ROI registration, and ROI post-processing to obtain the final thalamus ROI of the subject's individual brain. The method is:

Perform HCP minimal preprocessing on the structural MRI images and diffusion tensor MRI images of the subject's individual brain to obtain preprocessed structural MRI images and preprocessed diffusion tensor MRI images;

Based on the preprocessed structural magnetic resonance image and the preprocessed diffusion tensor nuclear magnetic resonance image, the diffusion magnetic resonance space and the structural magnetic resonance space, the structural magnetic resonance space and the standard space are aligned through the ROI registration method. Accurately, obtain individual thalamus ROI;

Use FSL to calculate the anisotropy fraction corresponding to each voxel point in the diffusion magnetic resonance space of the subject's individual brain;

Use SPM to calculate the cerebrospinal fluid probability value corresponding to each voxel point in the structural magnetic resonance space of the subject's individual brain;

In the individual thalamus ROI, voxel points whose anisotropy fraction value is greater than the set anisotropy fraction threshold or whose cerebrospinal fluid probability value is greater than the set cerebrospinal fluid probability threshold are removed, and the remaining voxel points are used as the final individual thalamus ROI.

In some preferred embodiments, the local diffusion characteristics of each voxel in the final thalamic ROI of the subject's individual brain are obtained by:

In the diffusion magnetic resonance space within the final thalamic ROI of the subject's individual brain, use the dhollander algorithm to calculate the response function of the brain tissue under different diffusion weighting factor b value parameters on the diffusion magnetic resonance data; combine the response of the brain tissue function, using the multi-tissue multi-spherical shell restricted spherical deconvolution method to calculate the 45-dimensional coefficients of the 8th order spherical harmonic function to quantify the local diffusion characteristics of each voxel.

In some preferred implementations, the similarity between voxels is calculated as follows:

Among them, S(i,j) represents the similarity between two voxels, E _pos (i, j) represents the Euclidean distance between the three-dimensional coordinates of two voxels in the diffusion magnetic resonance space, E _odf (i, j) represents the Euclidean distance between the sampling coefficients of the 45-dimensional spherical harmonic function of two voxels, w _pos and w _odf respectively represent the correspondence of E _pos (i, j) and E _odf (i, j) when calculating similarity. Weighting coefficient.

In some preferred embodiments, the clustering results of voxels in the final thalamic ROI of the subject's individual brain are obtained by:

The spectral clustering method is used to reduce the dimensionality of the local diffusion characteristics of each voxel in the final thalamic ROI of the subject's individual brain and the similarity between voxels;

Based on the local diffusion characteristics after dimensionality reduction and the similarity between voxels, K-means clustering is used to cluster each voxel into K categories, which is used as the clustering result of the voxels in the final thalamus ROI of the subject's individual brain. .

In some preferred embodiments, the sub-region label corresponding to each voxel in the final thalamus ROI of the subject's individual brain is obtained by:

Calculate the N-dimensional label vector of each voxel point on N subjects, and then cluster all voxel points in the final thalamus ROI of the subject's individual brain according to the similarity of their label vectors, and the clustering results as partition label;

Based on the partition labels, the clustering results of the voxels in the final thalamic ROI of each subject's individual brain are relabeled according to the spatial maximum overlap method, thereby obtaining the final thalamic ROI of each subject's individual brain. The subregion label corresponding to the voxel;

Among them, the N-dimensional label vector of each voxel point on N subjects is calculated, that is, the sub-region label corresponding to the voxel point of each subject is extracted, and one sub-region is extracted for each N subjects. labels, forming an N-dimensional label vector.

In some preferred embodiments, the training method of the individualized classification model is:

Through the methods of S100-S400, obtain individual features and input the individual features into the pre-built individualized classification model to obtain the predicted probability value vector of voxels in each undefined area;

Based on the predicted probability value vector, combined with the sub-region category corresponding to the group prior map, the loss value is obtained through the mean square error loss function, and the model parameters of the individualized classification model are updated;

Repeat the above steps until the trained individualized classification model is obtained.

The second aspect of the present invention proposes a group prior guided thalamic individualized map drawing system based on deep learning, including: a thalamic ROI acquisition module, a group prior map acquisition module, a elimination processing module, and an individualized classification Modules, personalized map generation modules;

The thalamic ROI acquisition module is configured to acquire the group thalamic probability map under a first threshold and perform binarization to obtain a mask of the group thalamic probability map as the first mask; convert the first mask Register to the diffusion magnetic resonance space of the subject's individual brain to obtain the thalamus ROI of the subject's individual brain;

The group prior map acquisition module is configured to obtain the group thalamic probability map under the second threshold, and register it to the diffusion magnetic resonance space of the subject's individual brain to obtain the group's individual brain. Group prior map;

The elimination processing module is configured to eliminate the group prior map in the thalamus ROI in the diffusion magnetic resonance space of the subject's individual brain, and obtain a mask of the undefined thalamus region as the second mask. ;

The individualized classification module is configured to calculate the coefficients of the 45-dimensional spherical harmonic function and the position coordinates of the 3-dimensional diffusion magnetic resonance space of each voxel in the group prior map in combination with the second mask, and combine the two or merged into a 48-dimensional feature vector as individual features; input the individual features into the pre-built individualized classification model to obtain the predicted probability value vector of voxels in each undefined area, and add the predicted probability value vector to The corresponding largest sub-region label is used as the final sub-region label of the voxel, and then a map of the undefined area is generated; the individualized classification model is built based on a deep learning neural network;

The individualized map generation module is configured to merge the group prior map and the map of the undefined area to generate a subject's individualized thalamic map;

Wherein, the construction method of the group thalamic probability map is:

A400, register the clustering results obtained by A300 to the standard space, and perform label remapping to obtain the sub-region label corresponding to each voxel in the final thalamus ROI of the individual subject's brain, that is, N individual subjects are obtained thalamic partitioning of the brain in standard space;

A third aspect of the present invention provides an electronic device, including: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions that can be executed by the processor. , the instructions are configured to be executed by the processor to implement the above-mentioned group prior guided deep learning-based individualized map drawing method of the thalamus.

A fourth aspect of the present invention provides a computer-readable storage medium that stores computer instructions, and the computer instructions are used to be executed by the computer to implement the above-mentioned group a priori guidance. A method for drawing individualized maps of the thalamus based on deep learning.

Beneficial effects of the present invention:

The present invention uses high-confidence group a priori to guide individualized thalamic partitioning and map drawing, thereby improving the accuracy, robustness and repeatability of the drawn individualized thalamic map.

1) The present invention draws an individualized thalamic atlas based on high-confidence group-level thalamic atlas and single subject data, which not only integrates the group consistency of the thalamic partitioning pattern, but also reflects the specificity of the individual partitioning pattern. When using the within-subject partitioning pattern consistency index to quantify the partitioning accuracy of this individualized thalamic atlas, the results show that this method is more effective than single-subject clustering and group atlas registration on the HCP-3T and HCP-7T data. High partitioning accuracy. When testing the reproducibility of this individualized method using Test-retest data, the results show that this method has high inter-scan partitioning pattern consistency on Test-retest data. When reproducibly tested using 7 and 12 as the number of subregions of the thalamus, the results show that the method is robust to different numbers of subregions, and has higher partitioning at the finer number of 12 subregions. accuracy.

2) The present invention partitions the thalamus at the group level based on high-order diffusion characteristics and spatial location, and discovers a more refined thalamic partitioning pattern; and shows higher intra-subject partitioning consistency on the more refined thalamic atlas, It can support the drawing of more fine-grained individualized thalamic maps; it is robust under different magnetic field scanning intensities, and higher magnetic field scanning intensities can provide higher intra-subject partition consistency gains; in the same subject's Repeatability across different scan batches.

Description of drawings

Other features, objects and advantages of the present application will become more apparent upon reading the detailed description of non-limiting embodiments taken with reference to the following drawings.

Figure 1 is a schematic flowchart of a method for drawing an individualized map of the thalamus based on deep learning guided by group priori according to an embodiment of the present invention;

Figure 2 is a schematic framework diagram of a group prior-guided deep learning-based individualized map drawing system for the thalamus according to an embodiment of the present invention;

Figure 3 is a schematic flow chart of the final thalamic ROI acquisition process of an individual subject's brain according to an embodiment of the present invention;

Figure 4 is a schematic flow chart of ODF estimation according to an embodiment of the present invention;

Figure 5 is a schematic flowchart of the construction process of a group thalamic probability map according to an embodiment of the present invention;

Figure 6 is a schematic flowchart of the construction process of a personalized thalamic map according to an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not All examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

The present application will be further described in detail below in conjunction with the accompanying drawings and examples. It can be understood that the specific embodiments described here are only used to explain the relevant invention, but not to limit the invention. It should also be noted that, for convenience of description, only the parts related to the invention are shown in the drawings.

It should be noted that, as long as there is no conflict, the embodiments and features in the embodiments of this application can be combined with each other.

The group prior-guided deep learning-based personalized map drawing method of the thalamus of the present invention, as shown in Figure 1, includes the following steps:

Wherein, the construction method of the group thalamic probability map is:

In order to more clearly explain the group a priori-guided deep learning-based personalized map drawing method of the thalamus of the present invention, each step in one embodiment of the method of the present invention will be described in detail below with reference to the accompanying drawings.

The group prior guided deep learning-based individualized map drawing method of the thalamus described in the present invention is divided into two main steps: the construction of the group probability map of the thalamus and the construction of the individualized thalamus map. In the construction stage of the thalamic group probability map, first based on a set of high-quality data, the direction distribution function and individual diffusion magnetic resonance space of each voxel in the thalamic region of interest (Region of Interest, ROI) are extracted in diffusion magnetic resonance imaging. coordinates, using this as a feature to construct a similarity matrix between voxels, using the spectral clustering algorithm to initially draw the initial partition of the individual, and register it to the MNI standard space to establish a group-level thalamic probability map and maximum probability map . In the individualized map drawing stage, map areas with high probability values at the group level of the thalamus are regarded as group priors, and are registered to the individual diffusion magnetic resonance space as high-confidence group priors for individual maps. At the same time, the maximum The template of the probability map was registered to the individual diffusion magnetic resonance space as the ROI of the individual thalamic map. Next, on the individual diffusion magnetic resonance data, the 45-dimensional coefficients of the 8th order spherical harmonic function in the ROI are extracted through Multi-Shell Multi-Tissue Constrained Spherical Deconvolution (MSMT-CSD) as ODF features, while calculating 3-dimensional individual diffusion magnetic resonance spatial coordinates as spatial position features, thereby constructing 48-dimensional individual features for voxels in each ROI. Using the sub-region categories of voxels within the ROI as labels, the group prior map as training data, and the unlabeled areas within the ROI as test data, an individualized classification model is constructed based on deep learning methods. Finally, the high-confidence group prior map and the predicted unlabeled map are combined to generate an individualized thalamic map.

In the following embodiments, the construction process of the group thalamic probability map is first described in detail, and then the process of generating the individualized map based on a group prior-guided deep learning-based individualized map drawing method of the thalamus is described in detail. narrate.

1. Construction process of group thalamic probability map

In this embodiment, a group (that is, N) structural MRI images and diffusion tensor MRI images of the brains of individual subjects are first obtained.

In this embodiment, minimum preprocessing of HCP data (i.e., HCP minimum preprocessing) is first used to perform minimum preprocessing on the structural NMR image (T1) and diffusion tensor NMR image (Diffusion Weighted Imaging, DWI) data respectively. The preprocessed structural NMR image and the preprocessed diffusion tensor NMR image are obtained.

Based on the preprocessed structural MRI image and the preprocessed diffusion tensor MRI image, in FMRIB's Software Library (FSL), the diffusion magnetic resonance space (B0, referred to as diffusion space) of the subject's individual brain is , as shown in Figure 3) and the structural magnetic resonance space (T1, referred to as the structural space, as shown in Figure 3) are linearly registered, and the structural magnetic resonance space (T1) and the standard space (Montreal Neurological Institute, MNI) are Nonlinear registration, and then combine the two registration matrices to generate a registration matrix from the diffusion magnetic resonance space (B0) of the subject's individual brain to the standard space (MNI) and transpose it to obtain the standard space to the individual diffusion space. registration matrix. Based on this registration matrix, the classic Morel thalamic atlas was registered to the individual diffusion magnetic resonance space, which was used as the individual thalamic ROI. As shown in (a) in Figure 3.

Then, ROI post-processing is performed, as shown in (b) in Figure 3: specifically: using FSL to calculate the FA map (i.e., the anisotropy distribution corresponding to each voxel point) in the diffusion magnetic resonance space of the individual brain of the subject. value); use SPM to calculate the CSF probability map (that is, the cerebrospinal fluid probability value corresponding to each voxel point) in the structural magnetic resonance space of the individual brain of the subject; remove the anisotropy fraction (FA, Fractional) in the individual thalamus ROI Anisotropy) value is greater than the set anisotropy fraction FA threshold (the present invention is preferably set to 0.6) or the cerebrospinal fluid (CSF, Cerebrospinal Fluid) probability value is greater than the set cerebrospinal fluid probability threshold (the present invention is preferably set to 0.05) voxel points, the The remaining voxel points serve as the final individual thalamic ROI (i.e., the ROI in the individual diffusion space in Figure 3).

Finally, ODF estimation is performed, as shown in Figure 4: specifically: in MRtrix3, the dhollander algorithm is used to first calculate the response function of the brain tissue under different diffusion weighting factor b value parameters on the diffusion magnetic resonance data, and then combine the response functions of the brain tissue The response function uses the multi-tissue multi-spherical shell restricted spherical deconvolution method (i.e., the abbreviated multi-tissue multi-spherical shell restricted deconvolution in Figure 4) to calculate the 45-dimensional coefficients of the 8th order spherical harmonic function (i.e., the multi-tissue multi-spherical shell restricted deconvolution in Figure 4). Spherical harmonic coefficient matrix) to represent the local diffusion characteristics of each voxel in the thalamus (that is, taking the 8th order spherical harmonic function as the sampling base, calculating the 45-dimensional spherical harmonic function coefficient of each voxel in the thalamus ROI to construct its ODF model, thus obtaining Spherical harmonic function coefficient matrix within the thalamus ROI).

In this embodiment, the similarity between voxels is first calculated based on the diffusion characteristics and spatial position characteristics of the voxels, and a similarity matrix is constructed. The specific calculation method of the similarity between voxels is as shown in formula (1):

Among them, S(i,j) represents the similarity between two voxels, E _pos (i, j) represents the Euclidean distance between the three-dimensional coordinates of two voxels in the diffusion magnetic resonance space, E _odf (i, j) represents the Euclidean distance between the sampling coefficients of the 45-dimensional spherical harmonic function of two voxels, w _pos and w _odf respectively represent the correspondence of E _pos (i, j) and E _odf (i, j) when calculating similarity. Weighting coefficient. In the present invention, w _pos and w _odf are preferably set to 0.5. In addition, in order to balance the distance bias in the clustering process, the 45-dimensional spherical harmonic function coefficient is multiplied by a scaling factor and then the Euclidean distance between ODFs is calculated. The scaling factor is 89 (3T data) and 98 (7T data). ). Among them, (Indi1...Indi100 in Figure 5 represents the thalamic ROI corresponding to the 100 subjects preferred by the present invention)

Then, clustering is performed, specifically: using the spectral clustering method to reduce the dimensionality of the local diffusion characteristics of each voxel in the final thalamic ROI of the subject's individual brain and the similarity between voxels; based on the dimensionality reduction Based on the local diffusion characteristics and the similarity between voxels, K-means clustering is used to cluster each voxel into K categories, which is used as the clustering result of the voxels within the final thalamus ROI of the subject's individual brain. In the present invention, in order to test the optimal number of thalamic partitions, the value of K ranges from 2 to 28.

In this embodiment, label remapping is performed first, specifically:

After registering the clustering results in the individual subject space to the MNI space, the sub-region labels between the initial partitions of the individual subject cannot correspond to the same. Label remapping first calculates the N-dimensional label vector of each voxel point on N (preferably set to N=100 in the present invention) subjects (i.e., subjects) (i.e., extracts the N-dimensional label vector of each subject at the voxel point). The corresponding sub-region labels are extracted from each of the N subjects to form an N-dimensional label vector). Next, all voxel points within the final thalamus ROI of the subject's individual brain are clustered according to the similarity of their label vectors, and the clustering result is a group-level partition label. Based on the partition labels, the clustering results of each subject are re-labeled according to the spatial maximum overlap method, so as to obtain the sub-region label corresponding to each voxel in the final thalamus ROI of the subject's individual brain, that is, N The thalamic partitions of the individual subject's brains in standard space are used to obtain consistent labels across individuals.

In this embodiment, based on the thalamic partitions in the MNI space of N subject individuals, the sub-region probability value of each voxel in the final thalamic ROI of the individual subject's brain is calculated, that is, the subject belonging to a certain sub-region is calculated. The ratio of the number of subjects to the total number of subjects. After calculating the sub-region probability values of all voxels, take a certain threshold (0.25) and remove the voxel points whose maximum sub-region probability value is lower than the threshold (that is, obtain the maximum sub-region probability corresponding to each voxel point, If the maximum sub-region probability is lower than the set first threshold, the voxel point is removed), and a group-level thalamic probability map is constructed based on the remaining voxel points, that is, the group thalamic probability map, that is, in Figure 5 Group map. Binarize the probability map to obtain the ROI (mask) of the group probability map. According to the winner-take-all principle, the sub-region label of each voxel in the probability map is set to the label corresponding to the maximum probability value, thereby constructing the maximum probability map.

In addition, post-A500 also includes verification of the optimal number of partitions: based on the maximum probability map generated above, calculate the partition consistency between individuals and groups and the topological consistency between cerebral hemispheres to determine the optimal number of thalamic sub-regions. The consistency of partitions between individuals and groups is verified using the overlap rate Dice value of individual partitions and group partitions. The closer the Dice value is to 1, the higher the consistency of partitioning between individuals and groups, and the more consistent the number of thalamic partitions here is with the internal partitioning pattern of the thalamus. Based on the leave-one-out method, the individual partitioned subjects and the group partitioned subject sets were divided, and the Dice average value of the consistency of individual and group partitions was calculated for 100 times. The topological consistency index between cerebral hemispheres is verified on the group map using topological distance TpD (Topological Distance, TpD). The closer the TpD value is to 0, the higher the degree of homology between the left and right hemispheres of the brain, and the number of thalamic divisions here is more consistent with the internal division pattern of the thalamus. Topological distance first calculates the number of adjacent voxels between each sub-region and other sub-regions in a single hemisphere, and generates a K×K-dimensional matrix. After unfolding this matrix, the similarity of this matrix between the two cerebral hemispheres is calculated, that is, the topological distance within the partition.

Among them, when the group maximum probability map is verified, the Dice indicator appears as a local maximum when the number of clusters is 7 and 12, and the TpD value under these two cluster numbers is basically 0, indicating that the number of partitions is 7 and 12. Partitioning patterns within the thalamus.

When the thalamus is divided into 7 subregions, the anterior thalamic nucleus, ventroanterior nucleus, dorsomedial nucleus, dorsolateral-posterolateral nucleus, lateral pulvinar nucleus, ventrolateral-ventroposterior nucleus and medial pulvinar nucleus can be complete according to the number of subregions. Split it out. There is strong homology between the left and right hemispheres of the thalamus. Each subregion of the thalamus shows strong structural homogeneity.

When the thalamus is divided into 12 subregions, the number of subdivisions is: anterior thalamic nucleus, ventroanterior nucleus, dorsomedial nucleus, posteromedial nucleus, ventrolateral nucleus, ventroposterior nucleus, anterior pulvinar nucleus, lateral pulvinar nucleus, medial pulvinar nucleus The nucleus, dorsoanterior dorsomedial nucleus, dorsal dorsomedial nucleus, and ventral dorsomedial nucleus can be completely segmented. There is strong homology between the left and right hemispheres of the thalamus. Each subregion of the thalamus shows strong structural homogeneity. Compared with the zoning pattern of 7 subregions, the thalamic map of 12 subregions is more refined.

2. Group prior guided thalamus individualized map drawing method based on deep learning, as shown in Figure 6

In this embodiment, the group thalamic probability map constructed above (ie, the group map in Figure 6(a) and (b)) is taken as a first threshold (preferably set to 0.25 in the present invention) and binarized Generate a mask mask for the group thalamic probability map, that is, the maximum probability mask in (a) in Figure 6. The maximum probability mask is registered to the diffusion magnetic resonance space of the subject's individual brain to obtain the thalamus ROI of the subject's individual brain (ie, the individual mask in (a) in Figure 6).

In this embodiment, the group thalamic probability map constructed above is set to a second threshold (preferably set to >0.75 in the present invention) to generate a high-probability thalamic map (ie, the high-confidence group in (b) in Figure 6 group map), and register it to the individual diffusion magnetic resonance space as a group prior map (i.e., individual prior map) of the subject's individual brain, as shown in (b) in Figure 6 . Among them, the first threshold and the second threshold are used to remove the voxel points with the largest sub-region probability value lower than the threshold (i.e. the first threshold or the second threshold), and then construct the group thalamic probability map (specifically, as shown in A500 Show).

In this embodiment, the group prior map in the thalamus ROI is eliminated to obtain a mask of the undefined thalamus region as the second mask, as shown in (c) of Figure 6 .

In this embodiment, combined with the second mask (during individualized classification model training, the area within the second mask is used, and during verification, the area outside the second mask is used), the group is first calculated. The coefficients of the 45-dimensional spherical harmonic function and the position coordinates of the 3-dimensional diffusion magnetic resonance space of each voxel in the prior map are combined, and the individual characteristics are obtained by merging the two. Input the individual characteristics (i.e. 48 dimensions) into the pre-built individualized classification model to obtain the predicted probability value vector of the voxels in each undefined area, and use the sub-region label with the largest predicted probability value vector as the voxel's The final sub-region label is generated, and a map of the undefined area is generated; the individualized classification model is built based on a deep learning neural network (ie, the deep learning model in (e) in Figure 6).

Among them, the individualized classification model is built based on the deep learning neural network, and its training process is as follows, as shown in (d) in Figure 6:

Through the methods of S100-S400, obtain individual characteristics and input the individual characteristics into the pre-built individualized classification model to obtain the K (number of K sub-regions) dimensional predicted probability value vector of voxels in each undefined area;

Based on the predicted probability value vector, combined with the sub-region category corresponding to the group prior map, the loss value is obtained through the mean square error loss function, and the model parameters of the individualized classification model are updated; the individualized classification model is trained in a loop until the Trained individualized classification model.

S500: Merge the group prior map and the map of the undefined area to generate a subject's individualized thalamus map. As shown in (f) in Figure 6.

Finally, the improved silhouette coefficient Silhouette value is used to verify the individualized thalamic atlas, specifically:

The Euclidean distance of the 45-dimensional spherical harmonic function coefficients is used as the distance metric. A quantitative comparison of three individuation methods, group prior guidance, group registration, and single-subject clustering, was performed using Silhouette's improvement ratio as the gain index. The gain index uses the Silhouette value of the individualized thalamic map clustered by a single subject as the denominator, and the difference between the Silhouette value of the individualized thalamic map guided by group prior guidance or group registration and its numerator as the numerator. A positive gain means that the accuracy of a certain individualized map drawing method is higher than that of single-subject clustering, and the greater the gain, the greater the accuracy improvement. A negative number means that it is lower than single-subject clustering. Four data sets, HCP 3T, HCP 7T, HCP Test and HCP Retest, were used to verify the robustness and repeatability of the personalized thalamus map drawing scheme under different magnetic field strengths and different scanning batches. Different partition numbers of 7 and 12 were used to verify the robustness of the personalized thalamic atlas drawing scheme under different numbers of thalamic subregions.

The intra-subject partition pattern consistency indicators of the individualized thalamic atlas on the four HCP data sets are HCP 3T (N=100), HCP 7T (N=100), HCP Test (N=30) and HCP Retest ( Validation metrics on N=30) data set. On the four data sets, the verification indicators were calculated when the number of partitions was 7 and 12, respectively. The results show that on the four data sets, the three individuation methods of group a priori guidance, group registration and single-subject clustering show similar consistency of within-subject partitioning patterns, but the group a priori guidance of individuals The thalamus atlas had higher within-subject consistency than the other two methods.

The gain index of the intra-subject partition consistency of the individualized map is the gain comparison of the intra-subject partition consistency on the HCP 3T and HCP 7T data, and the gain comparison of the intra-subject partition consistency on the HCP Test and HCP Retest data. . The within-subject partition consistency of the individualized thalamic atlas drawn by the group registration method is worse than that of single-subject clustering. However, the group-prior guided individualized thalamic atlas has stronger intra-subject partition consistency than single-subject clustering. When the number of partitions is 7, the gains of 3T and 7T data are the same. When the number of partitions is 12, the gain of 7T data is higher than that of 3T data. When the number of partitions is 7 and 12, the gains of Test and Retest are basically the same. On the four data sets, the group prior-guided individualized thalamic atlas has a higher gain when the number of partitions is 12 than when the number of partitions is 7.

When the group a priori guided individualized thalamic atlas was verified for consistency on HCP Test and HCP Retest, when the number of partitions was 7 and 12, the group a priori guided individualized thalamic atlas showed stable and reliable scanning batches inter-consistency.

A group prior-guided deep learning-based personalized map drawing system of the thalamus according to the second embodiment of the present invention, as shown in Figure 2, includes: a thalamus ROI acquisition module 100, a group prior map acquisition module 200, and elimination Processing module 300, individualized classification module 400, individualized map generation module 500;

The thalamic ROI acquisition module 100 is configured to acquire the group thalamic probability map under a first threshold and perform binarization to obtain a mask of the group thalamic probability map as the first mask; The membrane is registered to the diffusion magnetic resonance space of the subject's individual brain, and the thalamus ROI of the subject's individual brain is obtained;

The group prior map acquisition module 200 is configured to obtain the group thalamic probability map under the second threshold, and register it to the diffusion magnetic resonance space of the subject's individual brain to obtain the subject's individual brain. Group prior map;

The elimination processing module 300 is configured to eliminate the group prior map in the thalamus ROI in the diffusion magnetic resonance space of the subject's individual brain to obtain a mask of the undefined thalamus region as the second mask. membrane;

The individualized classification module 400 is configured to calculate the coefficients of the 45-dimensional spherical harmonic function and the position coordinates of the 3-dimensional diffusion magnetic resonance space of each voxel in the group prior map, and merge the two to obtain individual characteristics; Input the individual characteristics into the pre-built individualized classification model to obtain the predicted probability value vector of the voxel in each undefined area, and use the sub-region label with the largest predicted probability value vector as the final sub-region label of the voxel, And generate a map of undefined areas; the individualized classification model is built based on a deep learning neural network;

The personalized map generation module 500 is configured to merge the group prior map and the map of the undefined area to generate a subject's individualized thalamic map;

Wherein, the construction method of the group thalamic probability map is:

It should be noted that the group a priori-guided deep learning-based personalized map drawing system of the thalamus provided in the above embodiments is only illustrated by the division of each functional module mentioned above. In practical applications, the above mentioned functions can be used as needed. Function allocation is completed by different functional modules, that is, the modules or steps in the embodiment of the present invention are decomposed or combined. For example, the modules in the above embodiment can be combined into one module, or further divided into multiple sub-modules to complete All or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only used to distinguish each module or step and are not regarded as improper limitations of the present invention.

An electronic device according to a third embodiment of the present invention includes: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions that can be executed by the processor, so The instructions are used to be executed by the processor to implement the above-mentioned group prior-guided deep learning-based individualized map drawing method of the thalamus.

A computer-readable storage medium according to the fourth embodiment of the present invention, the computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by the computer to implement the above-mentioned group a priori guidance based on depth. A learning method for individualized mapping of the thalamus.

A device for drawing individualized maps of the thalamus based on deep learning guided by group prior guidance in the fifth embodiment of the present invention includes: a magnetic resonance image acquisition device and a central processing device,

The magnetic resonance image acquisition equipment includes magnetic resonance imaging equipment and a superconducting magnetic resonance instrument, and is configured to acquire structural MRI images and diffusion tensor MRI images of the subject's individual brain;

The central processing device includes a GPU and is configured to obtain a group thalamic probability map under a first threshold and perform binarization to obtain a mask of the group thalamic probability map as a first mask; converting the first mask The membrane is registered to the diffusion magnetic resonance space of the subject's individual brain, and the thalamus ROI of the subject's individual brain is obtained;

Obtain the group thalamic probability map under the second threshold, and register it to the diffusion magnetic resonance space of the subject's individual brain to obtain the group prior map of the subject's individual brain; in the subject's individual brain In the diffusion magnetic resonance space, the group prior map in the thalamus ROI is eliminated to obtain a mask of the undefined thalamus area as the second mask;

Combined with the second mask, calculate the coefficients of the 45-dimensional spherical harmonic function and the position coordinates of the 3-dimensional diffusion magnetic resonance space of each voxel in the group prior map, and merge the two into a 48-dimensional feature vector , as individual features; input the individual features into the pre-built individualized classification model to obtain the predicted probability value vector of voxels in each undefined area, and use the largest sub-region label corresponding to the predicted probability value vector as the volume The final sub-region label of the element is then generated to generate a map of the undefined area; the individualized classification model is constructed based on a deep learning neural network; the group prior map and the map of the undefined area are merged to generate a subject individualized thalamus map;

Wherein, the construction method of the group thalamic probability map is:

Those skilled in the technical field can clearly understand that, for the convenience and simplicity of description, the specific details of the above-described electronic equipment, computer-readable storage media, and group prior-guided deep learning-based personalized map drawing device of the thalamus are For the working process and related instructions, please refer to the corresponding process in the foregoing method examples, and will not be repeated here.

Those skilled in the art should be able to realize that the modules and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, computer software, or a combination of both, and the programs corresponding to the software modules and method steps Can be placed in random access memory (RAM), memory, read only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, hard disk, removable disk, CD-ROM, or any other device known in the art. any other form of storage media. In order to clearly illustrate the interchangeability of electronic hardware and software, the composition and steps of each example have been generally described according to function in the above description. Whether these functions are performed as electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementations should not be considered to be beyond the scope of the present invention.

The terms "first", "second", "third", etc. are used to distinguish similar objects, but are not used to describe or indicate a specific order or sequence.

So far, the technical solution of the present invention has been described with reference to the preferred embodiments shown in the drawings. However, those skilled in the art can easily understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or replacements to relevant technical features, and the technical solutions after these changes or replacements will fall within the protection scope of the present invention.

Claims

A group prior-guided deep learning-based personalized map drawing method for the thalamus, which is characterized in that the method includes the following steps:

S100, obtain the group thalamic probability map under the first threshold and perform binarization to obtain the mask of the group thalamic probability map as the first mask; register the first mask to the individual subject The diffusion magnetic resonance space of the brain is used to obtain the thalamus ROI of the subject's individual brain;

S200, obtain the group thalamic probability map under the second threshold, and register it to the diffusion magnetic resonance space of the subject's individual brain to obtain the group prior map of the subject's individual brain;

S300, in the diffusion magnetic resonance space of the subject's individual brain, eliminate the group prior map in the thalamus ROI and obtain a mask of the undefined thalamus area as the second mask;

S400, combined with the second mask, calculate the coefficients of the 45-dimensional spherical harmonic function and the position coordinates of the 3-dimensional diffusion magnetic resonance space of each voxel in the group prior map, and merge the two into a 48-dimensional Feature vector, as an individual feature; input the individual features into the pre-built individualized classification model to obtain the predicted probability value vector of the voxels in each undefined area, and assign the largest sub-region label corresponding to the predicted probability value vector As the final sub-region label of the voxel, a map of the undefined area is generated; the individualized classification model is built based on a deep learning neural network;

S500, merge the group prior map and the map of the undefined area to generate a subject's individualized thalamus map;

Wherein, the construction method of the group thalamic probability map is:

A100, obtain the structural MRI images and diffusion tensor MRI images of the individual brains of N subjects as input images; N is a positive integer;

A200, sequentially perform HCP minimum preprocessing, ROI registration, and ROI postprocessing on the input image to obtain the final thalamus ROI of the subject's individual brain, and obtain the final thalamus ROI of the subject's individual brain through the ODF estimation algorithm. Local diffusion characteristics of each voxel within the ROI;

A300, combined with the local diffusion characteristics obtained by A200, calculates the similarity between voxels and performs clustering to obtain the clustering results of voxels within the final thalamic ROI of the subject's individual brain;

A400, register the clustering results obtained by A300 to the standard space, and perform label remapping to obtain the sub-region label corresponding to each voxel in the final thalamus ROI of the individual subject's brain, that is, N individual subjects are obtained thalamic partitioning of the brain in standard space;

A500, calculate the sub-region probability value corresponding to each voxel in the final thalamus ROI of the individual brain of the subject, remove the voxel points with the largest sub-region probability value lower than the first threshold, and then construct the remaining voxel points The thalamic probability map at the group level is the group thalamic probability map; the sub-region probability value is the ratio of the number of subjects in each sub-region to the total number of subjects.
The group a priori-guided deep learning-based personalized map drawing method of the thalamus according to claim 1, characterized in that the input image is sequentially subjected to HCP minimum pre-processing, ROI registration, and ROI post-processing to obtain the subject The final thalamus ROI of the individual brain of the subject is determined as follows:

Perform HCP minimal preprocessing on the structural MRI images and diffusion tensor MRI images of the subject's individual brain to obtain preprocessed structural MRI images and preprocessed diffusion tensor MRI images;

Based on the preprocessed structural magnetic resonance image and the preprocessed diffusion tensor nuclear magnetic resonance image, the diffusion magnetic resonance space and the structural magnetic resonance space, the structural magnetic resonance space and the standard space are aligned through the ROI registration method. Accurately, obtain individual thalamus ROI;

Use FSL to calculate the anisotropy fraction corresponding to each voxel point in the diffusion magnetic resonance space of the subject's individual brain;

Use SPM to calculate the cerebrospinal fluid probability value corresponding to each voxel point in the structural magnetic resonance space of the subject's individual brain;

In the individual thalamus ROI, voxel points whose anisotropy fraction value is greater than the set anisotropy fraction threshold or whose cerebrospinal fluid probability value is greater than the set cerebrospinal fluid probability threshold are removed, and the remaining voxel points are used as the final individual thalamus ROI.
The group a priori-guided deep learning-based personalized map drawing method of the thalamus according to claim 1, characterized in that the local diffusion characteristics of each voxel in the final thalamus ROI of the subject's individual brain, The acquisition method is:

In the diffusion magnetic resonance space within the final thalamic ROI of the subject's individual brain, use the dhollander algorithm to calculate the response function of the brain tissue under different diffusion weighting factor b value parameters on the diffusion magnetic resonance data; combine the response of the brain tissue function, using the multi-tissue multi-spherical shell restricted spherical deconvolution method to calculate the 45-dimensional coefficients of the 8th order spherical harmonic function to quantify the local diffusion characteristics of each voxel.
The group prior-guided deep learning-based personalized map drawing method of the thalamus according to claim 1, characterized in that the similarity between voxels is calculated by:

Among them, S(i,j) represents the similarity between two voxels, E pos (i, j) represents the Euclidean distance between the three-dimensional coordinates of two voxels in the diffusion magnetic resonance space, E odf (i, j) represents the Euclidean distance between the sampling coefficients of the 45-dimensional spherical harmonic function of two voxels, w pos and w odf respectively represent the correspondence of E pos (i, j) and E odf (i, j) when calculating similarity. Weighting coefficient.
The group a priori-guided deep learning-based personalized map drawing method of the thalamus according to claim 4, characterized in that the clustering result of the voxels in the final thalamus ROI of the subject's individual brain is obtained. The method is:

The spectral clustering method is used to reduce the dimensionality of the local diffusion characteristics of each voxel in the final thalamic ROI of the subject's individual brain and the similarity between voxels;

Based on the local diffusion characteristics after dimensionality reduction and the similarity between voxels, K-means clustering is used to cluster each voxel into K categories, which is used as the clustering result of the voxels in the final thalamus ROI of the subject's individual brain. .
The group a priori-guided deep learning-based personalized map drawing method of the thalamus according to claim 1, characterized in that the sub-region label corresponding to each voxel in the final thalamus ROI of the subject's individual brain, The acquisition method is:

Calculate the N-dimensional label vector of each voxel point on N subjects, and then cluster all voxel points in the final thalamus ROI of the subject's individual brain according to the similarity of their label vectors, and the clustering results as partition label;

Based on the partition labels, the clustering results of the voxels in the final thalamic ROI of each subject's individual brain are relabeled according to the spatial maximum overlap method, thereby obtaining the final thalamic ROI of each subject's individual brain. The subregion label corresponding to the voxel;

Among them, the N-dimensional label vector of each voxel point on N subjects is calculated, that is, the sub-region label corresponding to the voxel point of each subject is extracted, and one sub-region is extracted for each N subjects. labels, forming an N-dimensional label vector.
The group prior-guided deep learning-based individualized map drawing method of the thalamus according to claim 6, characterized in that the training method of the individualized classification model is:

Through the methods of S100-S400, obtain individual features and input the individual features into the pre-built individualized classification model to obtain the predicted probability value vector of voxels in each undefined area;

Based on the predicted probability value vector, combined with the sub-region category corresponding to the group prior map, the loss value is obtained through the mean square error loss function, and the model parameters of the individualized classification model are updated;

Repeat the above steps until the trained individualized classification model is obtained.
A group prior guided thalamic individualized map drawing system based on deep learning, which is characterized in that the system includes: a thalamic ROI acquisition module, a group prior map acquisition module, a elimination processing module, an individualized classification module, and an individual Chemical map generation module;

The thalamic ROI acquisition module is configured to acquire the group thalamic probability map under a first threshold and perform binarization to obtain a mask of the group thalamic probability map as the first mask; convert the first mask Register to the diffusion magnetic resonance space of the subject's individual brain to obtain the thalamus ROI of the subject's individual brain;

The group prior map acquisition module is configured to obtain the group thalamic probability map under the second threshold, and register it to the diffusion magnetic resonance space of the subject's individual brain to obtain the group's individual brain. Group prior map;

The elimination processing module is configured to eliminate the group prior map in the thalamus ROI in the diffusion magnetic resonance space of the subject's individual brain, and obtain a mask of the undefined thalamus region as the second mask. , as a training set for the individualized model;

The individualized classification module is configured to calculate the coefficients of the 45-dimensional spherical harmonic function and the position coordinates of the 3-dimensional diffusion magnetic resonance space of each voxel in the group prior map in combination with the second mask, and combine the two or merged into a 48-dimensional feature vector as individual features; input the individual features into the pre-built individualized classification model to obtain the predicted probability value vector of voxels in each undefined area, and add the predicted probability value vector to The corresponding largest sub-region label is used as the final sub-region label of the voxel, and then a map of the undefined area is generated; the individualized classification model is built based on a deep learning neural network;

The individualized map generation module is configured to merge the group prior map and the map of the undefined area to generate a subject's individualized thalamic map;

Wherein, the construction method of the group thalamic probability map is:

A100, obtain the structural MRI images and diffusion tensor MRI images of the individual brains of N subjects as input images; N is a positive integer;

A200, sequentially perform HCP minimum preprocessing, ROI registration, and ROI postprocessing on the input image to obtain the final thalamus ROI of the subject's individual brain, and obtain the final thalamus ROI of the subject's individual brain through the ODF estimation algorithm. Local diffusion characteristics of each voxel within the ROI;

A300, combined with the local diffusion characteristics obtained by A200, calculates the similarity between voxels and performs clustering to obtain the clustering results of voxels within the final thalamic ROI of the subject's individual brain;

A400, register the clustering results obtained by A300 to the standard space, and perform label remapping to obtain the sub-region label corresponding to each voxel in the final thalamus ROI of the individual subject's brain, that is, N individual subjects are obtained thalamic partitioning of the brain in standard space;

A500, calculate the sub-region probability value corresponding to each voxel in the final thalamus ROI of the individual brain of the subject, remove the voxel points with the largest sub-region probability value lower than the first threshold, and then construct the remaining voxel points The thalamic probability map at the group level is the group thalamic probability map; the sub-region probability value is the ratio of the number of subjects in each sub-region to the total number of subjects.
An electronic device, characterized by including:

at least one processor; and a memory communicatively connected to at least one processor;

Wherein, the memory stores instructions that can be executed by the processor, and the instructions are used by the processor to implement the group a priori guided deep learning based on any one of claims 1-7. A method for drawing individualized maps of the thalamus.
A computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by the computer to implement the group of any one of claims 1-7 A priori-guided deep learning-based personalized mapping method for the thalamus.