Multi-level segmentation method of brain based on diffusion weighted imaging

Abstract

The application provides a brain multi-level segmentation method based on diffusion weighted imaging, acquires diffusion weighted imaging data and extracts fiber direction distribution function features, uses the fiber direction distribution function features to sequentially perform voxel-level tissue semantic segmentation, partial volume effect correction of tissue boundary regions, division and segmentation of brain anatomic structures of interest, and subregion division of thalamic regions based on the fiber direction distribution function features and thalamic probability atlas prior, thereby constructing a multi-level segmentation process based on diffusion weighted imaging data and sequentially progressing, effectively overcoming the technical defects of the prior art that excessively rely on structural magnetic resonance imaging, resulting in low segmentation accuracy of deep subcortical nuclei, and achieving the beneficial effects of improving the segmentation accuracy of deep subcortical nuclei and realizing multi-level integrated segmentation.

Description

Technical Field

[0001] This invention relates to the field of medical imaging processing technology, and in particular to a method for multi-level brain segmentation based on diffusion-weighted imaging. Background Technology

[0002] Diffusion-weighted imaging (DWI) is a non-invasive magnetic resonance imaging technique that utilizes the Brownian motion properties of water molecules in the brain to analyze tissue composition and brain structure. It quantifies diffusion anisotropy within tissues, thus non-invasively revealing the orientation and integrity of white matter fiber bundles in vivo. Based on DWI data, a constrained spherical deconvolution algorithm can reconstruct the fiber orientation distribution function. This function represents the distribution information of multiple fiber orientations within a voxel in the form of spherical probability density, overcoming the limitation of traditional tensor models in representing intersecting fibers. Brain segmentation is an automated process that assigns specific anatomical labels to each voxel in neuroimaging data. Depending on the segmentation granularity, it can be divided into two levels: tissue segmentation and anatomical partitioning. Tissue segmentation aims to divide intracranial structures into three main categories: gray matter, white matter, and cerebrospinal fluid. Anatomical partitioning further divides the brain into specific anatomical regions of interest using standard atlases, such as the globus pallidus, caudate nucleus, putamen, and thalamus. Brain structural segmentation has significant clinical value in the diagnosis, treatment, and prognostic assessment of neurological diseases and is widely used in the imaging assessment of neurodegenerative diseases such as Parkinson's disease and Alzheimer's disease.

[0003] Existing brain structure segmentation methods primarily rely on structural magnetic resonance imaging (SMRI), and their technical approaches can be broadly categorized into three types: first, atlas-based methods, which map anatomical labels from a standard brain space to individual brain spaces through nonlinear registration; second, traditional machine learning-based methods, such as using clustering algorithms to classify tissues based on voxel grayscale features; and third, deep learning-based methods, which utilize convolutional neural networks to extract image features end-to-end for high-precision automated segmentation. However, in deep subcortical nuclei such as the globus pallidus and thalamus, SMRI often exhibits low tissue contrast and blurred boundaries, significantly increasing the difficulty of segmentation. In contrast, diffusion-weighted imaging (DWI) not only contains rich anatomical information, but its average b1000 image and fractional anisotropy image show significantly better tissue contrast than SWI in regions such as the globus pallidus, providing a new technical approach for brain structure segmentation using DWI.

[0004] Therefore, how to utilize the high tissue contrast characteristics of diffusion-weighted imaging and the fiber bundle orientation information contained in the fiber orientation distribution function to get rid of the dependence of brain structure segmentation on structural magnetic resonance imaging and achieve multi-level and refined brain segmentation from the tissue level to the anatomical structure level and then to the subregion level is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This invention provides a brain multi-level segmentation method based on diffusion-weighted imaging, which solves the technical defects of existing technologies that rely on structural magnetic resonance imaging, resulting in low segmentation accuracy of deep subcortical nuclei and the inability to achieve multi-level segmentation from the tissue level to the anatomical structure level and then to the subregion level.

[0006] On one hand, the present invention provides a method for multi-level brain segmentation based on diffusion-weighted imaging, characterized in that it includes: Diffusion-weighted imaging data of the brain to be detected is acquired, and after preprocessing the diffusion-weighted imaging data, diffusion features are obtained, wherein the diffusion features include at least fiber orientation distribution function features. Based on the aforementioned diffusion characteristics, voxel-level tissue semantic segmentation of the brain to be detected is performed to obtain initial tissue segmentation results including three types of tissues: gray matter, white matter, and cerebrospinal fluid. Partial volume effect correction is applied to the tissue boundary regions in the initial tissue segmentation results to obtain brain tissue segmentation results; The preprocessed diffusion-weighted imaging data is divided into regions of whole-brain anatomical structures to obtain the regions of interest corresponding to the target anatomical structures. Based on the aforementioned diffusion characteristics, the region of interest is segmented into anatomical structures to obtain segmentation results of multiple brain anatomical structures; Based on the fiber orientation distribution function characteristics and the prior information of the predefined thalamic probability map, the segmented thalamic regions are divided into subregions to obtain segmentation results of multiple thalamic subregions.

[0007] Optionally, diffusion-weighted imaging data of the brain to be detected is acquired, and the diffusion-weighted imaging data is preprocessed to obtain diffusion features, including: Head motion correction, eddy current correction, and gradient nonlinear correction were performed on the raw diffusion-weighted imaging data of the brain to be examined to obtain the registered diffusion-weighted imaging data. The diffusion-weighted imaging data is fitted using a multi-voxel constrained spherical deconvolution algorithm to reconstruct the fiber orientation distribution function of each voxel, and stored in the form of spherical harmonic coefficients to obtain diffusion characteristics.

[0008] Optionally, based on the diffusion features, voxel-level tissue semantic segmentation is performed on the brain to be detected to obtain initial tissue segmentation results including three types of tissues: gray matter, white matter, and cerebrospinal fluid, including: The fiber orientation distribution function features are input into a pre-trained semantic segmentation neural network, which outputs probability maps of three types of tissues—gray matter, white matter, and cerebrospinal fluid—using voxels as units. An argmax operation is performed on the probability maps of the three types of tissues to assign the category with the highest probability value to each voxel, thus obtaining the initial tissue segmentation result.

[0009] Optionally, partial volume effect correction is applied to the tissue boundary regions in the initial tissue segmentation result to obtain the brain tissue segmentation result, including: From the initial tissue segmentation results, voxels located at the boundary between gray and white matter are identified, and a boundary voxel set is generated. The initial tissue segmentation results are reconstructed in three dimensions and resampled to a target space with a resolution higher than that of the original diffusion-weighted imaging data. By counting the number of high-resolution voxel categories included in each original voxel, the initial tissue volume ratio of each original voxel is calculated to obtain an initial partial volume estimation map. The initial partial volume estimation map and the boundary voxel set are input into the iterative optimization model, and the tissue volume ratio of each boundary voxel is expressed as the weighted sum of the tissue volume ratios of all voxels in the 26 neighborhoods surrounding the boundary voxel. Based on the weighted sum relationship, the tissue volume ratio of all boundary voxels is updated to obtain the updated partial volume estimation map. The updated partial volume estimation map is used as the input for the next iteration. The weighted sum update step is repeated until the difference between the two iteration results of all boundary voxels is less than a preset threshold, and the brain tissue segmentation result is output.

[0010] Optionally, the preprocessed diffusion-weighted imaging data is divided into regions representing whole-brain anatomical structures to obtain the regions of interest corresponding to the target anatomical structures, including: Extract the average diffusion-weighted image from the preprocessed diffusion-weighted imaging data; The average diffusion-weighted image is input into a pre-trained coarse segmentation network to generate coarse segmentation label maps including multiple anatomical regions; wherein the coarse segmentation network adopts the SynthSeg network, and the multiple anatomical regions include at least 23 brain anatomical structures; From the coarse segmentation label map, the minimum bounding cube is extracted for each target anatomical structure to be finely segmented, and a predetermined number of voxels are extended outward as the boundary to obtain the region of interest corresponding to each target anatomical structure.

[0011] Optionally, by combining the diffusion features, the region of interest is segmented into anatomical structures to obtain segmentation results of multiple brain anatomical structures, including: Based on the region of interest corresponding to the target anatomical structure, the corresponding local diffusion imaging data block is cropped from the diffusion-weighted imaging data, and the local fiber orientation distribution function feature block corresponding to the region of interest is cropped from the fiber orientation distribution function features; The local diffusion imaging data block and the local fiber orientation distribution function feature block are concatenated and input into a pre-trained 3D U-Net fine segmentation network to output a segmentation probability map of the target anatomical structure. The segmentation probability map is fused back into the whole-brain space according to the spatial location, and the whole-brain segmentation label map is obtained through argmax operation, which serves as the segmentation result of the multiple brain anatomical structures; the segmentation result includes at least the segmentation mask of the left thalamus and the right thalamus.

[0012] Optionally, the local diffusion imaging data block and the local fiber orientation distribution function feature block are concatenated and input into a pre-trained 3D U-Net fine segmentation network to output a segmentation probability map of the target anatomical structure, including: Nine independent 3D U-Net models were trained using a 9-fold cross-validation method. The local diffusion imaging data block and the local fiber orientation distribution function feature block are respectively input into 9 independent 3D U-Net models to obtain 9 independent segmentation probability maps; The nine segmentation probability maps are fused by majority voting at the voxel level to generate a fused segmentation probability map.

[0013] Optionally, when training the 3D U-Net fine segmentation network, a combined data augmentation strategy is applied, which includes at least one of the following: random flipping, random rotation, adding random Gaussian noise, random intensity shift, and random contrast adjustment.

[0014] Optionally, based on the fiber orientation distribution function characteristics and predefined thalamic probability map prior information, the segmented thalamic region is further subdivided to obtain segmentation results for multiple thalamic subregions, including: Obtain a predefined thalamic subregion probability map template. The thalamic subregion probability map template is stored in the form of a deformable grid. Each grid node of the thalamic subregion probability map template includes a prior probability vector corresponding to multiple thalamic subregions. The prior probability vector is used to characterize the initial probability that the grid node belongs to each thalamic subregion. From the segmentation results of multiple brain anatomical structures, the segmentation masks corresponding to the left and right thalamus were extracted respectively; Based on the segmentation mask of the left and right thalamus, a local fiber direction distribution function feature tensor corresponding only to the thalamic region is cropped from the whole brain fiber direction distribution function features. The local feature tensor contains only the fiber direction distribution information of each voxel in the thalamic region. The feature tensor of the local fiber orientation distribution function is modeled as a multivariate Gaussian distribution to obtain the likelihood model; Construct a Bayesian inference objective function, which includes a grid deformation prior term constraining the rationality of grid deformation, a model parameter prior term regularizing the model parameters, and a joint likelihood term integrating the thalamic subregion probability map prior information with the likelihood model. The objective function is iteratively solved using an alternating optimization strategy, alternately updating the mesh deformation parameters and the likelihood model parameters until the objective function converges, thus obtaining the likelihood model parameters. Based on the likelihood model parameters and the prior information of the thalamic subregion probability map template, the posterior probability map of each voxel in the thalamic region belonging to each thalamic subregion is calculated. Thresholding is performed on the posterior probability map, and the thalamic subregion category corresponding to the maximum posterior probability of each voxel is taken as the voxel's classification, thus obtaining the segmentation results of multiple thalamic subregions corresponding to the left and right thalamus respectively.

[0015] Optionally, an alternating optimization strategy is used to iteratively solve the objective function, alternately updating the mesh deformation parameters and the likelihood model parameters until the objective function converges, yielding the likelihood model parameters, including: The expectation-maximization algorithm is used for iterative optimization. In each iteration, based on the current mesh deformation parameters and likelihood model parameters, the joint posterior probability of each voxel belonging to each thalamic subregion and each hybrid component is calculated to obtain the soft assignment weight matrix. Based on the soft-assigned weight matrix, update the mean and variance parameters of the multivariate Gaussian distribution, as well as the weight coefficients of each thalamic subregion on different hybrid components. Repeat the above steps until the likelihood model parameters converge.

[0016] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the brain multi-level segmentation method based on diffusion-weighted imaging as described above.

[0017] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the brain multi-level segmentation method based on diffusion-weighted imaging as described above.

[0018] On the other hand, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the brain multi-level segmentation method based on diffusion-weighted imaging as described above.

[0019] This invention provides a multi-level brain segmentation method based on diffusion-weighted imaging. By acquiring diffusion-weighted imaging data and extracting fiber orientation distribution function features, the method sequentially performs voxel-level tissue semantic segmentation, partial volume effect correction of tissue boundary regions, and region of interest (ROI) division and segmentation of whole-brain anatomical structures using fiber orientation distribution function features and prior thalamic probability maps. Furthermore, based on the fiber orientation distribution function features and prior thalamic probabilistic maps, the method subdivides the thalamic region, thus constructing a multi-level segmentation process entirely based on diffusion-weighted imaging data and progressing sequentially. This effectively overcomes the low segmentation accuracy of deep subcortical nuclei caused by excessive reliance on structural magnetic resonance imaging in existing technologies, achieving the beneficial effects of improving the accuracy of deep subcortical nuclei segmentation and realizing multi-level integrated segmentation. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the brain multi-level segmentation method based on diffusion-weighted imaging provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the gray matter / white matter / cerebrospinal fluid segmentation results provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the 23 brain structure segmentation results provided in this embodiment of the invention; Figure 4 This is a schematic diagram of the thalamic subregion segmentation results provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] Figure 1 This is one of the flowcharts of the brain multi-level segmentation method based on diffusion-weighted imaging provided in the embodiments of the present invention.

[0024] like Figure 1As shown, the brain multi-level segmentation method based on diffusion-weighted imaging provided in this embodiment of the invention includes the following steps: 101. Obtain diffusion-weighted imaging data of the brain to be tested, and obtain diffusion features after preprocessing the diffusion-weighted imaging data.

[0025] The diffusion characteristics include at least the fiber orientation distribution function characteristics.

[0026] Specifically, diffusion-weighted imaging data of the brain to be examined is acquired, and after preprocessing the diffusion-weighted imaging data, diffusion features are obtained, including: Head motion correction, eddy current correction, and gradient nonlinear correction were performed on the raw diffusion-weighted imaging data of the brain to be examined to obtain registered diffusion-weighted imaging data.

[0027] The process involves preprocessing the raw diffusion-weighted imaging data of the brain to be examined to eliminate various artifacts and distortions introduced during acquisition. Specifically, the topup tool in the FSL software package is used to correct distortion in the diffusion-weighted imaging data acquired with bidirectional phase encoding. By estimating the B0 field map caused by magnetic field inhomogeneity, geometric distortion correction is performed on the image. Subsequently, the eddy tool is used to perform head motion correction, eddy current correction, and signal dropout correction caused by subject movement on the corrected data. The eddy tool corrects image misalignment and signal intensity anomalies caused by subject movement or eddy currents by registering each volume of diffusion-weighted image to the first volume of undiffusion-weighted B0 image and using Gaussian process prediction and replacement of outliers. Finally, spatially aligned and signal-intensity-consistent registered diffusion-weighted imaging data is obtained.

[0028] The diffusion-weighted imaging data were fitted using a multi-voxel constrained spherical deconvolution algorithm to reconstruct the fiber orientation distribution function of each voxel, and stored in the form of spherical harmonic coefficients to obtain the diffusion characteristics.

[0029] This study extracts fiber orientation distribution function features from registered diffusion-weighted imaging data to characterize the multi-directional distribution information of white matter fibers within each voxel. A multi-shell voxel constrained spherical deconvolution algorithm implemented in MRtrix3 software is used. This algorithm first fits the multi-shell response function using non-negative least squares based on the signal intensity of different b values ​​(b=0, 1000, 2000, 3000, etc.) in the diffusion-weighted imaging data, automatically estimating the characteristic signal responses of each tissue type (gray matter, white matter, and cerebrospinal fluid) from the data. Then, based on spherical harmonic basis functions and considering the joint constraints of multi-shell data, iterative deconvolution operations are used to solve for the fiber orientation distribution function within each voxel, i.e., the angular probability density distribution stored as a set of spherical harmonic coefficients. This algorithm ensures the non-negativity and sparsity of the solution through constraints, accurately resolving multiple fiber orientations in complex intersecting fiber regions, and finally outputting a set of spherical harmonic coefficients for each voxel, thus obtaining the fiber orientation distribution function features.

[0030] 102. Based on diffusion features, voxel-level tissue semantic segmentation is performed on the brain to be detected to obtain initial tissue segmentation results including three types of tissues: gray matter, white matter, and cerebrospinal fluid.

[0031] Specifically, based on diffusion features, voxel-level semantic segmentation of the brain to be tested is performed to obtain initial tissue segmentation results including three types of tissues: gray matter, white matter, and cerebrospinal fluid, including: The fiber orientation distribution function features are input into a pre-trained semantic segmentation neural network, which outputs probability maps of three types of tissues—gray matter, white matter, and cerebrospinal fluid—using voxels as units. An argmax operation is performed on the probability maps of the three types of tissues to assign the class with the highest probability value to each voxel, thus obtaining the initial tissue segmentation result.

[0032] The extracted fiber orientation distribution function features are input into a pre-trained semantic segmentation neural network to obtain the probability distribution of each voxel belonging to different tissue categories. The semantic segmentation neural network adopts a three-dimensional fully convolutional neural network architecture, specifically 3D U-Net. 3D U-Net includes an encoder path and a decoder path. The encoder path progressively downsamples through multiple three-dimensional convolutions, batch normalization, ReLU activation function, and max pooling operations to extract high-dimensional semantic features from the input feature map. The decoder path restores the feature map resolution through upsampling operations and concatenates and fuses the feature maps of the corresponding layers in the encoder path with the decoder feature maps through skip connections to preserve shallow spatial detail information. The last layer of 3D U-Net uses a 1×1×1 convolution to map the number of channels in the feature map to 3, and uses the Softmax activation function to normalize the output value of each voxel into a probability form, so that the sum of the probability values ​​of each voxel in the three categories of gray matter, white matter, and cerebrospinal fluid is 1, thereby outputting a three-dimensional probability map with the same size as the input image. The three-dimensional probability map contains three channels, corresponding to the tissue category probabilities of gray matter, white matter, and cerebrospinal fluid, respectively.

[0033] Secondly, the 3D probability map is processed to obtain a unique tissue category label for each voxel, resulting in the initial tissue segmentation result. Specifically, the argmax operation is used to compare the three probability values ​​of each voxel along the channel dimension. That is, for each voxel location, the category index corresponding to the maximum probability value among gray matter, white matter, and cerebrospinal fluid is selected as the final category of that voxel, generating a 3D label map of the same size as the input image. Each voxel in the label map takes a value of 0, 1, or 2, representing gray matter, white matter, or cerebrospinal fluid, respectively. The argmax operation is a deterministic decision-making process that does not introduce additional learnable parameters. It can transform the continuous probability values ​​output by the neural network into discrete tissue category labels, forming the initial tissue segmentation result and providing input for subsequent volumetric effect correction of tissue boundary regions.

[0034] 103. Perform partial volume effect correction on the tissue boundary regions in the initial tissue segmentation results to obtain the brain tissue segmentation results.

[0035] Specifically, partial volume effect correction is applied to the tissue boundary regions in the initial tissue segmentation results to obtain the brain tissue segmentation results, including: Identify voxels located at the boundary between gray and white matter from the initial tissue segmentation results, and generate a boundary voxel set.

[0036] Specifically, the marching cubes algorithm is used to reconstruct the three-dimensional surfaces of the gray and white matter labels in the initial tissue segmentation results, and extract the isosurfaces between gray and white matter. Then, the reconstructed triangular mesh is smoothed by Laplacian smoothing to eliminate surface irregularities caused by noise and make the boundary surface smoother and more continuous. All voxels in the whole brain are traversed. If a voxel is traversed by both gray matter isosurfaces and white matter isosurfaces, that is, the voxel contains both gray and white matter tissue types, it is marked as a boundary voxel. Finally, a set including the spatial coordinates of all boundary voxels is obtained.

[0037] The initial tissue segmentation results are reconstructed in three dimensions and resampled to a target space with a resolution higher than that of the original diffusion-weighted imaging data. By counting the number of high-resolution voxel categories included in each original voxel, the initial tissue volume ratio of each original voxel is calculated to obtain an initial partial volume estimation map.

[0038] Specifically, the discrete label maps from the initial tissue segmentation results are used to generate separate 3D surface models for gray matter, white matter, and cerebrospinal fluid using the marching cubes algorithm. Since the surface models consist of continuous triangular facets and lack a fixed resolution, they can be uniformly resampled to any level of detail in the target space as needed to obtain the category assignment for each high-resolution voxel. Subsequently, for each original resolution voxel, the number of gray matter, white matter, and cerebrospinal fluid cells in all high-resolution voxels within it is counted, and each is divided by the total number of high-resolution voxels in the original voxel to obtain the volume ratio of the three tissue types within the original voxel. This forms a three-channel initial partial volume estimation map, with each channel storing the ratio value of one tissue type, and the sum of all ratio values ​​is 1.

[0039] The initial partial volume estimation map and the boundary voxel set are input into the iterative optimization model, and the tissue volume ratio of each boundary voxel is expressed as the weighted sum of the tissue volume ratios of all voxels in the 26 neighborhoods surrounding the boundary voxel.

[0040] Specifically, for each boundary voxel, the original diffusion-weighted imaging feature vector of the boundary voxel is first extracted, and the initial weight coefficients are solved using a quadratic programming method. Then, all weight coefficients are normalized so that the sum of all weight coefficients equals 1, resulting in normalized weights. Finally, the updated tissue proportion of the boundary voxel is calculated as the weighted sum of the current tissue proportions of all voxels within its 26-neighborhood.

[0041] Based on the weighted sum relationship, the tissue volume proportions of all boundary voxels are updated to obtain the updated partial volume estimation map.

[0042] Specifically, based on weighted sum relationships, the tissue volume proportions of all boundary voxels are updated to obtain an updated partial volume estimate map. A synchronous update strategy is employed, meaning that the new proportions are calculated in parallel for all boundary voxels using the partial volume estimate map obtained from the previous iteration. Once all calculations are complete, the old values ​​are replaced uniformly to avoid bias introduced by the update order. After completing one full boundary update, a new partial volume estimate map is obtained.

[0043] The updated partial volume estimation map is used as the input for the next iteration. The weighted sum update step is repeated until the difference between the two iteration results of all boundary voxels is less than a preset threshold, and the brain tissue segmentation result is output.

[0044] The convergence condition is set as follows: for each boundary voxel, calculate the absolute value of the proportional change of the boundary voxel for the same tissue type in the old and new iterations, and take the maximum change among all boundary voxels and all tissue types. If the maximum change is less than a preset threshold, for example, the preset threshold is 0.01, then the iteration is considered to have converged, and the iteration stops. The partial volume estimation map obtained at this time is the final partial volume correction result, such as... Figure 2 As shown, Figure 2 This is a schematic diagram of the gray matter / white matter / cerebrospinal fluid segmentation results. Partial volume correction results are stored as a three-channel image, preserving the precise volume ratios of gray matter, white matter, and cerebrospinal fluid within each voxel.

[0045] 104. The preprocessed diffusion-weighted imaging data is divided into regions of whole-brain anatomical structure to obtain the region of interest corresponding to the target anatomical structure.

[0046] Specifically, the preprocessed diffusion-weighted imaging data is used to divide the whole brain anatomical structures into regions, obtaining the regions of interest corresponding to the target anatomical structures, including: Extract the average diffusion-weighted image from the preprocessed diffusion-weighted imaging data.

[0047] The average diffusion-weighted image is input into a pre-trained coarse segmentation network to generate coarse segmentation label maps that include multiple anatomical regions; the coarse segmentation network uses the SynthSeg network, and the multiple anatomical regions include at least 23 brain anatomical structures.

[0048] Specifically, the coarse segmentation network adopts the SynthSeg network architecture. Coarse segmentation is a deep learning-based image segmentation method capable of multi-class segmentation of new images without nonlinear registration. Specifically, the average diffusion-weighted image is preprocessed, including resampling to the spatial resolution required by the SynthSeg network, intensity normalization to zero mean, and unit variance. The preprocessed image is then input into the SynthSeg network, which outputs a 3D label map of the same size as the input image, where each voxel is assigned an anatomical structure label value. In this embodiment, the coarse segmentation label map covers 32 anatomical regions of the brain, including the left and right cerebellar gray matter, cerebellar white matter, caudate nucleus, putamen, globus pallidus, amygdala, hippocampus, thalamus, lateral ventricles, nucleus accumbens, brainstem, third ventricle, and fourth ventricle. Of these regions, 23 structures will be further processed by the subsequent fine segmentation module.

[0049] From the coarse segmentation label image, the minimum bounding cube is extracted for each target anatomical structure to be finely segmented, and a predetermined number of voxels are extended outward as the boundary to obtain the region of interest corresponding to each target anatomical structure.

[0050] Specifically, for each target structure to be finely segmented, a binary mask corresponding to the target structure is first extracted from the coarse segmentation label image; this is the set of voxels whose label values ​​are equal to the structure's encoding. Then, the minimum bounding cube of the binary mask is calculated, i.e., the minimum and maximum coordinate values ​​of the mask in the x, y, and z directions are obtained, forming a bounding box. To compensate for potential localization errors in coarse segmentation and ensure the target structure is completely contained within the region of interest (ROI), the bounding box boundary is extended outward by a predetermined number of voxels, for example, 5 voxels in each direction. The extended bounding box constitutes the ROI for that target structure. The ROI is cropped from the original diffusion-weighted imaging data as an image patch and used as input for the subsequent fine segmentation module. This process is repeated for all target structures to be finely segmented to obtain their respective ROIs.

[0051] 105. Combining diffusion characteristics, the region of interest is segmented into anatomical structures to obtain segmentation results of multiple brain anatomical structures.

[0052] Specifically, by combining diffusion characteristics, the region of interest is segmented into anatomical structures, resulting in segmentation results of multiple brain anatomical structures, including: Based on the region of interest corresponding to the target anatomical structure, the corresponding local diffusion imaging data block is cropped from the diffusion-weighted imaging data, and the local fiber orientation distribution function feature block corresponding to the region of interest is cropped from the fiber orientation distribution function features.

[0053] Specifically, for each target anatomical structure, the starting coordinates and size of the region of interest in the original image space are obtained. Then, all voxel values ​​within the cubic region are extracted from the preprocessed diffusion-weighted imaging data to form a local diffusion imaging data block. Simultaneously, the extracted whole-brain fiber orientation distribution function feature tensor is cropped according to the same spatial coordinate range to obtain a local fiber orientation distribution function feature block. The local diffusion imaging data block retains the original image intensity information, while the local fiber orientation distribution function feature block provides the distribution information of fiber orientations within each voxel. Together, they constitute the multimodal input of the fine segmentation network. To adapt to the network input requirements, the cropped data blocks are uniformly resampled or padded to a fixed size (e.g., 64×64×64 voxels) and intensity normalized to ensure a mean of 0 and a variance of 1.

[0054] The local diffusion imaging data block and the local fiber orientation distribution function feature block are concatenated and then input into the pre-trained 3D U-Net fine segmentation network to output the segmentation probability map of the target anatomical structure.

[0055] Specifically, the local diffusion imaging data block and the local fiber orientation distribution function feature block are concatenated along the channel dimension to form a multi-channel 3D data tensor, which is then input into a pre-trained 3D U-Net fine segmentation network. The 3D U-Net fine segmentation network outputs a segmentation probability map of the target anatomical structure. The 3D U-Net network consists of an encoder and a decoder. The encoder progressively downsamples through multiple 3D convolutions, batch normalization, ReLU activation, and max pooling operations to extract high-dimensional semantic features. The decoder restores spatial resolution through upsampling and uses skip connections to fuse the feature maps of the corresponding layers of the encoder to preserve detailed information. The last layer of the network uses a 1×1×1 convolution to map the number of channels in the feature map to the number of target categories (usually 2 for the segmentation of a single structure, i.e., target structure and background), and outputs the probability value of each voxel belonging to the target structure and background through the Softmax activation function, forming a 3D probability map with the same size as the input data block. During the training phase, the network is optimized using the Dice loss function to maximize the overlap between the predicted segment and the gold standard label.

[0056] In some embodiments, when training the 3D U-Net fine segmentation network, a combined data augmentation strategy is applied, which includes at least one of the following: random flipping, random rotation, adding random Gaussian noise, random intensity shift, and random contrast adjustment.

[0057] The model employs several techniques to enhance its generalization ability. These include random flipping in each direction with a 20% probability, random 90-degree rotation with a 20% probability, adding random Gaussian noise (mean 0, standard deviation 0.1, 20% probability), random intensity shifting within a -10% to 10% range with a 50% probability, and random contrast adjustment with a gamma range of 0.7 to 1.3 with a 30% probability. In this embodiment, the fine-grained segmentation network is trained with independent models for each of the 23 brain anatomical structures, each model specifically designed for segmenting a particular target structure.

[0058] The segmentation probability map is fused back into the whole-brain space based on spatial location, and the whole-brain segmentation label map is obtained through argmax operation, which serves as the segmentation result for multiple brain anatomical structures; the segmentation result includes at least the segmentation mask of the left thalamus and the right thalamus.

[0059] Specifically, for each target anatomical structure, the starting position of the region of interest of the target anatomical structure in the whole-brain coordinate system is recorded. The 3D segmentation probability map output by the network is then placed back into the corresponding region in the whole-brain coordinate system, forming a probability distribution map of the target anatomical structure within the whole brain. This process is repeated for all target structures to be segmented, resulting in multiple probability maps aligned in the whole-brain space. These probability maps are then stacked along the channel dimension to form a multi-channel whole-brain probability tensor, with each channel corresponding to an anatomical structure. An argmax operation is performed on each voxel along the channel dimension, selecting the category with the highest probability value among all structures as its final anatomical label, resulting in a whole-brain segmentation label map. The value of each voxel in the label map represents its corresponding anatomical structure. For example... Figure 3 As shown, Figure 3 The diagram shows the segmentation results of 23 brain structures. The whole-brain segmentation label map contains all 23 finely segmented anatomical structures, with the left and right thalamus being accurately segmented as independent structures, providing a precise spatial localization mask for subsequent thalamic subregion division.

[0060] In some embodiments, local diffusion imaging data blocks and local fiber orientation distribution function feature blocks are concatenated and input into a pre-trained 3D U-Net fine segmentation network to output a segmentation probability map of the target anatomical structure, including: Nine independent 3D U-Net models were trained using the 9-fold cross-validation method.

[0061] Specifically, the training dataset was randomly and uniformly divided into nine subsets, each containing approximately 1 / 9 of the data samples. In each training iteration, eight subsets were selected as the training set, and the remaining subset as the validation set. This process was repeated nine times, with each iteration yielding an independent model weight. During training, each model's input consisted of a stitched local diffusion imaging data block and a local fiber orientation distribution function feature block, with the output being a segmentation probability map of the target anatomical structure. The Dice loss function was used, the Adam optimizer was selected, and the training run consisted of 100 epochs with an early stopping strategy. Through this cross-validation approach, the nine models learned on slightly different training data, each with a different focus in fitting the data, thus complementing each other's strengths during ensemble testing.

[0062] The local diffusion imaging data block and the local fiber orientation distribution function feature block are respectively input into 9 independent 3DU-Net models to obtain 9 independent segmentation probability maps.

[0063] Specifically, local diffusion imaging data blocks and local fiber orientation distribution function feature blocks are input into nine independent 3D U-Net models, resulting in nine independent segmentation probability maps. During the inference phase, for the same region of interest (ROI) data block to be segmented, the weights of the nine trained models are loaded for forward propagation calculation. Each model independently outputs a 3D segmentation probability map, where the value of each voxel represents the probability that the voxel belongs to the target anatomical structure. Because the training data of the nine models are divided differently, their prediction results for the same input data may have slight differences, but these differences provide diversity information for subsequent fusion steps.

[0064] The nine segmentation probability maps are merged by majority voting at the voxel level to generate a merged segmentation probability map.

[0065] Specifically, for each voxel location, the class labels predicted by nine models at that location are collected. The probabilities are converted into binary labels using the argmax operation, where 0 represents background and 1 represents the target structure. The class with the most frequent occurrence among the nine labels is taken as the final class for that voxel. If continuous probability values ​​are required, the average or median of the nine probability maps can be taken to obtain a fused probability map. In this embodiment, a majority voting strategy is used, meaning the final label for each voxel is determined by a vote of the predictions from the nine models. The class receiving more than half the votes wins, resulting in the fused binary segmentation label map. This fused segmentation result integrates the prediction information from multiple models, effectively reducing the impact of overfitting or random errors from a single model, and improving the stability and accuracy of the segmentation results.

[0066] 106. Based on the fiber orientation distribution function characteristics and the prior information of the predefined thalamic probability map, the segmented thalamic regions are divided into subregions to obtain the segmentation results of multiple thalamic subregions.

[0067] Among them, such as Figure 4 As shown, Figure 4 This is a schematic diagram of the thalamic subregion segmentation results.

[0068] Specifically, based on the characteristics of the fiber orientation distribution function and the prior information of the predefined thalamic probability map, the segmented thalamic regions are divided into subregions, resulting in segmentation results for multiple thalamic subregions, including: Obtain a predefined probability map template for thalamic subregions.

[0069] The thalamic subregion probability map template is stored in the form of a deformable grid, and each grid node of the thalamic subregion probability map template includes a prior probability vector corresponding to multiple thalamic subregions. The prior probability vector is used to characterize the initial probability that the grid node belongs to each thalamic subregion.

[0070] The thalamic subregion probability map template is defined in the form of a three-dimensional tetrahedral mesh, with mesh nodes evenly distributed across the thalamic region in standard space. Each mesh node stores a probability vector, the dimension of which is equal to the number of thalamic subregions to be segmented. Each component value represents the prior probability that the node belongs to the corresponding thalamic subregion, and the sum of all components is 1. The thalamic subregion probability map template also includes a corresponding thalamic mask to define the spatial range of subregion segmentation.

[0071] From the segmentation results of multiple brain anatomical structures, the segmentation masks corresponding to the left and right thalamus were extracted.

[0072] Specifically, based on the segmentation results, voxel sets corresponding to the label values ​​of the left and right thalamus are extracted respectively, generating two binary mask images. The voxel value inside the mask is 1, and the voxel value outside the mask is 0. The two masks precisely define the spatial extent of the left and right thalamus, serving as regions of interest for subregion division.

[0073] Based on the segmentation mask of the left and right thalamus, the local fiber orientation distribution function feature tensor corresponding only to the thalamic region is extracted from the whole brain fiber orientation distribution function features. The local feature tensor only includes the fiber orientation distribution information of each voxel in the thalamic region.

[0074] Specifically, the left and right thalamus are processed separately. The minimum bounding box of the thalamic region in the original image space is determined based on its respective binary mask, and then extended outward by 2 voxels to ensure complete coverage. From the extracted whole-brain fiber orientation distribution function feature tensor, a local feature tensor is cropped according to the bounding box range. Simultaneously, the starting position of this bounding box in whole-brain coordinates is recorded so that the segmentation results can be mapped back to whole-brain space later. The size of the cropped local feature tensor is... ,in, For the dimensions of the enclosure, represents the characteristic dimension of the fiber orientation distribution function.

[0075] The feature tensor of the local fiber orientation distribution function is modeled as a multivariate Gaussian distribution to obtain the likelihood model.

[0076] Specifically, the modeling process assumes that the fiber orientation distribution function feature vector of each voxel follows a multivariate Gaussian distribution, and that different voxels are independent. To simplify the calculation, it is further assumed that each feature dimension is independent of each other, that is, each feature dimension has independent mean and variance parameters, and each thalamic subregion corresponds to a set of mean and variance parameters used to describe the statistical distribution characteristics of voxel features within that subregion.

[0077] The objective function of Bayesian inference is constructed, which includes a prior term for grid deformation to constrain the rationality of grid deformation, a prior term for model parameters to regularize model parameters, and a joint likelihood term that integrates prior information from the probability map of thalamic subregions with the likelihood model.

[0078] The objective function comprises three components. The first is a mesh deformation prior term, which constrains the deformable mesh to avoid excessive twisting or folding when adapting to individual anatomical structures, controlling the smoothness of deformation through stiffness parameters. The second is a model parameter prior term, which regularizes the mean and variance parameters of the fiber orientation distribution function model to prevent overfitting. The third is a joint likelihood term, which integrates prior information provided by the thalamic subregion probability map and a likelihood model based on image features. Specifically, the label probability of each voxel is determined by a weighted combination of map priors and feature likelihoods.

[0079] The objective function is: ; in, Indicates fODF features, Represents the spectral grid. Represents the stiffness parameter. It is the number of hybrid components. Indicates the location of the grid node. Indicates the parameters of the hybrid model. Indicates the first The first label in the The probability of a voxel. Indicates deformation parameters, Indicates the parameters of the fODF model. Indicates the first The hybrid component in the first The weight of each label.

[0080] An alternating optimization strategy is used to iteratively solve the objective function, alternately updating the mesh deformation parameters and likelihood model parameters until the objective function converges, thus obtaining the likelihood model parameters.

[0081] Specifically, the optimization process consists of two alternating steps. The first step fixes the fiber orientation distribution function model parameters and optimizes the grid node positions using gradient descent. This allows the probability map template to better match the individual thalamic anatomy, and the updated grid node positions enable a more accurate mapping of the prior probabilities of each node in the map to the individual's spatial distribution. The second step fixes the grid deformation parameters and updates the mean, variance, and mixing weights of the fiber orientation distribution function model using the expectation-maximization algorithm. This algorithm iteratively calculates the posterior probabilities of each voxel belonging to each thalamic subregion and each mixed component, and re-estimates the model parameters based on these posterior probabilities until the parameters converge. These two steps are executed alternately. After each grid deformation optimization and expectation-maximization optimization, the change in the objective function value is checked to ensure it is less than a preset threshold. If the threshold is met, the iteration stops, and the final model parameters are output.

[0082] Based on the likelihood model parameters and combined with the prior information of the thalamic subregion probability map template, the posterior probability map of each voxel in the thalamic region belonging to each thalamic subregion is calculated.

[0083] After obtaining the likelihood model parameters, the mean and variance parameters of the optimized fiber orientation distribution function model, the mixed weights, and the deformed atlas prior are used to calculate the posterior probability value of each voxel in the thalamic region by comprehensively considering the matching degree between its image features and the statistical model of each subregion, as well as the atlas prior information. This results in a multi-channel posterior probability map, with each channel corresponding to a thalamic subregion.

[0084] Thresholding is performed on the posterior probability map, and the thalamic subregion category corresponding to the maximum posterior probability of each voxel is taken as the voxel's classification, resulting in the segmentation results of multiple thalamic subregions corresponding to the left and right thalamus respectively.

[0085] Specifically, for each voxel, the category index corresponding to the maximum value among the posterior probabilities of all its thalamic subregions is found, and the label of the voxel is set to this index value, generating a subregion segmentation label map with the same size as the local thalamic region. Based on the previously recorded bounding box starting position, this local label map is mapped back to whole-brain space to obtain the subregion segmentation results corresponding to the left and right thalamus. Each thalamus is divided into multiple subregions, and different subregions are labeled with different integer values, thus achieving fine segmentation of thalamic subregions based on diffusion-weighted imaging data.

[0086] In some embodiments, an alternating optimization strategy is used to iteratively solve the objective function, alternately updating the mesh deformation parameters and likelihood model parameters until the objective function converges, thereby obtaining the likelihood model parameters, including: The expectation-maximization algorithm is used for iterative optimization. In each iteration, based on the current mesh deformation parameters and likelihood model parameters, the joint posterior probability of each voxel belonging to each thalamic subregion and each hybrid component is calculated to obtain the soft-assignment weight matrix.

[0087] Specifically, for each voxel within the thalamic region, based on the voxel's fiber orientation distribution function characteristics, the mean and variance parameters of each mixed component in each thalamic subregion, and the prior probability of the voxel belonging to each thalamic subregion provided by the deformed atlas, the posterior probability of the voxel being generated by each mixed component in each thalamic subregion is calculated. The posterior probability values ​​form a three-dimensional matrix, where each element corresponds to the joint posterior probability of a specific voxel, a specific thalamic subregion, and a specific mixed component. For each voxel, the sum of the posterior probabilities of all thalamic subregions and all mixed components is 1, resulting in a soft-assignment weight matrix, representing each voxel belonging to different thalamic subregions and different mixed components with different weights.

[0088] Based on the soft-assigned weight matrix, the mean and variance parameters of the multivariate Gaussian distribution, as well as the weight coefficients of each thalamic subregion on different mixed components, are updated.

[0089] Specifically, for each thalamic subregion and each hybrid component, the fiber orientation distribution function feature vectors of all voxels are weighted and averaged according to their corresponding soft assignment weights to obtain the updated mean parameter; at the same time, the weighted squared difference between the feature vector of each voxel and the updated mean vector is calculated to obtain the updated variance parameter; the weight coefficients of each thalamic subregion on different hybrid components are normalized and updated according to the total posterior probability of all voxels belonging to that subregion and that hybrid component.

[0090] Repeat the above steps until the likelihood model parameters converge.

[0091] Specifically, the expected and maximization steps described above are repeated. Each time, the soft-assignment weights are calculated based on the current parameters, and then the model parameters are updated based on these soft-assignment weights. This process is repeated alternately until the likelihood model parameters converge. Convergence is determined by the following criteria: the change in model parameters between two consecutive iterations is less than a preset threshold, the change in the objective function value is less than a preset threshold, or the preset maximum number of iterations is reached. When the convergence condition is met, iteration stops, and the final likelihood model parameters are output, including the mean, variance, and weight coefficients of each mixture component in each thalamic subregion. These parameters are used for subsequent posterior probability calculations and thalamic subregion label assignment.

[0092] Figure 5 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention.

[0093] like Figure 5 As shown, the electronic device may include a processor 510, a communications interface 520, a memory 530, and a communication bus 540, wherein the processor 510, communications interface 520, and memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a brain multi-level segmentation method based on diffusion-weighted imaging.

[0094] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0095] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the brain multi-level segmentation method based on diffusion-weighted imaging provided by the above methods.

[0096] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the brain multi-level segmentation method based on diffusion-weighted imaging provided by the methods described above.

[0097] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0098] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for brain multi-level segmentation based on diffusion weighted imaging, characterized in that, include: Diffusion-weighted imaging data of the brain to be detected is acquired, and after preprocessing the diffusion-weighted imaging data, diffusion features are obtained, wherein the diffusion features include at least fiber orientation distribution function features. Based on the aforementioned diffusion characteristics, voxel-level tissue semantic segmentation of the brain to be detected is performed to obtain initial tissue segmentation results including three types of tissues: gray matter, white matter, and cerebrospinal fluid. Partial volume effect correction is applied to the tissue boundary regions in the initial tissue segmentation results to obtain brain tissue segmentation results; The preprocessed diffusion-weighted imaging data is divided into regions of whole-brain anatomical structures to obtain the regions of interest corresponding to the target anatomical structures. Based on the aforementioned diffusion characteristics, the region of interest is segmented into anatomical structures to obtain segmentation results of multiple brain anatomical structures; Based on the fiber orientation distribution function characteristics and the prior information of the predefined thalamic probability map, the segmented thalamic regions are divided into subregions to obtain segmentation results of multiple thalamic subregions.

2. The method of claim 1, wherein the method is based on diffusion weighted imaging. Acquire diffusion-weighted imaging data of the brain to be examined, and after preprocessing the diffusion-weighted imaging data, obtain diffusion features, including: Head motion correction, eddy current correction, and gradient nonlinear correction were performed on the raw diffusion-weighted imaging data of the brain to be examined to obtain the registered diffusion-weighted imaging data. The diffusion-weighted imaging data is fitted using a multi-voxel constrained spherical deconvolution algorithm to reconstruct the fiber orientation distribution function of each voxel, and stored in the form of spherical harmonic coefficients to obtain diffusion characteristics.

3. The brain multi-level segmentation method based on diffusion-weighted imaging according to claim 1, characterized in that, Based on the aforementioned diffusion characteristics, voxel-level semantic segmentation of the brain to be detected is performed, yielding initial tissue segmentation results including three types of tissues: gray matter, white matter, and cerebrospinal fluid. The fiber orientation distribution function features are input into a pre-trained semantic segmentation neural network, which outputs probability maps of three types of tissues—gray matter, white matter, and cerebrospinal fluid—using voxels as units. An argmax operation is performed on the probability maps of the three types of tissues to assign the category with the highest probability value to each voxel, thus obtaining the initial tissue segmentation result.

4. The brain multi-level segmentation method based on diffusion-weighted imaging according to claim 1, characterized in that, Partial volume effect correction is applied to the tissue boundary regions in the initial tissue segmentation results to obtain brain tissue segmentation results, including: From the initial tissue segmentation results, voxels located at the boundary between gray and white matter are identified, and a boundary voxel set is generated. The initial tissue segmentation results are reconstructed in three dimensions and resampled to a target space with a resolution higher than that of the original diffusion-weighted imaging data. By counting the number of high-resolution voxel categories included in each original voxel, the initial tissue volume ratio of each original voxel is calculated to obtain an initial partial volume estimation map. The initial partial volume estimation map and the boundary voxel set are input into the iterative optimization model, and the tissue volume ratio of each boundary voxel is expressed as the weighted sum of the tissue volume ratios of all voxels in the 26 neighborhoods surrounding the boundary voxel. Based on the weighted sum relationship, the tissue volume ratio of all boundary voxels is updated to obtain the updated partial volume estimation map. The updated partial volume estimation map is used as the input for the next iteration. The weighted sum update step is repeated until the difference between the two iteration results of all boundary voxels is less than a preset threshold, and the brain tissue segmentation result is output.

5. The brain multi-level segmentation method based on diffusion-weighted imaging according to claim 1, characterized in that, The preprocessed diffusion-weighted imaging data was used to divide the whole brain anatomical structures into regions, obtaining the regions of interest corresponding to the target anatomical structures, including: Extract the average diffusion-weighted image from the preprocessed diffusion-weighted imaging data; The average diffusion-weighted image is input into a pre-trained coarse segmentation network to generate coarse segmentation label maps including multiple anatomical regions; wherein the coarse segmentation network adopts the SynthSeg network, and the multiple anatomical regions include at least 23 brain anatomical structures; From the coarse segmentation label map, the minimum bounding cube is extracted for each target anatomical structure to be finely segmented, and a predetermined number of voxels are extended outward as the boundary to obtain the region of interest corresponding to each target anatomical structure.

6. The brain multi-level segmentation method based on diffusion-weighted imaging according to claim 5, characterized in that, Based on the aforementioned diffusion features, the region of interest is segmented into anatomical structures to obtain segmentation results of multiple brain anatomical structures, including: Based on the region of interest corresponding to the target anatomical structure, the corresponding local diffusion imaging data block is cropped from the diffusion-weighted imaging data, and the local fiber orientation distribution function feature block corresponding to the region of interest is cropped from the fiber orientation distribution function feature. The local diffusion imaging data block and the local fiber orientation distribution function feature block are concatenated and input into a pre-trained 3D U-Net fine segmentation network to output a segmentation probability map of the target anatomical structure. The segmentation probability map is fused back into the whole-brain space according to the spatial location, and the whole-brain segmentation label map is obtained through argmax operation, which serves as the segmentation result of the multiple brain anatomical structures; the segmentation result includes at least the segmentation mask of the left thalamus and the right thalamus.

7. The brain multi-level segmentation method based on diffusion-weighted imaging according to claim 6, characterized in that, The local diffusion imaging data block and the local fiber orientation distribution function feature block are concatenated and input into a pre-trained 3D U-Net fine segmentation network to output a segmentation probability map of the target anatomical structure, including: Nine independent 3D U-Net models were trained using a 9-fold cross-validation method. The local diffusion imaging data block and the local fiber orientation distribution function feature block are respectively input into 9 independent 3DU-Net models to obtain 9 independent segmentation probability maps; The nine segmentation probability maps are fused by majority voting at the voxel level to generate a fused segmentation probability map.

8. The brain multi-level segmentation method based on diffusion-weighted imaging according to claim 7, characterized in that, When training the 3D U-Net fine segmentation network, a combined data augmentation strategy is applied, which includes at least one of the following: random flipping, random rotation, adding random Gaussian noise, random intensity shift, and random contrast adjustment.

9. The brain multi-level segmentation method based on diffusion-weighted imaging according to claim 1, characterized in that, Based on the fiber orientation distribution function characteristics and predefined thalamic probability map prior information, the segmented thalamic regions are subdivided to obtain segmentation results for multiple thalamic subregions, including: Obtain a predefined thalamic subregion probability map template. The thalamic subregion probability map template is stored in the form of a deformable grid. Each grid node of the thalamic subregion probability map template includes a prior probability vector corresponding to multiple thalamic subregions. The prior probability vector is used to characterize the initial probability that the grid node belongs to each thalamic subregion. From the segmentation results of multiple brain anatomical structures, the segmentation masks corresponding to the left and right thalamus were extracted respectively; Based on the segmentation mask of the left and right thalamus, a local fiber direction distribution function feature tensor corresponding only to the thalamic region is cropped from the whole brain fiber direction distribution function features; the local feature tensor contains only the fiber direction distribution information of each voxel in the thalamic region; The feature tensor of the local fiber orientation distribution function is modeled as a multivariate Gaussian distribution to obtain the likelihood model; Construct a Bayesian inference objective function, which includes a grid deformation prior term constraining the rationality of grid deformation, a model parameter prior term regularizing the model parameters, and a joint likelihood term integrating the thalamic subregion probability map prior information with the likelihood model. The objective function is iteratively solved using an alternating optimization strategy, alternately updating the mesh deformation parameters and the likelihood model parameters until the objective function converges, thus obtaining the likelihood model parameters. Based on the likelihood model parameters and the prior information of the thalamic subregion probability map template, the posterior probability map of each voxel in the thalamic region belonging to each thalamic subregion is calculated. Thresholding is performed on the posterior probability map, and the thalamic subregion category corresponding to the maximum posterior probability of each voxel is taken as the voxel's classification, thus obtaining the segmentation results of multiple thalamic subregions corresponding to the left and right thalamus respectively.

10. The brain multi-level segmentation method based on diffusion-weighted imaging according to claim 9, characterized in that, An alternating optimization strategy is used to iteratively solve the objective function, alternately updating the mesh deformation parameters and the likelihood model parameters until the objective function converges, yielding the likelihood model parameters, including: The expectation-maximization algorithm is used for iterative optimization. In each iteration, based on the current mesh deformation parameters and likelihood model parameters, the joint posterior probability of each voxel belonging to each thalamic subregion and each hybrid component is calculated to obtain the soft assignment weight matrix. Based on the soft-assigned weight matrix, update the mean and variance parameters of the multivariate Gaussian distribution, as well as the weight coefficients of each thalamic subregion on different hybrid components. Repeat the above steps until the likelihood model parameters converge.

Images