Intelligent recognition and typing method and system for spatial distribution pattern of brain iron deposition

Abstract

The present application relates to the technical field of medical image processing and artificial intelligence, and discloses a brain iron deposition spatial distribution pattern intelligent recognition and typing method and system, which comprises: spatial standardization and brain region segmentation of whole brain QSM images; generating a spatial distribution feature vector through a multi-scale spatial attention feature extraction network; dividing the iron deposition pattern into a cortical type, a deep nuclear group type, a periventricular type and a diffuse type by using a fuzzy clustering algorithm with spatial neighborhood constraints; establishing a mapping relationship between the iron deposition pattern and the etiological subtype of vascular cognitive impairment; and evaluating the bilateral hemisphere iron deposition asymmetry, wherein the present application realizes automatic recognition and typing of the brain iron deposition spatial distribution pattern, the pattern recognition accuracy reaches more than 85%, and provides an auxiliary diagnosis basis for etiological typing of vascular cognitive impairment.

Description

Technical Field

[0001] This invention relates to the technical fields of medical image processing and artificial intelligence, specifically to a method and system for intelligent recognition and classification of spatial distribution patterns of brain iron deposition. Background Technology

[0002] Iron, an important metallic element in the brain, plays a crucial role in axonal myelination, energy metabolism, and neurotransmitter synthesis. The iron content in specific brain regions gradually changes with age. Disruption of the iron homeostasis system can lead to a series of age-related neurodegenerative diseases. Previous studies have shown abnormal iron deposition in the brain in diseases such as Alzheimer's disease, Parkinson's disease, multiple sclerosis, and vascular cognitive impairment. Therefore, quantitatively assessing and detecting the spatial distribution characteristics of brain iron deposition is of great significance for understanding disease mechanisms and guiding clinical interventions.

[0003] Quantitative susceptibility mapping (QSM) is an advanced magnetic resonance imaging technique that non-invasively and quantitatively assesses the magnetic susceptibility distribution of tissues in vivo through complex image post-processing and inversion algorithms. Since paramagnetic iron is the primary source of gray matter magnetic susceptibility, QSM has become an ideal tool for non-invasive quantitative analysis of iron deposition in vivo and has been widely applied in research on neurological diseases.

[0004] Chinese Patent Publication No. CN116934662A discloses a data processing method and apparatus for whole-brain quantitative magnetic susceptibility magnetic resonance imaging (QSM). This method reconstructs the original gradient echo data using QSM, constructs a standardized brain template for individual space using structural images, registers the subject's QSM image with the brain template to achieve spatial standardization, and performs data analysis based on voxels or regions of interest. This method can automatically segment QSM images, enabling quantitative analysis of magnetic susceptibility values ​​and improving repeatability and comparability.

[0005] However, the above-mentioned technical solutions have the following shortcomings: First, the method only achieves standardized processing and basic statistical analysis of QSM images, and fails to perform in-depth feature mining and pattern recognition on the spatial distribution patterns of brain iron deposition, thus failing to effectively distinguish different spatial distribution types of iron deposition; second, the method lacks an association analysis mechanism between iron deposition patterns and disease etiological subtypes, and cannot provide auxiliary diagnostic support for etiological classification of vascular cognitive impairment; third, the method does not consider the asymmetric assessment of bilateral hemisphere iron deposition, and cannot identify asymmetric iron metabolism abnormalities related to unilateral vascular lesions; finally, the method lacks a collaborative optimization mechanism among its various processing modules, and cannot improve overall analytical performance through end-to-end learning.

[0006] Therefore, there is an urgent need for an intelligent analysis method that can automatically extract the spatial distribution characteristics of brain iron deposition, identify different iron deposition patterns, establish the association between iron deposition patterns and etiological subtypes of vascular cognitive impairment, and assess the asymmetry of bilateral hemisphere iron deposition. Summary of the Invention

[0007] To address the technical problems of strong heterogeneity in the spatial distribution patterns of brain iron deposition and the difficulty in analyzing their correlation with the etiology of vascular cognitive impairment in existing technologies, this invention provides an intelligent identification and classification method and system for the spatial distribution patterns of brain iron deposition.

[0008] The first aspect of the present invention provides a method for intelligent identification and classification of spatial distribution patterns of brain iron deposition, comprising the following steps:

[0009] Step S1, QSM image preprocessing step: Obtain the whole brain quantitative magnetic susceptibility map image of the subject, perform spatial normalization processing on the whole brain quantitative magnetic susceptibility map image, register the whole brain quantitative magnetic susceptibility map image to the standard brain template space, and perform brain region segmentation on the registered image to generate a brain region segmentation mask, wherein the brain region segmentation mask includes the cortical region, the deep nucleus region and the periventricular region.

[0010] Step S2, Spatial Distribution Feature Extraction Step: Input the preprocessed quantitative magnetic susceptibility map image and the brain region segmentation mask into the multi-scale spatial attention feature extraction network. Extract the spatial distribution features of iron deposition within different receptive fields through multi-scale convolutional layers. Use the spatial attention mechanism to weightedly fuse the features of different brain regions to generate a spatial distribution feature vector. Based on the brain region segmentation mask, statistically analyze the magnetic susceptibility value distribution features of each region to generate regional magnetic susceptibility statistical features.

[0011] Step S3, iron deposition pattern clustering and classification step: Input the spatial distribution feature vector and the regional magnetic susceptibility statistical features into the adaptive iron deposition pattern clustering model, and use the fuzzy clustering algorithm based on spatial neighborhood constraints to classify the spatial distribution pattern of iron deposition into cortical type, deep nucleus type, periventricular type and diffuse type, and output the iron deposition pattern label and pattern confidence score.

[0012] Step S4, asymmetry index analysis step: Based on the brain region segmentation mask, the quantitative magnetic susceptibility map image is divided into the left hemisphere region and the right hemisphere region. The magnetic susceptibility value distribution characteristics of the left and right hemisphere regions are statistically analyzed respectively. The asymmetry index of the two hemispheres is calculated to quantify the degree of difference in iron deposition between the left and right hemispheres. The asymmetry index is used to identify asymmetric iron metabolism abnormalities related to unilateral vascular lesions. The asymmetry index is passed to step S5 for etiological subtype association mapping.

[0013] Step S5, Etiological Subtype Association Mapping Step: Based on the iron deposition pattern label, the pattern confidence score, and the asymmetry index output in Step S4, an association relationship between iron deposition patterns and vascular cognitive impairment etiological subtypes is established through an etiological subtype probability mapping network, and an etiological subtype probability vector is output. The vascular cognitive impairment etiological subtypes include multi-infarction type, small vessel disease type, hypoperfusion type, and mixed type. Etiological subtyping auxiliary prompt information is generated based on the etiological subtype probability vector.

[0014] Preferably, in step S1, the spatial normalization process includes: rigidly registering the whole-brain quantitative magnetic susceptibility map image with a high-resolution structural image to obtain a first registration transformation matrix; nonlinearly registering the high-resolution structural image with a standard brain template to obtain a second registration transformation matrix; and transforming the whole-brain quantitative magnetic susceptibility map image to the standard brain template space based on the first registration transformation matrix and the second registration transformation matrix.

[0015] Preferably, in step S2, the multi-scale spatial attention feature extraction network includes a first convolutional branch, a second convolutional branch, and a third convolutional branch in parallel. The kernel size of the first convolutional branch is 3×3×3, the kernel size of the second convolutional branch is 5×5×5, and the kernel size of the third convolutional branch is 7×7×7.

[0016] Preferably, the regional magnetic susceptibility statistical characteristics include the mean magnetic susceptibility value, standard deviation of magnetic susceptibility value, skewness of magnetic susceptibility value, and kurtosis of magnetic susceptibility value for each brain region.

[0017] Preferably, in step S3, the adaptive iron deposition pattern clustering model is implemented based on a fuzzy clustering algorithm with spatial neighborhood constraints, and the algorithm introduces a spatial neighborhood regularization term in the optimization objective function.

[0018] Preferably, the method further includes a closed-loop feedback optimization step: calculating the feedback error signal based on the cross-entropy loss between the etiological subtype probability vector and the labeled true etiological subtype label, and backpropagating the feedback error signal to the multi-scale spatial attention feature extraction network and the adaptive iron deposition pattern clustering model.

[0019] The second aspect of the present invention provides an intelligent identification and classification system for spatial distribution patterns of brain iron deposition, comprising: a QSM image preprocessing module, a spatial distribution feature extraction module, an iron deposition pattern clustering and classification module, an etiological subtype association mapping module, and an asymmetric index analysis module, the functions of each module corresponding one-to-one with the steps of the above method.

[0020] The beneficial effects of this invention include: First, it automatically extracts the spatial distribution features of brain iron deposition through a multi-scale spatial attention feature extraction network, avoiding the limitations of traditional methods that rely on manual feature design, and exhibits strong feature expression capabilities and good generalization performance; Second, it achieves automatic identification and classification of iron deposition spatial distribution patterns through an adaptive iron deposition pattern clustering algorithm, which can summarize complex iron deposition distribution patterns into four typical patterns: cortical, deep nucleus, periventricular, and diffuse, with a pattern recognition accuracy rate of over 85%; Third, it establishes a correlation mapping relationship between iron deposition patterns and vascular cognitive impairment etiological subtypes, providing objective auxiliary diagnostic basis for clinical etiological classification; Fourth, it assesses the difference in iron deposition between the two hemispheres through an asymmetric index analysis module, which can identify asymmetric iron metabolism abnormalities caused by unilateral vascular lesions; Fifth, it adopts a closed-loop feedback optimization mechanism to achieve end-to-end collaborative learning, forming a deeply coupled closed-loop collaborative relationship between modules, and the overall analysis performance is better than the effect of each module processing independently. Attached Figure Description

[0021] Figure 1 This is a flowchart of the intelligent identification and classification method for spatial distribution patterns of brain iron deposition provided in an embodiment of the present invention.

[0022] Figure 2 This is an architecture diagram of the intelligent recognition and classification system for spatial distribution patterns of brain iron deposition provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] like Figure 1 As shown, the intelligent identification and classification method for spatial distribution patterns of brain iron deposition provided in this embodiment of the invention includes steps S1 to S5, and a closed-loop collaborative relationship of forward data transmission and reverse feedback optimization is formed between each step. Each step is described in detail below.

[0025] Step S1: QSM image preprocessing step.

[0026] This step is used to standardize the input whole-brain QSM images for preprocessing, providing a unified and standardized data foundation for subsequent feature extraction and pattern analysis.

[0027] First, quantitative susceptibility mapping (QSM) images of the whole brain are acquired from a magnetic resonance scanner. In a preferred embodiment of the invention, a 3.0T magnetic resonance scanner with a 32-channel head coil is used to acquire multi-echo three-dimensional gradient echo sequence images. The QSM images are obtained through phase unwinding, background field removal, and inversion calculation. Preferably, the acquisition parameters include: echo time of 6.8 ms for the first echo, echo interval of 4.2 ms, and a total of 8 echoes acquired; repetition time of 50 ms; flip angle of 15 degrees; and voxel size of 1.0 mm × 1.0 mm × 1.0 mm.

[0028] Next, the whole-brain QSM images are spatially standardized. Due to individual differences in brain morphology among different subjects, all images need to be registered to a unified standard brain template space for subsequent group analysis and pattern comparison. Spatial standardization specifically includes a two-step registration process: First, the whole-brain QSM images are rigidly registered with high-resolution T1-weighted structural images of the same subject. Since the QSM images and T1 structural images originate from the same scan and the head position is fixed, a six-parameter rigid transformation model (including three translation parameters and three rotation parameters) can achieve accurate registration, resulting in the first registration transformation matrix. Second, the high-resolution T1 structural images are nonlinearly registered with the standard brain template, using a symmetric normalization algorithm for multi-resolution iterative registration, resulting in the second registration transformation matrix. Finally, based on the composite transformation of the first and second registration transformation matrices, a third-order B-spline interpolation algorithm is used to transform the whole-brain QSM images to the standard brain template space.

[0029] Then, brain region segmentation is performed on the registered standardized QSM images to generate brain region segmentation masks. In a preferred embodiment of the present invention, a pre-trained brain region segmentation model is used to automatically segment images in the standard brain template space. The brain region segmentation mask includes the following brain regions: cortical regions, further subdivided into the frontal cortex, parietal cortex, temporal cortex, and occipital cortex; deep nuclei regions, including the caudate nucleus, putamen, globus pallidus, thalamus, red nucleus, substantia nigra, and hippocampus; and periventricular regions, including the periventricular white matter of the lateral ventricles and the periventricular region of the third ventricle. Each brain region corresponds to a binary mask, with an internal voxel value of 1 and an external voxel value of 0.

[0030] In a preferred embodiment of the present invention, the specific implementation process of brain region segmentation is as follows: First, a standardized high-resolution T1 structural image is input into a pre-trained 3D U-Net segmentation network. This network adopts an encoder-decoder structure. The encoder contains four downsampling stages, each containing two 3×3×3 convolutional layers and one 2×2×2 max-pooling layer; the decoder contains four upsampling stages, each containing one 2×2×2 transposed convolutional layer and two 3×3×3 convolutional layers. The feature maps of the corresponding layers of the encoder and the feature maps of the decoder are concatenated through skip connections. The network finally outputs a multi-channel segmentation probability map with the same size as the input, with each channel corresponding to the segmentation probability of a brain region. The brain region attribution label of each voxel is determined by taking the maximum value of each channel, generating a brain region segmentation mask.

[0031] The quality of the brain region segmentation mask directly affects the accuracy of subsequent extraction of iron deposition spatial distribution features and asymmetry analysis. To ensure segmentation quality, this invention employs the following quality control strategies: First, during the training phase, the segmentation network is trained using a dataset containing 500 manually labeled brain regions to ensure good segmentation performance. Second, during the inference phase, morphological post-processing is performed on the segmentation results, including hole filling, small connected component removal, and edge smoothing, to eliminate segmentation noise. Third, volume consistency checks are performed on key brain regions such as the basal ganglia and hippocampus; if the segmented volume exceeds the normal range, a manual review mechanism is triggered.

[0032] In addition, this step involves intensity normalization of the standardized QSM images, normalizing the magnetic susceptibility values ​​to a fixed range to eliminate intensity differences caused by different scanning devices and acquisition parameters. Specifically, a cerebrospinal fluid-based zero-reference correction method is used to correct the mean magnetic susceptibility value of the lateral ventricle region to zero, while the magnetic susceptibility values ​​of other brain regions are adjusted relative to the cerebrospinal fluid reference value. After intensity normalization, the QSM images of different subjects are comparable, facilitating subsequent group analysis and pattern comparison.

[0033] The preprocessing steps described above convert the original QSM image into standardized spatial distribution data and generate a detailed brain region segmentation mask, laying the foundation for subsequent spatial distribution feature extraction and asymmetry analysis. The output of this step includes the standardized QSM image and the brain region segmentation mask, which are then passed to steps S2 and S4, respectively.

[0034] Step S2: Spatial distribution feature extraction step.

[0035] This step is used to automatically extract high-dimensional feature vectors representing the spatial distribution pattern of brain iron deposition from the preprocessed QSM image.

[0036] The preprocessed standardized QSM image and brain region segmentation mask are used as inputs and fed into a multi-scale spatial attention feature extraction network. This network is one of the core innovations of this invention, capable of adaptively extracting iron deposition distribution features at different spatial scales and highlighting key brain regions associated with iron deposition patterns through an attention mechanism.

[0037] The architecture of the multi-scale spatial attention feature extraction network is as follows: The network adopts an encoder structure design, including three parallel multi-scale convolutional branches and a spatial attention module. Specifically, the first convolutional branch uses a 3×3×3 three-dimensional convolutional kernel with a small receptive field, capable of capturing local iron deposition distribution details, such as focal iron deposition within deep nuclei; the second convolutional branch uses a 5×5×5 three-dimensional convolutional kernel with a moderate receptive field, capable of capturing iron deposition distribution patterns at the brain region level; the third convolutional branch uses a 7×7×7 three-dimensional convolutional kernel with a large receptive field, capable of capturing the spatial distribution patterns of iron deposition across brain regions. The three branches process the input image in parallel, with each branch containing two convolutional layers, followed by a batch normalization layer and a modified linear unit activation function after each convolutional layer.

[0038] The output feature maps of the three convolutional branches are concatenated along the channel dimension to form a multi-scale fused feature map. Then, the concatenated feature map is fed into the spatial attention module. This module generates a spatial attention weight map based on a brain region segmentation mask, enabling the network to adaptively focus on key brain regions related to iron deposition pattern recognition. The calculation process for the spatial attention weights is as follows:

[0039] First, the multi-scale fused feature map and the brain region segmentation mask are concatenated along the channel dimension. Then, attention features are extracted using a two-layer convolutional network. Finally, the feature values ​​are normalized to the range of 0 to 1 using the sigmoid activation function to obtain the spatial attention weight map. The multi-scale fused feature map and the spatial attention weight map are multiplied element-wise to obtain the weighted feature map.

[0040] The spatial attention weights are calculated using the following formula:

[0041] ,

[0042] in: For position Spatial attention weights at each location; Use the Sigmoid activation function; and These are the weight matrices for the first and second convolutional layers, respectively. and These are the corresponding bias vectors; For multi-scale fusion feature maps at location The eigenvector at that location; For brain region segmentation mask at location The encoded vector at that location; Represents a vector concatenation operation; ReLU is the modified linear unit activation function.

[0043] The weighted feature map is then subjected to global average pooling to compress the three-dimensional feature map into a one-dimensional feature vector, i.e., a spatially distributed feature vector. In a preferred embodiment of the present invention, the spatially distributed feature vector has a dimension of 512.

[0044] The training process of the multi-scale spatial attention feature extraction network is as follows: An end-to-end supervised learning approach is adopted, using iron deposition pattern labels as supervisory signals to train the network parameters. The training dataset contains QSM image samples labeled with iron deposition pattern categories. The training set size is expanded using data augmentation techniques, including random rotation (rotation angle range from -15 degrees to +15 degrees), random translation (translation range from -10 voxels to +10 voxels), random scaling (scaling ratio range from 0.9 to 1.1), and random noise addition (Gaussian noise with a standard deviation of 0.01). The training process uses the Adam optimizer with an initial learning rate of 0.0001, employing a cosine annealing learning rate scheduling strategy, with 200 training epochs and a batch size of 16.

[0045] The physical meaning of the spatial distribution feature vectors can be interpreted as follows: The features extracted by the first convolutional branch mainly encode the intensity and texture information of local iron deposition, enabling the capture of iron deposition hotspots within deep nuclei; the features extracted by the second convolutional branch mainly encode the distribution patterns of iron deposition at the brain region level, enabling the differentiation between cortical and deep nucleus-type iron deposition patterns; the features extracted by the third convolutional branch mainly encode the spatial gradient information of iron deposition across brain regions, enabling the identification of the whole-brain distribution features of diffuse iron deposition patterns. The spatial attention mechanism allows the network to adaptively emphasize key brain regions relevant to pattern recognition and suppress background regions unrelated to pattern recognition, thereby improving the discriminative power of the features.

[0046] During feature extraction, the network simultaneously outputs an attention weight map, which visualizes the brain regions the network focuses on. By analyzing the attention weight map, clinicians can intuitively understand which brain regions' iron deposition plays a decisive role in the final pattern classification, enhancing the model's interpretability.

[0047] Furthermore, this step also generates regional magnetic susceptibility statistical features based on the distribution characteristics of magnetic susceptibility values ​​in each region according to the brain region segmentation mask. For each brain region, the following statistics are calculated: mean magnetic susceptibility value, representing the average level of iron deposition in that brain region; standard deviation of magnetic susceptibility value, representing the dispersion of iron deposition distribution in that brain region; skewness of magnetic susceptibility value, representing the asymmetry of iron deposition distribution in that brain region; and kurtosis of magnetic susceptibility value, representing the sharpness of iron deposition distribution in that brain region.

[0048] The statistical characteristics of magnetic susceptibility in different zones are calculated using the following formula:

[0049] The formula for calculating the mean is:

[0050] ,

[0051] Formula for calculating standard deviation:

[0052] ,

[0053] Skewness calculation formula:

[0054] ,

[0055] Kurtosis calculation formula:

[0056] ,

[0057] in: brain region The average magnetic susceptibility value; brain region The standard deviation of the magnetic susceptibility value; brain region Magnetic susceptibility deviation; brain region Magnetic susceptibility kurtosis; brain region The number of voxels within; brain region The corresponding set of voxels; For position The magnetic susceptibility value at that location.

[0058] In a preferred embodiment of the present invention, four statistical measures are calculated for each of the 10 brain regions (frontal cortex, parietal cortex, temporal cortex, occipital cortex, caudate nucleus, putamen, globus pallidus, thalamus, hippocampus and periventricular white matter) to generate a total of 40-dimensional regional magnetic susceptibility statistical feature vectors.

[0059] The output of this step includes a 512-dimensional spatial distribution feature vector and a 40-dimensional regional magnetic susceptibility statistical feature vector. The two are concatenated and passed to step S3 for iron deposition pattern clustering and typing.

[0060] Step S3: Iron deposition pattern clustering and classification step.

[0061] This step is used to perform cluster analysis on the extracted spatial distribution features to identify and classify different spatial distribution patterns of brain iron deposition.

[0062] The spatial distribution feature vector and the regional magnetic susceptibility statistical feature vector output from step S2 are concatenated into a 552-dimensional comprehensive feature vector, which is then input into the adaptive iron deposition pattern clustering model. This model is based on a fuzzy clustering algorithm with spatial neighborhood constraints and can classify the spatial distribution patterns of iron deposition in subjects into four typical types: cortical, deep nucleus, periventricular, and diffuse.

[0063] The four iron deposition patterns are defined as follows: Cortical type: Iron deposition is mainly distributed in the cerebral cortex, with the mean magnetic susceptibility value of the cortex being significantly higher than that of the deep nuclei and periventricular regions; Deep nucleus type: Iron deposition is mainly distributed in the basal ganglia (caudate nucleus, putamen, globus pallidus) and thalamus, with the mean magnetic susceptibility value of these regions being significantly higher than that of the cortex and periventricular regions; Periventricular type: Iron deposition is mainly distributed in the periventricular white matter region, with the mean magnetic susceptibility value of this region being significantly higher than that of the cortex and deep nuclei; Diffuse type: Iron deposition is distributed in the cortex, deep nuclei, and periventricular regions, with the mean magnetic susceptibility values ​​of each region being similar and showing no obvious regional bias.

[0064] The objective function of the adaptive iron deposition pattern clustering algorithm is as follows:

[0065] ,

[0066] in: The objective function value; The total number of samples; The number of clusters, in this invention Corresponding to four iron deposition modes; For the first The sample belongs to the first The membership degree of each cluster satisfies and ; The fuzziness index controls the degree of fuzziness in clustering, with a preferred value of 2. For the first Feature vectors of each sample; For the first The center vectors of each cluster; The distance is Euclidean. The spatial neighborhood regularization coefficient controls the strength of spatial constraints, and its preferred value range is 0.1 to 0.5. For the first The spatial neighborhood sample set of each sample.

[0067] The first term of the objective function is the objective term of classical fuzzy C-means clustering, which encourages samples to move closer to their respective cluster centers; the second term is the spatial neighborhood regularization term, which encourages spatially adjacent samples to have similar cluster membership, thereby ensuring the spatial continuity of iron deposition pattern classification.

[0068] The update formula for the cluster center vector is as follows:

[0069] ,

[0070] The membership update formula is as follows:

[0071] ,

[0072] The algorithm employs an alternating iterative optimization strategy: first, the membership matrix is ​​fixed, and the cluster center vector is updated; then, the cluster center vector is fixed, and the membership matrix is ​​updated. The iteration terminates when the relative change in the objective function value is less than a preset threshold (preferably 0.001) or the number of iterations reaches a preset upper limit (preferably 100 times).

[0073] After the algorithm converges, it outputs the iron deposition pattern label and pattern confidence score for each sample. The iron deposition pattern label is taken from the cluster category with the highest membership degree.

[0074] ,

[0075] The pattern confidence score is the corresponding maximum membership value:

[0076] ,

[0077] The output of this step includes an iron deposition pattern label (cortical, deep nucleus, periventricular, or diffuse) and a pattern confidence score (a real number between 0 and 1), which is passed to step S5 for etiological subtype association mapping.

[0078] The initialization strategy of the adaptive iron deposition pattern clustering algorithm is as follows: the initialization of cluster centers has a significant impact on the convergence speed and final result of the algorithm. This invention employs a cluster center initialization method based on prior knowledge, pre-setting the feature distributions of four typical iron deposition patterns according to clinical experience: the initial cluster centers for the cortical pattern are set as feature vectors with high magnetic susceptibility values ​​in the cortical region and low magnetic susceptibility values ​​in the deep nuclei and periventricular regions; the initial cluster centers for the deep nucleus pattern are set as feature vectors with high magnetic susceptibility values ​​in the basal ganglia region and low magnetic susceptibility values ​​in the cortex and periventricular regions; the initial cluster centers for the periventricular pattern are set as feature vectors with high magnetic susceptibility values ​​in the periventricular white matter and low magnetic susceptibility values ​​in the cortex and deep nuclei; and the initial cluster centers for the diffuse pattern are set as feature vectors with uniformly distributed magnetic susceptibility values ​​in each region. This prior knowledge-based initialization method can accelerate algorithm convergence and avoid getting trapped in undesirable local optima.

[0079] The choice of spatial neighborhood regularization coefficient has a significant impact on clustering results. When the regularization coefficient is too small, the clustering results may exhibit spatially discontinuous and noisy classifications; when the regularization coefficient is too large, the clustering results may become overly smooth, losing local details. This invention determines the optimal regularization coefficient through cross-validation experiments, evaluates the clustering consistency index and pattern recognition accuracy under different regularization coefficients on the validation set, and selects the regularization coefficient value with the best overall performance. In a preferred embodiment of this invention, the optimal regularization coefficient is 0.25.

[0080] Pattern confidence scores are used to assess the reliability of pattern classification. A pattern confidence score close to 1 indicates very strong evidence that the sample belongs to a specific iron deposition pattern, and the classification result is highly reliable. A pattern confidence score close to a critical value (e.g., 0.4) indicates that the sample characteristics fall between multiple patterns, and the classification result needs careful interpretation. In clinical applications, for samples with low pattern confidence scores, the system will prompt clinicians to make a comprehensive judgment based on other imaging and clinical information.

[0081] Step S4: Asymmetric index analysis steps.

[0082] This step is used to assess the asymmetry of iron deposition distribution in the bilateral cerebral hemispheres and to identify asymmetric iron metabolism abnormalities associated with unilateral vascular lesions.

[0083] Based on the brain region segmentation mask output in step S1, the brain regions of the standardized QSM image are divided into the left hemisphere and the right hemisphere regions along the midline. For each brain region... Extracting the left brain region separately and the right brain region A set of voxels.

[0084] The distribution characteristics of magnetic susceptibility values ​​in each brain region of the left and right hemispheres were statistically analyzed separately. In this step, the mean magnetic susceptibility value was mainly calculated for asymmetry assessment.

[0085] Calculate the whole-brain asymmetry index and the regional asymmetry index vector. The whole-brain asymmetry index quantifies the degree of difference in the overall level of iron deposition between the left and right hemispheres, and the calculation formula is as follows:

[0086] ,

[0087] in: The whole-brain asymmetry index; This represents the average magnetic susceptibility value of all voxels in the left hemisphere; This represents the average magnetic susceptibility value of all voxels in the right hemisphere. The whole-brain asymmetry index ranges from -2 to 2. A positive value indicates that the iron deposition level in the left hemisphere is higher than that in the right hemisphere, while a negative value indicates that the iron deposition level in the right hemisphere is higher than that in the left hemisphere. The larger the absolute value, the greater the degree of asymmetry.

[0088] The partition asymmetry index vector quantifies the difference in iron deposition levels between the left and right sides of each corresponding brain region. The calculation formula is as follows:

[0089] ,

[0090] in: brain region The asymmetry index; Left brain region The average magnetic susceptibility value; Right brain region The average magnetic susceptibility value.

[0091] In a preferred embodiment of the present invention, the partition asymmetry index is calculated for each of the five pairs of corresponding brain regions (frontal cortex, parietal cortex, temporal cortex, caudate nucleus, and putamen-thalamus) to form a 10-dimensional partition asymmetry index vector.

[0092] The criteria for diagnosing asymmetric iron metabolism abnormalities are as follows: when the absolute value of the whole-brain asymmetry index exceeds a preset asymmetry threshold (preferably 0.15), it is considered that there is a significant whole-brain asymmetric iron metabolism abnormality; when the absolute value of the asymmetry index of any region exceeds a preset region asymmetry threshold (preferably 0.20), it is considered that there is a localized asymmetric iron metabolism abnormality in the brain region. Asymmetric iron metabolism abnormalities suggest the possible presence of unilateral vascular lesions and require further evaluation in conjunction with imaging and clinical information.

[0093] The output of this step includes the whole-brain asymmetry index and the regional asymmetry index vector, which are passed to step S5 as one of the input features of the etiological subtype probability mapping network.

[0094] The clinical significance of asymmetric iron metabolism abnormalities is explained as follows: In healthy individuals, iron deposition distribution in both cerebral hemispheres typically exhibits a basically symmetrical pattern, with minimal difference in magnetic susceptibility values ​​between corresponding brain regions in the left and right hemispheres. Significant asymmetric iron deposition may indicate the following pathological conditions: First, local hypoperfusion caused by unilateral large vessel stenosis or occlusion, leading to abnormal iron metabolism in the corresponding blood supply area; second, hemosiderin deposition following unilateral cerebral infarction, resulting in a significantly elevated magnetic susceptibility value on the lesion side; third, increased local iron deposition due to unilateral small vessel lesions. Therefore, asymmetric index analysis provides important imaging evidence for identifying unilateral vascular lesions.

[0095] When calculating the asymmetry index, the following points should be noted: First, brain regions that cross the midline, such as the corpus callosum and midline structures, are not included in the calculation of the regional asymmetry index; second, for smaller structures such as the hippocampus, the interpretation of their asymmetry index needs to be more cautious because segmentation errors have a significant impact on the statistics of magnetic susceptibility values; third, age has a certain influence on iron deposition asymmetry, and age-related physiological changes need to be considered when judging asymmetry abnormalities.

[0096] The asymmetry threshold is set based on statistical analysis of a large-scale healthy population. In a preferred embodiment of the invention, a dataset containing 200 healthy volunteers is used to calculate the asymmetry index distribution of healthy individuals in each age group. The mean plus or minus two standard deviations is taken as the normal range, and values ​​exceeding this range are considered abnormal. For the whole-brain asymmetry index, the upper and lower bounds of the normal range are ±0.15, respectively. For the regional asymmetry index, due to the different physiological variability of different brain regions, the normal range thresholds for each brain region vary, ranging from ±0.15 to ±0.25.

[0097] Step S5: Etiological subtype association mapping step.

[0098] This step is used to establish the association between the spatial distribution pattern of brain iron deposition and the etiological subtypes of vascular cognitive impairment, and to output etiological subtyping auxiliary information.

[0099] Clinical studies have revealed a correlation between different spatial distribution patterns of iron deposition and different etiological subtypes of vascular cognitive impairment: cortical iron deposition patterns are commonly seen in multi-infarct vascular cognitive impairment caused by large vessel disease; deep nucleus iron deposition patterns are commonly seen in small vessel disease-related vascular cognitive impairment caused by perforator artery disease; periventricular iron deposition patterns are commonly seen in watershed white matter damage caused by hypoperfusion; and diffuse iron deposition patterns are common in mixed vascular cognitive impairment. Furthermore, the asymmetry of iron deposition in both hemispheres is also an important reference indicator for etiological classification.

[0100] The etiological subtype probability mapping network receives the following inputs: the iron deposition pattern label output in step S3 (encoded as an isolated heat vector with a dimension of 4), the pattern confidence score output in step S3 (scalar), and the whole-brain asymmetry index (scalar) and partition asymmetry index vector (with a dimension of 10) output in step S4. The input features are concatenated to form a 15-dimensional input vector.

[0101] The architecture of the etiology subtype probability mapping network is as follows: The network contains three fully connected layers. The first fully connected layer maps the 15-dimensional input vector to a 64-dimensional hidden layer representation, using ReLU as the activation function; the second fully connected layer maps the 64-dimensional hidden layer representation to a 32-dimensional hidden layer representation, also using ReLU as the activation function; the third fully connected layer maps the 32-dimensional hidden layer representation to a 4-dimensional output vector, corresponding to the original scores of the four etiology subtypes. The output vector is normalized using the Softmax function to obtain the etiology subtype probability vector.

[0102] The formula for calculating the probability vector of etiological subtypes is as follows:

[0103] ,

[0104] in: Given input features The time prediction is the first The probability of a specific etiological subtype; The output of the third fully connected layer One component; The etiological subtype variable ranges from 1 to 4, corresponding to multiple infarction type, small vessel disease type, hypoperfusion type and mixed type, respectively.

[0105] The category corresponding to the largest component in the etiological subtype probability vector is used as the predicted etiological subtype:

[0106] ,

[0107] Based on the predicted etiological subtype and its corresponding probability value, etiological subtyping auxiliary prompts are generated. In a preferred embodiment of the present invention, when the predicted probability exceeds a preset high confidence threshold (preferably 0.7), the prompt is marked as high confidence; when the predicted probability is in the medium confidence range (preferably 0.4 to 0.7), the prompt is marked as medium confidence; when the predicted probability is below a low confidence threshold (preferably 0.4), the prompt is marked as low confidence, and it is recommended to combine it with other clinical information for comprehensive judgment.

[0108] The output of this step includes a pathological subtype probability vector and pathological classification auxiliary prompts, which are presented to clinicians as the final auxiliary diagnostic results.

[0109] The training process of the etiology subtype probability mapping network is as follows: Supervised learning is employed, using clinically diagnosed vascular cognitive impairment etiology subtype labels as the training target. The training dataset needs to contain patient samples whose etiology subtypes have been determined through comprehensive evaluation by clinicians. Each sample includes QSM image data and the corresponding etiology subtype label. The training process uses the cross-entropy loss function, and the network parameters are optimized through the backpropagation algorithm. To prevent overfitting, Dropout regularization layers are added between fully connected layers, with the Dropout probability set to 0.5.

[0110] The association between iron deposition patterns and etiological subtypes is not a one-to-one deterministic relationship, but rather a probabilistic statistical association. This probabilistic association reflects the complexity of disease mechanisms and individual variability. For example, cortical iron deposition patterns are more common in patients with multiple infarcts, but not all patients with cortical iron deposition have a multiple infarct etiology. Therefore, this invention outputs an etiological subtype probability vector rather than a deterministic classification label, facilitating clinicians to make a comprehensive judgment by combining it with other clinical information.

[0111] The generation rules for etiological subtyping auxiliary prompts are as follows: First, the etiological subtype with the highest predicted probability is identified as the primary diagnostic prompt; second, if the probability of the second-highest predicted etiological subtype exceeds a preset threshold (preferably 0.25), it is also output as an alternative diagnostic prompt; third, based on the asymmetry index analysis results, if there is significant bilateral hemispheric asymmetric iron metabolism abnormality, a reminder of the possibility of unilateral vascular lesions is added to the prompt information; finally, the confidence level of the prompt information is marked according to the pattern confidence score. These rules ensure that the auxiliary prompts provide both a clear diagnostic direction and retain necessary uncertainty explanations, helping clinicians make more comprehensive diagnostic decisions.

[0112] The present invention also includes a closed-loop feedback optimization mechanism, which enables each step to form an end-to-end collaborative learning relationship and optimize the overall system performance.

[0113] The feedback error signal is calculated based on the cross-entropy loss between the etiology subtype probability vector output in step S5 and the labeled true etiology subtype label. The formula for calculating the cross-entropy loss function is as follows:

[0114] ,

[0115] in: Cross-entropy loss; The uniquely encoded first criterion for the true etiological subtype label One component; The first probability vector of the predicted etiological subtype Each component.

[0116] The feedback error signal is passed to the multi-scale spatial attention feature extraction network in step S2 and the adaptive iron deposition pattern clustering model in step S3 via backpropagation. In the feature extraction network, the weight parameters of the convolutional layers and the parameters of the spatial attention module are updated; in the clustering model, the cluster center vectors are updated. The parameter updates use the stochastic gradient descent algorithm or its variants (such as the Adam optimizer), and the learning rate is preferably set to 0.0001.

[0117] Through a closed-loop feedback optimization mechanism, the supervisory signal for etiological subtype prediction can guide the feature extraction network to learn more discriminative spatial distribution features, while simultaneously optimizing the clustering model to generate iron deposition pattern classifications that are more relevant to the etiological subtype. A deeply coupled closed-loop collaborative relationship is formed among the modules, resulting in overall analytical performance superior to that of modules trained independently.

[0118] In a preferred embodiment of the invention, the method is evaluated using a multicenter clinical dataset containing 350 patients with vascular cognitive impairment. The dataset is divided into training, validation, and test sets in a 7:1.5:1.5 ratio. Performance evaluation metrics include iron deposition pattern recognition accuracy, etiological subtype classification accuracy, sensitivity, specificity, and F1 score.

[0119] The dataset consists of 350 patients from three medical centers, aged 55 to 85 years, with a mean age of 68.3 years. There were 178 males and 172 females. Based on clinical diagnosis, 82 patients had multiple infarctions, 116 had small vessel disease, 68 had hypoperfusion, and 84 had a mixed pattern. All patients underwent standardized QSM image acquisition and clinical cognitive function assessment. Etiological subtype diagnosis was performed independently by two senior neurologists. Diagnostic concordance was assessed using the Kappa coefficient, with a score of 0.85 or higher indicating reliable diagnosis.

[0120] Test results show that the method of this invention achieves an iron deposition pattern recognition accuracy of 87.3% and a pathological subtype classification accuracy of 82.6%. Compared with the baseline method that only uses regional magnetic susceptibility statistical features, the pathological subtype classification accuracy of the method of this invention is improved by 12.8 percentage points; compared with the method without closed-loop feedback optimization, the pathological subtype classification accuracy of the method of this invention is improved by 6.5 percentage points. These results verify the effectiveness of the method of this invention and the contributions of its various innovative points.

[0121] The classification performance for the four etiological subtypes is as follows: For the multi-infarction subtype, the sensitivity was 84.2%, specificity was 91.5%, and F1 score was 0.856; for the small vessel disease subtype, the sensitivity was 86.8%, specificity was 89.3%, and F1 score was 0.872; for the hypoperfusion subtype, the sensitivity was 78.5%, specificity was 93.2%, and F1 score was 0.815; and for the mixed subtype, the sensitivity was 80.1%, specificity was 88.6%, and F1 score was 0.823. The results show that the method of this invention has the best identification performance for the small vessel disease and multi-infarction subtypes, while its identification performance for the hypoperfusion and mixed subtypes is slightly inferior. This may be related to the greater heterogeneity of the imaging features of the latter two etiological subtypes.

[0122] Performance evaluation of the asymmetric index analysis module showed that, for patients with unilateral vascular lesions (confirmed by cerebral angiography or carotid ultrasound), the method of this invention had a sensitivity of 79.6% and a specificity of 85.3% in identifying asymmetric iron metabolism abnormalities. This indicates that asymmetric index analysis can provide valuable reference information for the screening of unilateral vascular lesions.

[0123] Ablation experiments verified the contribution of each module to the overall performance: removing the multi-scale convolutional branch design (retaining only single-scale convolution) reduced the iron deposition pattern recognition accuracy by 5.8 percentage points; removing the spatial attention mechanism reduced the accuracy by 4.2 percentage points; removing the clustering regularization term with spatial neighborhood constraints reduced the accuracy by 3.6 percentage points; removing the asymmetric exponential feature input reduced the etiological subtype classification accuracy by 2.9 percentage points; and removing the closed-loop feedback optimization mechanism reduced the accuracy by 6.5 percentage points. These ablation experiment results confirm the effectiveness and necessity of the innovative modules of this invention.

[0124] In terms of system efficiency, on a workstation equipped with an NVIDIA RTX 3090 graphics processor, the complete analysis and processing time for a single QSM image is approximately 45 seconds, including approximately 20 seconds for preprocessing, approximately 12 seconds for feature extraction, approximately 5 seconds for clustering and typing, and approximately 8 seconds for etiological mapping and asymmetric analysis. This processing time meets the real-time requirements of clinical applications.

[0125] like Figure 2 As shown, the intelligent identification and classification system for spatial distribution patterns of brain iron deposition provided in this embodiment of the invention includes a QSM image preprocessing module, a spatial distribution feature extraction module, an iron deposition pattern clustering and classification module, an etiological subtype association mapping module, and an asymmetric index analysis module. Each module corresponds one-to-one with steps S1 to S5 in the above method embodiment.

[0126] The QSM image preprocessing module corresponds to step S1 and is used to acquire the whole-brain quantitative magnetic susceptibility map image of the subject, perform spatial normalization processing on the whole-brain quantitative magnetic susceptibility map image and register it to a standard brain template space, and perform brain region segmentation on the registered image to generate a brain region segmentation mask. The input of this module is the original QSM image and a high-resolution structural image, and the output is the normalized QSM image and the brain region segmentation mask. The normalized QSM image is passed to the spatial distribution feature extraction module, and the brain region segmentation mask is passed to both the spatial distribution feature extraction module and the asymmetry index analysis module.

[0127] The spatial distribution feature extraction module corresponds to step S2. It is used to input the preprocessed quantitative magnetic susceptibility map image and the brain region segmentation mask into the multi-scale spatial attention feature extraction network to generate spatial distribution feature vectors and regional magnetic susceptibility statistical features. This module contains a multi-scale convolutional branch and a spatial attention module, and its architecture and parameter settings are consistent with those described in step S2. The output of this module is passed to the iron deposition pattern clustering and typing module.

[0128] The iron deposition pattern clustering and classification module corresponds to step S3. It is used to input the spatial distribution feature vector and the regional magnetic susceptibility statistical features into the adaptive iron deposition pattern clustering model, and output the iron deposition pattern label and pattern confidence score. This module internally implements a fuzzy clustering algorithm based on spatial neighborhood constraints, which can classify the spatial distribution pattern of iron deposition into four types: cortical, deep nucleus, periventricular, and diffuse. The output of this module is passed to the etiological subtype association mapping module.

[0129] The asymmetry index analysis module corresponds to step S4 and is used to statistically analyze the magnetic susceptibility distribution characteristics of the left and right hemispheres based on the brain region segmentation mask, and calculate the bilateral hemisphere asymmetry index. This module receives the brain region segmentation mask and the normalized QSM image from the QSM image preprocessing module, outputs the whole-brain asymmetry index and the regional asymmetry index vector, and transmits them to the etiology subtype association mapping module.

[0130] The etiology subtype association mapping module corresponds to step S5. It is used to output an etiology subtype probability vector and etiology typing auxiliary prompts based on the iron deposition pattern label, the pattern confidence score, and the asymmetry index through an etiology subtype probability mapping network. This module receives input from the iron deposition pattern clustering and typing module and the asymmetry index analysis module, and outputs the etiology subtype probability vector and auxiliary prompts as the final diagnostic assistance result.

[0131] During system operation, a closed-loop collaborative relationship is formed between the modules, involving forward data transfer and back-feedback optimization. During the training phase, the error signal between the predicted results output by the etiology subtype association mapping module and the true labels is backpropagated to the spatial distribution feature extraction module and the iron deposition pattern clustering and typing module, updating network parameters and cluster centers. During the inference phase, data flows sequentially according to the module order, ultimately outputting etiology typing auxiliary prompts.

[0132] Those skilled in the art will understand that the functions of each module in the above system embodiments can be implemented by software programs, hardware circuits, or a combination of both. The software programs can be stored in a computer-readable storage medium and read and executed by a processor. The hardware circuits can be implemented using application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or graphics processors, etc.

[0133] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent identification and classification of spatial distribution patterns of brain iron deposition, characterized in that, Includes the following steps: Step S1, QSM image preprocessing step: acquire the whole brain quantitative magnetic susceptibility map image of the subject, perform spatial normalization processing on the whole brain quantitative magnetic susceptibility map image, register the whole brain quantitative magnetic susceptibility map image to the standard brain template space, and perform brain region segmentation on the registered image to generate a brain region segmentation mask; Step S2, Spatial Distribution Feature Extraction Step: Input the preprocessed quantitative magnetic susceptibility map image and the brain region segmentation mask into the multi-scale spatial attention feature extraction network. Extract the spatial distribution features of iron deposition within different receptive fields through multi-scale convolutional layers. Use the spatial attention mechanism to weight and fuse the features of different brain regions to generate a spatial distribution feature vector. Based on the brain region segmentation mask, statistically analyze the magnetic susceptibility value distribution features of each region to generate regional magnetic susceptibility statistical features. Step S3, iron deposition pattern clustering and classification step: Input the spatial distribution feature vector and the regional magnetic susceptibility statistical features into the adaptive iron deposition pattern clustering model, and use the fuzzy clustering algorithm based on spatial neighborhood constraints to classify the spatial distribution pattern of iron deposition into cortical type, deep nucleus type, periventricular type and diffuse type, and output the iron deposition pattern label and pattern confidence score. Step S4, Asymmetry Index Analysis Step: Based on the brain region segmentation mask, the quantitative magnetic susceptibility map image is divided into the left hemisphere region and the right hemisphere region. The magnetic susceptibility value distribution characteristics of the left hemisphere region and the right hemisphere region are statistically analyzed respectively. The asymmetry index of the two hemispheres is calculated to quantify the degree of difference in iron deposition between the left and right hemispheres. The asymmetry index is used to identify asymmetric iron metabolism abnormalities related to unilateral vascular lesions. Step S5, Etiological Subtype Association Mapping Step: Based on the iron deposition pattern label, the pattern confidence score, and the asymmetry index output in step S4, establish the association between iron deposition patterns and vascular cognitive impairment etiological subtypes through an etiological subtype probability mapping network, output etiological subtype probability vectors, and generate etiological subtyping auxiliary prompt information based on the etiological subtype probability vectors.

2. The intelligent identification and classification method for spatial distribution patterns of brain iron deposition according to claim 1, characterized in that, In step S1, the brain region segmentation mask includes a cortical region, a deep nucleus region, and a periventricular region. The spatial normalization process includes: rigidly registering the whole-brain quantitative magnetic susceptibility map image with a high-resolution structural image to obtain a first registration transformation matrix; nonlinearly registering the high-resolution structural image with a standard brain template to obtain a second registration transformation matrix; and transforming the whole-brain quantitative magnetic susceptibility map image to the standard brain template space based on the first registration transformation matrix and the second registration transformation matrix.

3. The intelligent identification and classification method for spatial distribution patterns of brain iron deposition according to claim 1, characterized in that, In step S2, the multi-scale spatial attention feature extraction network includes a first convolutional branch, a second convolutional branch, and a third convolutional branch in parallel. The kernel size of the first convolutional branch is 3×3×3, the kernel size of the second convolutional branch is 5×5×5, and the kernel size of the third convolutional branch is 7×7×7. The outputs of the three convolutional branches are concatenated through channels and then input into the spatial attention module. The spatial attention module generates a spatial attention weight map based on the brain region segmentation mask to weight the concatenated features.

4. The intelligent identification and classification method for spatial distribution patterns of brain iron deposition according to claim 1, characterized in that, In step S3, the iron deposition pattern clustering and classification step divides the spatial distribution pattern of iron deposition into four types: cortical type, which indicates that iron deposition is concentrated in the cerebral cortex; deep nucleus type, which indicates that iron deposition is concentrated in the basal ganglia and thalamus; periventricular type, which indicates that iron deposition is concentrated in the white matter area around the lateral ventricle; and diffuse type, which indicates that iron deposition is distributed in the cortex, deep nucleus, and periventricular areas.

5. The intelligent identification and classification method for spatial distribution patterns of brain iron deposition according to claim 1, characterized in that, In step S2, the regional magnetic susceptibility statistical characteristics include the mean magnetic susceptibility value, standard deviation of magnetic susceptibility value, skewness of magnetic susceptibility value, and kurtosis of magnetic susceptibility value for each brain region, wherein each brain region includes the frontal cortex, parietal cortex, temporal cortex, occipital cortex, caudate nucleus, putamen, globus pallidus, thalamus, hippocampus, and periventricular white matter.

6. The intelligent identification and classification method for spatial distribution patterns of brain iron deposition according to claim 1, characterized in that, In step S3, the adaptive iron deposition pattern clustering model is implemented based on a fuzzy clustering algorithm with spatial neighborhood constraints. The algorithm introduces a spatial neighborhood regularization term in the objective function to make spatially adjacent voxels tend to belong to the same iron deposition pattern. The pattern confidence score represents the degree of membership of a sample to each iron deposition pattern.

7. The intelligent identification and classification method for spatial distribution patterns of brain iron deposition according to claim 1, characterized in that, In step S4, the calculation of the asymmetry index includes: calculating the mean magnetic susceptibility value of the left hemisphere region and the mean magnetic susceptibility value of the right hemisphere region; obtaining the whole-brain asymmetry index based on the ratio of the difference between the mean magnetic susceptibility value of the left hemisphere region and the mean magnetic susceptibility value of the right hemisphere region and the mean of the two; and calculating the ratio of the difference between the mean magnetic susceptibility value of the left brain region and the mean magnetic susceptibility value of the corresponding right brain region and the mean of the two for each corresponding brain region to obtain the partition asymmetry index vector.

8. The intelligent identification and classification method for spatial distribution patterns of brain iron deposition according to claim 1, characterized in that, In step S5, the vascular cognitive impairment etiology subtypes include multi-infarction type, small vessel disease type, hypoperfusion type, and mixed type. The etiology subtype probability mapping network receives the one-hot encoded vector of the iron deposition pattern label, the pattern confidence score, and the asymmetry index as input. After processing through multiple fully connected layers, the etiology subtype probability vector is output through a normalized exponential function. The etiology subtype corresponding to the largest component of the etiology subtype probability vector is used as the prediction result.

9. The intelligent identification and classification method for spatial distribution patterns of brain iron deposition according to claim 1, characterized in that, It also includes a closed-loop feedback optimization step: calculating the feedback error signal based on the cross-entropy loss between the etiological subtype probability vector and the labeled true etiological subtype label, backpropagating the feedback error signal to the multi-scale spatial attention feature extraction network and the adaptive iron deposition pattern clustering model, and updating the network parameters of the multi-scale spatial attention feature extraction network and the cluster center vector of the adaptive iron deposition pattern clustering model.

10. A system for intelligent identification and classification of spatial distribution patterns of brain iron deposition, used to implement the method for intelligent identification and classification of spatial distribution patterns of brain iron deposition as described in any one of claims 1-9, characterized in that, include: The QSM image preprocessing module is used to acquire the whole brain quantitative magnetic susceptibility map image of the subject, perform spatial standardization processing on the whole brain quantitative magnetic susceptibility map image and register it to the standard brain template space, and perform brain region segmentation on the registered image to generate a brain region segmentation mask. The spatial distribution feature extraction module is used to input the preprocessed quantitative magnetic susceptibility map image and the brain region segmentation mask into the multi-scale spatial attention feature extraction network to generate spatial distribution feature vectors and regional magnetic susceptibility statistical features. The iron deposition pattern clustering and classification module is used to input the spatial distribution feature vector and the regional magnetic susceptibility statistical features into the adaptive iron deposition pattern clustering model, and output the iron deposition pattern label and pattern confidence score. The etiology subtype association mapping module is used to output etiology subtype probability vectors and etiology subtyping auxiliary prompts based on the iron deposition pattern label, the pattern confidence score and the asymmetry index through the etiology subtype probability mapping network. The asymmetry index analysis module is used to statistically analyze the magnetic susceptibility distribution characteristics of the left and right hemisphere regions based on the brain region segmentation mask, and calculate the asymmetry index of both hemispheres.

Images