Anatomical correlation prior driven intracranial hematoma segmentation method and system

Abstract

This invention discloses an anatomically-associated prior-driven method and system for intracranial hematoma segmentation. By identifying anatomical anchor points in the input image, and utilizing a global feature modeling layer guided by anatomical association priors, the lexical units of the input feature map are assigned to matching anatomical anchor points, forming anatomical regions corresponding to each anchor point. Outlier regions within these anatomical regions are then corrected. Furthermore, lexical units within each anatomical region are fused to extract the corresponding anatomical attribute features. Based on these anatomical attribute features, the correlation between each lexical unit and each anatomical region in the feature map is modeled. This allows for the fusion of anatomical attribute features from different anatomical regions into individual lexical units, fully utilizing prior knowledge of brain anatomy. This significantly improves the segmentation accuracy and robustness of intracranial hematoma regions while reducing computational complexity.

Description

Technical Field

[0001] This invention belongs to the field of image segmentation, and more specifically, relates to a method and system for intracranial hematoma segmentation driven by anatomical association priors. Background Technology

[0002] Intracranial hematoma is one of the major subtypes of stroke, characterized by rapid onset, high mortality, and high disability rates. Precise segmentation of the intracranial hematoma region based on plain CT images of the brain is a crucial step in rapidly assessing the condition and developing personalized treatment plans in clinical practice. The brain's anatomical structure is complex, with clear inherent anatomical connections between different anatomical regions such as the ventricles, brain parenchyma, and skull. Furthermore, the image characteristics of each region differ to some extent. Fully utilizing this anatomical prior can effectively improve the accuracy and robustness of intracranial hematoma segmentation models.

[0003] In recent years, Transformer has been widely used in medical image segmentation due to its global self-attention mechanism. It can establish long-distance dependencies between pixels, making up for the limitations of the local receptive field of traditional convolutional neural networks (CNNs). However, the dense self-attention mechanism of traditional Transformer has the problem of high computational complexity. Its computational cost is quadratic with the number of image terms, making it difficult to deploy on portable clinical diagnostic devices. At the same time, brain CT images have problems such as small data volume, low signal-to-noise ratio, and diverse morphology and blurred boundaries of hematoma regions. Dense self-attention is prone to introducing a large amount of noise interference from irrelevant background terms, causing the attention matrix to tend to be uniform and failing to fully capture the feature differences and inherent anatomical relationships between intracranial hematoma and surrounding normal anatomical structures.

[0004] Among existing intracranial hematoma segmentation methods, some rely solely on CNNs to extract local features, making it difficult to model the global relationships between different anatomical regions of the brain, and easily leading to undersegmentation or oversegmentation of the hematoma region. Other methods introduce the dense self-attention mechanism of Transformer, which learns global feature modeling solely through data-driven learning, and is prone to overfitting in small sample scenarios, resulting in low segmentation accuracy for small-volume, vaguely defined intracranial hematomas. Still other attention sparsity methods use manually designed window attention, axial attention, and other methods, which lack adaptive adaptation to the anatomical features of the brain, making it difficult to establish stable anatomical dependencies between intracranial hematomas and regions such as ventricles, brain parenchyma, and skull, resulting in poor robustness of the segmentation results. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides an anatomically correlated prior-driven method and system for intracranial hematoma segmentation, thereby solving the technical problems of existing intracranial hematoma segmentation methods, such as difficulty in modeling brain anatomical correlations, high computational complexity of global modeling, and low segmentation accuracy in small samples.

[0006] To achieve the above objectives, according to a first aspect of the present invention, an anatomically correlated prior-driven method for intracranial hematoma segmentation is provided, comprising: The brain CT scan image to be segmented is received, and after preprocessing, the brain CT scan image is transmitted as the input image to the intracranial hematoma segmentation network driven by anatomical association prior, so as to obtain the segmentation feature map output by the intracranial hematoma segmentation network. The segmentation feature map is classified by an activation function to obtain a segmentation result map of intracranial hematoma and background; The intracranial hematoma segmentation network includes an anatomical anchor point localization module, an encoder, a bottleneck network, and a decoder that skips links to the encoder. The encoder includes a convolutional block and multiple encoding blocks, each encoding block consisting of a convolutional block, a downsampling layer, and a global feature modeling layer guided by anatomical association priors. The bottleneck network includes multiple global feature modeling layers. The decoder includes multiple decoding blocks, with the last decoding block consisting of a convolutional block containing channel connections, an upsampling layer, and another convolutional block. The remaining decoding blocks consist of a convolutional block containing channel connections, an upsampling layer, and the global feature modeling layer. The anatomical anchor point localization module is used to identify anatomical anchor points in the input image; The global feature modeling layer includes an anatomical anchor-guided region partitioning module, an anatomical region partitioning verification module, a region-level anatomical attribute modeling module, and an anatomical region selective interaction module. The anatomical anchor-guided region partitioning module divides the lexical units of the feature map input to the current global feature modeling layer into matching anatomical anchors, forming anatomical regions corresponding to each anchor. Each anatomical region is composed of image patches corresponding to lexical units matching the same anchor. The anatomical region partitioning verification module corrects outlier sub-regions within the anatomical regions by partitioning them into other connected anatomical regions. The region-level anatomical attribute modeling module fuses lexical units within each anatomical region to extract anatomical attribute features. The anatomical region selective interaction module models the correlation between lexical units and anatomical regions in the feature map based on the anatomical attribute features, and fuses the anatomical attribute features of different anatomical regions into each lexical unit based on the correlation between lexical units and anatomical regions in the feature map, obtaining and outputting a new feature map.

[0007] Based on the aforementioned anatomically-associated prior-driven intracranial hematoma segmentation method, anatomical anchor points in the input image are identified, specifically including: The input image is matched with all template images in the template library at the feature level, and the global feature similarity between the input image and each template image is calculated. The template images cover multiple standard brain CT images without intracranial hematoma and are marked with corresponding normal anatomical anchor points, which correspond to the center of each anatomical region of the brain. Select the image with the highest global feature similarity to the input image. k A template image is used to construct the input image and the template image. k The mapping relationship between template images, and based on the mapping relationship, the... k A template image m Mapping normal anatomical anchor points to the input image yields the following results in the input image: One candidate normal anchor point; The candidate normal anchor points are deduplicated to obtain the normal anchor points in the input image. m A normal anatomical anchor point; Calculate the feature similarity between each word in the deep feature map of the input image and m normal anatomical anchor points in the template image, and filter out words in the deep feature map of the input image whose feature similarity is lower than a preset threshold to form an abnormal feature word set; Spatial location clustering is performed on the set of abnormal feature lexical units to obtain n Cluster centers, and the n Cluster centers serve as anatomical anchors for individual differences; The normal anatomical anchor points and the individual-difference anatomical anchor points are determined as the anatomical anchor points in the input image.

[0008] Based on the aforementioned intracranial hematoma segmentation method driven by anatomical association priors, the terms in the feature map input to the current global feature modeling layer are divided into matching anatomical anchor points, forming the anatomical regions corresponding to each anchor point, specifically including: Calculate each word in the feature map With each anatomical anchor point Spatial European distance With feature similarity ; Based on each word With each anatomical anchor point Spatial European distance With feature similarity construct lexical units using the following formula Anatomical anchor points Multi-dimensional fusion score :

[0009] in, and These are the weighting coefficients. This is the spatial distance attenuation coefficient; For each lexical unit, the anatomical anchor point with the highest multi-dimensional fusion score is selected as its matching anatomical anchor point, and the corresponding anatomical region is generated based on the lexical units that match the same anatomical anchor point.

[0010] Based on the aforementioned anatomically-associated prior-driven intracranial hematoma segmentation method, outlier sub-regions within the anatomical regions are modified by assigning them to other connected anatomical regions. Specifically, this includes: Connectivity analysis is performed on all subregions contained in the same anatomical region to determine whether each subregion is an outlier subregion that is not connected to the subregion where the corresponding anatomical anchor point is located. For each outlier subregion, calculate the spatial distance between the outlier subregion and the subregions containing other anatomical anchors; Each outlier sub-region is assigned to the anatomical region to which the sub-region corresponding to the shortest spatial distance belongs; where, if there are multiple sub-regions corresponding to the shortest spatial distance, the corresponding outlier sub-region is assigned to the anatomical region to which the feature similarity between the corresponding anatomical anchor point and the word in the outlier sub-region is the highest.

[0011] Based on the above-mentioned intracranial hematoma segmentation method driven by anatomical association priors, the lexical units within each anatomical region are fused to extract the anatomical attribute features of the corresponding anatomical region, specifically including: For each anatomical region, calculate each word within that region. The partition consistency score, grouping determinism score, and feature significance score; Among them, any word element The partition consistency score is determined based on the mean of the feature similarity between the word and other words within the same anatomical region; any word The grouping deterministic score is determined based on the difference between the multidimensional fusion score between the word and the anatomical region and the highest multidimensional fusion score between the word and other anatomical regions; any word Feature salience score It is calculated by the following formula: ,in, It is a learnable mapping matrix; For each anatomical region, for each word within that anatomical region The partition consistency score, grouping determinism score, and feature saliency score are normalized and then summed to obtain the individual word units. The weighted fusion weights are calculated based on each lexical unit within the anatomical region. The weighted fusion weight for each word unit By performing a weighted summation, the anatomical characteristics of the anatomical region are obtained.

[0012] Based on the aforementioned anatomical association prior-driven intracranial hematoma segmentation method, the correlation degree between each word in the feature map and each anatomical region is modeled based on the anatomical attribute features of each anatomical region. Then, based on the correlation degree between each word in the feature map and each anatomical region, the anatomical attribute features of different anatomical regions are fused into each word to obtain and output a new feature map. Specifically, this includes: Calculate the semantic relevance coefficient matrix between each word in the feature map and the anatomical attribute features of each anatomical region. ; Calculate the semantic similarity matrix between the anatomical attribute features of each anatomical region. ; Based on the symmetry of lexical units in different anatomical regions relative to the brain midline, the symmetry association weights between various anatomical regions are determined. The greater the number of lexical units that are symmetrical about the midline of the brain between any two anatomical regions, the higher the symmetry association weight between the two anatomical regions. The closer it is to 1; Based on the adjacency boundary dimensions between different anatomical regions, determine the adjacency association weights between each anatomical region. The longer the adjacency boundary between any two anatomical regions, the greater the adjacency association weight between them. use and right Enhancement, and through For the enhanced We perform weighted analysis to obtain the degree of correlation between each word and each anatomical region; For each lexical unit, the anatomical attribute features of each anatomical region are weighted and summed based on the degree of association between the lexical unit and each anatomical region, and then added to the lexical unit to obtain a new lexical unit for the image block corresponding to the lexical unit. Based on the new lexical units of all image patches, generate and output a new feature map.

[0013] Based on the above-mentioned anatomically-associated prior-driven intracranial hematoma segmentation method, the loss function used in the training process of the intracranial hematoma segmentation network is:

[0014] in, , and For loss weights; The segmentation loss is used to measure the difference between the segmentation result map and the pixel-level annotation; This is a region consistency loss, used to constrain the consistency of lexical units within the same anatomical region; This is a connectivity loss used to constrain the spatial connectivity between sub-regions within the anatomical region generated by the region segmentation module guided by the anatomical anchors.

[0015] According to a second aspect of the invention, an anatomically correlated prior-driven intracranial hematoma segmentation system is provided, comprising: The image segmentation unit is used to receive the brain CT scan image to be segmented, preprocess the brain CT scan image and then transmit it as the input image to the intracranial hematoma segmentation network driven by anatomical association prior, so as to obtain the segmentation feature map output by the intracranial hematoma segmentation network. The segmentation result generation unit is used to classify the segmentation feature map through an activation function to obtain a segmentation result map of intracranial hematoma and background; The intracranial hematoma segmentation network includes an anatomical anchor point localization module, an encoder, a bottleneck network, and a decoder that skips links to the encoder. The encoder includes a convolutional block and multiple encoding blocks, each encoding block consisting of a convolutional block, a downsampling layer, and a global feature modeling layer guided by anatomical association priors. The bottleneck network includes multiple global feature modeling layers. The decoder includes multiple decoding blocks, with the last decoding block consisting of a convolutional block containing channel connections, an upsampling layer, and another convolutional block. The remaining decoding blocks consist of a convolutional block containing channel connections, an upsampling layer, and the global feature modeling layer. The anatomical anchor point localization module is used to identify anatomical anchor points in the input image; The global feature modeling layer includes an anatomical anchor-guided region partitioning module, an anatomical region partitioning verification module, a region-level anatomical attribute modeling module, and an anatomical region selective interaction module. The anatomical anchor-guided region partitioning module divides the lexical units of the feature map input to the current global feature modeling layer into matching anatomical anchors, forming anatomical regions corresponding to each anchor. Each anatomical region is composed of image patches corresponding to lexical units matching the same anchor. The anatomical region partitioning verification module corrects outlier sub-regions within the anatomical regions by partitioning them into other connected anatomical regions. The region-level anatomical attribute modeling module fuses lexical units within each anatomical region to extract anatomical attribute features. The anatomical region selective interaction module models the correlation between lexical units and anatomical regions in the feature map based on the anatomical attribute features, and fuses the anatomical attribute features of different anatomical regions into each lexical unit based on the correlation between lexical units and anatomical regions in the feature map, obtaining and outputting a new feature map.

[0016] According to a third aspect of the present invention, an electronic device is provided, comprising: a computer-readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method as described in the first aspect.

[0017] According to a fourth aspect of the invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to perform the method as described in the first aspect.

[0018] According to a fifth aspect of the invention, a computer program product is provided, comprising a computer program or instructions that, when executed by a processor, implement the method as described in the first aspect.

[0019] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: An anatomically-associated prior-driven intracranial hematoma segmentation network was constructed. This network identifies anatomical anchor points in the input image and uses a global feature modeling layer guided by anatomical association priors to segment the lexical units of the feature map input to the current global feature modeling layer to match the anatomical anchor points, forming anatomical regions corresponding to each anchor point. Outlier regions within these anatomical regions are corrected by reclassifying them to other connected anatomical regions. Furthermore, lexical units within each anatomical region are fused to extract the anatomical attribute features of the corresponding region, achieving representative anatomical attribute modeling for each region. Based on the anatomical attribute features of each anatomical region, the correlation between each lexical unit in the feature map and each anatomical region is modeled. This allows for the fusion of anatomical attribute features of different anatomical regions into each lexical unit. Based on the inherent anatomical correlations between different anatomical regions, an anatomically-associated prior-driven dependency relationship is established between image lexical units at each location and different anatomical regions. This fully utilizes prior knowledge of brain anatomy, significantly improving the segmentation accuracy and robustness of intracranial hematoma regions while reducing the computational complexity of global modeling.

[0020] Furthermore, the anatomical anchor point localization module selects highly similar templates from a normal brain anatomical structure template library through template matching, obtains stable normal anatomical anchor points through spatial mapping and deduplication, and obtains individual-differentiated anchor points by analyzing abnormal feature words. This not only ensures the anatomical rationality of the anchor points, but also adaptively captures individual anatomical differences and hematoma-related abnormal features, solving the problem that traditional methods cannot balance universality and personalization.

[0021] Furthermore, the anatomical anchor-guided region segmentation module combines spatial distance and feature similarity to construct a multi-dimensional criterion, assigning each lexical unit to the anatomical region represented by the most relevant anatomical anchor. This enables accurate clustering of regions with similar features, such as intracranial hematoma, ventricles, brain parenchyma, and skull, avoiding segmentation bias caused by a single criterion. This makes the region segmentation more closely match the actual distribution of brain anatomical structures and improves the segmentation accuracy between intracranial hematoma and normal anatomical regions.

[0022] Furthermore, the anatomical region division verification module detects non-connected regions through a connected region labeling algorithm, classifying outlier sub-regions into the nearest connected anatomical regions. This effectively avoids the problem of region fragmentation, improves the clinical rationality of anatomical region division, and solves the defect of existing methods that ignore the spatial continuity of regions.

[0023] Furthermore, the regional anatomical attribute modeling module adopts a multi-dimensional score weighted fusion strategy, which comprehensively considers the consistency of partitions, the determinism of grouping, and the saliency of features. This enables it to more accurately and comprehensively represent the anatomical attributes of regions, providing high-quality feature support for inter-regional interactions and improving the model's ability to represent complex anatomical structures.

[0024] Furthermore, the anatomical region selective interaction module combines semantic similarity between lexical units and regions, semantic similarity between regions, symmetry association, and adjacency association to construct global interaction relationships. This not only better models anatomical relationships but also effectively solves the problem of insufficient segmentation accuracy caused by existing methods not incorporating anatomical priors. In particular, it improves the segmentation accuracy of small-volume, vaguely defined intracranial hematomas and reduces undersegmentation and oversegmentation. Simultaneously, it reduces the computational complexity of global interaction from... Reduce to ,in N The number of lexical units, For the number of regions, This enables more efficient and reasonable global feature modeling.

[0025] Furthermore, the constructed total loss function combines segmentation loss, region consistency loss, and connectivity loss, which not only ensures the accuracy of pixel-level classification but also constrains the consistency of lexical features and spatial connectivity of region partitioning, making network training more stable and segmentation results more robust, and still achieving excellent segmentation results on small sample brain CT datasets. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the anatomically-associated prior-driven intracranial hematoma segmentation method provided in an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of an intracranial hematoma segmentation network based on anatomical association prior modeling, provided for an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram of the structure of the global feature modeling layer provided in an embodiment of the present invention.

[0029] Figure 4 This is a schematic diagram of the anatomical region division results provided in an embodiment of the present invention.

[0030] Figure 5 This is a schematic diagram of the intracranial hematoma segmentation results provided in an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0032] This invention provides a method for intracranial hematoma segmentation driven by anatomical association priors, such as... Figure 1 As shown, it includes: S1, Receive the brain CT scan image to be segmented, preprocess the brain CT scan image and then transmit it as the input image to the intracranial hematoma segmentation network driven by anatomical association prior, and obtain the segmentation feature map output by the intracranial hematoma segmentation network. S2, the segmentation feature map is classified by an activation function to obtain a segmentation result map of intracranial hematoma and background; The intracranial hematoma segmentation network includes an anatomical anchor point localization module, an encoder, a bottleneck network, and a decoder that skips links to the encoder. The encoder includes a convolutional block and multiple encoding blocks, each encoding block consisting of a convolutional block, a downsampling layer, and a global feature modeling layer guided by anatomical association priors. The bottleneck network includes multiple global feature modeling layers. The decoder includes multiple decoding blocks, with the last decoding block consisting of a convolutional block containing channel connections, an upsampling layer, and another convolutional block. The remaining decoding blocks consist of a convolutional block containing channel connections, an upsampling layer, and the global feature modeling layer. The anatomical anchor point localization module is used to identify anatomical anchor points in the input image; The global feature modeling layer includes an anatomical anchor-guided region partitioning module, an anatomical region partitioning verification module, a region-level anatomical attribute modeling module, and an anatomical region selective interaction module. The anatomical anchor-guided region partitioning module divides the lexical units of the feature map input to the current global feature modeling layer into matching anatomical anchors, forming anatomical regions corresponding to each anchor. Each anatomical region is composed of image patches corresponding to lexical units matching the same anchor. The anatomical region partitioning verification module corrects outlier sub-regions within the anatomical regions by partitioning them into other connected anatomical regions. The region-level anatomical attribute modeling module fuses lexical units within each anatomical region to extract anatomical attribute features. The anatomical region selective interaction module models the correlation between lexical units and anatomical regions in the feature map based on the anatomical attribute features, and fuses the anatomical attribute features of different anatomical regions into each lexical unit based on the correlation between lexical units and anatomical regions in the feature map, obtaining and outputting a new feature map.

[0033] Specifically, clinically acquired brain CT scan images can be selected as the dataset, covering cases from patients of different ages and genders. This includes positive images of intracranial hematomas with varying volumes, locations, and morphologies, as well as negative images without intracranial hematomas. All images are accompanied by pixel-level gold-standard segmentation maps annotated by clinicians (for model validation). The dataset is randomly divided into training, validation, and test sets in a 7:1:2 ratio to ensure representative case distribution in each set. Furthermore, with the assistance of clinical experts, a template library of normal brain anatomy structures is constructed. Thirty standard brain CT images without hematomas and with clear anatomical structures are selected. Clinicians annotate the centers of different anatomical regions of normal anatomical structures such as the ventricles, brain parenchyma, and skull as normal anatomical anchor points, and record the coordinates of each anchor point for subsequent anatomical anchor point localization.

[0034] When receiving a brain CT scan image to be segmented (which can be a brain CT scan image from the training set, validation set, or test set), the brain CT scan image is normalized, and the original pixel values ​​are linearly mapped to the range [0,1] to eliminate feature distribution deviations caused by differences in different scanning devices and scanning parameters. A rectangular cropping operation is used to remove invalid background areas outside the head (such as the scanning bed, air, etc.), retaining the effective area containing complete brain anatomical structures (skull, brain parenchyma, ventricles, etc.) to ensure that subsequent feature extraction focuses on key anatomical structures. The cropped image is scaled to a uniform size (e.g., 512×512 pixels) using bilinear interpolation to obtain a standardized input image, and this input image is transmitted to an intracranial hematoma segmentation network driven by anatomical association priors for segmentation processing to obtain the segmentation feature map output by the intracranial hematoma segmentation network.

[0035] Among them, such as Figure 2 As shown, the intracranial hematoma segmentation network includes an anatomical anchor localization module, an encoder, a bottleneck network, and a decoder that skips links to the encoder. The anatomical anchor localization module identifies anatomical anchors in the input image. The encoder performs deep feature modeling on the input image and consists of concatenated convolutional blocks and multiple encoding blocks. Each encoding block comprises a convolutional block, a downsampling layer (e.g., a 2×2 downsampling layer with a stride of 2), and a global feature modeling layer guided by anatomical association priors. The convolutional blocks extract local texture features from the image and output feature maps, extracting local texture features at different scales to deepen the feature capture of local anatomical structures in the brain. The downsampling layer uses max pooling downsampling to reduce the feature map size to half its original size, reducing computation while expanding the receptive field, laying the foundation for global feature modeling. The global feature modeling layer guided by anatomical association priors is used to establish stable anatomical dependencies between different anatomical regions of the brain. In some embodiments, a convolutional block may consist of two networks containing convolutional layers (e.g., a 3×3 convolutional layer with a stride of 1 and padding of 1), a batch normalization layer, and a ReLU activation layer. Taking a convolutional block with 32 output channels and 4 coding blocks as an example, the number of output channels for the 4 coding blocks are 64, 128, 256, and 512 respectively, to progressively improve the semantic expressiveness of the features.

[0036] The bottleneck network is used to further enhance the model's global perception by modeling global anatomical associations in deep feature maps. It consists of multiple (e.g., 5) global feature modeling layers, without downsampling layers. The input to the bottleneck network is the feature map (512 channels) output from the last encoding block of the encoder. By modeling global anatomical associations through global feature modeling layers guided by anatomical association priors, it achieves the fusion and enhancement of deep features, and outputs a feature map (512 channels) that incorporates global semantic features, which is then transmitted to the decoder.

[0037] The decoder, similar in structure to the encoder, is used to progressively restore the feature map resolution and predict the final segmentation result through jump links, feature concatenation, local feature modeling, and global feature modeling. It comprises multiple sequentially concatenated decoding blocks. The last decoding block consists of a convolutional block with channel connections, an upsampling layer, and another convolutional block. The remaining decoding blocks consist of convolutional blocks with channel connections, upsampling layers, and the aforementioned global feature modeling layer. The convolutional block with channel connections is composed of a feature concatenation layer and the aforementioned convolutional block. This feature concatenation layer concatenates the feature map of the decoding block with the corresponding scale feature map of the encoder through jump links, fusing shallow detail features (such as hematoma edge details and small blood vessel textures) with deep semantic features (such as anatomical structure associations and the overall morphology of the hematoma). The convolutional block is used to enhance the local features of the concatenated features. The upsampling layer enlarges the feature map size to twice its original size through transposed convolution operations, progressively restoring image resolution and providing a foundation for accurate segmentation. The global feature modeling layer is used to perform global anatomical association modeling, ensuring the consistency of anatomical dependencies during feature restoration and avoiding feature distortion during resolution restoration. Taking a coding block of 4 as an example, the decoder also has 4 decoding blocks. The output channels of the 4 decoding blocks are 256, 128, 64 and 32 respectively, and the final output channel is a segmentation feature map (512×512×2) with 2 channels (corresponding to the intracranial hematoma area and the background area respectively).

[0038] In some embodiments, the anatomical anchor point positioning module is specifically used for: The input image is matched against all template images in the constructed template library at the feature level, and the global feature similarity between the input image and each template image is calculated. This can be achieved by using an additional encoding network (e.g., Unet) to capture the global feature vectors of the input and template images. For example, the input and template images are fed into this encoding network, and global average pooling is performed on the deep feature maps output from the last network layer to obtain the global feature vectors of the input and template images, respectively. Then, the global feature similarity between the global feature vectors of the input and template images is calculated based on these vectors. : .in, The global feature vector of the input image. For the first i The global feature vector of a template image.

[0039] Then, select the image with the highest global feature similarity to the input image. k (For example, 5) template images, constructing the input image with these... k The mapping relationship between template images, and based on this mapping relationship, the image is... k Annotations in the template image m(For example, 18) normal anatomical anchor points are mapped to the input image to obtain the values ​​in the input image. There are several candidate normal anchor points. Among them, the mapping of normal anatomical anchor points can be achieved using a thin-plate spline interpolation algorithm. For the input image... To remove duplicate candidate normal anchor points, for example, K-Medoids clustering can be performed on the candidate normal anchor points, and the cluster centers can be used as the normal anatomical anchor points after deduplication, thereby obtaining the normal anatomical anchor points in the input image. m There are m normal anatomical anchor points. Then, the feature similarity between each word in the deep feature map of the input image (output by the last layer of the additionally configured encoding network) and the m normal anatomical anchor points in the template image can be calculated. This allows for the selection of words in the global feature map of the input image whose feature similarity to all m normal anatomical anchor points is below a preset threshold, forming an abnormal feature word set. Here, a word refers to an image patch in the corresponding feature map. The feature similarity between any word in the deep feature map of the input image and any normal anatomical anchor point in the template image can be obtained by calculating the cosine similarity between the word and the word corresponding to the image patch in the deep feature map of the corresponding template image (output by the last layer of the additionally configured encoding network).

[0040] Furthermore, spatial location clustering is performed on the aforementioned set of anomalous feature lexical units, resulting in n (e.g., 6) cluster centers. For example, K-Medoids clustering analysis can be performed on the spatial locations of all anomalous feature lexical units, with the number of clusters set to [number missing]. n ,get n There are cluster centers. Then, n Each cluster center serves as an individual-specific anatomical anchor point. Finally, the aforementioned normal anatomical anchor points and individual-specific anatomical anchor points are identified as all anatomical anchor points in the input image, thereby constructing a set of anchor point parameters for the input image that includes anatomical anchor point location information.

[0041] A global feature modeling layer guided by anatomical association priors serves as the core layer for global feature modeling, such as... Figure 3 As shown, its built-in anatomical anchor point-guided region division module, anatomical region division verification module, region-level anatomical attribute modeling module, and anatomical region selective interaction module work as follows: (1) Anatomical anchor point guided region division module: used to divide the word units of the feature map input to the current global feature modeling layer into matching anatomical anchor points to form multiple anatomical regions, wherein the anatomical regions are composed of image blocks corresponding to the word units that match the same anatomical anchor point.

[0042] In some embodiments, each word in the feature map input to the current global feature modeling layer can be computed. With each anatomical anchor point Spatial European distance With feature similarity Among them, word elements Anatomical anchor points Feature similarity It can be based on this word element Anatomical anchor points It is calculated using the cosine similarity of the feature embeddings at the corresponding position in the current input feature map. Based on each word... With each anatomical anchor point Spatial European distance With feature similarity construct lexical units using the following formula Anatomical anchor points Multi-dimensional fusion score :

[0043] in, and These are the weighting coefficients. This is the spatial distance attenuation coefficient. For example, it can be set to... , Spatial distance attenuation coefficient .

[0044] For each lexical unit, the anatomical anchor point with the highest multi-dimensional fusion score is selected as its matching anatomical anchor point. Based on lexical units matching the same anatomical anchor point, corresponding anatomical regions are generated. The resulting anatomical regions are shown below. Figure 3 As shown.

[0045] (2) Anatomical region division verification module: It is used to correct the outlier sub-regions in the anatomical region output by the anatomical anchor point guided region division module, and divide the outlier sub-regions into other anatomical regions connected to them, thereby realizing the correction of the anatomical region.

[0046] In some embodiments, connected component analysis can be performed on all sub-regions contained in the same anatomical region to determine whether each sub-region is an outlier that is not connected to the sub-region containing the corresponding anatomical anchor point. For example, an 8-neighborhood connected component labeling algorithm can be used to determine whether each anatomical region is a single connected region. If an anatomical region is a disconnected region, i.e., some sub-regions are not connected to other sub-regions, then the sub-regions in that anatomical region that are not connected to the main body of the anatomical region (i.e., the sub-region containing the anatomical anchor point) are extracted as outlier sub-regions. For each outlier sub-region, the spatial distance between the outlier sub-region and the sub-regions containing other anatomical anchor points is calculated, and each outlier sub-region is assigned to the anatomical region to which the sub-region corresponding to the shortest spatial distance belongs. Where there are multiple sub-regions corresponding to the shortest spatial distance, the corresponding outlier sub-region is assigned to the anatomical region to which the sub-region with the highest feature similarity between the corresponding anatomical anchor point and the word in the outlier sub-region belongs. Perform a second connectivity analysis on the corrected anatomical regions to ensure that there are no outlier subregions. If outlier subregions still exist in the anatomical regions, repeat the above correction steps until all anatomical regions meet the connectivity requirements (i.e., do not contain outlier subregions), and obtain the final anatomical regions.

[0047] (3) Regional-level anatomical attribute modeling module: used to fuse the lexical units in each anatomical region and extract the anatomical attribute features of the corresponding anatomical region.

[0048] In some embodiments, for each anatomical region, each lexical unit within that anatomical region can be calculated. The scores are: partition consistency score, grouping determinism score, and feature significance score.

[0049] Among them, any word element The partition consistency score is determined based on the mean of the feature similarity between the word and other words within the same anatomical region; any word The grouping deterministic score is determined based on the difference between the multidimensional fusion score between the word and the anatomical region and the highest multidimensional fusion score between the word and other anatomical regions; any word Feature salience score It is calculated by the following formula: ,in, It is a learnable mapping matrix.

[0050] For each anatomical region, for each word within that anatomical region Partition consistency score Grouping Deterministic Score and feature significance score After normalization, the individual lexical units are summed. The weighted fusion weights can be normalized and accumulated using, for example, the following formula: Subsequently, based on each lexical unit within this anatomical region... The weighted fusion weight for each word unit The anatomical attributes of the anatomical region are obtained by performing a weighted summation, which is used to represent the core features of the anatomical region.

[0051] (4) Anatomical Region Selective Interaction Module: It is used to model the degree of association between each word in the feature map and each anatomical region based on the anatomical attribute features of each anatomical region, and to fuse the anatomical attribute features of different anatomical regions into each word based on the degree of association between each word in the feature map and each anatomical region, so as to obtain and output a new feature map.

[0052] In some embodiments, a semantic correlation coefficient matrix between each word in the feature map and the anatomical attribute features of each anatomical region can be calculated. , The number of lexical units is used to calculate the semantic similarity matrix between the anatomical attribute features of each anatomical region. .in, The semantic relevance coefficient between a word and each anatomical region is initialized to 0. The semantic relevance coefficient between a word and its corresponding anatomical region is then updated to the semantic similarity between the word and its corresponding anatomical attribute features. The semantic similarity between anatomical attribute features of any two anatomical regions is included. Furthermore, based on the symmetry of lexical units within different anatomical regions relative to the brain midline, symmetry association weights are determined between each anatomical region. The greater the number of lexical units that are symmetrical about the midline of the brain between any two anatomical regions, the higher the symmetry association weight between the two anatomical regions. The closer it gets to 1. Here, the anatomical area... Anatomical region Symmetric correlation weights between It can be calculated based on the following formula:

[0053] in, Anatomical region Anatomical region The number of paired word pairs symmetrical about the brain's midline. and Anatomical regions Anatomical region The number of lexical units included.

[0054] In addition, the adjacency association weights between different anatomical regions are determined based on the adjacency boundary dimensions between them. In this context, the longer the adjacency boundary between any two anatomical regions, the greater the adjacency association weight between them. Here, the anatomical regions... Anatomical region Adjacency association weights between It can be calculated based on the following formula:

[0055] in, Anatomical region Anatomical region The length of the adjacent boundary, and Anatomical regions Anatomical region The total length of the boundary.

[0056] Subsequently, using and right Enhancements can be made. For example, the following formula can be used to achieve this. Enhancement:

[0057] use For the enhanced Weighted summaries are applied to determine the degree of association between each lexical unit and each anatomical region. This includes... Compared with the enhanced After performing matrix multiplication, the data is processed by the Softmax function to obtain the degree of association between each word and each anatomical region, thus realizing the dependency modeling driven by the prior of anatomical association.

[0058] For each lexical unit, the anatomical attribute features of each anatomical region can be weighted and summed based on the degree of correlation between the lexical unit and each anatomical region, and then added to the lexical unit to obtain a new lexical unit for the corresponding image patch, thus realizing selective feature interaction between lexical units and anatomical regions. Alternatively, a feedforward neural network can be used to achieve weighted fusion of the anatomical attribute features of each anatomical region and the lexical unit; this embodiment of the invention does not specifically limit this approach. Based on the new lexical units of all image patches, a new feature map is generated and output.

[0059] In some embodiments, the loss function used during the training of the intracranial hematoma segmentation network is:

[0060] in, , and For example, the loss weights can be set to 0.8, 0.1, and 0.1 respectively; The segmentation loss is used to measure the difference between the segmentation result map and the pixel-level annotation; This is a region consistency loss, used to constrain the consistency of lexical units within the same anatomical region; This is the connectivity loss, used to constrain the spatial connectivity between sub-regions within the anatomical region generated by the region partitioning module guided by anatomical anchors.

[0061] In other embodiments, segmentation loss The Dice loss is used to measure the difference between the segmentation result map and the pixel-level annotation. and cross-entropy loss The sum of the components is composed of:

[0062] in The formula used to measure the overlap between the segmentation result image and the pixel-level annotation is as follows:

[0063] in, H and W Given the length and width of the input image, The segmentation network predicts the location of intracranial hematoma. (i, j) The probability that the pixels on the screen belong to an intracranial hematoma. The label is for the corresponding pixel (1 for intracranial hematoma, 0 for background area). To prevent tiny constants with a denominator of 0.

[0064] The formula used to measure pixel-level classification error is as follows:

[0065] Regional consistency loss The consistency of lexical features within the same anatomical region is constrained by the mean of feature similarity within the region:

[0066] in Avg This indicates the mean operation. For the first l Anatomical characteristics of each anatomical region Indicates that it is located at the th l The lexical units of an anatomical region.

[0067] Connectivity loss Used to constrain the connectivity of region partitioning, and to penalize the generation of disconnected regions:

[0068] in, For the first l The number of connected regions within each anatomical region.

[0069] During training, the Adam optimizer can be used to optimize the model. The initial learning rate is set to 0.0001, the batch size is set to 8, and the weight decay is set to 1e-5 to avoid overfitting during model training. The model is trained for a total of 400 rounds. At the same time, the early stopping method is used to prevent overfitting. When the Dice score on the validation set does not improve for 20 consecutive rounds, training is stopped and the current optimal model parameters are saved.

[0070] After obtaining the segmentation feature map output by the intracranial hematoma segmentation network, an activation function (e.g., Sorfmax) can be used to classify the segmentation feature map, obtaining the probability value of each pixel belonging to the intracranial hematoma region and the background region (the sum of the two probability values ​​is 1). Pixel-level classification is performed using the maximum probability method, that is, the category with the highest probability value for each pixel is taken as the segmentation label for that pixel (label 1 for intracranial hematoma region, label 0 for background region), generating a segmentation result map with the same size as the original brain image, such as... Figure 4 As shown. In addition, the segmentation results can be visualized (e.g., using specific colors to mark the outline and extent of the hematoma area), providing clinicians with intuitive diagnostic references.

[0071] Based on the above embodiments, this application can also be implemented for small-sample scenarios where clinical data annotation is scarce, suitable for scenarios lacking a large amount of pixel-level labeled data. Specifically, data preparation involves constructing a small-sample dataset by selecting a portion of images from the dataset used in the above embodiments. This dataset includes a certain number of positive and negative images of intracranial hematoma, ensuring coverage of typical cases with different hematoma morphologies and locations. The small-sample dataset is divided into training, validation, and test sets in a 7:1:2 ratio. Model training uses the same network structure as the above embodiments for the intracranial hematoma segmentation network. Hyperparameter configuration is based on the above embodiments, with a slight reduction in the number of anatomical anchor points to decrease the difficulty of fine-grained anatomical region segmentation. The training parameter configuration remains consistent with the above embodiments. Segmentation process follows the complete segmentation steps of the above embodiments (image preprocessing, feature extraction, anatomical dependency modeling, feature fusion, and segmentation result generation) to complete intracranial hematoma segmentation in small-sample scenarios, fully utilizing prior knowledge of brain anatomy to compensate for insufficient labeled data.

[0072] In summary, the method provided in this invention constructs an intracranial hematoma segmentation network driven by anatomical association priors. This network identifies anatomical anchor points in the input image and uses a global feature modeling layer guided by anatomical association priors to divide the lexical units of the feature map input to the current global feature modeling layer into matching anatomical anchor points, forming anatomical regions corresponding to each anchor point. Outlier regions within the anatomical regions are corrected and divided into other connected anatomical regions. Furthermore, lexical units within each anatomical region are fused to extract the anatomical attribute features of the corresponding anatomical region, achieving representative anatomical attribute modeling for each region. Then, based on the anatomical attribute features of each anatomical region, the correlation between each lexical unit in the feature map and each anatomical region is modeled. Thus, based on the correlation between each lexical unit in the feature map and each anatomical region, the anatomical attribute features of different anatomical regions are fused into each lexical unit. Based on the inherent anatomical correlations of different anatomical regions, an anatomical association prior-driven dependency relationship is established between image lexical units at each location and different anatomical regions. This fully utilizes prior knowledge of brain anatomy, significantly improving the segmentation accuracy and robustness of intracranial hematoma regions while reducing the computational complexity of global modeling.

[0073] The anatomical association prior-driven intracranial hematoma segmentation system provided by the present invention will be described below. The anatomical association prior-driven intracranial hematoma segmentation system described below can be referred to in correspondence with the anatomical association prior-driven intracranial hematoma segmentation method described above.

[0074] This invention provides an anatomically-associated prior-driven intracranial hematoma segmentation system, comprising: The image segmentation unit is used to receive the brain CT scan image to be segmented, preprocess the brain CT scan image and then transmit it as the input image to the intracranial hematoma segmentation network driven by anatomical association prior, so as to obtain the segmentation feature map output by the intracranial hematoma segmentation network. The segmentation result generation unit is used to classify the segmentation feature map through an activation function to obtain a segmentation result map of intracranial hematoma and background; The intracranial hematoma segmentation network includes an anatomical anchor point localization module, an encoder, a bottleneck network, and a decoder that skips links to the encoder. The encoder includes a convolutional block and multiple encoding blocks, each encoding block consisting of a convolutional block, a downsampling layer, and a global feature modeling layer guided by anatomical association priors. The bottleneck network includes multiple global feature modeling layers. The decoder includes multiple decoding blocks, with the last decoding block consisting of a convolutional block containing channel connections, an upsampling layer, and another convolutional block. The remaining decoding blocks consist of a convolutional block containing channel connections, an upsampling layer, and the global feature modeling layer. The anatomical anchor point localization module is used to identify anatomical anchor points in the input image; The global feature modeling layer includes an anatomical anchor-guided region partitioning module, an anatomical region partitioning verification module, a region-level anatomical attribute modeling module, and an anatomical region selective interaction module. The anatomical anchor-guided region partitioning module divides the lexical units of the feature map input to the current global feature modeling layer into matching anatomical anchors, forming anatomical regions corresponding to each anchor. Each anatomical region is composed of image patches corresponding to lexical units matching the same anchor. The anatomical region partitioning verification module corrects outlier sub-regions within the anatomical regions by partitioning them into other connected anatomical regions. The region-level anatomical attribute modeling module fuses lexical units within each anatomical region to extract anatomical attribute features. The anatomical region selective interaction module models the correlation between lexical units and anatomical regions in the feature map based on the anatomical attribute features, and fuses the anatomical attribute features of different anatomical regions into each lexical unit based on the correlation between lexical units and anatomical regions in the feature map, obtaining and outputting a new feature map.

[0075] This invention provides an electronic device, including: a computer-readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method as described in any of the above embodiments.

[0076] This invention provides a computer-readable storage medium storing computer instructions that cause a processor to perform the method described in any of the above embodiments.

[0077] This invention provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the method described in any of the above embodiments.

[0078] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intracranial hematoma segmentation driven by anatomical association priors, characterized in that, include: The brain CT scan image to be segmented is received, and after preprocessing, the brain CT scan image is transmitted as the input image to the intracranial hematoma segmentation network driven by anatomical association prior, so as to obtain the segmentation feature map output by the intracranial hematoma segmentation network. The segmentation feature map is classified by an activation function to obtain a segmentation result map of intracranial hematoma and background; The intracranial hematoma segmentation network includes an anatomical anchor point localization module, an encoder, a bottleneck network, and a decoder that skips links to the encoder. The encoder includes a convolutional block and multiple encoding blocks, each encoding block consisting of a convolutional block, a downsampling layer, and a global feature modeling layer guided by anatomical association priors. The bottleneck network includes multiple global feature modeling layers. The decoder includes multiple decoding blocks, with the last decoding block consisting of a convolutional block containing channel connections, an upsampling layer, and another convolutional block. The remaining decoding blocks consist of a convolutional block containing channel connections, an upsampling layer, and the global feature modeling layer. The anatomical anchor point localization module is used to identify anatomical anchor points in the input image; The global feature modeling layer includes an anatomical anchor-guided region segmentation module, an anatomical region segmentation verification module, a region-level anatomical attribute modeling module, and an anatomical region selective interaction module. The anatomical anchor-guided region segmentation module is used to segment the lexical units of the feature map input to the current global feature modeling layer to matching anatomical anchors, forming anatomical regions corresponding to each anatomical anchor, wherein the anatomical region is composed of image patches corresponding to lexical units matching the same anatomical anchor. The anatomical region segmentation verification module is used to correct outlier sub-regions in the anatomical regions and segment the outlier sub-regions to other anatomical regions connected to them. The region-level anatomical attribute modeling module is used to fuse lexical units in each anatomical region and extract the anatomical attribute features of the corresponding anatomical region. The anatomical region selective interaction module is used to model the degree of association between each word in the feature map and each anatomical region based on the anatomical attribute features of each anatomical region, and to fuse the anatomical attribute features of different anatomical regions into each word based on the degree of association between each word in the feature map and each anatomical region, so as to obtain and output a new feature map.

2. The intracranial hematoma segmentation method driven by anatomical association priors as described in claim 1, characterized in that, Identifying anatomical anchor points in the input image specifically includes: The input image is matched with all template images in the template library at the feature level, and the global feature similarity between the input image and each template image is calculated. The template images cover multiple standard brain CT images without intracranial hematoma and are marked with corresponding normal anatomical anchor points, which correspond to the center of each anatomical region of the brain. Select the image with the highest global feature similarity to the input image. k A template image is used to construct the input image and the template image. k The mapping relationship between template images, and based on the mapping relationship, the... k A template image m Mapping normal anatomical anchor points to the input image yields the following results in the input image: One candidate normal anchor point; The candidate normal anchor points are deduplicated to obtain the normal anchor points in the input image. m A normal anatomical anchor point; Calculate the feature similarity between each word in the deep feature map of the input image and m normal anatomical anchor points in the template image, and filter out words in the deep feature map of the input image whose feature similarity is lower than a preset threshold to form an abnormal feature word set; Spatial location clustering is performed on the abnormal feature word set to obtain n cluster centers, and the n cluster centers are used as individual difference anatomy anchor points; The normal anatomical anchor points and the individual-difference anatomical anchor points are determined as the anatomical anchor points in the input image.

3. The intracranial hematoma segmentation method driven by anatomical association priors as described in claim 1, characterized in that, The lexical units of the feature map input to the current global feature modeling layer are divided into matching anatomical anchors, forming anatomical regions corresponding to each anatomical anchor, specifically including: Calculate each word in the feature map With each anatomical anchor point Spatial European distance With feature similarity ; Based on each word With each anatomical anchor point Spatial European distance With feature similarity construct lexical units using the following formula Anatomical anchor points Multi-dimensional fusion score : in, and These are the weighting coefficients. This is the spatial distance attenuation coefficient; For each lexical unit, the anatomical anchor point with the highest multi-dimensional fusion score is selected as its matching anatomical anchor point, and the corresponding anatomical region is generated based on the lexical units that match the same anatomical anchor point.

4. The intracranial hematoma segmentation method driven by anatomical association priors as described in claim 1, characterized in that, The outlier sub-regions in the anatomical region are modified by dividing the outlier sub-regions into other anatomical regions connected to them, specifically including: Connectivity analysis is performed on all subregions contained in the same anatomical region to determine whether each subregion is an outlier subregion that is not connected to the subregion where the corresponding anatomical anchor point is located. For each outlier subregion, calculate the spatial distance between the outlier subregion and the subregions containing other anatomical anchors; Each outlier sub-region is assigned to the anatomical region to which the sub-region corresponding to the shortest spatial distance belongs; where, if there are multiple sub-regions corresponding to the shortest spatial distance, the corresponding outlier sub-region is assigned to the anatomical region to which the feature similarity between the corresponding anatomical anchor point and the word in the outlier sub-region is the highest.

5. The intracranial hematoma segmentation method driven by anatomical association priors as described in claim 3, characterized in that, The lexical units within each anatomical region are fused to extract the anatomical attribute features of the corresponding region, specifically including: For each anatomical region, calculate each word within that region. The partition consistency score, grouping determinism score, and feature significance score; Among them, any word element The partition consistency score is determined based on the mean of the feature similarity between the word and other words within the same anatomical region; any word The grouping deterministic score is determined based on the difference between the multidimensional fusion score between the word and the anatomical region and the highest multidimensional fusion score between the word and other anatomical regions; any word Feature salience score It is calculated by the following formula: ,in, It is a learnable mapping matrix; For each anatomical region, for each word within that anatomical region The partition consistency score, grouping determinism score, and feature saliency score are normalized and then summed to obtain the individual word units. The weighted fusion weights are calculated based on each lexical unit within the anatomical region. The weighted fusion weight for each word unit By performing a weighted summation, the anatomical characteristics of the anatomical region are obtained.

6. The intracranial hematoma segmentation method driven by anatomical association priors as described in claim 1, characterized in that, Based on the anatomical attribute features of each anatomical region, the correlation between each word in the feature map and each anatomical region is modeled. Then, based on the correlation between each word in the feature map and each anatomical region, the anatomical attribute features of different anatomical regions are fused into each word to obtain and output a new feature map. Specifically, this includes: Calculate the semantic relevance coefficient matrix between each word in the feature map and the anatomical attribute features of each anatomical region. ; Calculate the semantic similarity matrix between the anatomical attribute features of each anatomical region. ; Based on the symmetry of lexical units in different anatomical regions relative to the brain midline, the symmetry association weights between various anatomical regions are determined. The greater the number of lexical units that are symmetrical about the midline of the brain between any two anatomical regions, the higher the symmetry association weight between the two anatomical regions. The closer it is to 1; Based on the adjacency boundary dimensions between different anatomical regions, determine the adjacency association weights between each anatomical region. The longer the adjacency boundary between any two anatomical regions, the greater the adjacency association weight between them. use and right Enhancement, and through For the enhanced We perform weighted analysis to obtain the degree of correlation between each word and each anatomical region; For each lexical unit, the anatomical attribute features of each anatomical region are weighted and summed based on the degree of association between the lexical unit and each anatomical region, and then added to the lexical unit to obtain a new lexical unit for the image block corresponding to the lexical unit. Based on the new lexical units of all image patches, generate and output a new feature map.

7. The anatomically-related prior-driven intracranial hematoma segmentation method according to any one of claims 1 to 6, characterized in that, The loss function used in the training process of the intracranial hematoma segmentation network is: in, , and For loss weights; The segmentation loss is used to measure the difference between the segmentation result map and the pixel-level annotation; This is a region consistency loss, used to constrain the consistency of lexical units within the same anatomical region; This is a connectivity loss used to constrain the spatial connectivity between sub-regions within the anatomical region generated by the region segmentation module guided by the anatomical anchors.

8. An anatomically-associated prior-driven intracranial hematoma segmentation system, characterized in that, include: The image segmentation unit is used to receive the brain CT scan image to be segmented, preprocess the brain CT scan image and then transmit it as the input image to the intracranial hematoma segmentation network driven by anatomical association prior, so as to obtain the segmentation feature map output by the intracranial hematoma segmentation network. The segmentation result generation unit is used to classify the segmentation feature map through an activation function to obtain a segmentation result map of intracranial hematoma and background; The intracranial hematoma segmentation network includes an anatomical anchor point localization module, an encoder, a bottleneck network, and a decoder that skips links to the encoder. The encoder includes a convolutional block and multiple encoding blocks, each encoding block consisting of a convolutional block, a downsampling layer, and a global feature modeling layer guided by anatomical association priors. The bottleneck network includes multiple global feature modeling layers. The decoder includes multiple decoding blocks, with the last decoding block consisting of a convolutional block containing channel connections, an upsampling layer, and another convolutional block. The remaining decoding blocks consist of a convolutional block containing channel connections, an upsampling layer, and the global feature modeling layer. The anatomical anchor point localization module is used to identify anatomical anchor points in the input image; The global feature modeling layer includes an anatomical anchor-guided region segmentation module, an anatomical region segmentation verification module, a region-level anatomical attribute modeling module, and an anatomical region selective interaction module. The anatomical anchor-guided region segmentation module is used to segment the lexical units of the feature map input to the current global feature modeling layer to matching anatomical anchors, forming anatomical regions corresponding to each anatomical anchor, wherein the anatomical region is composed of image patches corresponding to lexical units matching the same anatomical anchor. The anatomical region segmentation verification module is used to correct outlier sub-regions in the anatomical regions and segment the outlier sub-regions to other anatomical regions connected to them. The region-level anatomical attribute modeling module is used to fuse lexical units in each anatomical region and extract the anatomical attribute features of the corresponding anatomical region. The anatomical region selective interaction module is used to model the degree of association between each word in the feature map and each anatomical region based on the anatomical attribute features of each anatomical region, and to fuse the anatomical attribute features of different anatomical regions into each word based on the degree of association between each word in the feature map and each anatomical region, so as to obtain and output a new feature map.

9. An electronic device, characterized in that, include: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a processor to perform the method as described in any one of claims 1-7.

Images