Method and device for automatic segmentation of brain white matter fiber bundles based on anatomical priors

Abstract

The present invention provides a method and device for automatic segmentation of brain white matter fiber bundles fused with anatomical priors, the method comprising: obtaining whole-brain fiber point coordinates and structural T1 magnetic resonance images, determining superficial white matter fibers and deep white matter fibers based on the whole-brain fiber point coordinates, generating an anatomical brain partition map based on the structural T1 magnetic resonance image; determining individual-level anatomical feature descriptors of each fiber based on superficial white matter fibers, deep white matter fibers and the anatomical brain partition map, respectively determining cluster-level anatomical feature descriptors corresponding to each fiber based on individual-level anatomical feature descriptors of each superficial white matter fiber and individual-level anatomical feature descriptors of each deep white matter fiber; inputting whole-brain fiber point coordinates, individual-level anatomical feature descriptors and cluster-level anatomical feature descriptors into a trained fiber bundle segmentation model, and obtaining classification results of each fiber bundle based on the trained fiber bundle segmentation model. The present invention can improve the accuracy of whole-brain white matter fiber bundle classification results.

Description

Technical Field

[0001] The invention relates to the technical field of magnetic resonance imaging, and in particular to a method and a device for automatically segmenting brain white matter fiber bundles integrating anatomical priors. Background Art

[0002] Diffusion magnetic resonance technology uses a pulsed magnetic field gradient to generate a non-uniform linearly distributed magnetic field, so that the diffusion of water molecules will cause a sensitive change in the signal intensity along the gradient direction, thereby measuring the diffusion coefficient of water molecules in the gradient direction and obtaining a set of diffusion-weighted image data; therefore, it can depict the structure of living tissue and reconstruct fibers.

[0003] Fiber bundle segmentation is to divide the whole brain fibers into different clusters according to the fiber's geometric shape, anatomical characteristics or functional characteristics; fiber bundle segmentation is the basis of brain science research. The more accurate the segmentation, the more helpful it is for downstream research tasks. With the development of artificial intelligence technology, neural network segmentation models are more efficient for fiber bundle segmentation.

[0004] There are two existing fiber bundle segmentation methods: voxel-based semantic segmentation and fiber streamline-based segmentation. Voxel-based semantic segmentation requires voxel-level masks of fiber bundles, and its annotation is complex and time-consuming. Fiber streamline-based segmentation regards fibers as streamlines connected by a series of points. The annotation of each streamline can be obtained through clustering and other methods, and it contains shape information such as point coordinates and streamline curvature, which is not available in voxel-based semantic segmentation methods. In the fiber streamline-based segmentation method, there are generally two steps: obtaining the feature descriptor of the fiber and constructing the corresponding segmentation method using the feature descriptor. The feature descriptor of the fiber often comes from geometric information, such as construction based on point coordinates and curvature.

[0005] Although the existing fiber bundle segmentation methods can achieve efficient fiber bundle segmentation, the fiber feature descriptions only consider the geometric features of the fiber streamlines, without considering the characteristics of the fibers in the anatomical brain region, resulting in low accuracy of the fiber bundle segmentation results. Therefore, how to improve the accuracy of the whole brain white matter fiber bundle classification results is a technical problem that needs to be solved urgently. Summary of the invention

[0006] In view of this, an embodiment of the present invention provides a method and device for automatically segmenting brain white matter fiber bundles integrating anatomical priors, so as to eliminate or improve one or more defects existing in the prior art.

[0007] One aspect of the present invention provides a method for automatically segmenting brain white matter fiber bundles by integrating anatomical priors, the method comprising the following steps:

[0008] Acquire whole-brain fiber point coordinates and a structural T1 magnetic resonance image, determine superficial white matter fibers and deep white matter fibers based on the whole-brain fiber point coordinates, and generate an anatomical brain partition map based on the structural T1 magnetic resonance image;

[0009] Determine the individual level anatomical feature descriptor of each fiber based on the superficial white matter fibers, the deep white matter fibers and the anatomical brain partition map, and determine the cluster level anatomical feature descriptor corresponding to each fiber based on the individual level anatomical feature descriptor of each superficial white matter fiber and the individual level anatomical feature descriptor of each deep white matter fiber;

[0010] The whole-brain fiber point coordinates, individual-level anatomical feature descriptors, and cluster-level anatomical feature descriptors are input into a trained fiber bundle segmentation model, and the classification results of each fiber bundle are obtained based on the trained fiber bundle segmentation model.

[0011] In some embodiments of the present invention, determining the individual level anatomical feature descriptor of each fiber based on the superficial white matter fibers, the deep white matter fibers and the anatomical brain partition map comprises:

[0012] Performing 3D affine transformation on each of the superficial white matter fibers and the deep white matter fibers, mapping each fiber point of each fiber to the brain partition of the anatomical brain partition map, and obtaining the anatomical distribution of each fiber point of each fiber;

[0013] An individual level anatomical feature descriptor of each fiber is obtained based on the anatomical distribution of the fiber points of each of the fibers.

[0014] In some embodiments of the present invention, the cluster-level anatomical feature descriptors corresponding to the respective fibers are determined based on the individual-level anatomical feature descriptors of the respective superficial white matter fibers and the individual-level anatomical feature descriptors of the respective deep white matter fibers, including:

[0015] hierarchically clustering the superficial white matter fibers and the deep white matter fibers based on the individual level anatomical feature descriptors of the superficial white matter fibers and the individual level anatomical feature descriptors of the deep white matter fibers;

[0016] Determining the number of clusters corresponding to the superficial white matter fibers and the number of clusters corresponding to the deep white matter fibers;

[0017] The cluster to which each fiber belongs is determined based on the number of clusters corresponding to the superficial white matter fibers, the number of clusters corresponding to the deep white matter fibers, and the hierarchical clustering result, and the cluster-level anatomical feature descriptor corresponding to each fiber is determined based on the cluster to which each fiber belongs.

[0018] In some embodiments of the present invention, the fiber bundle segmentation model includes a point cloud encoder module, a first embedding layer, a second embedding layer and a decoder module; and / or,

[0019] The anatomical brain parenchyma map includes 286 brain parenchyma regions.

[0020] In some embodiments of the present invention, the whole-brain fiber point coordinates, individual-level anatomical feature descriptors, and cluster-level anatomical feature descriptors are input into a trained fiber bundle segmentation model, and the classification results of each fiber bundle are obtained based on the trained fiber bundle segmentation model, including:

[0021] Performing local feature encoding and global feature encoding on the whole-brain fiber point coordinates based on a point cloud encoder, and fusing the local feature encoding and the global feature encoding to obtain a feature encoding;

[0022] Inputting the individual level anatomical feature descriptor into the first embedding layer to obtain a first embedding code;

[0023] Inputting the cluster-level anatomical feature descriptor into the second embedding layer to obtain a second embedding code;

[0024] Fusing the feature code, the first embedded code and the second embedded code to obtain a fused code;

[0025] The fusion code is input into the decoder to obtain the classification result of each fiber bundle.

[0026] In some embodiments of the present invention, the method further comprises:

[0027] Construct an initial fiber bundle segmentation model;

[0028] Construct model loss function and sample data set;

[0029] The initial fiber bundle segmentation model is pre-trained based on the model loss function and the sample data set to obtain a trained fiber bundle segmentation model.

[0030] In some embodiments of the present invention, the loss function is:

[0031]

[0032] Where M represents the number of categories into which the fiber bundles are divided; y ic is 0 or 1. When the predicted category of fiber i is the same as the true category, y ic is 1, when the category prediction result of fiber i is different from the true category result, y ic is 0; N represents the total number of fibers, p ic represents the probability that fiber i belongs to fiber bundle category c.

[0033] In some embodiments of the present invention, determining superficial white matter fibers and deep white matter fibers based on the whole-brain fiber point coordinates includes:

[0034] Inputting the whole-brain fiber point coordinates into a trained fiber classification model for pre-classification to obtain a superficial white matter fiber pre-classification result and a deep white matter fiber pre-classification result;

[0035] Inputting each superficial white matter fiber in the superficial white matter fiber pre-classification result into a fiber filtering model to obtain pseudo superficial white matter fibers;

[0036] The pseudo superficial white matter fibers were labeled as deep white matter fibers.

[0037] According to another aspect of the present invention, a system for automatic segmentation of brain white matter fiber bundles integrating anatomical priors is also disclosed, comprising a processor, a memory, and a computer program stored in the memory, wherein the processor is used to execute the computer program, and when the computer program is executed, the system implements the steps of the method described in any of the above embodiments.

[0038] According to another aspect of the present invention, a computer-readable storage medium is disclosed, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in any of the above embodiments are implemented.

[0039] The present invention integrates anatomical priors to automatically segment the brain white matter fiber bundles. The method and device generate anatomical brain partition maps based on individual structural T1 magnetic resonance images and construct feature descriptors with fiber anatomical information; that is, through the acquired white matter fiber annotation data and the corresponding individual anatomical brain partition map, the fiber anatomical feature descriptors of the sample are calculated and established to achieve accurate segmentation of the white matter fiber bundles. Compared with the traditional method of fiber segmentation based only on the geometric characteristics of the fibers, the anatomically guided fiber bundle segmentation method of the present invention has more detailed constraints and more accurate segmentation results, providing a basis for downstream brain science research.

[0040] Additional advantages, purposes, and features of the present invention will be described in part in the following description, and will become apparent to those skilled in the art after studying the following, or may be learned from the practice of the present invention. The purposes and other advantages of the present invention may be achieved and obtained by the structures specifically indicated in the specification and the accompanying drawings.

[0041] Those skilled in the art will appreciate that the objectives and advantages that can be achieved with the present invention are not limited to the above specific description, and the above and other objectives that can be achieved by the present invention will be more clearly understood from the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings described herein are used to provide a further understanding of the present invention, constitute a part of this application, and do not constitute a limitation of the present invention. The components in the drawings are not drawn to scale, but are only for illustrating the principles of the present invention. In order to facilitate the illustration and description of some parts of the present invention, the corresponding parts in the drawings may be enlarged, that is, they may become larger relative to other components in the exemplary device actually manufactured according to the present invention. In the drawings:

[0043] Figure 1 Schematic diagram of the process of automatic segmentation of brain white matter fiber bundles integrating anatomical priors in one embodiment of the present invention.

[0044] Figure 2 Schematic diagram of the flow of an automatic brain white matter fiber bundle segmentation method integrating anatomical priors in another embodiment of the present invention.

[0045] Figure 3 Schematic diagram of the architecture of a system for automatic segmentation of white matter fiber bundles integrating anatomical priors in one embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0047] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, only structures and / or processing steps closely related to the solutions according to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.

[0048] It should be emphasized that the term “include / comprises” when used herein refers to the presence of features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0049] It should also be noted that, unless otherwise specified, the term “connection” herein may refer not only to a direct connection but also to an indirect connection with an intermediate.

[0050] The existing white matter fiber bundle segmentation method uses a fiber tracking algorithm to obtain the whole brain fiber map of each sample, and then uses a clustering method to generate multiple white matter fiber bundles when annotating the fibers. Each fiber in each sample has an independent annotation. Although the existing white matter fiber bundle segmentation method can complete the segmentation of the fiber bundle, the fiber bundle obtained does not conform to anatomical knowledge and is still a long way from fine and accurate segmentation. The inventor found in the study that the existing white matter fiber bundle segmentation method cannot achieve fine and accurate segmentation. The main reasons are as follows: First, the whole brain fiber streamlines of a single sample obtained by fiber tracking often have millions, and each streamline contains a large number of sampling points. The huge amount of data and relying on geometric information such as the position and shape of the fiber streamline for segmentation are a major challenge faced by traditional machine learning methods; second, the introduction of anatomical knowledge is a difficult process. Each fiber has its own starting, passing and terminating brain area, and multiple fibers pass through the same brain area. How to give anatomical feature descriptions to whole brain fibers is an unresolved problem.

[0051] In order to achieve fine and accurate segmentation of fiber bundles, the present invention discloses a method and device for automatic segmentation of white matter fiber bundles integrating anatomical priors. The method obtains the individual brain anatomical partition map of each sample T1 structural magnetic resonance image according to the anatomical brain atlas, and constructs a new feature descriptor and neural network model based on the individual fiber annotation data and the brain anatomical partition map.

[0052] In this article, the terms fiber tracing, anatomical brain atlas, superficial white matter, and deep white matter are explained as follows:

[0053] Fiber tracking: The process of reconstructing brain white matter fibers from diffusion magnetic resonance images. This process is based on the characteristics of diffusion magnetic resonance imaging, which can detect the degree and direction of anisotropy of water molecules diffusing in tissues in vivo, estimate the distribution of fiber orientations in each voxel in the image, and further select the most likely orientation based on the probability of connection between the selected seed voxel and other voxels, thereby obtaining whole-brain fiber connections. Fiber tracking includes two categories: deterministic and probabilistic. Deterministic fiber tracking assumes that there is a unique fiber orientation estimate in each voxel, providing a single path from each seed voxel; probabilistic fiber tracking takes into account imaging noise, artifacts, and model errors, and generates a large set of possible trajectories from seed voxels. Trajectories with higher density are considered to have a higher probability of connection.

[0054] Anatomical brain atlas: a brain partition map used to analyze structural T1-weighted magnetic resonance images. Brain regions are often divided according to structure, function, etc., and different brain maps have different division rules. Among the commonly used brain maps, the Desikan-Killiany atlas contains 68 cortical partitions, which retains the anatomical positions of the macroscopic sulci and gyri; the Brainnetome atlas contains 246 cortical partitions, which is a very detailed cortical anatomical partition map, providing detailed anatomical patterns at the subregional level; the subcortical partition map often uses the atlas proposed by Fischl et al., which contains 44 partitions; the white matter partition map often uses the JHU atlas or ICBM atlas.

[0055] Superficial white matter and deep white matter: The cerebral white matter is divided into superficial and deep parts according to whether it is close to the cortex. The superficial white matter is close to the cortex, and it contains a large number of short-range U-shaped fibers that connect adjacent or nearby gyri; it is the last myelinated white matter area, and the oligodendrocytes in it are more sensitive to metabolic damage; in addition, the superficial white matter contains the highest density of interstitial cells in the white matter, which may have neurofibrillary tangles. The deep white matter is the white matter area far away from the cortex, in which the fibers are divided into three types: communication fibers, commissural fibers, and projection fibers. Communication fibers connect different anatomical areas of the ipsilateral hemisphere; commissural fibers connect the cortices of the left and right hemispheres, mainly forming the corpus callosum; projection fibers connect the cerebral cortex and subcortical structures, and most of the fibers pass through the internal capsule and are projected radially to different functional areas of the cerebral cortex.

[0056] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the accompanying drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0057] Figure 1 FIG. 1 is a flow chart of a method for automatically segmenting white matter fiber bundles by integrating anatomical priors in one embodiment of the present invention. Figure 1 As shown, the method for automatic segmentation of brain white matter fiber bundles at least includes steps S10 to S30.

[0058] Step S10: Acquire the whole-brain fiber point coordinates and the structural T1 magnetic resonance image, determine the superficial white matter fibers and the deep white matter fibers based on the whole-brain fiber point coordinates, and generate an anatomical brain partition map based on the structural T1 magnetic resonance image.

[0059] In this step, the whole-brain fiber point coordinates and structural T1 magnetic resonance images of the subject are first obtained, and then the fibers are classified based on the obtained whole-brain fiber point coordinates to obtain the superficial white matter fibers and deep white matter fibers of the subject; and the anatomical brain partition map of the subject is generated based on the obtained structural T1 magnetic resonance image of the subject.

[0060] In one embodiment, superficial white matter fibers and deep white matter fibers are determined based on the whole-brain fiber point coordinates, including: inputting the whole-brain fiber point coordinates into a trained fiber classification model for pre-classification to obtain superficial white matter fiber pre-classification results and deep white matter fiber pre-classification results; inputting each superficial white matter fiber in the superficial white matter fiber pre-classification results into a fiber filtering model to obtain pseudo superficial white matter fibers; and marking the pseudo superficial white matter fibers as deep white matter fibers.

[0061] In this embodiment, the classification of superficial white matter fibers and deep white matter fibers of the subject can be completed based on the trained fiber classification model and fiber filtering model, and in the pre-training process of the fiber classification model and the fiber filtering model, a sample data set needs to be constructed. The exemplary sample data set includes 100 sample data, each of which includes fiber data and annotation data, and the 100 sample data are further divided into 5 groups, each of which contains 20 sample data. Specifically, each sample data includes fiber point coordinates, fiber labels, label names, and structural T1 magnetic resonance images.

[0062] When classifying fibers based on the fiber classification model, the first stage is to divide the superficial white matter fibers into deep white matter fibers. In the second stage, the abnormal fibers in the superficial white matter fibers obtained in the first stage are filtered out and classified into deep white matter fibers. Finally, superficial white matter fiber data with 198 labels and deep white matter fiber data with 602 labels are obtained.

[0063] Furthermore, the fiber classification model is a neural network used in the first stage to divide superficial white matter fibers and deep white matter fibers. It includes an encoder and a decoder composed of a multi-layer perceptron. The input of the network is the fiber coordinate points of the whole brain, and the output is the category to which each fiber belongs (0 represents superficial white matter, and 1 represents deep white matter). The fiber filtering model is a neural network used in the second stage to filter false positive superficial white matter fibers. It is similar to the neural network in the first stage, but the input of the neural network in the second stage is the fiber coordinate points corresponding to the superficial white matter fibers obtained in the first stage, and the output of the neural network in the second stage is the category to which the fiber belongs (0 represents real superficial white matter fibers, and 1 represents false positive superficial white matter fibers that need to be filtered out).

[0064] For the structural T1 magnetic resonance images in the 100 sample data, they can be processed first to obtain the average T1 image, and the Brainnetome brain atlas is used for cerebral cortical partitioning, and the atlas proposed by Fischl et al. is used for subcortical partitioning to anatomize brain partitioning, and an anatomical segmentation map containing a total of 290 brain regions is obtained. Since the brain atlas also contains non-gray matter anatomical areas such as the left white matter, right white matter, and white matter high signals of the brain, the corresponding labels are set to 0 when writing the program again, and finally a partition map containing 286 brain regions is obtained, which is saved as gm_parc.nii.gz. This partition map provides a very detailed brain anatomical brain region division, which can provide important anatomical guidance for fiber bundle segmentation and greatly improve the accuracy of fiber bundle segmentation.

[0065] In addition, when implementing this step based on a computer, you can first configure the environment, for example, you can use the python3 programming language and the pytorch library that supports GPU acceleration; first download and install the ubuntu18.04 system, and install cuda and cudnn related to GPU acceleration, install anaconda3, create a virtual environment, and install the gpu version of pytorch and other necessary third-party libraries such as numpy, antspy, etc. in the virtual environment. In addition, Mrtrix3 and Freesurfer need to be installed for the generation of anatomical brain partition maps.

[0066] Step S10 is a data preprocessing step. Before calculating the fiber anatomical feature descriptor, the whole-brain fibers with an appropriate number of fibers and their annotations, as well as the anatomical brain region segmentation map of each sample are obtained, and the foundation for calculating the anatomical feature descriptor is laid.

[0067] Step S20: Determine the individual-level anatomical feature descriptor of each fiber based on the superficial white matter fibers, the deep white matter fibers and the anatomical brain partition map, and determine the cluster-level anatomical feature descriptor corresponding to each fiber based on the individual-level anatomical feature descriptor of each superficial white matter fiber and the individual-level anatomical feature descriptor of each deep white matter fiber.

[0068] In this step, anatomical feature descriptors are obtained based on the superficial white matter fibers, deep white matter fibers, and anatomical brain partition maps obtained in step S10. The anatomical feature descriptors include individual-level anatomical feature descriptors and cluster-level anatomical feature descriptors, which realize the digital representation of the relationship between anatomical brain partitions and fibers. Specifically, the anatomical feature descriptors of superficial white matter fibers are obtained based on superficial white matter fibers and anatomical brain partition maps, and the anatomical feature descriptors of deep white matter fibers are obtained based on deep white matter fibers and anatomical brain partition maps.

[0069] Exemplarily, the individual-level anatomical feature descriptor of each fiber is determined based on the superficial white matter fibers, deep white matter fibers and the anatomical brain partition map, including: performing a 3D affine transformation on each of the superficial white matter fibers and deep white matter fibers, mapping each fiber point of each fiber to the brain partition of the anatomical brain partition map, and obtaining the anatomical distribution of each fiber point of each fiber; and obtaining the individual-level anatomical feature descriptor of each fiber based on the anatomical distribution of the fiber points of each fiber.

[0070] Based on the individual-level anatomical feature descriptors of each superficial white matter fiber and the individual-level anatomical feature descriptors of each deep white matter fiber, the cluster-level anatomical feature descriptors corresponding to each fiber are determined respectively, including: hierarchically clustering the superficial white matter fibers and the deep white matter fibers based on the individual-level anatomical feature descriptors of each superficial white matter fiber and the individual-level anatomical feature descriptors of each deep white matter fiber; determining the number of clusters corresponding to the superficial white matter fibers and the number of clusters corresponding to the deep white matter fibers; determining the cluster to which each fiber belongs based on the number of clusters corresponding to the superficial white matter fibers, the number of clusters corresponding to the deep white matter fibers and the hierarchical clustering result, and determining the cluster-level anatomical feature descriptors corresponding to each fiber based on the cluster to which each fiber belongs.

[0071] In the following embodiment, the process of obtaining the anatomical feature descriptor of the superficial white matter fiber is taken as an example to explain this step in detail:

[0072] When obtaining the individual-level anatomical feature descriptor, firstly, the whole brain fiber is subjected to 3D affine transformation, and the fiber streamlines are mapped to the anatomical brain partition map; in this embodiment, each fiber has 15 points, each point is mapped to a specific brain partition (also called anatomical brain region), and each brain partition has a different number of points. Therefore, the anatomical distribution of the points of each fiber can be obtained, which is recorded as the individual-level anatomical feature descriptor I.

[0073] When obtaining cluster-level anatomical feature descriptors, hierarchical clustering is performed on the obtained individual-level anatomical feature descriptors I. The hierarchical clustering results can be used as rough fiber bundles clustered according to anatomical information, so that the following step S30 can obtain a fine classification result of white matter fiber bundles based on the rough classification of fiber bundles obtained in this step.

[0074] Exemplarily, in order to complete the pre-training of the model, 100 sample data are determined in step S10. At this time, the individual-level anatomical feature descriptors of the superficial white matter fibers corresponding to each sample data can be determined respectively, and then the cluster-level anatomical feature descriptors are determined based on the individual-level anatomical feature descriptors corresponding to the 100 sample data. Specifically, since the number of brain regions in the anatomical brain region map sample is 286 and the number of points per fiber is 15, the individual-level anatomical feature descriptor I is a sparse matrix; at this time, the individual-level anatomical feature descriptors I of each sample are added to obtain the overall anatomical feature descriptor S of all 100 sample data. Exemplarily, I n Represents the individual-level anatomical feature descriptor corresponding to the sample data n. Further calculate the Euclidean distance between each row in S: S i,k represents the data corresponding to the kth brain region in the i-th row in S, S j,k represents the data corresponding to the kth brain region in the jth row of S; d ij represents the distance between the i-th row and the j-th row in S; based on this formula, two initial clusters C can be obtained i and C j Further calculations C i and C j The distance between the two clusters is used for the next step of clustering. In the hierarchical clustering process, the non-weighted group average method in hierarchical clustering is adopted to recursively merge the rows in two clusters into a new cluster according to the distance between the rows in the two clusters; C i and C j The distance between |C i | and |C j | respectively represent C i and C j The number of rows in D ik and D jk Respectively represent C i The kth row and C j The distance between the kth row in . Repeat the above steps until all fibers belong to a single cluster, generating a clustering hierarchical tree. After the clustering hierarchical tree is generated, specify the threshold to cut and generate the required number of clusters. In the above-mentioned step of obtaining the anatomical feature descriptor of the superficial white matter fibers, the threshold is set to 8, and the superficial white matter fibers are divided into 8 clusters, which roughly correspond to the fiber bundles in the frontal lobe, frontal-parietal lobe, parietal lobe, parietal-occipital lobe, parietal-temporal lobe, occipital lobe, occipital-temporal lobe and temporal lobe regions; at this time, the cluster to which each fiber belongs is marked to obtain the cluster-level anatomical feature descriptor G of each fiber.

[0075] Further, the anatomical feature descriptors of the deep white matter fibers are obtained by a method similar to the method for obtaining the anatomical feature descriptors of the superficial white matter fibers. After the clustering hierarchical tree corresponding to the deep white matter fibers is obtained, the number of clusters corresponding to the deep white matter fibers is determined. For example, the deep white matter fibers can be divided into 34 clusters.

[0076] For example, in the rough classification result of the white matter fiber bundle concerned in one embodiment, the superficial white matter fibers are divided into 8 clusters, and the deep white matter fibers are divided into 34 clusters. The specific clusters are shown in the following table:

[0077]

[0078]

[0079]

[0080] Step S30: inputting the whole-brain fiber point coordinates, individual-level anatomical feature descriptors, and cluster-level anatomical feature descriptors into a trained fiber bundle segmentation model, and obtaining classification results of each fiber bundle based on the trained fiber bundle segmentation model.

[0081] After obtaining the individual-level anatomical feature descriptors and cluster-level anatomical feature descriptors corresponding to each fiber based on the above step S20, the whole-brain fiber point coordinates, individual-level anatomical feature descriptors and cluster-level anatomical feature descriptors are further input into the trained fiber bundle segmentation model to obtain a detailed classification result of the fiber bundle.

[0082] like Figure 3 As shown, the fiber bundle segmentation model includes a point cloud encoder module, a first embedding layer, a second embedding layer and a decoder module. That is, the fiber bundle segmentation model first performs local feature encoding and global feature encoding on the whole brain fiber point coordinates based on the point cloud encoder, and fuses the local feature encoding and the global feature encoding to obtain feature encoding; inputs the individual level anatomical feature descriptor into the first embedding layer to obtain the first embedded code; inputs the cluster level anatomical feature descriptor into the second embedding layer to obtain the second embedded code; fuses the feature code, the first embedded code and the second embedded code to obtain a fused code; and inputs the fused code into the decoder to obtain the classification results of each fiber bundle. In this embodiment, the point cloud segmentation network is combined with the fiber geometry information "point coordinates" input, and the anatomical feature descriptor is combined with the point cloud feature through the embedding layer, and finally passes through the decoder together to obtain a fine segmentation result. Among them, the shallow white matter fibers are finally segmented into 198 categories, and the deep white matter fibers are finally segmented into 602 categories, realizing the segmentation from coarse to fine.

[0083] In the above embodiment, the encoder and decoder in the fiber bundle segmentation model can be composed of a variety of neural network modules, wherein the encoder uses an advanced point cloud processing module, which is different from the fiber classification model in the preprocessing, and directly reduces the dimension of all points on a fiber into a global feature. In the present application, a more advanced point cloud convolution operation is adopted, which can consider the spatial correlation of points on a fiber, so that the feature of a fiber is represented as a splicing of a global feature and each point feature. In addition, for the anatomical feature descriptors at the individual and cluster levels, embedding layers of different dimensions are used to map the high-dimensional sparse matrix into a low-dimensional continuous vector, and the feature vector is spliced ​​with the feature obtained by the point cloud convolution. In the decoder, a multi-layer perceptron is used for classification, and the features obtained by the encoder are sent to the multi-layer perceptron, and finally through a probability regression layer, the probability of each fiber belonging to the category is output; finally, the maximum probability is selected to obtain the category to which each fiber belongs, so as to achieve the purpose of fine and accurate segmentation of the fiber bundle.

[0084] In order to obtain a trained fiber bundle segmentation model, the automatic segmentation method of white matter fiber bundles may also include the following steps: constructing an initial fiber bundle segmentation model; constructing a model loss function and a sample data set; and pre-training the initial fiber bundle segmentation model based on the model loss function and the sample data set to obtain a trained fiber bundle segmentation model.

[0085] Furthermore, the loss function is:

[0086]

[0087] Where M represents the number of categories into which the fiber bundles are divided. For superficial white matter fibers, M is 198, and for deep white matter fibers, M is 602; ic is 0 or 1. When the predicted category of fiber i is the same as the true category, y ic is 1, when the category prediction result of fiber i is different from the true category result, y ic is 0; N represents the total number of fibers, p ic represents the probability that fiber i belongs to fiber bundle category c, which is obtained from the decoder output of the fiber bundle segmentation model.

[0088] For example, during the model training process, the fiber and anatomical feature descriptor data of 100 subjects were obtained to obtain a white matter fiber bundle segmentation data set, which was divided into 5 equal parts, and the network training was performed using a five-fold cross-validation method to obtain a segmentation model with strong generalization and good robustness. Specifically, the learning rate was set to 0.0001, the training batch size was set to 2048, the Adam optimizer was used without setting weight decay, and the network training algebra was 500 times.

[0089] For the automatic segmentation method of brain white matter fiber bundles integrated with anatomical priors disclosed in the above embodiment, the accuracy and F1 score of each fiber bundle can be calculated based on the following method to evaluate the method:

[0090]

[0091]

[0092]

[0093]

[0094] Among them, TP is the number of positive examples predicted correctly by the model, TN is the number of negative examples predicted correctly by the model, FP is the number of negative examples predicted incorrectly by the model, and FN is the number of positive examples predicted incorrectly by the model.

[0095] Based on the model evaluation indicators calculated by the above method, it can be found that the accuracy and F1 score of the fiber bundle segmentation method proposed in the present invention can be improved by 2%-7% compared with other existing methods. In order to further test the effectiveness of the anatomical feature descriptors used in the method of the present invention, ablation experiments were conducted. The results showed that the individual-level anatomical feature descriptors and cluster-level anatomical feature descriptors proposed in this method are both effective, and the model with individual and cluster-level anatomical feature descriptions can improve the accuracy by about 2%.

[0096] Figure 2 FIG. 4 is a flow chart of a method for automatically segmenting brain white matter fiber bundles by integrating anatomical priors in another embodiment of the present invention. Figure 2 As shown, the system corresponding to the method may specifically include a data preprocessing module, an individual level anatomical feature descriptor acquisition module, a cluster level anatomical feature descriptor acquisition module, a fiber bundle segmentation model module, and a parameter and result export module.

[0097] In the data preprocessing module, the fiber data is divided into superficial white matter fibers and deep white matter fibers of the brain based on the pre-trained fiber classification neural network model, and the outliers of the divided superficial white matter fibers are further filtered, and the sample average structural T1 magnetic resonance image is processed to obtain the anatomical brain partition map. In the individual-level anatomical feature descriptor acquisition module, the affine transformation method is used to map each fiber to the corresponding anatomical brain partition map, and the number of fiber points in each anatomical brain region is calculated to obtain the individual-level anatomical feature descriptor. In the cluster-level anatomical feature descriptor acquisition module, the hierarchical clustering algorithm is used to cluster the individual-level anatomical feature descriptors, so that each fiber is clustered into a cluster to obtain the cluster-level anatomical feature descriptor. In the fiber bundle segmentation model module, based on the point cloud neural network model, a neural network is designed to input the fiber point coordinate geometry information and the individual-level and cluster-level anatomical feature descriptors to obtain the specific fiber classification results. For the parameter and result export module, the pre-trained segmentation model can be exported for fiber bundle segmentation of different test data; in addition, the accuracy, F1 score, cross entropy loss function, etc. during the training process can also be output simultaneously.

[0098] Through the above embodiments, it can be found that the automatic segmentation method of white matter fiber bundles integrating anatomical priors of the present invention includes three parts: image preprocessing, anatomical feature descriptor calculation and segmentation model training. These three parts are connected and form a complete fiber bundle segmentation scheme in sequence. It is a difficult problem to construct feature descriptors from the anatomical brain partition map and fiber data of a single sample. The present invention uses affine transformation to obtain the distribution of sampling points on a single fiber in the brain area, that is, the distribution at the individual level. Secondly, hierarchical clustering is used to preliminarily classify the fibers on the individual level distribution to obtain the anatomical features at the cluster level; that is, the distribution at the individual level and the cluster level constitute the anatomical feature descriptor. In addition, the present application also regards the fiber geometry information as a point cloud, relying on the advanced artificial intelligence algorithm-point cloud neural network to accelerate the segmentation, and at the same time embeds the obtained anatomical feature descriptors, which solves the problem of large computing resource occupation and slow computing speed caused by the huge amount of data in traditional methods. In addition, the present application fuses the anatomical brain partition map based on the structural T1 magnetic resonance image and the whole-brain fiber data obtained by diffusion magnetic resonance, and proposes a description of the anatomical brain region distribution of single fibers and cluster-level fibers, which can accurately and effectively represent the anatomical characteristics of the fibers; and the present application embeds the anatomical distribution of each fiber and the anatomical distribution of the cluster level into a neural network based on the fiber geometric point cloud, respectively, and trains the network to obtain a white matter fiber bundle segmentation network model, thereby greatly improving the automatic segmentation accuracy of the fiber bundle. That is, the automatic segmentation method of white matter fiber bundles disclosed in the present application can accurately segment fiber bundles connecting different brain regions, and further segment large fiber bundles into small clusters, providing a basis for subsequent brain research and medical diagnosis, and has broad biomedical application prospects; the automatic segmentation method of white matter fiber bundles in the present application can segment whole-brain fibers of different data sets without modifying the underlying training parameters, and is universal for whole-brain fiber data of patients such as Alzheimer's disease and mild cognitive impairment.

[0099] Correspondingly, the present invention also discloses an automatic brain white matter fiber bundle segmentation system integrating anatomical priors, comprising a processor, a memory and a computer program stored in the memory, wherein the processor is used to execute the computer program, and when the computer program is executed, the system implements the steps of the method described in any of the above embodiments.

[0100] The embodiment of the present invention also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the aforementioned edge computing server deployment method are implemented. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the technical field.

[0101] It should be understood by those skilled in the art that the exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.

[0102] It should be clear that the present invention is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present invention.

[0103] In the present invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with features of other embodiments or replace features of other embodiments.

[0104] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the embodiments of the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for automatic segmentation of brain white matter fiber bundles integrating anatomical priors, characterized in that: The method comprises the following steps: Acquire whole-brain fiber point coordinates and a structural T1 magnetic resonance image, determine superficial white matter fibers and deep white matter fibers based on the whole-brain fiber point coordinates, and generate an anatomical brain partition map based on the structural T1 magnetic resonance image; Determine the individual level anatomical feature descriptor of each fiber based on the superficial white matter fibers, the deep white matter fibers and the anatomical brain partition map, and determine the cluster level anatomical feature descriptor corresponding to each fiber based on the individual level anatomical feature descriptor of each superficial white matter fiber and the individual level anatomical feature descriptor of each deep white matter fiber; Inputting the whole-brain fiber point coordinates, individual-level anatomical feature descriptors, and cluster-level anatomical feature descriptors into a trained fiber bundle segmentation model, and obtaining classification results of each fiber bundle based on the trained fiber bundle segmentation model; The individual level anatomical feature descriptors of each fiber are determined based on the superficial white matter fibers, deep white matter fibers and anatomical brain partition map, including: Performing 3D affine transformation on each of the superficial white matter fibers and the deep white matter fibers, mapping each fiber point of each fiber to the brain partition of the anatomical brain partition map, and obtaining the anatomical distribution of each fiber point of each fiber; Obtaining an individual-level anatomical feature descriptor of each fiber based on the anatomical distribution of the fiber points of each fiber; Determining the cluster level anatomical feature descriptors corresponding to each fiber based on the individual level anatomical feature descriptors of each superficial white matter fiber and the individual level anatomical feature descriptors of each deep white matter fiber, respectively, includes: hierarchically clustering the superficial white matter fibers and the deep white matter fibers based on the individual level anatomical feature descriptors of the superficial white matter fibers and the individual level anatomical feature descriptors of the deep white matter fibers; Determining the number of clusters corresponding to the superficial white matter fibers and the number of clusters corresponding to the deep white matter fibers; The cluster to which each fiber belongs is determined based on the number of clusters corresponding to the superficial white matter fibers, the number of clusters corresponding to the deep white matter fibers, and the hierarchical clustering result, and the cluster-level anatomical feature descriptor corresponding to each fiber is determined based on the cluster to which each fiber belongs.

2. The method for automatic segmentation of brain white matter fiber bundles integrating anatomical priors according to claim 1, characterized in that: The fiber bundle segmentation model comprises a point cloud encoder module, a first embedding layer, a second embedding layer and a decoder module; and / or, The anatomical brain parenchyma map includes 286 brain parenchyma regions.

3. The method for automatic segmentation of brain white matter fiber bundles integrating anatomical priors according to claim 2, characterized in that: The whole-brain fiber point coordinates, individual-level anatomical feature descriptors, and cluster-level anatomical feature descriptors are input into the trained fiber bundle segmentation model, and the classification results of each fiber bundle are obtained based on the trained fiber bundle segmentation model, including: Performing local feature encoding and global feature encoding on the whole-brain fiber point coordinates based on a point cloud encoder, and fusing the local feature encoding and the global feature encoding to obtain a feature encoding; Inputting the individual level anatomical feature descriptor into the first embedding layer to obtain a first embedding code; Inputting the cluster-level anatomical feature descriptor into the second embedding layer to obtain a second embedding code; Fusing the feature code, the first embedded code and the second embedded code to obtain a fused code; The fusion code is input into the decoder to obtain the classification result of each fiber bundle.

4. The method for automatic segmentation of brain white matter fiber bundles integrating anatomical priors according to claim 2, characterized in that: The method further comprises: Construct an initial fiber bundle segmentation model; Construct model loss function and sample data set; The initial fiber bundle segmentation model is pre-trained based on the model loss function and the sample data set to obtain a trained fiber bundle segmentation model.

5. The method for automatic segmentation of brain white matter fiber bundles integrating anatomical priors according to claim 4, characterized in that: The loss function is: Where M represents the number of categories into which the fiber bundles are divided; y ic is 0 or 1. When the predicted category of fiber i is the same as the true category, y ic is 1, when the category prediction result of fiber i is different from the true category result, y ic is 0; N represents the total number of fibers, p ic represents the probability that fiber i belongs to fiber bundle category c.

6. The method for automatic segmentation of brain white matter fiber bundles integrating anatomical priors according to claim 1, characterized in that: Determining superficial white matter fibers and deep white matter fibers based on the whole-brain fiber point coordinates includes: Inputting the whole-brain fiber point coordinates into a trained fiber classification model for pre-classification to obtain a superficial white matter fiber pre-classification result and a deep white matter fiber pre-classification result; Inputting each superficial white matter fiber in the superficial white matter fiber pre-classification result into a fiber filtering model to obtain pseudo superficial white matter fibers; The pseudo superficial white matter fibers were labeled as deep white matter fibers.

7. An automatic brain white matter fiber bundle segmentation system integrating anatomical priors, comprising a processor, a memory, and a computer program stored in the memory, characterized in that: The processor is used to execute the computer program. When the computer program is executed, the system implements the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images