Multi-source reference map adaptive dispersion MRI brain region fine segmentation method without label dependence

Abstract

The invention provides a multi-source reference map adaptive diffusion magnetic resonance imaging (MRI) brain region fine segmentation method without label dependence, which comprises the following steps: acquiring a reference map of a target brain region and diffusion magnetic resonance (dMRI) data of a subject, preprocessing the dMRI data and then calculating a multi-dimensional diffusion feature map; establishing a mapping relation between a reference map and a dMRI feature space through a nonlinear registration algorithm, and outputting an initial coarse segmentation result as supervision prior; constructing an integrated convolutional neural network and a Mama module, embedding an attention fusion U-Mama deep learning network of a channel attention mechanism, and processing a class imbalance problem through a dichotomy subtask decomposition strategy; in a prediction stage, a target region is registered to an individual space to obtain a coarse segmentation result, the coarse segmentation result and the dMRI multi-modal feature map are input into the trained network, and a high-precision segmentation result of the target brain region is output. The method is used for realizing cross-age group and high-precision individualized white matter brain region segmentation only by depending on the standard map label in an actual clinical application environment.

Description

Technical Field

[0001] This invention belongs to the field of magnetic resonance imaging (MRI) and medical image processing technology, specifically relating to a label-free, multi-source reference atlas-adapted fine segmentation method for diffusion-weighted magnetic resonance imaging (dMRI) brain regions, applicable to the analysis and application of diffusion-weighted magnetic resonance imaging (dMRI) data. Background Technology

[0002] Brain regions refer to specific areas within the brain that share similarities in one or more attributes, such as cellular structure, myelin sheath structure, connectivity patterns, functional activity, or topological representation. Brain region anatomical segmentation, a core component of neuroimaging analysis, provides fundamental support for mapping brain structure and function by precisely labeling anatomically separated cortical and subcortical regions. It also plays an irreplaceable role in clinical applications such as the microstructural representation of brain regions. In large-scale, multimodal neuroimaging studies integrating multicenter data, segmentation methods based on standard spatial maps have become the mainstream choice due to the difficulty in obtaining high-quality manual annotations for each individual. However, achieving high-precision, individualized brain region segmentation for heterogeneous multimodal data with limited annotation information remains a critical technical challenge that urgently needs to be addressed.

[0003] Diffusion magnetic resonance imaging (dMRI), as an advanced neuroimaging technique, can quantify the microstructural features of tissues and reconstruct white matter fiber tracts through fiber tractography. The microstructural details it provides are crucial for distinguishing subtle differences between brain regions, especially between the cortex and subcortical areas. This technological advantage makes it highly promising for the diagnosis and monitoring of neurological diseases such as Alzheimer's and Parkinson's, as well as for lesion localization after traumatic brain injury, providing a new perspective for basic neuroscience research and precision clinical diagnosis and treatment. With the deepening application of artificial intelligence technology in the field of medical imaging, dMRI-based brain region segmentation technology is expected to overcome the limitations of traditional methods, providing more reliable imaging evidence for the early diagnosis and precision treatment of brain diseases.

[0004] To address the challenge of brain region segmentation with limited annotations, researchers have proposed various techniques, including data augmentation, generative models, self-supervised learning, few-shot learning, and semi-supervised learning. Among these, synthetic data generation and deep ensemble models have shown promise in improving the robustness and accuracy of segmentation models. However, these methods generally still rely on a certain number of high-quality individual-level annotations for training supervision. When only a single atlas label from the standard space is available, the model cannot learn individual anatomical variations, severely limiting its ability to achieve personalized and anatomically accurate segmentation.

[0005] Current mainstream brain region segmentation methods are mostly based on anatomical MRI (magnetic resonance imaging) data such as T1 or T2 weighted images, relying on the rich anatomical details provided by their high spatial resolution to define brain regions. These methods typically use standard spatial atlases such as the MNI, mapping the atlas to individual brain images through nonlinear registration. However, due to significant anatomical differences between individuals, errors easily occur during the registration process, especially in subcortical regions involving differences in brain microstructure, where segmentation accuracy is significantly reduced. In recent years, various direct brain region segmentation methods based on atlases, clustering, graph theory, and surfaces have emerged. Breakthroughs in deep learning technology have further propelled a leap in segmentation performance. Among them, convolutional neural networks (CNNs), with their ability to automatically learn from low-level to high-level spatial feature layers of images, have become the most widely used model architecture. Deep learning methods such as U-Net, nnU-Net, and Swin-Unet have achieved excellent performance in brain region segmentation based on anatomical MRI, the SLANT method reduces computational overhead through spatial localization design, and AssemblyNet achieves coarse-to-fine segmentation optimization through ensemble learning.

[0006] However, existing technologies still have many limitations that urgently need to be overcome: First, they are heavily reliant on high-quality manual annotations. Obtaining such annotations in medical imaging tasks such as brain region segmentation is costly and time-consuming, resulting in insufficient generalization ability of the model on multi-center or cross-age datasets. This problem is particularly prominent in image segmentation of special conditions such as traumatic brain injury. Second, mainstream methods are based on structural MRI, which effectively captures macroscopic anatomical information but neglects the rich microscopic structural features contained in dMRI, making it difficult to accurately identify brain regions with blurred boundaries or similar structures. Third, existing dMRI-based segmentation... The segmentation methods mostly rely on a single diffusion feature map, and the simple fusion strategy used cannot effectively filter redundant or conflicting information. In addition, nonlinear registration is usually required to map the segmentation results of anatomical MRI to dMRI space. However, the inherent low spatial resolution, signal noise and geometric distortion of dMRI can easily lead to registration errors, which in turn affect the segmentation accuracy and reliability. In the absence of high-quality anatomical images, they will lose their application basis. Fourth, the existing methods are mostly end-to-end frameworks for fixed brain region label sets, which lack adaptability to emerging anatomical definitions or more refined segmentation needs, further limiting their clinical application scenarios.

[0007] From a research perspective, developing a high-precision, personalized brain region segmentation technology that relies solely on standard atlas labels and integrates multimodal diffusion features will provide a novel tool for analyzing brain functional networks, driving breakthroughs in the research of early diagnostic mechanisms for neurodegenerative diseases. In terms of social value, this technology can improve the accuracy of brain disease screening and diagnosis, promoting the implementation of early intervention measures. Especially in the diagnosis and treatment of diseases such as traumatic brain injury and dementia, which are prevalent among adolescents and the elderly, it can effectively reduce the social burden of these diseases and improve public health services. In terms of economic value, by reducing reliance on expensive manual annotation and multimodal image data, it can lower medical costs, improve the efficiency of medical resource utilization, and simultaneously promote the innovative development of medical image analysis equipment and algorithm industries, injecting new impetus into economic growth.

[0008] Therefore, in order to address the shortcomings of existing brain region segmentation technologies in terms of annotation dependence, modality limitation, feature utilization and generalization ability, developing a high-precision segmentation framework based on dMRI and incorporating multiple features is of great significance for promoting the clinical translation of neuroimaging technology and improving the level of diagnosis and treatment of brain diseases. It is also the core demand for the current technological development in this field. Summary of the Invention

[0009] To overcome the challenges of existing technologies, this invention proposes a label-free, multi-source reference atlas-adaptive, diffusion-weighted MRI brain region fine segmentation method, which can achieve high-precision, individualized white matter brain region segmentation across age groups by relying solely on standard atlas labels in actual clinical applications.

[0010] The technical solution to achieve the purpose of this invention is as follows:

[0011] A label-independent, multi-source reference map-adapted, diffusion-weighted MRI brain region fine segmentation method includes:

[0012] Step 1: Obtain reference atlases of the target brain region and multi-b-value diffusion magnetic resonance imaging (dMRI) data of the subjects. Preprocess the diffusion MRI data and calculate multi-dimensional diffusion feature maps using a diffusion model.

[0013] Step 2: Based on the binary sub-task decomposition strategy, the reference map is decomposed into multiple independent anatomical region masks in the standard space;

[0014] Step 3: For each anatomical region mask obtained in Step 2, a nonlinear registration algorithm is used to independently map it to the individual diffusion feature space of the subject. The spatial overlap after multi-region registration is fused to output the initial coarse segmentation map.

[0015] Step 4: Construct an attention-fusion U-Mamba deep learning network, which integrates a convolutional neural network and a Mamba module, and embeds a channel attention mechanism;

[0016] Step 5: Obtain the individual white matter partition map generated by the subject using FreeSurfer software, decompose it into binary classification masks corresponding to each target brain region as supervision signals; use the initial coarse segmentation map output in Step 3 as the prior signal, fuse multi-dimensional diffusion feature maps to construct network input, train the attention fusion U-Mamba deep learning network through weak supervision learning, fit the supervision signal, and complete the model parameter tuning;

[0017] Step 6: Acquire the initial coarse segmentation map of the subject in real time, fuse it with the multidimensional diffusion feature map of the subject's dMRI, input it into the attention fusion U-Mamba deep learning network trained in Step 5, and output the high-precision segmentation result of the target brain region.

[0018] Further, in step 1, the reference spectrum is derived from a regional spectrum of a standard space or any sample; the preprocessing of the diffuse magnetic resonance data includes image axial alignment, centering, Gibbs artifact removal, intensity non-uniformity correction, denoising, eddy current distortion correction, and rotation correction; the multidimensional diffusion feature map includes average diffusion rate, anisotropy fraction, first eigenvalue, second eigenvalue, linear anisotropy coefficient, planar anisotropy coefficient, and spherical anisotropy coefficient; the b-value of the diffuse magnetic resonance data includes b=0 s / mm² and other selectable b-value combinations; the multidimensional diffusion feature map is calculated using the MRtrix3 tool.

[0019] Furthermore, other possible combinations of b values ​​include b = 1000 s / mm², 1500 s / mm², 2000 s / mm², and 3000 s / mm².

[0020] Furthermore, in step 2, the reference atlas is used to generate a white matter partition map WMPARC using FreeSurfer software. The white matter partition map WMPARC divides the brain into 182 independent anatomical regions in the standard MNI space.

[0021] Furthermore, the nonlinear registration algorithm in step 3 is a registration algorithm based on symmetric differential homeomorphism transformation. The registration process uses mutual information as a similarity metric function and introduces Gaussian smoothing constraints. When spatial overlap occurs after registration of different brain regions, the maximum probability voting strategy is used to fuse the overlapping voxels.

[0022] Furthermore, in step 4, the attention fusion U-Mamba deep learning network adopts an encoder-decoder structure; the encoder module includes multiple progressive downsampling stages, each stage consisting of residual convolutional blocks, and embeds a channel attention mechanism; a Mamba module based on a state space model is introduced at the end of the encoder; the decoder gradually restores the spatial resolution through upsampling, and uses skip connections to fuse deep and shallow features, and finally maps to a single-channel segmentation output through a 1×1 convolutional layer.

[0023] Furthermore, the channel attention mechanism generates channel attention weights by performing global average pooling on the feature map and utilizing the bottleneck structure. To achieve feature calibration, specifically:

[0024] ;

[0025] ;

[0026] in, These are the feature descriptors after pooling. Represents the ReLU activation function. This represents the sigmoid function. , For weight parameters, This is the original feature map.

[0027] Furthermore, the Mamba module is used to model long-range spatial dependencies of three-dimensional feature sequences, as follows:

[0028] ;

[0029] ;

[0030] Where A and B are the learning parameters, Representation layer normalization, This indicates input.

[0031] Furthermore, step 5 trains the network's joint loss function using a weakly supervised learning approach, which includes binary cross-entropy loss, Top-K cross-entropy loss, and Dice loss.

[0032] Furthermore, in step 5, the multi-dimensional diffusion feature map and the initial coarse segmentation map are concatenated along the channel dimension to construct an 8-channel multimodal input data. Before inputting the attention to the U-Mamba deep learning network, the multimodal input data is uniformly resampled, and the channels corresponding to the multi-dimensional diffusion feature map are normalized using Z-score, while the spatial prior channels corresponding to the initial coarse segmentation map are not normalized.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] (1) The method of the present invention uses the coarse segmentation map generated by nonlinear registration of standard map as the only reference signal, which eliminates the need for high-quality individual manual annotation data and completely solves the industry pain points of high cost and long time consumption in medical image annotation.

[0035] (2) The method of the present invention supports multi-source reference map input. Regardless of whether the reference map comes from the standard space or the local region map of any sample, it can be effectively mapped through nonlinear registration, which can meet the diverse clinical needs.

[0036] (3) The method of the present invention integrates multi-dimensional diffusion features of dMRI, dynamically allocates the weight of each feature through the channel attention mechanism, automatically suppresses redundant information and strengthens effective features, and significantly improves the recognition accuracy of brain regions with blurred boundaries or similar structures.

[0037] (4) The attention fusion U-Mamba network of the present invention innovatively combines the advantages of CNNs and Mamba modules. CNNs accurately capture local spatial details, while Mamba modules efficiently model long-distance feature associations. The two work together to improve the accuracy and consistency of segmentation boundaries.

[0038] (5) The method of the present invention decomposes the white matter segmentation task into multiple independent binary classification sub-tasks, which not only effectively alleviates the class imbalance problem in multi-class segmentation, but also supports the flexible expansion of new brain regions. It can adapt to emerging anatomical definitions or more refined segmentation needs without retraining the entire network, breaks through the fitting limitations of traditional methods on specific datasets, and has excellent cross-scene generalization ability.

[0039] (6) The collaborative architecture of nonlinear registration and deep learning in the method of this invention realizes a two-level processing flow of "coarse segmentation prior - fine correction", which improves the segmentation accuracy of complex microstructure regions such as white matter by 15%-25% compared with traditional methods, while the inference speed meets the needs of real-time clinical applications.

[0040] (7) The method of the present invention does not rely on structural MRI data and can complete high-precision segmentation only through dMRI. It can still be stably applied in scenarios where the quality of structural MRI images is poor or missing, which greatly expands the clinical applicability of the technology. Attached Figure Description

[0041] Figure 1 This is a flowchart of the label-free, multi-source reference map-adapted diffusion MRI brain region fine segmentation method (RefParcel) of the present invention.

[0042] Figure 2 This describes the overall workflow of the method of the present invention.

[0043] Figure 3 This is a schematic diagram of the network structure and modules of the present invention (RefParcel).

[0044] Figure 4 This is a performance comparison chart of the method of the present invention (RefParcel) and the state-of-the-art methods.

[0045] Figure 5 This is a diagram showing the results of testing the method (RefParcel) of the present invention on the right cerebellar cortex and left thalamus, which are not previously seen in the model.

[0046] Figure 6 This is a qualitative comparison of the partitioning results and MD value distribution within the right thalamus proprioceptive region using the method (RefParcel) and the registration method (Reference) of the present invention.

[0047] Figure 7 This is a visualization of the effect of the method (RefParcel) of the present invention on randomly delineating brain regions. Detailed Implementation

[0048] This embodiment provides a label-independent, multi-source reference map-adapted, diffusion-weighted MRI brain region fine segmentation method, combined with... Figure 1 and Figure 2 The method involves clinicians acquiring and delineating target brain regions from arbitrary samples, including but not limited to standard templates in the MNI space. The MNI space templates are then nonlinearly registered onto a diffusion magnetic resonance imaging (dMRI) feature map of the individual patient. This invention uses a FA feature map to map the target brain region to the individual dMRI space, obtaining a coarse segmentation of the brain region in the individual space. The coarse segmentation is then fused with the patient's dMRI feature map, input into a pre-trained attention-fusion U-Mamba network, and outputting a fine-scale anatomical partitioning result of the target brain region in the individual dMRI space. Specifically, this method includes:

[0049] Step 1: Obtain reference atlases of the target brain region and multi-b-value diffusion magnetic resonance imaging (dMRI) data of the subjects. Preprocess the diffusion MRI data, including motion correction and eddy current correction, to eliminate data acquisition interference. Then, calculate multi-dimensional diffusion feature maps, such as fractional anisotropy (FA) and mean diffusion coefficient (MD), using a diffusion model.

[0050] First, the subject undergoes diffusion magnetic resonance imaging (dMRI) scanning on a magnetic resonance imaging (MRI) device. Data is acquired according to a preset diffusion weighting factor b value, resulting in multi-directional, multi-b-valued diffusion MRI data. The b-values ​​include at least b = 0 s / mm², and different combinations of b-values ​​can be set according to the specific acquisition protocol, such as b = 1000, 2000, and 3000 s / mm². After acquisition, the diffusion MRI data undergoes preprocessing. The preprocessing steps include: image axial alignment, centering, Gibbs artifact removal, intensity non-uniformity correction based on the N4 algorithm, denoising based on MP-PCA, eddy current distortion correction, and rotation correction to ensure the accuracy of subsequent diffusion parameter calculations.

[0051] Based on the preprocessed diffusion magnetic resonance imaging (DTI) data, the MRtrix3 tool was used to calculate DTI feature maps for diffusion-weighted images with b=0 and b=1500. The dwi2tensor was used to calculate the tensor map, and various diffusion parameters were extracted using tensor2metric to characterize the microscopic diffusion properties of brain tissue. This invention extracts the following seven diffusion tensor imaging parameters as microstructural feature inputs: 1) Average diffusivity (MD), reflecting the average diffusion intensity of water molecules in all directions; 2) Fraction of anisotropy (FA), characterizing the consistency of diffusion directions and the integrity of white matter structure; 3) First eigenvalue E1 and second eigenvalue E2, describing the diffusion intensity in the primary and secondary diffusion directions; 4) Linear anisotropy coefficient cl, planar anisotropy coefficient cp, and spherical anisotropy coefficient cs, used to characterize the geometric morphology of the diffusion tensor ellipsoid. Among them, cl, cp, and cs reflect unidirectional diffusion, planar diffusion, and isotropic diffusion characteristics, respectively, and their calculation formulas are as follows:

[0052]

[0053]

[0054]

[0055] The method does not require high-quality manually labeled individual data; it can achieve segmentation solely based on reference atlases. It also supports multi-source reference atlas input, adapts to heterogeneous dMRI data from different age groups and multiple centers, and can be extended to segmentation of any anatomical region.

[0056] Step 2: Based on the binary subtask decomposition strategy, the reference map is decomposed into multiple independent anatomical region masks in the standard space.

[0057] To provide anatomical constraints to the neural network, this invention introduces a white matter partitioning prior. The white matter partitioning atlas WMPARC, generated using FreeSurfer software, divides the brain into 182 independent anatomical regions in the standard MNI space.

[0058] Step 3: For each anatomical region mask obtained in Step 2, a nonlinear registration algorithm is used to independently map it to the individual diffusion feature space of the subject. The spatial overlap after multi-region registration is fused to output the initial coarse segmentation map.

[0059] For each white matter region, a nonlinear registration algorithm based on symmetric homeomorphism (SyN) is used to independently register the region from the standard space to the subject's diffusion feature space. During registration, mutual information is used as the similarity metric, and Gaussian smoothing constraints are introduced to limit the spatial deformation amplitude, thus obtaining a coarse brain region segmentation result located in the subject's original space. When different brain regions spatially overlap after registration, a maximum probability voting strategy is used to fuse overlapping voxels, ensuring that each voxel corresponds to only one brain region label. The coarse segmentation result obtained above serves as spatial prior information, providing anatomical guidance for subsequent neural network segmentation.

[0060] The seven diffusion microstructure feature parameter maps (FA, MD, E1, E2, cl, cp, cs) obtained above are concatenated with the corresponding coarse segmentation priors of white matter brain regions along the channel dimension to construct an 8-channel three-dimensional multimodal input volume data. Before inputting into the neural network, the multimodal input data is uniformly resampled to achieve isotropic voxel resolution; the diffusion microstructure feature channels are Z-score normalized, while the spatial prior channels are not normalized to preserve their discrete label attributes.

[0061] Step 4: Construct an attention-fusion U-Mamba deep learning network, which integrates a convolutional neural network and a Mamba module, and embeds a channel attention mechanism.

[0062] This invention constructs a convolutional neural network for fine-grained brain region partitioning. This network employs an encoder-decoder structure and combines a channel attention mechanism with a state-space modeling module. Figure 3 , Figure 3(a) shows the overall network framework, which adopts an encoder-decoder structure and introduces a Mamba layer and a channel attention mechanism. The input data passes through the Stem module and the feature extraction and downsampling processes from stage 0 to stage 4. Taking an 8-channel input as an example, the number of output channels in each stage is as follows: Stem has 32 channels, stage 0 has 32 channels, stage 1 has 64 channels, stage 2 has 128 channels, stage 3 has 256 channels, and stage 4 has 320 channels. The Mamba layer is based on the State Space Model (SSM) and models and captures long-range dependencies without increasing the number of channels. In the decoding stage, the encoded features are reconstructed through upsampling and a segmentation head to generate the final fine-grained brain region partitioning results. The number of output channels is as follows: upsampling 0 has 256 channels, stage 0 has 128 channels, stage 1 has 64 channels, stage 2 has 32 channels, stage 3 has 32 channels, and the segmentation head has 1 channel. Figure 3 Image (b) shows a detailed structural diagram of the BasicBlockD module, BasicResBlock module, and MambaLayer in the network. The encoder module includes multiple progressive downsampling stages, each consisting of a residual convolutional block (BasicResBlock) used to extract local spatial features and progressively reduce the spatial resolution while gradually increasing the number of feature channels to alleviate the gradient vanishing problem in deep networks. A channel attention mechanism is introduced into the encoder to adaptively weight the multi-channel diffuse microstructure features. This mechanism performs global average pooling on the feature map and then uses the bottleneck structure to generate channel attention weights. This enhances the focus on discriminative diffusion characteristics and suppresses redundant or noisy information, as defined below:

[0063]

[0064] in These are the feature descriptors after pooling. Represents the ReLU activation function. This represents the sigmoid function. Each channel of the original feature map X is multiplied by its corresponding weight to achieve feature calibration.

[0065]

[0066] A Mamba module based on a state-space model (SSM) is introduced at the end of the encoder to model the three-dimensional feature sequence to capture long-range spatial dependencies, thereby making up for the shortcomings of traditional convolutional neural networks in global modeling capabilities. The module is defined as follows.

[0067]

[0068]

[0069] Where A and B are learning parameters used to model long-range dependencies. The representation layer is normalized to ensure training stability. The decoder then gradually restores the spatial resolution through upsampling and fuses deep and shallow features using skip connections. Finally, a 1×1 convolutional layer maps the final feature map into a single-channel segmentation output.

[0070] Step 5: Obtain the individual white matter partition map generated by the subject using FreeSurfer software, decompose it into binary classification masks corresponding to each target brain region as supervision signals; use the initial coarse segmentation map output in Step 3 as the prior signal, fuse multi-dimensional diffusion feature maps to construct network input, train the attention fusion U-Mamba deep learning network through weak supervision learning, fit the supervision signal, and complete the model parameter tuning.

[0071] During model training, each brain region was treated as an independent binary classification segmentation task, with the goal of distinguishing the target brain region from the background region. Individual WMPARC segmentation results generated by FreeSurfer software were used as the gold standard supervision signal. A joint loss function was constructed as the model optimization objective, comprising binary cross-entropy loss, Top-K cross-entropy loss, and Dice loss. Top-K cross-entropy loss only involved backpropagation on voxels with the largest loss values ​​to enhance the model's ability to learn about regions with blurred boundaries and complex structures; Dice loss was used to alleviate class imbalance and improve the overall consistency of the segmentation results.

[0072] For background regions where the anisotropy score (FA) is below a preset threshold, a label-ignoring strategy is adopted during the loss calculation process to make its gradient zero, thereby avoiding interference from invalid regions to the model training.

[0073] Step 6: In the prediction stage, the initial coarse segmentation map of the subject to be tested is obtained using the same method as described in Step 3. It is then fused with the multidimensional diffusion feature map of the subject's dMRI and input into the attention fusion U-Mamba deep learning network trained in Step 5 to output the high-precision segmentation result of the target brain region.

[0074] In practical applications, the diffusion magnetic resonance imaging data of the test subjects first undergoes the same preprocessing, diffusion feature extraction, and multimodal input construction process as in the training phase, and then is input into the trained neural network model to obtain the fine partitioning results of the target brain region.

[0075] Figure 4This section compares the performance of the method of this invention with existing baseline methods such as nnU-Net, Swin-UNet, Mamba-UNet, and UNet on the HCP, HCPA, and HCPD datasets. The first column shows the reference map obtained through registration for visual comparison, indicated by a blue outline; the remaining columns show the partitioning results of different methods, indicated by red outlines; and green outlines represent the ground truth labeled brain regions. The results show that the partitioning results of this invention best match the ground truth outlines on each dataset, with the smallest boundary deviation, demonstrating that this invention has high accuracy in fine-grained brain region partitioning and rapid generalization ability across different datasets.

[0076] Figure 5 The method (RefParcel) of this invention was trained on brain region data from healthy subjects in the HCP, HCPA, and HCPD datasets, and then tested on the right cerebellar cortex and left thalamus, areas not previously seen by the model. Green outlines represent ground truth labeled boundaries, red outlines represent the partitioning results predicted by RefParcel, and blue outlines represent the partitioning results based on the registration-based reference method. The results show that the method of this invention has better partitioning accuracy than the registration method in untrained anatomical regions, while also demonstrating rapid generalization ability across datasets.

[0077] Figure 6 This paper presents a qualitative comparison of the partitioning results and MD value distribution within the right thalamus proprioceptive region using the method of this invention (RefParcel) and the registration method (Reference). The MD distribution corresponding to the RefParcel partitioning results is more concentrated and centered, while the MD distribution of the registration method is significantly right-skewed and contains more voxels with high MD values, resulting in a higher relative standard deviation (RSD). A two-tailed t-test confirmed that the difference in MD distribution concentration between the two methods was statistically significant (p<0.001). The results show that the method of this invention can obtain partitioning results that are more consistent in anatomical structure and more compact in structural features.

[0078] Figure 7The method (RefParcel) of this invention is visualized to demonstrate the partitioning effect of randomly delineated brain regions. The left column shows three manually delineated regions (Region 1, Region 2, and Region 3) in the MNI standard space, indicated by green outlines. The middle column shows the partitioning results obtained after registering the manually delineated regions to samples from the HCP, HCPA, and HCPD datasets, serving as a reference method based on registration, indicated by blue outlines. The right column shows the final partitioning results output by the RefParcel method of this invention, indicated by red outlines. The results show that, compared with the registration method, the method of this invention exhibits better anatomical consistency and structural coherence in the partitioning of randomly delineated brain regions, further illustrating its applicability and robustness in the task of fine partitioning of arbitrary brain regions.

[0079] This invention enables high-precision brain region segmentation of dMRI data without annotation dependence and across various scenarios, lowering the technical threshold for brain image analysis and promoting the widespread application of brain region segmentation technology in primary hospitals and multi-center studies, providing a reliable tool for neuroscience research. The binary subtask decomposition architecture of this invention supports flexible expansion to new brain regions. Combined with interpretable attention feature weight allocation, it can quickly adapt to segmentation tasks of specific lesion regions according to clinical diagnostic and treatment needs, showing great application potential in clinical scenarios such as brain injury localization and white matter disease assessment.

[0080] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, 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. All should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A label-free, multi-source reference map-adapted, diffusion-weighted MRI brain region fine segmentation method, characterized in that, include: Step 1: Obtain reference atlases of the target brain region and multi-b-value diffusion magnetic resonance imaging (dMRI) data of the subjects. Preprocess the diffusion MRI data and calculate multi-dimensional diffusion feature maps using a diffusion model. Step 2: Based on the binary sub-task decomposition strategy, the reference map is decomposed into multiple independent anatomical region masks in the standard space; Step 3: For each anatomical region mask obtained in Step 2, a nonlinear registration algorithm is used to independently map it to the individual diffusion feature space of the subject. The spatial overlap after multi-region registration is fused to output the initial coarse segmentation map. Step 4: Construct an attention-fusion U-Mamba deep learning network, which integrates a convolutional neural network and a Mamba module, and embeds a channel attention mechanism; Step 5: Obtain the individual white matter partition map generated by the subject using FreeSurfer software, and decompose it into binary classification masks corresponding to each target brain region as supervision signals; Using the initial coarse segmentation map output in step 3 as the prior signal, the network input is constructed by fusing multi-dimensional diffusion feature maps. The attention fusion U-Mamba deep learning network is trained through weakly supervised learning, and the supervision signal is fitted to complete the model parameter tuning. Step 6: Acquire the initial coarse segmentation map of the subject in real time, fuse it with the multidimensional diffusion feature map of the subject's dMRI, input it into the attention fusion U-Mamba deep learning network trained in Step 5, and output the high-precision segmentation result of the target brain region.

2. The multi-source reference atlas-adapted diffusion MRI brain region fine segmentation method according to claim 1, characterized in that, In step 1, the reference spectrum is derived from a regional spectrum of a standard space or any sample; the preprocessing of the diffuse magnetic resonance data includes image axial alignment, centering, Gibbs artifact removal, intensity non-uniformity correction, denoising, eddy current distortion correction, and rotation correction; the multidimensional diffusion feature map includes average diffusion rate, anisotropy fraction, first eigenvalue, second eigenvalue, linear anisotropy coefficient, planar anisotropy coefficient, and spherical anisotropy coefficient; the b-value of the diffuse magnetic resonance data includes b=0 s / mm² and other selectable b-value combinations; the multidimensional diffusion feature map is calculated using the MRtrix3 tool.

3. The multi-source reference atlas-adapted diffusion MRI brain region fine segmentation method according to claim 2, characterized in that, Other possible combinations of b values ​​include b = 1000 s / mm², 1500 s / mm², 2000 s / mm², and 3000 s / mm².

4. The multi-source reference atlas-adapted diffusion MRI brain region fine segmentation method according to claim 1, characterized in that, In step 2, the reference map is used to generate the white matter partition map WMPARC using FreeSurfer software. The white matter partition map WMPARC divides the brain into 182 independent anatomical regions in the standard MNI space.

5. The multi-source reference atlas-adapted diffusion MRI brain region fine segmentation method according to claim 1, characterized in that, In step 3, the nonlinear registration algorithm is a registration algorithm based on symmetric differential homeomorphism transformation. The registration process uses mutual information as the similarity measurement function and introduces Gaussian smoothing constraints. When spatial overlap occurs after registration of different brain regions, the maximum probability voting strategy is used to fuse the overlapping voxels.

6. The multi-source reference atlas-adapted diffusion MRI brain region fine segmentation method according to claim 1, characterized in that, In step 4, the attention fusion U-Mamba deep learning network adopts an encoder-decoder structure. The encoder module includes multiple progressive downsampling stages, each consisting of residual convolutional blocks and embedding a channel attention mechanism. A Mamba module based on a state space model is introduced at the end of the encoder. The decoder gradually restores the spatial resolution through upsampling and uses skip connections to fuse deep and shallow features. Finally, it is mapped to a single-channel segmentation output through a 1×1 convolutional layer.

7. The multi-source reference atlas-adapted diffusion MRI brain region fine segmentation method according to claim 6, characterized in that, The channel attention mechanism generates channel attention weights by performing global average pooling on the feature map and utilizing the bottleneck structure. To achieve feature calibration, specifically: ； ； in, These are the feature descriptors after pooling. Represents the ReLU activation function. This represents the sigmoid function. , For weight parameters, This is the original feature map.

8. The multi-source reference atlas-adapted diffusion MRI brain region fine segmentation method according to claim 6, characterized in that, The Mamba module is used to model long-range spatial dependencies of three-dimensional feature sequences, as follows: ； ； Where A and B are the learning parameters, Representation layer normalization, This indicates input.

9. The multi-source reference atlas-adapted diffusion MRI brain region fine segmentation method according to claim 6, characterized in that, Step 5 trains the network using a weakly supervised learning approach, employing a joint loss function that includes binary cross-entropy loss, Top-K cross-entropy loss, and Dice loss.

10. The multi-source reference atlas-adapted diffusion MRI brain region fine segmentation method according to claim 2, characterized in that, In step 5, the multi-dimensional diffusion feature map and the initial coarse segmentation map are concatenated along the channel dimension to construct an 8-channel multimodal input data. Before the input attention is fused to the U-Mamba deep learning network, the multimodal input data is uniformly resampled, and the channels corresponding to the multi-dimensional diffusion feature map are normalized using Z-score, while the spatial prior channels corresponding to the initial coarse segmentation map are not normalized.

Images