A cerebral microbleed image segmentation method and system based on time sequence feature aggregation

Abstract

The application discloses a kind of cerebral microbleed image segmentation method and system based on timing feature aggregation, it is related to medical image processing and depth learning technical field, including the following steps, obtaining to be segmented slice, constitute two-dimensional slice sequence;Two-dimensional slice is inputed into encoder, and multi-level feature map is extracted;Multi-level feature map is inputed into timing feature aggregation module, and the aggregation of multiple slice features is realized;After the aggregation of features, detail recovery and refining module is inputed, and the fine timing details in feature are recovered and enhanced;Enhanced feature is inputed into decoder, and the cerebral microbleed segmentation prediction graph of to-be-segmented slice is generated;By combining timing feature fusion, the way of detail recovery and refining, excellent segmentation effect is realized on self-built dataset, reduces the subjectivity of artificial judgment and observer difference and saves clinical diagnosis time, provides reliable technical support for cerebral microbleed clinical detection and segmentation.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and deep learning technology, specifically to a method and system for segmenting brain microbleed images based on temporal feature aggregation. Background Technology

[0002] Existing technologies related to cerebral microbleeds (CMBs) can be mainly divided into two categories: traditional medical diagnostic methods and automated methods based on deep learning, as detailed below:

[0003] Traditional medical diagnostic methods mainly rely on magnetic resonance imaging (MRI) examinations, combined with imaging features of special sequences such as gradient echo (GRE), susceptibility-weighted imaging (SWI), and T2*-weighted imaging, as well as the distribution patterns of lesions in specific locations in the brain. CMBs are identified through comprehensive manual judgment without the assistance of automated algorithms.

[0004] Deep learning-based automated methods are divided into two-stage detection and segmentation strategies and single-stage end-to-end strategies. The two-stage strategy, as exemplified by Al-Masni MA, Kim WR, Kim EY, and Wei Z, Chen X, Huang J, first uses a model (such as YOLO) to detect potential CMB candidate units (including CMBs and simulated objects), and then uses another classification or segmentation model (such as 3D-CNN, a modified U-Net, and the full-resolution network FRN) to accurately identify, distinguish, or segment the candidate regions, thereby reducing the false positive rate. Single-stage end-to-end strategies employ an end-to-end model architecture to directly process CMB-related tasks. They often utilize the three-dimensional information (axial, sagittal, and coronal planes) of brain images to construct networks (such as the TPE-Det proposed by Lee H, Kim JH, Lee S, and others, and the detector integrating 3D U-Net and Faster R-CNN region proposal networks). Kim JH, Noh Y, Lee H, and others introduced loss terms such as hard sample prototype learning (HSPL) to optimize performance. These methods often input 3D data from SWI sequences. To reduce computational resource consumption, the original image data is preprocessed by cropping, interpolation, etc., and the data voxel size is adjusted before training and detection. For example, Kim JH, Noh Y, Lee H, and others first increased the number of slices to 224 from the original data of 512×448×72 and 448×392×52 size through interpolation, and then cropped them into voxels of 128×128×128 size for detection studies, and used 64×64×16 for anatomical localization tasks; Lee H, Kim JH, Lee S, and others cropped the original 512×448×72 data into voxels of 360×360×72, and then used interpolation to transform the samples into 360×360×360 data before conducting studies.

[0005] Traditional diagnostic methods are inefficient and highly subjective, relying on manual interpretation of MRI image sequences. This is not only time-consuming and labor-intensive, but also subject to significant variability among observers, making it difficult to guarantee the reliability and repeatability of diagnostic results. Two-stage automation strategies have certain drawbacks: the training of the second stage heavily depends on the detection results of the first stage, and there is a lack of effective feature information exchange between the two stages. Features extracted in the first stage cannot be effectively transferred to the second stage, affecting overall recognition and segmentation accuracy. Furthermore, single-stage 3D correlation methods are prone to information loss and have high computational costs: utilizing 3D information requires processing large amounts of 3D image data, leading to high computational resource consumption; operations such as cropping and interpolation of the original data to reduce resource requirements can easily result in the loss of lesion details and complete brain anatomical structure information, affecting the accuracy of CMBs segmentation. Meanwhile, CMBs segmentation research still faces many challenges, and related work is relatively scarce: due to the extremely small size of the lesions, usually less than 10 mm in diameter, and the difficulty in obtaining labeled data, current research on CMBs segmentation is quite difficult. Existing research mostly focuses on automatic detection of CMBs, while segmentation research is relatively scarce. Furthermore, existing models lack effective feature fusion and detail recovery mechanisms, making it difficult to accurately segment small lesions and resulting in low running and training efficiency and high computational complexity. Existing automated methods often require complex preprocessing steps (such as data cropping and interpolation), which are cumbersome and limit inference speed. Summary of the Invention

[0006] The purpose of this invention is to provide a brain microbleed image segmentation method based on temporal feature aggregation, in order to solve the problems in the existing technology of brain microbleeds (CMBs) such as small lesions, difficulty in annotation, lack of segmentation research, high computational cost, easy loss of information, low efficiency of two-stage models, poor information transmission, and difficulty in preserving details of small targets.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for segmenting brain microbleed images based on temporal feature aggregation, comprising the following steps:

[0008] Obtain the slice to be segmented and at least one adjacent slice before and after it to form a sequence of multiple consecutive two-dimensional slices;

[0009] The multiple consecutive two-dimensional slices are input into a shared encoder to extract multi-level feature maps of each slice;

[0010] The multi-level feature map is input into the temporal feature aggregation module, and the feature representation of the slice to be segmented is enhanced by the feature information of adjacent slices, so as to realize the aggregation of features of multiple slices;

[0011] The aggregated features are input into the detail recovery and refinement module, which restores and enhances the fine temporal details in the features through temporal context modeling and channel-level feature recalibration.

[0012] The enhanced features are input into the decoder, and a segmentation prediction map of brain microbleeds is generated through multi-level upsampling.

[0013] The encoder is responsible for extracting multi-level feature representations of each slice and outputting multiple feature maps of different resolutions.

[0014] Furthermore, the temporal feature aggregation module includes an inter-slice temporal attention module and a cross-channel temporal module;

[0015] The inter-slice temporal attention module is used to model the one-to-one relationship between similar feature groups in different slices, generate feature weights, and multiply them element-wise with the original features.

[0016] The cross-channel temporal module is used to extract temporal dimension information through 3D convolution and generate a weight vector through convolution, which is then used to perform channel-level enhancement with the features of the slice to be segmented.

[0017] Furthermore, the processing procedure of the inter-slice temporal attention module includes,

[0018] The corresponding channel features of the current slice and the adjacent slices are concatenated, input into the convolutional layer, and activated by the Sigmoid function to generate weights.

[0019] The weights are element-wise multiplied with the corresponding feature groups, and then the features are aggregated through a 1×1 convolution.

[0020] Furthermore, the processing procedure of the cross-channel timing module includes,

[0021] Perform 3D convolution processing on feature groups of the same type to extract time-dimensional features;

[0022] The extracted features are concatenated with the features of the slice to be segmented, and channel-level weights are generated through convolution.

[0023] The weights are element-wise multiplied with the features of the slice to be segmented, and the enhanced features are output.

[0024] Furthermore, the processing steps of the detail recovery and refinement module include:

[0025] The original multi-frame features are subjected to two 3D convolutions to form a bottleneck structure, and a compact temporal representation is learned.

[0026] Average pooling is performed along the time dimension to compress time information;

[0027] The compressed features are added element-wise to the aggregated features, and then attention weights are generated through convolution.

[0028] Attention weights are used to adjust aggregated features at the channel level, and the output features are enhanced through residual connections.

[0029] Furthermore, the multi-level feature maps extracted by the encoder include feature levels of different resolutions, namely P1, P2, P3 and P4; wherein, shallow feature maps P1 and P2 are used to preserve spatial detail information, and deep feature maps P3 and P4 are used to provide semantic context information, providing multi-scale feature representation for the subsequent temporal feature aggregation module and detail recovery and refinement module.

[0030] Furthermore, the decoder gradually restores the resolution of the feature map through multi-level upsampling blocks, and receives enhanced features from the detail recovery and refinement module at each upsampling stage. It then fuses the encoder features of the corresponding level using a skip connection mechanism to achieve refined segmentation of the boundaries of cerebral microbleed lesions.

[0031] A brain microbleed image segmentation method based on temporal feature aggregation includes an input module for acquiring the slice to be segmented and at least one adjacent slice before and after it, forming a sequence of multiple consecutive two-dimensional slices.

[0032] An encoder is used to extract shared multi-level feature maps from the multiple slices;

[0033] The temporal feature aggregation module is used to enhance the feature representation of the slice to be segmented by utilizing the feature information of adjacent slices;

[0034] The detail recovery and refinement module is used to recover and enhance fine temporal details in features through temporal context modeling and channel-level feature recalibration;

[0035] A decoder is used to generate a segmentation prediction map of brain microbleeds from the slices to be segmented through multi-level upsampling.

[0036] Furthermore, the time-series feature aggregation module includes,

[0037] The inter-slice temporal attention module is used to model the one-to-one relationship between similar feature groups in different slices;

[0038] The cross-channel time series module is used to extract time dimension information and perform channel-level feature enhancement.

[0039] Furthermore, the detail restoration and refinement module includes,

[0040] 3D convolutional units are used to learn compact temporal representations;

[0041] Temporal pooling units are used to compress temporal dimension information;

[0042] Attention weight generation unit, used to generate channel-level attention weights;

[0043] The feature adjustment unit is used to enhance aggregated features based on attention weights.

[0044] Compared with existing technologies, the brain microbleed image segmentation method based on temporal feature aggregation provided by this invention has the following advantages:

[0045] (1) No step-by-step processing is required, which overcomes the shortcomings of low efficiency and discontinuous feature propagation in the two-stage framework.

[0046] (2) By using continuous multi-dimensional slice data instead of traditional three-dimensional data as the input of the network, and combining temporal feature fusion, detail restoration and refinement, excellent segmentation results were achieved on the self-built dataset, with a recall rate of 86.05%, an average intersection-union ratio of 75.07% and an F2 score of 85.93%. Automatic segmentation reduces the subjectivity of manual judgment and inter-observer differences and saves clinical diagnosis time, providing reliable technical support for the clinical detection and segmentation of cerebral microbleeds.

[0047] (3) A Temporal Feature Aggregation Module (TFA) is proposed to enhance the feature details of the slice to be segmented. The module divides the feature map sets of different slices into feature groups according to feature type, and applies two different feature aggregation methods: Inter-slice Temporal Attention Module (ISA) and Cross-Channel Temporal Module (CCT). By using the temporal information of adjacent slices, the feature details of the slice to be segmented are enhanced, making the model pay more attention to the CMBs region.

[0048] (4) By combining the detail recovery and refinement module, the feature recalibration is achieved by integrating the temporal context modeling information and channel-level features. This helps to more accurately delineate boundaries and enhance the detection capability of brain microbleeds, enabling us to selectively recover and enhance features that are conducive to accurate boundary delineation and small target segmentation.

[0049] (5) Constructing a finely labeled CMBs segmentation dataset not only provides high-quality data support for my own network training, but also provides data reference for subsequent clinical CMBs fine segmentation research. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0051] Figure 1 This is a schematic diagram of the overall structure of TDAR-Net provided in an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of the overall architecture of the time-series feature aggregation module provided in an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of the overall architecture of the detail restoration and refinement module provided in an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram of the overall process provided for an embodiment of the present invention. Detailed Implementation

[0055] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0056] As attached Figure 1 To be continued Figure 4 As shown:

[0057] Example 1:

[0058] This invention provides a method for segmenting brain microbleed images based on temporal feature aggregation, comprising the following steps:

[0059] Obtain the slice to be segmented and at least one adjacent slice before and after it to form a sequence of multiple consecutive two-dimensional slices;

[0060] Multiple consecutive two-dimensional slices are input into a shared encoder to extract multi-level feature maps of each slice;

[0061] The multi-level feature map is input into the temporal feature aggregation module, and the feature representation of the slice to be segmented is enhanced by the feature information of adjacent slices, so as to realize the aggregation of features of multiple slices;

[0062] The aggregated features are input into the detail recovery and refinement module, which restores and enhances the fine temporal details in the features through temporal context modeling and channel-level feature recalibration.

[0063] The enhanced features are input into the decoder, and a segmentation prediction map of brain microbleeds is generated through multi-level upsampling.

[0064] The encoder is responsible for extracting multi-level feature representations of each slice and outputting multiple feature maps of different resolutions.

[0065] Furthermore, the temporal feature aggregation module includes an inter-slice temporal attention module and a cross-channel temporal module;

[0066] Among them, the inter-slice temporal attention module is used to model the one-to-one relationship between the same type of feature groups in different slices, generate feature weights and multiply them element-wise with the original features;

[0067] The cross-channel temporal module is used to extract temporal dimension information through 3D convolution and generate weight vectors through convolution, which are then used to enhance the features of the slice to be segmented at the channel level.

[0068] The processing steps of the inter-slice temporal attention module include:

[0069] The corresponding channel features of the current slice and the adjacent slices are concatenated, input into the convolutional layer, and activated by the Sigmoid function to generate weights.

[0070] The weights are element-wise multiplied with the corresponding feature groups, and then the features are aggregated through a 1×1 convolution.

[0071] Specifically, the processing procedure of the cross-channel timing module includes,

[0072] Perform 3D convolution processing on feature groups of the same type to extract time-dimensional features;

[0073] The extracted features are concatenated with the features of the slice to be segmented, and channel-level weights are generated through convolution.

[0074] The weights are element-wise multiplied with the features of the slice to be segmented, and the enhanced features are output.

[0075] It should be noted that the processing steps of the detail restoration and refinement module include:

[0076] The original multi-frame features are subjected to two 3D convolutions to form a bottleneck structure, and a compact temporal representation is learned.

[0077] Average pooling is performed along the time dimension to compress time information;

[0078] The compressed features are added element-wise to the aggregated features, and then attention weights are generated through convolution.

[0079] Attention weights are used to adjust aggregated features at the channel level, and the output features are enhanced through residual connections.

[0080] Furthermore, the multi-level feature maps extracted by the encoder include feature levels of different resolutions, namely P1, P2, P3 and P4. Among them, shallow feature maps P1 and P2 are used to preserve spatial detail information, while deep feature maps P3 and P4 are used to provide semantic context information, providing multi-scale feature representation for the subsequent temporal feature aggregation module and detail restoration and refinement module.

[0081] The decoder gradually restores the resolution of the feature map through multi-level upsampling blocks, and receives enhanced features from the detail recovery and refinement module at each upsampling stage. It then combines the encoder features of the corresponding level with a skip connection mechanism to achieve fine segmentation of the boundary of cerebral microbleed lesions.

[0082] The working process involves providing a single-stage end-to-end TDAR-Net network architecture for accurate segmentation of brain microbleeds (CMBs) based on this method. It takes the slice to be segmented and its two adjacent slices (a total of three slices) as input, and finally outputs a CMBs segmentation prediction map of the slice to be segmented, as shown below. Figure 1 As shown, in TDAR-Net, the three input images share the same encoder structure. The encoder is responsible for extracting multi-level feature representations of each slice and outputting four feature maps (P1, P2, P3, P4) at different resolutions. These feature maps fuse the semantic information of deep features and the contextual information of shallow features, providing rich feature representations for the subsequent decoding process.

[0083] Before the feature maps (P1, P2, P3, P4) are input into the decoder, they pass through a Temporal Feature Aggregation (TFA) module and a Details Recovery and Refinement (DRR) module. The TFA module utilizes feature information from adjacent slices to enhance the feature representation of the slice to be segmented, enriching its semantic and fine-grained information, ultimately achieving effective aggregation of the features from the three slices. The DRR module receives the output from the TFA module and the feature map outputs from each level of the encoder, preserving temporal details that may be lost during the TFA aggregation process, supplementing the aggregated features with fine-grained temporal information. Subsequently, the obtained features are input into the decoder, which generates the segmentation result through multi-level upsampling blocks, such as... Figure 2 As shown, in the temporal feature aggregation module, two different feature aggregation methods are applied: the inter-slice temporal attention module (ISA) and the cross-channel temporal module (CCT). The adjacent slices are used to enhance the feature details of the slice to be segmented, so that the model pays more attention to the CMBs region and finally obtains the feature aggregation result.

[0084] The current slice feature map Where T=3 represents the time dimension, and features are divided according to different feature types, i.e., according to channel order. Features of the same type from different feature maps are combined together to obtain C sizes. Feature groups A single feature group is n=1,2,3···C;

[0085] Next, two methods are used to perform feature fusion enhancement operations. ISA is used to enhance similar features between slices, and CCT is used to enhance features between slice feature groups. After the features are fused and enhanced using the ISA and CCT modules, the results obtained by the two modules are added together and the result is output.

[0086] The ISA module performs the same processing flow for each feature group of the same type, as detailed below.

[0087] First, each channel of a single feature group is compared with the corresponding feature of the slice to be segmented. Perform concat concatenation, where t represents the corresponding time dimension value, forming T values. The features are input into a 7×7 convolutional layer, and then the result is processed by the sigmoid function to obtain T feature relation weights of size H×W. This process is carried out according to the following formula.

[0088]

[0089] Where σ represents the sigmoid function, i.e. Where e is the base of the natural logarithm, with a value of 2.718. This represents a set of T feature relation weights, which respectively contain the relation weights between adjacent slices and the slice to be segmented, as well as the self-attention weights of the slice to be segmented.

[0090] Then, the obtained feature weights are multiplied element-wise with the feature group, and the features are aggregated using 1×1 convolution to finally obtain the result H×W;

[0091] Finally, the results obtained from processing the C feature groups separately are concatenated together to obtain... ,

[0092]

[0093] By fully considering the positional information of the image, one-to-one relationships between similar features of different slices are extracted to supplement the information in the time dimension and to increase the self-attention mechanism of the features of the slices to be segmented. The final 1×1 convolution can adaptively select and combine the three attention results to achieve effective feature aggregation.

[0094] The CCT module first processes features using 3D convolution to learn information in the time dimension. Then, it performs uniform processing on each feature group of the same type using a 3×3 convolution, integrating and learning the features to fully extract time-series information and obtain a feature relationship vector. ;

[0095] Next, all feature maps corresponding to the slice to be segmented are concatenated channel by channel. A weight relation vector of size C×H×W is obtained through a 7×7 convolutional layer. Then, it is element-wise multiplied with the feature map of the slice to be segmented, adding both temporal and channel dimension information to the features, and finally outputting the enhanced result. This process is carried out according to the following formula.

[0096]

[0097]

[0098] Among them, F t These are the original features corresponding to the slice to be segmented;

[0099] By extracting the deep relationships between similar features between slices, the obtained relationship matrix is ​​convolved with the features of the slice to be segmented to calculate the relationship weights, and then superimposed on the features of the slice to be detected, effectively enhancing its spatiotemporal information representation capability.

[0100] Finally, the results of ISA and CCT are added together and output to obtain the aggregated features. ;

[0101] The temporal feature aggregation module effectively aggregates temporal information from consecutive slices through inter-layer attention and cross-channel temporal mechanisms. However, the aggregation process inevitably leads to information compression, causing the input decoder to lose fine temporal details crucial for detecting microhemorrhagic brain hemorrhage. To alleviate this limitation, the detail recovery and refinement module integrates temporal context modeling and channel-level feature recalibration, using auxiliary temporal information flow to compensate for and enhance the globally aggregated features. This effectively compensates for potential information loss during feature aggregation, enabling fine temporal features to be more effectively transmitted to the decoder.

[0102] like Figure 3 As shown, the detail restoration and refinement module preprocesses the original multi-frame features before adaptive fusion. It uses two consecutive 3D convolutional layers with a kernel size of 3×3×3, that is,

[0103]

[0104] Where BN represents batch normalization, F sup This represents the temporal augmentation feature obtained after 3D convolution processing. ReLU means that the normalized result is passed through the ReLU function, which turns all negative values ​​into 0 and keeps positive values ​​unchanged.

[0105] The first 3D convolutional layer reduces the channel dimension from C to C / 2, while the second convolutional layer restores it to C, thus forming a bottleneck structure, which is beneficial for learning a compact temporal representation and obtaining a feature matrix containing temporal information.

[0106] Subsequently, pooling was used in the time dimension to enhance the time-enhanced feature F. sup Compression in the time dimension,

[0107]

[0108] This formula expresses averaging the feature maps over the time dimension; this aggregation step compresses the temporal information into the aggregated feature F. TFA Dimensionality-compatible feature representation;

[0109] Next, the two feature streams are combined element-wise. This combination is then processed through two convolutional blocks with a kernel size of 3×3. A sigmoid activation function is applied to generate normalized attention weights with values ​​ranging from [0, 1]. The importance of these weights is adjusted based on the contribution of different feature channels to the segmentation task. This process is performed according to the following formula.

[0110]

[0111] Attention weight W DRR Aggregated features from the TFA module are modulated using channel-level multiplication, and then a stable gradient flow is ensured through residual connections.

[0112]

[0113] This allows the network to selectively enhance the contribution of temporal supplementary information based on local features. Regions with high attention weights will receive stronger temporal enhancement, while regions with low weights will mainly rely on aggregated features.

[0114] By explicitly combining temporal context modeling information and channel-level feature recalibration, the DRR module addresses the inherent information loss problem in feature aggregation operations, enabling the decoder to simultaneously utilize global temporal consistency from multiple slices and local spatial precision from the target slice. This helps to delineate boundaries more accurately and improves the detection capability of microhemorrhagic brain hemorrhages when subtle intensity changes are easily masked during multi-layer feature fusion.

[0115] By replacing traditional 3D data with continuous 2D slice data, and combining temporal feature fusion with detail restoration and refinement, efficient and accurate segmentation of cerebral microbleeds is achieved. This eliminates the need for complex preprocessing steps such as cropping and interpolation, saving computational resources and allowing them to be used to maintain the anatomical integrity of the brain slices. The temporal feature aggregation module applies two different feature aggregation methods for each feature group: inter-slice temporal attention and cross-channel temporal modules, significantly improving the model's ability to detect small cerebral microbleeds and thus greatly increasing recall. The detail restoration and refinement module compensates for potential information loss in feature aggregation operations by integrating temporal context modeling information and channel-level feature recalibration. This facilitates more accurate boundary delineation and enhances the detection capability of cerebral microbleed foci.

[0116] Example 2:

[0117] This embodiment is basically the same as the previous embodiment, except that it provides a brain microbleed image segmentation system based on temporal feature aggregation, including:

[0118] The input module is used to obtain the slice to be segmented and at least one adjacent slice before and after it, forming a sequence of multiple consecutive two-dimensional slices.

[0119] An encoder is used to extract shared multi-level feature maps from multiple slices;

[0120] The temporal feature aggregation module is used to enhance the feature representation of the slice to be segmented by utilizing the feature information of adjacent slices;

[0121] The detail recovery and refinement module is used to recover and enhance fine temporal details in features through temporal context modeling and channel-level feature recalibration;

[0122] A decoder is used to generate a segmentation prediction map of brain microbleeds from the slices to be segmented through multi-level upsampling.

[0123] Furthermore, the time-series feature aggregation module includes,

[0124] The inter-slice temporal attention module is used to model the one-to-one relationship between similar feature groups in different slices;

[0125] The cross-channel time series module is used to extract time dimension information and perform channel-level feature enhancement.

[0126] Furthermore, the detail restoration and refinement module includes,

[0127] 3D convolutional units are used to learn compact temporal representations;

[0128] Temporal pooling units are used to compress temporal dimension information;

[0129] Attention weight generation unit, used to generate channel-level attention weights;

[0130] The feature adjustment unit is used to enhance aggregated features based on attention weights.

[0131] The input module acquires brain MRI slice images to be segmented. This module selects the slice to be segmented and at least one adjacent slice, forming a continuous two-dimensional slice sequence, providing input data containing temporal information for subsequent processing.

[0132] Subsequently, this set of consecutive two-dimensional slices is input into the encoder. The encoder adopts a shared weight structure and extracts features from multiple slices simultaneously, outputting multiple hierarchical feature maps of different resolutions, including shallow feature maps (to preserve spatial details) and deep feature maps (to provide semantic context information), forming rich multi-scale feature representations.

[0133] Next, these multi-level feature maps are fed into the temporal feature aggregation module. This module contains two core sub-units:

[0134] The inter-slice temporal attention module is used to model the one-to-one relationship between similar features in different slices, generate feature weights, and enhance the expression of features related to cerebral microbleeds in the slices to be segmented.

[0135] The cross-channel temporal module is used to extract information in the time dimension through three-dimensional convolution and generate channel-level weight vectors to further enhance the spatiotemporal representation capability of features.

[0136] The outputs of the two sub-modules are fused to achieve effective aggregation of multi-slice features.

[0137] The aggregated features are then fed into the detail restoration and refinement module. This module compensates for and enhances the features through the following steps:

[0138] First, three-dimensional convolutional units are used to process the original multi-frame features to learn a compact temporal representation;

[0139] Then, the features are compressed in the time dimension using time pooling units to make them compatible with the main feature stream dimension;

[0140] Next, the compressed temporal features are fused with the aggregated features, and channel-level attention weights are calculated through the attention weight generation unit.

[0141] Then, the feature adjustment unit weights and adjusts the aggregated features according to the attention weights, and outputs the enhanced features through residual connections.

[0142] The design of this module effectively compensates for the fine temporal details that may be lost during feature aggregation, and helps to delineate the boundaries of cerebral microbleed lesions more accurately.

[0143] Finally, the enhanced features are fed into the decoder. The decoder gradually restores the resolution of the feature map through multi-level upsampling blocks, and combines a skip connection mechanism at each upsampling stage to fuse the encoder features of the corresponding level, ultimately generating a brain microbleed segmentation prediction map of the slice to be segmented.

[0144] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for segmenting brain microbleed images based on temporal feature aggregation, characterized in that, Includes the following steps, Obtain the slice to be segmented and at least one adjacent slice before and after it to form a sequence of multiple consecutive two-dimensional slices; The multiple consecutive two-dimensional slices are input into a shared encoder to extract multi-level feature maps of each slice; The multi-level feature map is input into the temporal feature aggregation module, and the feature representation of the slice to be segmented is enhanced by the feature information of adjacent slices, so as to realize the aggregation of features of multiple slices; The aggregated features are input into the detail recovery and refinement module, which restores and enhances the fine temporal details in the features through temporal context modeling and channel-level feature recalibration. The enhanced features are input into the decoder, and a segmentation prediction map of brain microbleeds is generated through multi-level upsampling. The encoder is responsible for extracting multi-level feature representations of each slice and outputting multiple feature maps of different resolutions.

2. The brain microbleed image segmentation method based on temporal feature aggregation according to claim 1, characterized in that, The temporal feature aggregation module includes an inter-slice temporal attention module and a cross-channel temporal module; The inter-slice temporal attention module is used to model the one-to-one relationship between similar feature groups in different slices, generate feature weights, and multiply them element-wise with the original features. The cross-channel temporal module is used to extract temporal dimension information through 3D convolution and generate a weight vector through convolution, which is then used to perform channel-level enhancement with the features of the slice to be segmented.

3. The brain microbleed image segmentation method based on temporal feature aggregation according to claim 2, characterized in that, The processing procedure of the inter-slice temporal attention module includes, The corresponding channel features of the current slice and the adjacent slices are concatenated, input into the convolutional layer, and activated by the Sigmoid function to generate weights. The weights are element-wise multiplied with the corresponding feature groups, and then the features are aggregated through a 1×1 convolution.

4. The brain microbleed image segmentation method based on temporal feature aggregation according to claim 3, characterized in that, The processing procedure of the cross-channel timing module includes, Perform 3D convolution processing on feature groups of the same type to extract time-dimensional features; The extracted features are concatenated with the features of the slice to be segmented, and channel-level weights are generated through convolution. The weights are element-wise multiplied with the features of the slice to be segmented, and the enhanced features are output.

5. The brain microbleed image segmentation method based on temporal feature aggregation according to claim 4, characterized in that, The processing steps of the detail restoration and refinement module include: The original multi-frame features are subjected to two 3D convolutions to form a bottleneck structure, and a compact temporal representation is learned. Average pooling is performed along the time dimension to compress time information; The compressed features are added element-wise to the aggregated features, and then attention weights are generated through convolution. Attention weights are used to adjust aggregated features at the channel level, and residual connections are used to output enhanced features.

6. A method for segmenting brain microbleed images based on temporal feature aggregation according to claim 4 or 5, characterized in that, The encoder extracts multi-level feature maps including feature layers of different resolutions, namely P1, P2, P3 and P4. Among them, shallow feature maps P1 and P2 are used to preserve spatial detail information, while deep feature maps P3 and P4 are used to provide semantic context information, providing multi-scale feature representation for subsequent temporal feature aggregation module and detail recovery and refinement module.

7. The brain microbleed image segmentation method based on temporal feature aggregation according to claim 6, characterized in that, The decoder gradually restores the resolution of the feature map through multi-level upsampling blocks, and receives enhanced features from the detail recovery and refinement module at each upsampling stage. It then combines the encoder features of the corresponding level with a skip connection mechanism to achieve fine segmentation of the boundary of cerebral microbleed lesions.

8. A brain microbleed image segmentation system based on temporal feature aggregation, characterized in that, include, The input module is used to obtain the slice to be segmented and at least one adjacent slice before and after it, forming a sequence of multiple consecutive two-dimensional slices. An encoder is used to extract shared multi-level feature maps from the multiple slices; The temporal feature aggregation module is used to enhance the feature representation of the slice to be segmented by utilizing the feature information of adjacent slices; The detail recovery and refinement module is used to recover and enhance fine temporal details in features through temporal context modeling and channel-level feature recalibration; A decoder is used to generate a segmentation prediction map of brain microbleeds from the slices to be segmented through multi-level upsampling.

9. A brain microbleed image segmentation system based on temporal feature aggregation according to claim 8, characterized in that, The time-series feature aggregation module includes, The inter-slice temporal attention module is used to model the one-to-one relationship between similar feature groups in different slices; The cross-channel time series module is used to extract time dimension information and perform channel-level feature enhancement.

10. A brain microbleed image segmentation system based on temporal feature aggregation according to claim 9, characterized in that, The detailed restoration and refinement module includes, 3D convolutional units are used to learn compact temporal representations; Temporal pooling units are used to compress temporal dimension information; Attention weight generation unit, used to generate channel-level attention weights; The feature adjustment unit is used to enhance aggregated features based on attention weights.

Images