A brain region segmentation method based on multi-sequence magnetic resonance imaging collaborative feature fusion
By constructing the CIA-Net deep learning model that integrates collaborative features from multiple magnetic resonance imaging sequences, the limitations of registration accuracy and image contrast in existing brain region segmentation methods have been addressed. This model enables precise and automated segmentation of multiple brain regions, improving segmentation accuracy and stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUIZHOU PROVINCIAL PEOPLES HOSPITAL
- Filing Date
- 2026-03-13
- Publication Date
- 2026-06-09
Smart Images
Figure CN122176307A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical imaging technology, and in particular to a brain region segmentation method based on multi-sequence magnetic resonance imaging collaborative feature fusion, which is applicable to the precise segmentation of 115 brain regions in cranial magnetic resonance imaging. Background Technology
[0002] Fine-grained brain region segmentation is a fundamental technique in cranial magnetic resonance imaging (MRI) analysis, widely used in quantitative analysis of brain structure, research on brain diseases, and clinical auxiliary diagnosis. Existing brain region segmentation methods mainly include label propagation methods based on template registration and automatic segmentation methods based on single-sequence MRI images. However, both have shortcomings in practical applications: template registration methods are highly sensitive to registration accuracy in fine-grained, multi-region segmentation scenarios, easily leading to label misalignment; single-sequence segmentation methods are limited by image contrast, making it difficult to accurately characterize complex brain region boundaries, easily causing confusion between adjacent brain regions; and existing multi-sequence segmentation methods mostly employ simple feature stitching, failing to fully utilize the complementary information between different sequences, and their segmentation accuracy and stability remain limited under conditions of a large number of brain regions. Therefore, it is necessary to propose a brain region segmentation method based on collaborative feature fusion of multi-sequence MRI to achieve accurate and automatic segmentation of multiple brain regions and improve its practical application value. Summary of the Invention
[0003] The purpose of this invention is to provide a brain region segmentation method based on the fusion of collaborative features from multi-sequence magnetic resonance imaging (MRI). Addressing the problems of existing template registration methods being susceptible to registration errors, single-sequence segmentation methods having limited contrast, and multi-sequence methods relying solely on simple feature splicing and failing to fully exploit complementary information between sequences, this invention constructs a collaborative feature modeling and adaptive fusion mechanism among multiple sequences. This achieves deep correlation and complementary expression of brain structural information from different MRI sequences, enabling automated and precise segmentation of cranial MRI images under fine-grained conditions and with a large number of brain regions. This improves the accuracy, structural consistency, and stability of the segmentation results, enhances adaptability to complex brain region boundaries and individual differences, and provides reliable technical support for quantitative analysis of brain structure and clinical auxiliary diagnosis.
[0004] To achieve the above objectives, the present invention provides the following solution: A brain region segmentation method based on multi-sequence magnetic resonance imaging collaborative feature fusion includes: Brain amplitude map data, phase map data, and T1 structural image data were collected; The amplitude map data and the phase map data are preprocessed and reconstructed to obtain a quantitative magnetic susceptibility image; Spatial registration is performed between the T1 structural image data and the quantitative magnetic susceptibility image to obtain paired image data; The paired image data is segmented and labeled to construct a brain region segmentation label dataset; The CIA-Net deep learning model is trained using the training set of the brain region segmentation label dataset to obtain the trained CIA-Net deep learning model. The CIA-Net deep learning model is constructed by fusing an adaptive data domain partitioning module and a common and individual feature attention module. The adaptive data domain partitioning module partitions the training set to obtain multimodal decoupled features, and the common and individual feature attention module fuses the decoupled features to obtain collaborative fusion features. The quantitative magnetic susceptibility image and T1 structural image data to be segmented are input into the trained CIA-Net deep learning model, which outputs the automatic segmentation results of brain regions.
[0005] Optionally, the preprocessing and reconstruction of the amplitude map data and the phase map data include: The phase map data is processed using a phase unwrapping algorithm based on the Laplacian operator to obtain a spatially continuous phase map; Based on the amplitude map data, a brain tissue region mask is extracted. Based on the brain tissue region mask, the V-SHARP algorithm is used to remove the background field of the spatially continuous phase map, and the msQSM algorithm is used to perform magnetic susceptibility inversion to obtain the quantitative magnetic susceptibility image.
[0006] Optionally, training the CIA-Net deep learning model using the training set of the brain region segmentation label dataset includes: The training set data is input into the adaptive data domain partitioning module, which outputs the decoupled features. The decoupled features are input into the encoder part, and the output of the encoder part is input into the boundary attention feedback sub-network and the common and individual feature attention module, respectively. The boundary attention feedback sub-network is used to extract boundary features. The brain region segmentation is achieved by fusing features from the boundary attention feedback subnetwork and the skip connection input decoder part of the common and individual feature attention module.
[0007] Optionally, the encoder part comprises a CNRD module and a convolution module, wherein the CNRD module consists of a convolution block, a normalization block, an activation block, a ReLU block, and a random deactivation block, and the convolution module consists of a plurality of the CNRD modules; The decoder consists of a three-dimensional deconvolutional layer and a three-dimensional convolutional layer.
[0008] Optionally, the adaptive data domain partitioning module performs domain partitioning on the training set to obtain multimodal decoupled features, including: All modal data, semantic segmentation labels, and domain similarity thresholds in the training set are input into the adaptive data domain partitioning module. For each modality, the median, 0.05 percentile, and 99.5 percentile are calculated based on each label, and the domain partitioning of each modality is initialized by combining the corresponding label groups. If the difference between the medians of any two domains is less than the domain similarity threshold, the two domains are merged until the domains converge and stabilize to obtain the multimodal decoupled features. The merging operation includes averaging the medians, 0.05 percentiles, and 99.5 percentiles of the two domains and merging the corresponding label sets.
[0009] Optionally, the common and individual feature attention module fuses the decoupled features by: calculating individual attention for each decoupled feature to obtain individual features of each modality; The attention matrices of the decoupled features are aggregated to extract cross-modal common features; The individual and common features are fused using a multilayer perceptron to obtain the collaborative fusion feature.
[0010] Optionally, calculating individual attention (SA) includes: ; Where A represents the attention weight matrix, d represents the feature dimension, and Q represents the query. The key is represented by the superscript T, which indicates transpose, and X represents the input feature. These are the eigenvalues.
[0011] Optionally, extracting cross-modal common features (CA) includes: in, Indicates the first One modality, Indicates the number of modes; For each modality attention matrix, These are the characteristic values for each modality.
[0012] The beneficial effects of this invention are as follows: By constructing an adaptive data domain partitioning and common-individual feature attention fusion mechanism, this invention achieves collaborative modeling of T1 and QSM multimodal features, significantly improving segmentation accuracy and structural consistency in a fine-grained segmentation task of 115 brain regions. This method enhances the ability to identify small-volume brain regions and complex adjacent regions, reduces label confusion, and improves the stability and reliability of segmentation results. Attached Figure Description
[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0014] Figure 1 In the comparison chart of the segmentation results of 115 brain regions by the CIA-Net method and other segmentation methods in this embodiment of the invention, different colors represent different brain region locations, and Ground Truth is the baseline label. Figure 2 This is a visualization of individual and common features in the multimodal feature collaborative fusion of this invention. Figure 3 This is a multi-angle 3D view of the segmentation result in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the collaborative fusion calculation of common and individual features across multiple sequences and modalities according to an embodiment of the present invention. Figure 5 This is a diagram of the CIA-Net model architecture according to an embodiment of the present invention; Figure 6 This is a flowchart of a brain region segmentation method based on multi-sequence magnetic resonance imaging collaborative feature fusion, according to an embodiment of the present invention. Detailed Implementation
[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0017] like Figure 6 As shown, this embodiment proposes a brain region segmentation method based on multi-sequence magnetic resonance imaging collaborative feature fusion, including: Brain amplitude map data, phase map data, and T1 structural image data were collected; The amplitude map data and the phase map data are preprocessed and reconstructed to obtain a quantitative magnetic susceptibility image; Spatial registration is performed between the T1 structural image data and the quantitative magnetic susceptibility image to obtain paired image data; The paired image data is segmented and labeled to construct a brain region segmentation label dataset; The CIA-Net deep learning model is trained using the training set of the brain region segmentation label dataset to obtain the trained CIA-Net deep learning model. The CIA-Net deep learning model is constructed by fusing an adaptive data domain partitioning module and a common and individual feature attention module. The adaptive data domain partitioning module partitions the training set to obtain multimodal decoupled features, and the common and individual feature attention module fuses the decoupled features to obtain collaborative fusion features. The quantitative magnetic susceptibility image and T1 structural image data to be segmented are input into the trained CIA-Net deep learning model, which outputs the automatic segmentation results of brain regions.
[0018] Further multi-sequence magnetic resonance imaging data acquisition includes: acquiring brain amplitude and phase map data based on multi-echo gradient echo sequences, and acquiring high-resolution T1 structural image data based on three-dimensional fast gradient echo T1 weighted sequences, providing the original data foundation for subsequent quantitative magnetic susceptibility reconstruction and multimodal segmentation.
[0019] In one embodiment, based on the acquisition parameters recommended by QSM expert consensus, a multi-echo gradient echo sequence is configured on the magnetic resonance imaging (MRI) machine to acquire amplitude and phase maps. The key QSM acquisition parameters are: resolution: 1×1×1 mm. 3 FOV: 224×224×144 mm 3 .
[0020] Based on 3D fast gradient echo, high-resolution T1 structural image data can be acquired using a 3D T1-weighted sequence. Examples include SPGR / BRAVO on GE equipment, MPRAGE on Siemens equipment, and TFE on Philips equipment. Alternatively, MPRAGE can be used to configure a 3D T1-weighted sequence.
[0021] Furthermore, the preprocessing and reconstruction of the amplitude map data and the phase map data include: The phase map data is processed using a phase unwrapping algorithm based on the Laplacian operator to obtain a spatially continuous phase map; Based on the amplitude map data, a brain tissue region mask is extracted. Based on the brain tissue region mask, the V-SHARP algorithm is used to remove the background field of the spatially continuous phase map, and the msQSM algorithm is used to perform magnetic susceptibility inversion to obtain the quantitative magnetic susceptibility image.
[0022] Specifically, a phase unwinding algorithm based on the Laplacian operator is used to unwind the 2π jumps in the original phase map, obtaining a spatially continuous phase map. Then, the BET tool in FSL is used to extract a brain tissue region mask based on the amplitude map, which is used to limit the subsequent calculation range to the brain tissue region. Furthermore, the V-SHARP algorithm is used to remove the background field component in the phase map, extracting pure local tissue field information. Based on this, the msQSM algorithm is applied for magnetic susceptibility inversion, converting the local field map into a quantitative magnetic susceptibility (QSM) image, thereby more clearly revealing the magnetic susceptibility distribution characteristics of structures such as the basal ganglia and deep cerebellar nuclei.
[0023] Furthermore, image spatial registration includes: using the ANTs tool to perform a rigid transformation on the T1 image, accurately registering it to the QSM data space, obtaining spatially consistent T1-QSM paired image data, laying a spatial alignment foundation for subsequent multi-sequence feature fusion.
[0024] Specifically, the T1 structural image was spatially transformed using the ANTs (Advanced Normalization Tools) registration tool. First, the QSM image was used as the reference space, and the T1 image was treated as the moving image. Registration optimization was performed using a rigid transformation model (including translation and rotation parameters) to achieve spatial alignment between the two modalities while ensuring that the overall brain structure remained unchanged. Mutual information was used as the similarity metric during registration to enhance the matching accuracy between images under different imaging contrast mechanisms. After registration, the transformation matrix was applied and resampled to ensure that the T1 image resolution, voxel size, and spatial coordinate system were consistent with the QSM data. This resulted in spatially strictly corresponding T1-QSM paired image data, providing a reliable spatial consistency foundation for subsequent multi-sequence feature extraction and collaborative fusion.
[0025] Furthermore, the construction of fine brain region segmentation labels includes: based on T1 data, combined with the MONAI platform and uAIDiscover-Brain tool, to complete the semi-automatic delineation of 109 brain regions based on the DK template; based on QSM data, using the ITK-SNAP tool to complete the fine delineation of 6 brain regions, including the bilateral red nucleus, substantia nigra and dentate nucleus, and finally obtaining a segmentation result containing 115 brain regions, which serves as the standard label for model training.
[0026] Specifically, based on high-resolution T1 structural images, the MONAI deep learning platform and uAI Discover-Brain intelligent segmentation tool were used to semi-automatically segment 109 brain regions according to the DK (Desikan–Killiany) brain region anatomical template. Based on the automatic segmentation results, layer-by-layer correction and manual revision were performed by personnel with neuroimaging experience to ensure consistency between brain region boundaries and anatomical structures, thereby improving the accuracy and reliability of the labels.
[0027] Subsequently, for deep nuclei structures that are relatively weak in T1 contrast but clearly visible on QSM images, the ITK-SNAP tool was used to perform fine manual delineation of six brain regions, including the bilateral red nucleus, substantia nigra, and dentate nucleus, based on QSM data, in order to fully leverage the advantages of QSM in displaying deep iron deposition structures.
[0028] Finally, 109 brain regions from T1 were integrated with 6 deep nuclei from QSM to construct a complete segmentation label dataset containing 115 brain regions. This label set serves as a standard reference for deep learning model training and performance evaluation, providing a high-quality annotation foundation for achieving fine-grained automatic brain region segmentation under multimodal collaborative fusion.
[0029] Furthermore, training the CIA-Net deep learning model using the training set of the brain region segmentation label dataset includes: The training set data is input into the adaptive data domain partitioning module, which outputs the decoupled features. The decoupled features are input into the encoder part, and the output of the encoder part is input into the boundary attention feedback sub-network and the common and individual feature attention module, respectively. The boundary attention feedback sub-network is used to extract boundary features. The brain region segmentation is achieved by fusing features from the boundary attention feedback subnetwork and the skip connection input decoder part of the common and individual feature attention module.
[0030] Specifically, Figure 5 This paper presents the overall architecture of a segmentation network that integrates Common and Individual Attention (CIA) mechanisms with adaptive data domain partitioning (DDP). A boundary-attention feedback (BaF) subnetwork is embedded in the network to learn boundary information, thereby improving segmentation accuracy and structural contour depiction capabilities. The collaborative relationship between the CIA, DDP, and BaF modules is explained below.
[0031] The encoder consists of four stages. First, a CNRD module (Convolution + Norm + ReLU + Dropout) expands the input features to 16 channels without changing the spatial resolution. Then, the input features pass through three ConvBlock modules, each containing two CNRD modules. The first CNRD doubles the number of channels while halving the 3D spatial resolution, and the output of the second CNRD is fed into a BaF module to extract boundary features. The output of the final ConvBlock is then fed into a CIA module for multimodal feature co-modeling. The BaF module and the two CIA modules generate two high-dimensional features and two low-dimensional features, respectively, which are then passed to the decoder as skip connections.
[0032] The decoder consists of three ConvTranspose3d (3D deconvolutional) layers and four Conv3d (3D convolutional) layers. Each ConvTranspose3d layer doubles the spatial resolution while halving the number of feature channels. Skip connections from the BaF and CIA modules are fused at each stage to preserve detail and enhance feature representation. Each Conv3d layer maps features to the label space for deep supervision, thereby improving model training efficiency and segmentation accuracy.
[0033] Furthermore, the adaptive data domain partitioning module partitions the training set into domains to obtain multimodal decoupled features, including: All modal data, semantic segmentation labels, and domain similarity thresholds in the training set are input into the adaptive data domain partitioning module. For each modality, the median, 0.05 percentile, and 99.5 percentile are calculated based on each label, and the domain partitioning of each modality is initialized by combining the corresponding label groups. If the difference between the medians of any two domains is less than the domain similarity threshold, the two domains are merged until the domains converge and stabilize to obtain the multimodal decoupled features. The merging operation includes averaging the medians, 0.05 percentiles, and 99.5 percentiles of the two domains and merging the corresponding label sets.
[0034] Specifically, single-sequence modality data can reflect multiple anatomical structural features of the whole brain. Fusing these features into a single modality increases the burden on feature extraction for deep learning models. Therefore, this embodiment designs an adaptive data domain partitioning algorithm to decouple features for each modality. Table 1 describes the computation process of the adaptive data domain partitioning algorithm on the training set.
[0035] Table 1 The input to the adaptive data domain partitioning algorithm includes all modal data in the training set. Semantic segmentation tags and domain similarity threshold (Line 1). First, for the first... i Modality Based on the first j Tag ,calculate exist The median of the values ( median ), 0.05 percentile ( ) and the 99.5th percentile ( ), and combined with the corresponding tag groups Initialize this mode Domain partitioning (Line 4), is of size l A set of.
[0036] Then, repeat the following process until... Convergence and stability: Initialize the set of medians for each partition ( (line 6); if Any number in the middle j and the k Subdomain and ( The median difference and Less than the threshold (Line 8) will then subdomain and Perform the merge (lines 9 and 10). The merge operation is performed by modifying the two fields. median , and The values are averaged, and their corresponding tag sets are merged.
[0037] Hyperparameters Used to control the number of final partitioned domains: larger This results in fewer domains, potentially leading to insufficient decoupling of intramodal features; smaller... This can generate more domains, potentially leading to over-segmentation of features that originally belonged to the same semantic group. In this embodiment, the threshold... It was set to 0.1 based on experience.
[0038] Furthermore, the common and individual feature attention module fuses the decoupled features by: calculating individual attention for each decoupled feature to obtain individual features of each modality; The attention matrices of the decoupled features are aggregated to extract cross-modal common features; The individual and common features are fused using a multilayer perceptron to obtain the collaborative fusion feature.
[0039] Specifically, this embodiment constructs a common-and-individual attention (CIA) mechanism to decouple common features from modality-specific features in multimodal data. Based on the Transformer self-attention paradigm, this mechanism captures long-range dependencies by modeling pairwise similarities between features, thereby generating context-aware feature representations.
[0040] Given input features Through linear projection function Generate a matrix of queries, keys, and values, i.e. , The formula for calculating self-attention (SA) is as follows: ; in, This represents the attention weight matrix, where each element is used to characterize a feature. The pairwise similarity between elements in the text; Representing feature dimension, As a scaling factor, it is used to prevent gradient explosion caused by excessively large values during the calculation process.
[0041] Figure 4 This paper demonstrates a multimodal feature pattern fusion process based on the CIA mechanism. For each modality, intramodal correlations can be calculated using the SA mechanism to learn individual attention (IA) and identify modality-specific feature patterns. Attention matrices for different modalities are then applied. Aggregation allows for the extraction of common attention (CA). Simultaneously, the feature values from all modalities are pooled to form a unified content representation. Based on this, content-based message aggregation is performed to identify common feature patterns across modalities. ; in, Indicates the first One modality, Indicates the number of modes; attention matrix for each mode. Element-wise multiplication can highlight regions of common interest across different modalities, while also highlighting the modal eigenvalues. Summation achieves the aggregation of multimodal information. Finally, feature fusion is achieved through two layers of a multilayer perceptron (MLP), expressed as: .
[0042] Model training and performance validation include: The collected T1–QSM paired data and corresponding segmentation labels for 115 brain regions were divided into training, validation, and test sets. The training set was used for model parameter learning and optimization; the validation set was used to monitor the training process, adjust hyperparameters, and suppress overfitting; and the test set was used to evaluate the model's segmentation accuracy and generalization ability on independent data. After model training, the segmentation results underwent systematic performance validation, specifically including the following two aspects: 1. Multi-indicator quantitative evaluation: To comprehensively evaluate the model's performance on a fine-grained segmentation task of 115 brain regions, this embodiment employs multiple evaluation metrics for quantitative analysis, including Dice correlation coefficient (DSC), recall (Rec), precision (Pre), 95% Hausdorff distance (HD95), and average surface distance (ASD). Let the predicted segmentation result be... P The real label is G True positive, false positive, and false negative are respectively TP , FP , FN The corresponding definitions are as follows: 1) Dice correlation coefficient (DSC): ; Used to measure the degree of overlap between the predicted region and the actual region, with a value range of [value range missing]. A larger value indicates a better segmentation effect.
[0043] 2) Recall (Rec): ; Used to measure the model's ability to identify true positive regions.
[0044] 3) Precision (Pre): ; Used to measure the proportion of regions that were predicted to be positive but were actually positive.
[0045] 4) 95% Hausdorf Distance (HD95): set up and These are the boundary point sets for the predicted and actual segmentation, respectively. to set The distance is defined as: ; The Hausdorff distance is: ; The 95% Hausdorff distance is defined as: ; This is used to reduce the impact of extreme outliers on boundary distances, thus reflecting boundary errors more stably.
[0046] 5) Average Surface Distance (ASD): ; Used to measure the average distance error between the predicted boundary and the true boundary.
[0047] 2. Comparative experimental verification: To systematically verify the effectiveness and advancement of the CIA-Net model proposed in this embodiment, several representative deep learning segmentation methods for medical images were selected as comparison models, including nnU-Net, nnFormer, UNesT, DDParcel, and UNetR++. These methods cover different technical approaches such as convolutional neural networks (CNN), Transformer-based structures, and hybrid architectures, comprehensively reflecting the current level of mainstream brain region segmentation technologies.
[0048] Under the same data partitioning, training strategy, and evaluation criteria, all comparison models were trained and tested uniformly to ensure the fairness and comparability of the experimental results. Segmentation performance was comprehensively evaluated using multiple metrics, including Dice correlation coefficient (DSC), recall (Rec), precision (Pre), 95% Hausdorff distance (HD95), and average surface distance (ASD). Statistical analysis was performed using the mean and 95% confidence interval, and statistical significance tests were used to verify performance differences.
[0049] Experimental results show that, in the fine-grained segmentation task of 115 brain regions, the CIA-Net proposed in this embodiment outperforms the comparative models such as nnU-Net, nnFormer, UNesT, DDParcel and UNetR++ in terms of overall segmentation accuracy, ability to locate complex structural boundaries and model stability. This fully demonstrates the technical advantages and innovation of the adaptive data domain partitioning mechanism and the common-individual feature attention fusion mechanism in multimodal collaborative modeling.
[0050] Figure 1This paper presents the segmentation results of the CIA-Net method from this embodiment and various deep learning models on 115 brain regions (the comparison models were trained directly on the original QSM+T1 images). Each color in the figure represents a brain region location, with the ground truth as the baseline label. As shown in the figure, the CIA-Net method from this embodiment achieves the most accurate segmentation results, with only a small number of scattered errors. In contrast, UNetR++, nnFormer, and nnU-Net exhibit more significant boundary offsets in the segmentation results, especially in the basal ganglia and cerebrospinal fluid regions. These results demonstrate the significant challenge of achieving accurate boundary localization in anatomically complex regions, and also highlight the significant advantages of CIA-Net in maintaining consistency in segmentation results and the realism of anatomical structures.
[0051] Table 2 The bolded values represent the optimal results for that metric.
[0052] Table 2 presents the quantitative evaluation results of each method. In this embodiment, the CIA-Net method outperforms the comparative methods on all evaluation metrics, with an average performance improvement of 44.81%. Furthermore, CIA-Net exhibits a narrower 95% confidence interval (CI) on the independent test set, indicating higher accuracy and lower volatility; in contrast, other methods show wider confidence intervals, suggesting higher uncertainty and poorer performance stability. These results collectively demonstrate that CIA-Net possesses superior accuracy and robustness in fine-grained brain region segmentation tasks.
[0053] The adaptive domain partitioning algorithm proposed in this embodiment can decouple features for each sequence modality data, dividing multi-sequence features into data domains with common and unique expressions. This reduces redundant interference between different modalities and enhances the independent modeling ability of their respective advantageous information. The adaptive domain partitioning algorithm can be applied to other deep learning segmentation models to improve their segmentation performance.
[0054] Table 3 Table 3 shows the performance metrics of the models trained on the data processed by the adaptive data domain partitioning algorithm proposed in this embodiment. Compared with the data without data domain partitioning, the model's average DSC improved by 11.81% after introducing adaptive domain partitioning, and the overall average performance improvement across the five evaluation metrics reached 74.94%, with all improvements being statistically significant. This indicates that the algorithm can significantly improve the multimodal feature modeling effect and enhance segmentation accuracy. Furthermore, although adaptive data domain partitioning significantly improved the performance of the comparison models, CIA-Net's average performance across all metrics was still 26.61% higher than other methods, demonstrating stronger segmentation ability, boundary accuracy, and result stability. This fully reflects the key role and practical value of adaptive domain partitioning in multi-sequence brain region segmentation tasks.
[0055] The common and individual feature attention fusion mechanism proposed in this embodiment can effectively integrate multimodal features, learn the individual and common features of different modal data, and provide feature visualization results, thus providing interpretability for model segmentation.
[0056] Figure 2 This section presents a visualization of individual and common features in the multimodal feature fusion process. QSM features and T1 features represent the contributions of single-modal features learned by the model from QSM and T1 data to the segmentation results, respectively. Common features represent the contributions of features extracted from both modalities. In the heatmap, each value represents the sum of probabilities for the model's 115 output label channels. For QSM features, T1 features, and common features, only the feature components corresponding to the CIA calculation are retained, while the remaining components are set to zero.
[0057] Visualization results of the contributions of common and individual features show that the common and individual feature fusion mechanism proposed in this embodiment can effectively learn common and individual features from multimodal data, thereby improving the accuracy of brain region segmentation and anatomical consistency.
[0058] Table 4 lists the names and corresponding numbers of the 115 brain regions segmented by CIA-Net. Among them, 51 brain regions are distributed bilaterally in the left and right hemispheres, totaling 102; 13 brain regions are located at the junctions of the left and right hemispheres or the junctions of the cerebellum and cerebellum; for a total of 115 brain regions.
[0059] Figure 3A three-dimensional visualization of the segmentation results for 115 brain regions is presented, including lateral cortical views, medial cortical views, and deep nucleus segmentation views. The spatial distribution and anatomical hierarchy of each brain region within the overall cranial structure are intuitively presented through 3D reconstruction. Different colors correspond to individual brain regions, distinguishing different functional areas and anatomical regions. The lateral and medial cortical views clearly demonstrate the continuity and boundary morphology of each cortical region, while the nucleus segmentation map highlights the spatial location and volumetric morphology of structures such as the basal ganglia, thalamus, and deep cerebellar nuclei. This 3D visualization intuitively reflects the overall consistency, spatial integrity, and structural recognizability of the method in this embodiment for fine-grained segmentation of multiple brain regions, further validating the anatomical rationality and accuracy of the segmentation results.
[0060] Table 4 The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A brain segmentation method based on multi-sequence magnetic resonance imaging collaborative feature fusion, characterized in that, include: Brain amplitude map data, phase map data, and T1 structural image data were collected; The amplitude map data and the phase map data are preprocessed and reconstructed to obtain a quantitative magnetic susceptibility image; Spatial registration is performed between the T1 structural image data and the quantitative magnetic susceptibility image to obtain paired image data; The paired image data is segmented and labeled to construct a brain region segmentation label dataset; The CIA-Net deep learning model is trained using the training set of the brain region segmentation label dataset to obtain the trained CIA-Net deep learning model. The CIA-Net deep learning model is constructed by fusing an adaptive data domain partitioning module and a common and individual feature attention module. The adaptive data domain partitioning module partitions the training set to obtain multimodal decoupled features, and the common and individual feature attention module fuses the decoupled features to obtain collaborative fusion features. The quantitative magnetic susceptibility image and T1 structural image data to be segmented are input into the trained CIA-Net deep learning model, which outputs the automatic segmentation results of brain regions.
2. The brain region segmentation method based on multi-sequence magnetic resonance imaging collaborative feature fusion according to claim 1, characterized in that, The preprocessing and reconstruction of the amplitude map data and the phase map data include: The phase map data is processed using a phase unwrapping algorithm based on the Laplacian operator to obtain a spatially continuous phase map; Based on the amplitude map data, a brain tissue region mask is extracted. Based on the brain tissue region mask, the V-SHARP algorithm is used to remove the background field of the spatially continuous phase map, and the msQSM algorithm is used to perform magnetic susceptibility inversion to obtain the quantitative magnetic susceptibility image.
3. The brain region segmentation method based on multi-sequence magnetic resonance imaging collaborative feature fusion according to claim 1, characterized in that, Training the CIA-Net deep learning model using the training set of the brain region segmentation label dataset includes: The training set data is input into the adaptive data domain partitioning module, which outputs the decoupled features. The decoupled features are input into the encoder part, and the output of the encoder part is input into the boundary attention feedback sub-network and the common and individual feature attention module, respectively. The boundary attention feedback sub-network is used to extract boundary features. The brain region segmentation is achieved by fusing features from the boundary attention feedback subnetwork and the skip connection input decoder part of the common and individual feature attention module.
4. The brain region segmentation method based on multi-sequence magnetic resonance imaging collaborative feature fusion according to claim 3, characterized in that, The encoder part comprises a CNRD module and a convolution module. The CNRD module consists of a convolution block, a normalization block, an activation block, a ReLU block, and a random deactivation block, and the convolution module consists of several CNRD modules. The decoder consists of a three-dimensional deconvolutional layer and a three-dimensional convolutional layer.
5. The brain region segmentation method based on multi-sequence magnetic resonance imaging collaborative feature fusion according to claim 1, characterized in that, The adaptive data domain partitioning module partitions the training set into domains to obtain multimodal decoupled features, including: All modal data, semantic segmentation labels, and domain similarity thresholds in the training set are input into the adaptive data domain partitioning module. For each modality, the median, 0.05 percentile, and 99.5 percentile are calculated based on each label, and the domain partitioning of each modality is initialized by combining the corresponding label groups. If the difference between the medians of any two domains is less than the domain similarity threshold, the two domains are merged until the domains converge and stabilize to obtain the multimodal decoupled features. The merging operation includes averaging the medians, 0.05 percentiles, and 99.5 percentiles of the two domains and merging the corresponding label sets.
6. The brain region segmentation method based on multi-sequence magnetic resonance imaging collaborative feature fusion according to claim 1, characterized in that, The common and individual feature attention module fuses the decoupled features by: calculating individual attention for each decoupled feature to obtain individual features of each modality; The attention matrices of the decoupled features are aggregated to extract cross-modal common features; The individual and common features are fused using a multilayer perceptron to obtain the collaborative fusion feature.
7. The brain region segmentation method based on multi-sequence magnetic resonance imaging collaborative feature fusion according to claim 6, characterized in that, Calculating Personal Attention (SA) includes: ； Where A represents the attention weight matrix, d represents the feature dimension, and Q represents the query. The key is represented by the superscript T, which indicates transpose, and X represents the input feature. These are the eigenvalues.
8. The brain region segmentation method based on multi-sequence magnetic resonance imaging collaborative feature fusion according to claim 6, characterized in that, Extracting cross-modal common features (CA) includes: in, Indicates the first One modality, Indicates the number of modes; For each modality attention matrix, These are the characteristic values for each modality.