Cerebral hemorrhage volume measurement method and related equipment

Abstract

The invention discloses a cerebral hemorrhage volume measurement method and related equipment. The method comprises the following steps: acquiring brain medical image data to be measured; inputting the brain medical image data into a three-dimensional segmentation network for cerebral hemorrhage region segmentation to obtain a cerebral hemorrhage region image; identifying a cerebral hemorrhage stage of the cerebral hemorrhage area image according to the cerebral hemorrhage area image; constructing a standard hemorrhage template based on the cerebral hemorrhage stage and the lesion anatomical position in the cerebral hemorrhage area image; performing gradient response analysis based on the feature map of the last convolutional layer in the three-dimensional segmentation network to generate an interpretation map; performing spatial similarity analysis based on the standard bleeding template and the interpretation graph to obtain a consistency score between the standard bleeding template and the interpretation graph; according to the cerebral hemorrhage area image, calculating an initial cerebral hemorrhage volume through a preset dual-path estimation model; and adjusting the fusion weight of the dual-path model according to the consistency score, and determining the target cerebral hemorrhage volume according to the adjusted fusion weight and the initial cerebral hemorrhage volume.

Description

Technical Field

[0001] This application relates to the field of medical imaging technology, and in particular to a method and related equipment for measuring the volume of cerebral hemorrhage. Background Technology

[0002] In clinical practice, the measurement of intracerebral hemorrhage (ICH) volume is crucial for patient treatment planning and prognostic assessment. Currently, two main methods are used clinically for measuring hemorrhage volume: the traditional geometric estimation method (i.e., the ABC / 2 method) and the semi-automatic threshold segmentation method.

[0003] (1) Traditional geometric estimation method (ABC / 2 method): Based on the maximum long axis A, maximum short axis B and number of involved layers C of the hemorrhage foci on the CT cross section, the volume of cerebral hemorrhage is calculated by the formula V = (A × B × C) / 2.

[0004] (2) Semi-automatic threshold segmentation method: Utilizing the high density features of the hemorrhage foci in CT images, the hemorrhage area is segmented by a preset CT value threshold (e.g., 40–90 HU), and the area is calculated layer by layer to accumulate the volume, and finally the brain hemorrhage volume is obtained.

[0005] However, traditional geometric estimation methods are only applicable to regular elliptical hemorrhage foci, and have a large error for irregularly shaped hemorrhage foci (such as lobulated or multifocal foci); in addition, this method relies on manual measurement by doctors, which is time-consuming and has low clinical efficiency.

[0006] Secondly, while the semi-automatic threshold segmentation method offers a degree of automation in identifying acute hemorrhage, it is prone to missed detections or misidentifications in the subacute phase due to the decreased CT value of the hemorrhage and its overlap with surrounding edema areas. Furthermore, this method struggles to accurately identify areas with blurred edges or partial volume effects, and adjacent hemorrhage lesions are easily misidentified, leading to inflated volume estimates.

[0007] In summary, existing methods for measuring intracranial hemorrhage volume suffer from poor accuracy, reliance on manual labor, and low efficiency, making it difficult to meet the needs for precise and rapid assessment in complex clinical scenarios. Therefore, there is an urgent need for a novel method for measuring intracranial hemorrhage volume that possesses high precision, broad applicability, and a high degree of automation to improve the scientific rigor and efficiency of clinical diagnosis and treatment decisions. Summary of the Invention

[0008] This application provides a method for measuring the volume of cerebral hemorrhage, which solves the problems in the prior art where traditional geometric estimation methods have large measurement errors and are cumbersome to operate for irregular hemorrhage foci, and semi-automatic threshold segmentation methods are inaccurate in identifying hemorrhage morphology changes and CT value reductions, resulting in poor volume measurement accuracy, reliance on manual labor, low efficiency, and difficulty in meeting the needs of rapid and accurate assessment in complex clinical scenarios.

[0009] This application also provides a brain hemorrhage volume measuring device, an electronic device, a computer-readable storage medium, and a computer program product.

[0010] The embodiments of this application adopt the following technical solutions: In a first aspect, this application provides a method for measuring the volume of cerebral hemorrhage, comprising: Acquire medical image data of the brain to be measured; Brain medical image data is input into a 3D segmentation network to segment the brain hemorrhage area, resulting in an image of the brain hemorrhage area; Identify the stage of brain hemorrhage based on images of the brain hemorrhage area; Based on the stage of cerebral hemorrhage and the anatomical location of the lesion in the image of the cerebral hemorrhage region, a standard hemorrhage template is constructed; the standard hemorrhage template is used to characterize the distribution characteristics of different stages of cerebral hemorrhage in the corresponding anatomical region of the lesion. Gradient response analysis is performed based on the feature map of the last convolutional layer in the 3D segmentation network to generate an interpretation map. The interpretation map is used to characterize the contribution of each image voxel in the brain medical image data to the segmentation result of the brain hemorrhage region. Spatial similarity analysis was performed based on the standard bleed template and the interpretation diagram to obtain a consistency score between the standard bleed template and the interpretation diagram; Based on the image of the brain hemorrhage area, the initial brain hemorrhage volume is calculated using a pre-defined dual-path estimation model. The first estimation path of the dual-path estimation model estimates the initial brain hemorrhage volume using the voxel number statistics method, while the second estimation path estimates the initial brain hemorrhage volume using an image feature regression neural network. The fusion weights of the dual-path model are adjusted based on the consistency score, and the final target cerebral hemorrhage volume is determined based on the adjusted fusion weights and the initial cerebral hemorrhage volume.

[0011] Optionally, a standard hemorrhage template is constructed based on the stage of cerebral hemorrhage and the anatomical location of the lesion in the cerebral hemorrhage region image, including: Obtain the standard CT hemorrhage density distribution characteristics corresponding to the different stages of cerebral hemorrhage; Identify the anatomical location of lesions in images of brain hemorrhage areas. The anatomical location of lesions is used to characterize the spatial location of lesions in the midline of the brain, basal ganglia, lobes, cerebellum and / or brainstem. Extract target 3D reference images corresponding to the stage and anatomical location of the cerebral hemorrhage from a preset template library; the preset template library contains a set of 3D reference images constructed by statistically analyzing multiple historical cerebral hemorrhage image samples with cerebral hemorrhage stage labels and anatomical region annotations; The standard CT hemorrhage density distribution features are fused and registered with the target three-dimensional reference image to obtain a standard hemorrhage template.

[0012] Optionally, gradient response analysis is performed based on the feature map of the last convolutional layer in the 3D segmentation network to generate an interpretive map, including: Gradient backpropagation is performed based on the feature map of the last convolutional layer in the 3D segmentation network to obtain the contribution of each image voxel to the segmentation result of the brain hemorrhage region. Generate a response weight map based on the degree of contribution; The response weight graph is normalized to generate an explanatory graph.

[0013] Optionally, based on the image of the brain hemorrhage area, the initial brain hemorrhage volume is calculated using a preset dual-path estimation model, including: Using the first estimation path in the dual-path estimation model, target voxels with a probability greater than or equal to a preset probability threshold are statistically analyzed based on the brain hemorrhage region image in the hemorrhage mask. The initial cerebral hemorrhage volume of the first estimation path is calculated based on the target voxel and the unit voxel volume of the target voxel. Hemorrhage feature information is determined based on images of the brain hemorrhage region, including the maximum diameter, density peak, and texture directionality of the hemorrhage region. Based on the hemorrhage characteristics and the initial cerebral hemorrhage volume of the first estimation path, the initial cerebral hemorrhage volume of the second estimation path is obtained through the second estimation path in the dual-path estimation model.

[0014] Optionally, the 3D segmentation network is a dual-branch heterogeneous encoder structure, which includes two feature extraction branches: a structural path and a semantic path. The structural path adopts a convolutional structure based on 3D residual U-Net to extract low-level image features of the brain hemorrhage region in brain medical image data. The low-level image features include the density distribution, edge morphology and local texture of the brain hemorrhage region. The semantic path uses a multi-scale window attention module to perform global modeling of brain medical image data, which is used to extract the spatial context and global dependencies of the brain hemorrhage region; A cross-attention mechanism module is set between the structural path and the semantic path to combine the anatomical location of the lesion in the brain hemorrhage area image with the distribution features in the standard hemorrhage template constructed according to different brain hemorrhage stages, and to perform feature interaction fusion on the feature information extracted by the structural path and the semantic path.

[0015] Optionally, the 3D segmentation network includes a density-driven dynamic scaling module; The density-driven dynamic scaling module is used to adjust the size of the convolution window during the feature extraction process in the image segmentation process of the 3D segmentation network, based on the density changes and texture direction information of different lesion anatomical locations in the brain hemorrhage region image.

[0016] Optionally, brain medical image data can be input into a 3D segmentation network to segment the brain hemorrhage region, obtaining an image of the brain hemorrhage region, including: Brain medical image data is input into a 3D segmentation network to extract low-level image features of the brain hemorrhage area from the brain medical image data through the structural path of the 3D segmentation network; Furthermore, the spatial context and global dependencies of the brain hemorrhage region are extracted through the semantic path of the 3D segmentation network; Based on low-level image features, contextual relationships, and global dependencies, feature interaction and fusion are performed through the cross-attention mechanism module of a 3D segmentation network to output an image of the brain hemorrhage region.

[0017] Secondly, this application provides a device for measuring the volume of cerebral hemorrhage, comprising an acquisition module, a segmentation module, an identification module, a construction module, a generation module, an analysis module, an estimation module, and a determination module, wherein: The acquisition module is used to acquire the brain medical image data to be measured; The segmentation module is used to input brain medical image data into a 3D segmentation network to segment the brain hemorrhage area and obtain an image of the brain hemorrhage area; The identification module is used to identify the stage of brain hemorrhage based on the brain hemorrhage area image; A module is used to construct a standard hemorrhage template based on the stage of cerebral hemorrhage and the anatomical location of the lesion in the image of the cerebral hemorrhage region; the standard hemorrhage template is used to characterize the distribution characteristics of different stages of cerebral hemorrhage in the corresponding anatomical region of the lesion. The generation module is used to perform gradient response analysis based on the feature map of the last convolutional layer in the 3D segmentation network and generate an interpretation map. The interpretation map is used to characterize the contribution of each image voxel in the brain medical image data to the segmentation result of the brain hemorrhage region. The analysis module is used to perform spatial similarity analysis based on the standard bleed template and the interpretation diagram to obtain a consistency score between the standard bleed template and the interpretation diagram; The estimation module is used to calculate the initial brain hemorrhage volume based on the brain hemorrhage area image using a preset dual-path estimation model. The first estimation path of the dual-path estimation model estimates the initial brain hemorrhage volume using the voxel number statistics method, and the second estimation path estimates the initial brain hemorrhage volume using an image feature regression neural network. The determination module is used to adjust the fusion weights of the dual-path model based on the consistency score, and to determine the final target cerebral hemorrhage volume based on the adjusted fusion weights and the initial cerebral hemorrhage volume.

[0018] Thirdly, this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the brain hemorrhage volume measurement method as described above.

[0019] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the brain hemorrhage volume measurement method as described above.

[0020] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the brain hemorrhage volume measurement method as described above.

[0021] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: The method provided in this application introduces a three-dimensional segmentation network to accurately segment the brain hemorrhage region in brain medical image data. A standard hemorrhage template is constructed by combining the stage of brain hemorrhage and anatomical location. Then, an interpretation map is generated using gradient response analysis. Spatial similarity analysis is performed in combination with the standard template. The fusion weight of the dual-path estimation model is dynamically adjusted in the form of a score. This method can fully integrate the advantages of voxel statistics and image feature regression, two volume estimation paths, thereby effectively improving the accuracy of hemorrhage volume estimation. It solves the problems of existing technologies, such as reliance on manual work, large errors for irregular hemorrhage foci, weak ability to identify different hemorrhage stages, and estimation results that are too high or too low. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A schematic diagram illustrating the implementation process of a method for measuring the volume of cerebral hemorrhage provided in an embodiment of this application; Figure 2 This application provides a schematic diagram of the specific structure of a brain hemorrhage volume measuring device according to an embodiment of the present application; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0025] Example 1 To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] As will be known to those skilled in the art, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0027] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application.

[0028] To address the problems in existing technologies, such as large measurement errors and cumbersome operation of traditional geometric estimation methods for irregular hemorrhage foci, and inaccurate identification of hemorrhage morphology changes and CT value reduction in semi-automatic threshold segmentation methods, which result in poor volume measurement accuracy, reliance on manual labor, and low efficiency, making it difficult to meet the needs of rapid and accurate assessment in complex clinical scenarios, this application provides a method for measuring the volume of cerebral hemorrhage.

[0029] The execution subject of this method can be various types of computing devices, or it can be an application or app installed on the computing device. The computing device can be a user terminal such as a mobile phone, tablet computer, or smart wearable device, or it can be a server.

[0030] For ease of description, this application uses a server as the execution subject of the method in its embodiments to illustrate the method. Those skilled in the art will understand that this embodiment uses a server as an example to describe the method, which is merely an illustrative example and does not limit the scope of protection of the corresponding claims.

[0031] Specifically, the implementation flow of the method provided in this application embodiment is as follows: Figure 1 As shown, it includes the following steps: Step 11: Obtain the brain medical image data to be measured.

[0032] Brain medical image data can be understood as medical imaging data related to brain tissue structure or disease state used for medical diagnosis, treatment planning, or artificial intelligence analysis.

[0033] In this embodiment of the application, brain medical image data can include not only various imaging modalities such as computed tomography (CT), positron emission tomography (PET), magnetic resonance imaging (MRI), and cranial ultrasound, but also information such as three-dimensional spatial structure (voxels), density information (HU value) of CT images, and morphological features under specific pathologies (such as the morphology of cerebral hemorrhage).

[0034] Optionally, considering that brain medical image data often suffers from inconsistent slice thickness, which can easily lead to non-uniformity in the spatial structure of the image and thus affect subsequent analysis of brain hemorrhage volume, this embodiment of the application can resample the brain medical images in the brain medical image data and use trilinear interpolation to uniformly convert the brain medical images into three-dimensional voxel images with a spatial resolution of 1mm*1mm*1mm, thereby improving the consistency of the spatial structure and ensuring the accuracy of subsequent calculation results such as brain hemorrhage volume.

[0035] Optionally, to further ensure the accuracy of subsequent calculations of the brain hemorrhage volume, CT value linear normalization can be performed on the resampled brain medical images. Then, a skull dissection operation is performed on the CT value-linearly normalized brain medical images to preserve the soft tissue structures of the brain. Finally, the skull mask is subtracted from the complete brain medical image to obtain a ROI image region containing only the brain parenchyma. This ROI image region preserves key tissue structures such as hemorrhage, ventricles, and brain parenchyma, while excluding high-density bony interference.

[0036] Step 12: Input the brain medical image data into a three-dimensional segmentation network to segment the brain hemorrhage area and obtain an image of the brain hemorrhage area.

[0037] In this embodiment of the application, considering that related technologies often face problems such as insufficient image feature expression, difficulty in boundary determination, and poor adaptability to anatomical spatial structures when segmenting brain hemorrhage regions, especially in the clinical context of significant differences in the onset of brain hemorrhage, complex anatomical structures of brain hemorrhage, drastic density changes, wide distribution of lesion locations, and large differences in morphological characteristics, traditional segmentation methods cannot accurately identify the edge morphology of lesions in high-incidence anatomical regions, or easily miss hemorrhage signals in structurally complex regions, resulting in inaccurate images of brain hemorrhage regions, which in turn affects subsequent volume measurement and clinical diagnostic decisions.

[0038] To avoid the aforementioned problems, this application provides a three-dimensional segmentation network, which is a dual-branch heterogeneous encoder structure. This dual-branch heterogeneous encoder includes two feature extraction branches: a structural path and a semantic path. The structural path adopts a convolutional structure based on a three-dimensional residual U-Net to extract low-level image features of the brain hemorrhage region in brain medical image data. The low-level image features include the density distribution, edge morphology, and local texture of the brain hemorrhage region. The semantic path performs global modeling of the brain medical image data by integrating a multi-scale window attention module to extract the spatial context and global dependencies of the brain hemorrhage region.

[0039] In this embodiment, a cross-attention mechanism module is set between the structural path and the semantic path. This module is used to combine the anatomical location of the lesion in the brain hemorrhage region image with the distribution features in the standard hemorrhage template constructed according to different stages of brain hemorrhage. It performs feature interaction fusion on the low-level image features of the brain hemorrhage region extracted by the structural path and the spatial context relationship and global dependency of the brain hemorrhage region extracted by the semantic path.

[0040] In this embodiment, the cross-attention mechanism module combines the anatomical location of the lesion in the brain hemorrhage region image with the distribution features in the standard hemorrhage template constructed according to different stages of brain hemorrhage, and performs feature interaction fusion on the feature information extracted from the structural path and semantic path in the following way: (1) The cross-attention mechanism module receives low-level image features of the brain hemorrhage region extracted by the structural path, spatial context and global dependencies of the brain hemorrhage region extracted by the semantic path, and distribution feature information in the standard hemorrhage template constructed according to different brain hemorrhage stages.

[0041] (2) After receiving the above information, the cross-attention mechanism module uses the distribution feature information in the standard bleeding template as a weighting factor in spatial location and applies it to all channels of the semantic path to realize channel-by-channel spatial weighting operation.

[0042] The purpose of step (2) above is to enhance the feature response corresponding to the spatial location of the high-bleeding region in the semantic path, while suppressing the low-relevance region, so as to obtain the spatial context relationship and global dependency after the standard bleeding template is guided, and ensure that the spatial context relationship and global dependency are consistent with the prior distribution feature information in the standard bleeding template in the spatial dimension.

[0043] (3) Based on the spatial context and global dependency information of the weighted semantic path, and the low-level image features of the brain hemorrhage area extracted by the structural path, first calculate the similarity weight between each spatial location in the structural path and all locations in the semantic path to obtain the attention distribution from the structural path to the semantic path; then, calculate the similarity between each spatial location in the semantic path and all locations in the structural path in reverse to obtain the attention distribution from the semantic path to the structural path.

[0044] (4) Based on the attention distribution obtained in (3) above, the spatial context relationship and global dependency information of the semantic path are weighted and sampled and injected into the low-level image features of the structural path; at the same time, the fine-grained low-level features in the structural path are guided to be fused into the semantic path through the attention mechanism to form complementary and enhanced fusion features.

[0045] (5) Finally, the cross-attention mechanism module performs residual connection and feature mapping (including convolution, normalization and activation function processing) on ​​the fused features to output cross-fused features with uniform scale and semantic level, so as to improve the three-dimensional segmentation network's perception and discrimination accuracy of lesion sites and their stage changes.

[0046] Optionally, the cross-attention mechanism module in this application embodiment can be implemented by a cross-attention mechanism, a gated cross-attention mechanism, an attention-guided residual fusion mechanism, or a multi-scale cross-attention mechanism.

[0047] It should be noted that in this embodiment, a cross-attention mechanism module is set between the structural path and the semantic path. This is to combine the anatomical location of the lesion in the brain hemorrhage area image with the distribution characteristics in the standard hemorrhage template constructed according to different stages of brain hemorrhage, and to perform feature interaction fusion on the feature information extracted by the structural path and the semantic path. This is to improve the recognition accuracy and boundary determination ability of the three-dimensional segmentation network in high-incidence anatomical areas, thereby enhancing the integrity of the hemorrhage area segmentation and ensuring the accuracy of subsequent volume estimation.

[0048] Optionally, considering that in CT images of cerebral hemorrhage, especially in the acute phase or when multifocal lesions are present, the hemorrhage area usually appears as high density with certain well-defined boundaries. However, due to the density transition zone, partial volume effect, and artifact interference often existing between the hemorrhage area and the brain parenchyma, such medical images may experience problems during segmentation, such as blurred edges, difficulty in identifying small lesion areas, and misidentification of normal tissue as hemorrhage areas, thus affecting the accuracy of image segmentation. To solve this problem, in the embodiments of this application, the convolutional structure of the three-dimensional residual U-Net may include an encoder path, a decoder path, four skip connection paths, a feature fusion module, and an artifact suppression module.

[0049] The encoder path comprises four downsampling levels, with each downsampling module numbered sequentially from Encoder 1 to Encoder 4. Each encoder level includes two cascaded boundary-aware residual blocks, an artifact suppression module, and a 2×2×2 three-dimensional max-pooling (MaxPooling3D) downsampling operation. The artifact suppression module is used to reduce artifact effects after the output of the current level's residual block; the three-dimensional max-pooling (MaxPooling3D) downsampling operation is used to halve the spatial resolution of the feature map.

[0050] In this embodiment, after brain medical image data is input into the encoder path of the convolutional structure of the three-dimensional residual U-Net, four sets of multi-scale feature maps can be extracted through step-by-step encoding.

[0051] The decoder path includes four upsampling levels, with each level's decoding modules numbered sequentially from decoder 4 to decoder 1, symmetrical to the encoder levels. Each decoding module includes a 3D deconvolution upsampling module (with a 2×2×2 kernel and a stride of 2), a skip connection input channel, a feature fusion module, two cascaded 3D convolutional layers (with a 3×3×3 kernel), and a ReLU activation function.

[0052] A 3D deconvolutional upsampling module is used to restore the spatial dimensions to the size of the previous layer; a skip connection input channel receives the encoder feature map corresponding to the current level; a density-driven dynamic scaling module is used to adjust the convolution window size during feature extraction based on density changes and texture direction information at different anatomical locations of lesions in the brain hemorrhage region image during image segmentation of the 3D segmentation network; two cascaded 3D convolutional layers and a ReLU activation function are used for feature reconstruction and enhancement in the decoding path.

[0053] In this embodiment, the output of the last layer of the decoder path, decoder 1, is transformed by a 1×1×1 convolution to reduce the number of channels to 1, and then connected to a Sigmoid function to output an image of the brain hemorrhage region.

[0054] In this embodiment, the four skip connection paths can be numbered from skip connection path 1 to skip connection path 4. Skip connection path 1 connects to encoder 1 and decoder 1; skip connection path 2 connects to encoder 2 and decoder 2; skip connection path 3 connects to encoder 3 and decoder 3; and skip connection path 4 connects to encoder 4 and decoder 4. The features in each skip connection are modulated by a density-driven dynamic scaling module before fusion and then concatenated with the current layer of the decoder. The concatenated feature map is then subjected to convolutional fusion processing to form the input of the current layer decoder.

[0055] The boundary-aware residual block refers to a module that adds an edge modulation branch to the standard residual structure to enhance the model's perception capability in edge regions. Optionally, in this embodiment, the boundary-aware residual block includes a main branch: two cascaded 3×3×3 three-dimensional convolutional layers, each followed by batch normalization and ReLU activation; an auxiliary edge branch: applying a three-dimensional Sobel operator to the input feature map to extract edge images, and generating an edge weight map after Sigmoid activation; the output of the main branch and the edge weight map are modulated element-wise and added to the input residual to form the final output.

[0056] The artifact suppression module is used to suppress directional artifacts in CT images caused by equipment or tissue structures. In this embodiment, the artifact suppression module may include three separable directional convolutional branches with kernels of 1×3×3, 3×1×3, and 3×3×1, respectively. The artifact suppression module can perform mean pooling on each directional feature to generate a directional response map; then, the directional response map is activated by Sigmoid to form suppression weights; finally, the suppression weights are used to perform directional weighted suppression processing on the input feature map to reduce artifact responses.

[0057] Optionally, the density-driven dynamic scaling module can use a larger convolution window for regions with large density changes and a smaller convolution window for regions with obvious structural details but small density changes during the image segmentation process of the 3D segmentation network. This enhances the adaptability of the 3D segmentation network to the morphological features of bleeding regions, especially in the identification of small bleeding regions, thereby improving the accuracy of boundary segmentation.

[0058] Optionally, the density-driven dynamic scaling module may include multi-scale dilated convolution branches with kernel sizes of 3×3×3 and dilation rates of 1, 2, and 3, respectively; and a Softmax normalization operation to inversely weight the density variances at each scale to improve the response in density anomaly regions.

[0059] The convolutional structure of the three-dimensional residual U-Net provided in this application embodiment has the following advantages. On the one hand, since a boundary-aware residual block structure is introduced into the encoder path, and a three-dimensional Sobel gradient extraction channel is embedded in the residual unit, the edge response map is used as a feature modulation factor to realize the dynamic enhancement of edge position features. This can strengthen the gradient expression capability of the model in the edge blur region, thereby improving the localization accuracy of the bleeding boundary region and avoiding the problem of inaccurate bleeding region recognition in the gray-scale continuous transition region due to the unclear boundary features.

[0060] On the other hand, this application incorporates a density-driven dynamic scale adjustment module into the skip connection path to introduce prior information about density fluctuations during feature alignment. This module constructs a multi-scale density attention mechanism by performing local density variance analysis on encoder features at different scales. This guides the model to pay more attention to anomalous density regions during feature fusion, enhancing its density boundary modeling capability and effectively mitigating misclassification and omissions caused by continuous changes in tissue density.

[0061] In addition, an artifact suppression module is introduced at the encoder output. The response of the feature map in the axial, sagittal and coronal directions is extracted by spatially separable convolution. The normalized directional response map is combined to dynamically suppress the features, thereby avoiding the systematic artifact interference problem commonly seen in brain CT.

[0062] Finally, the four-level skip connection mechanism, by density-guided fusion of high-resolution spatial features in the shallow layer and semantic information in the deep layer, can prevent the structural information of small lesions from being erased by the downsampling compression process. At the same time, the decoder restores regional details through multi-channel reconstruction convolution, thereby avoiding the problem that small-volume hemorrhage lesions are easily missed in image segmentation.

[0063] The multi-scale window attention module for semantic paths can include a multi-scale window partitioning unit, a window self-attention sub-module, a cross-window interaction module, and a scale fusion unit. The multi-scale window partitioning unit divides the input brain medical image data into windows according to a preset scale, forming several overlapping or non-overlapping spatial sub-blocks. Window scales can include three categories: small scale (e.g., 4×4×4 voxels), medium scale (e.g., 8×8×8 voxels), and large scale (e.g., 16×16×16 voxels), used to extract structural and semantic information within different spatial ranges respectively.

[0064] Each partitioned window is input into its corresponding window self-attention submodule. A local multi-head attention mechanism is employed to calculate the correlation of spatial features within the window and extract local contextual relationships. The attention mechanism integrates relative position encoding in three-dimensional space to enhance the model's ability to recognize spatial relationships of brain structures.

[0065] The cross-window interaction module is mainly for associating semantic relationships between windows. It includes channels for feature interaction between adjacent windows and uses a gating mechanism to dynamically control the flow of information between windows.

[0066] In this embodiment, the brain medical image data input to the semantic path is first segmented into multiple sub-regions of different sizes by a multi-scale window segmentation unit, corresponding to small, medium, and large scales respectively. Each window at each scale independently executes a window self-attention mechanism within its own scope, extracting the relative relationships and semantic relevance between voxels within the local spatial range. During the window attention calculation process, relative position encoding is introduced to preserve three-dimensional spatial location information, thereby enhancing the model's ability to capture the geometric features of the brain hemorrhage region relative to surrounding tissues. This mechanism significantly improves the accuracy of region identification, especially when the density of the brain hemorrhage structure changes drastically or its location is complex (e.g., near the ventricles, insula, or brainstem).

[0067] Considering the potential for boundary continuity breaks between windows, a cross-window interaction mechanism is further introduced. By establishing a gated information flow channel between adjacent windows, the 3D segmentation network can dynamically decide whether to share feature information based on semantic similarity and the continuity of boundary structure. This enhances the modeling ability of lesion edge regions and reduces the risk of misjudgment caused by boundary ambiguity or structural discontinuity.

[0068] Finally, to ensure that the 3D segmentation network can integrate feature information from different scales, a scale fusion unit can be used to perform channel alignment and spatial fusion on feature maps extracted from multiple scales, thereby forming a multi-scale feature representation that combines local details and global structure, so as to improve the adaptability of the 3D segmentation network to brain hemorrhage regions of different sizes, shapes and distribution locations.

[0069] In this embodiment of the application, when performing step 12, brain medical image data can be input into a three-dimensional segmentation network to extract low-level image features of the brain hemorrhage region in the brain medical image data through the structural path of the three-dimensional segmentation network. The low-level image features include feature information such as density distribution, edge morphology, and local texture of the brain hemorrhage region. In addition, the spatial context and global dependencies of the brain hemorrhage region are extracted through the semantic path of the three-dimensional segmentation network. Based on the low-level image features, context and global dependencies, feature interaction and fusion are performed through the cross-attention mechanism module of the three-dimensional segmentation network to output the brain hemorrhage region image.

[0070] In an optional implementation, a feature filtering module can be included in the skip connections between the encoder and decoder of the 3D segmentation network to restrict the transmission of intermediate feature streams that do not conform to anatomical structures or are affected by artifacts. Optionally, this feature filtering module can determine the spatial location of regions with a high probability of hemorrhage in the brain hemorrhage region image output by the 3D segmentation network, combined with the brain parenchyma range identified in the brain medical image data. If a high-probability region does not fall within the brain parenchyma region, or its structural features deviate significantly from the preset anatomical distribution features, the intermediate features in the corresponding skip connection channel will be attenuated or masked to reduce the impact of unreliable information on the final segmentation result.

[0071] In one alternative implementation, the 3D segmentation network can employ a combination mechanism consisting of multiple loss functions during training to improve the prediction accuracy of segmented contours and boundary regions.

[0072] In one optional implementation, the combined mechanism may include: a main loss function based on the degree of regional overlap to optimize the consistency of the overall segmented regions; introducing a boundary-related penalty term in regions where there is a significant deviation in the predicted boundaries of the 3D segmentation network to improve the prediction effect of local edges; and further constraining the consistency of the predicted brain hemorrhage region image and the real brain hemorrhage region image in terms of overall distribution features by calculating the difference between the predicted brain hemorrhage region image and the training label in image grayscale or density distribution.

[0073] In one optional implementation, the three-dimensional segmentation network may further include a structure verification and correction mechanism based on the stage information of cerebral hemorrhage, used to judge and adjust whether the structure of the hemorrhage area is reasonable after segmentation; when the predicted cerebral hemorrhage area has a large difference in morphology or density distribution from the standard hemorrhage template, the predicted cerebral hemorrhage area image is subjected to range reduction or density reweighting processing according to the structural rules corresponding to the stage of cerebral hemorrhage, so as to improve the anatomical consistency of the final segmentation result.

[0074] Step 13: Identify the stage of brain hemorrhage based on the brain hemorrhage area image.

[0075] Among them, the stages of cerebral hemorrhage refer to the process from the onset of cerebral hemorrhage to its gradual absorption and repair, which are divided into several stages (stages) according to the time progression and / or brain medical imaging characteristics. Each stage reflects a different pathophysiological state of the hemorrhage.

[0076] Correspondingly, in the embodiments of this application, the cerebral hemorrhage stage is used to describe the progression stage of cerebral hemorrhage in the cerebral hemorrhage area image.

[0077] Optionally, in this embodiment of the application, cerebral hemorrhage can be classified into different stages such as hyperacute, acute, subacute, chronic, and late absorption stages according to the characteristics of brain medical imaging.

[0078] Optionally, the brain medical imaging features mentioned here include the mean CT value, standard deviation, gray-level histogram features, GLCM texture entropy, contrast, uniformity, etc. of the hemorrhage area in the brain hemorrhage area image.

[0079] Among them, GLCM texture entropy is used to describe the degree of uncertainty or disorder in the gray-level distribution of the hemorrhage area in the image of the brain hemorrhage region. Contrast is used to describe the magnitude of gray-level value changes in the hemorrhage region in the image of the brain hemorrhage region, that is, the degree of difference between dark and bright areas in the hemorrhage region. Uniformity is used to describe the degree of gray-level diagonal concentration in the hemorrhage region in the image of the brain hemorrhage region; a higher value indicates a more consistent gray-level distribution in the image.

[0080] In the hyperacute phase, the structure is a homogeneous or slightly high-density, indistinct-boundary, iso-dense structure, usually without obvious central or ring-like features; in the acute phase, the structure is a centrally symmetrical high-density structure; in the subacute phase, the structure is a ring-like structure with a high-density center and low-density edges; in the chronic phase, the structure is a diffuse structure with low overall density and blurred edges; in the late absorption phase, the structure is a low-density, cavity-like change, with the edges connected to or ablated by the cerebrospinal fluid, and signs of ventricular displacement and atrophy can be seen.

[0081] Optionally, if classified according to the time progression, the period from 0 to 6 hours after bleeding can be classified as the hyperacute phase; the period from 6 hours to 3 days after bleeding can be classified as the acute phase; the period from 3 days to 2 weeks after bleeding can be classified as the subacute phase; the period from 2 weeks to several months after bleeding can be classified as the chronic phase; and the period from several months or more after bleeding can be classified as the late absorption phase.

[0082] It should be noted that numerous factors influence the accuracy of cerebral hemorrhage volume measurement during CT image analysis, including changes in hemorrhage density, degree of boundary blurring, and anatomical location. The stage of cerebral hemorrhage (e.g., acute, subacute, chronic) directly determines the density distribution, edge clarity, and tissue structure complexity of the CT image, thus becoming a key factor affecting the differences in CT image representation and measurement stability. Furthermore, different hemorrhage stages exhibit significantly different appearances in CT images; for example, the acute stage typically has high density and clear boundaries, while the chronic stage is characterized by low density, diffusion, and structural blurring. Therefore, to ensure the accuracy and clinical reliability of cerebral hemorrhage volume measurement, this embodiment considers first identifying the stage of cerebral hemorrhage in the cerebral hemorrhage region image based on the cerebral hemorrhage region image during the measurement process.

[0083] Furthermore, in this embodiment of the application, the purpose of identifying the stage of brain hemorrhage in the image of the brain hemorrhage area is not only to label and classify the lesions, but mainly to use the identification result of the stage of brain hemorrhage to drive the parameter adjustment and structural adaptation in subsequent steps 14 to 16. The relevant steps will be explained in detail later.

[0084] Step 14: Construct a standard hemorrhage template based on the stage of cerebral hemorrhage and the anatomical location of the lesion in the cerebral hemorrhage area image.

[0085] The standard hemorrhage template is used to characterize the distribution characteristics of different stages of cerebral hemorrhage in the corresponding anatomical regions of the lesion.

[0086] Optionally, in this embodiment of the application, when constructing a standard hemorrhage template, the standard CT hemorrhage density distribution features corresponding to the stage of cerebral hemorrhage can be obtained; the anatomical location of the lesion in the cerebral hemorrhage area image can be identified, and the anatomical location of the lesion is used to characterize the spatial location of the lesion in the midline of the brain, basal ganglia, lobes, cerebellum and / or brainstem; a target three-dimensional reference image corresponding to the stage of cerebral hemorrhage and the anatomical location of the lesion can be extracted from a preset template library; the preset template library contains a set of three-dimensional reference images constructed by statistically analyzing multiple historical cerebral hemorrhage image samples with cerebral hemorrhage stage labels and anatomical region annotations; the standard CT hemorrhage density distribution features are fused and registered with the target three-dimensional reference image to obtain the standard hemorrhage template.

[0087] Step 15: Perform gradient response analysis based on the feature map of the last convolutional layer in the 3D segmentation network to generate an interpretation map.

[0088] The interpretation map is used to characterize the contribution of each image voxel in the brain medical image data to the segmentation results of the brain hemorrhage region.

[0089] Optionally, in this embodiment of the application, when generating the interpretation map, gradient backpropagation can be performed based on the feature map of the last convolutional layer in the 3D segmentation network to obtain the contribution of each image voxel to the segmentation result of the brain hemorrhage region; a response weight map is generated according to the contribution; and the interpretation map is generated by normalization processing according to the response weight map.

[0090] Step 16: Perform spatial similarity analysis based on the standard bleed template and the interpretation diagram to obtain the consistency score between the standard bleed template and the interpretation diagram.

[0091] In this embodiment, the consistency score between the standard hemorrhage template and the interpretation map can be understood as follows: under a unified reference space, the standard hemorrhage template and the interpretation map are divided according to brain anatomical regions. Within each anatomical region, similarity analysis and quantitative comparison are performed on the two from multiple dimensions such as voxel distribution pattern, shape and structural features, spatial position and orientation, voxel importance ranking, and adaptive region overlap, to obtain the local consistency score of the region. Then, based on the preset anatomical region weights, the local scores of each region are weighted and fused to form a numerical index used to characterize the degree of conformity between the interpretation map and the standard hemorrhage template in terms of spatial distribution and semantic features. The numerical range is limited to 0 to 100, with a higher value indicating a higher degree of conformity.

[0092] In this embodiment of the application, the following implementation method can be used when calculating the consistency score between the standard bleeding template and the interpretation diagram: (1) Divide the standard hemorrhage template and interpretation diagram according to the preset brain anatomical regions, such as the midline region, basal ganglia region, lobular region, cerebellum region, and brainstem region. Then, assign weights to each anatomical region. Optionally, when assigning weights to each anatomical region, the size of each anatomical region and the intensity of the stage characteristics of the standard hemorrhage template in that region can be considered to ensure that the score reflects both anatomical balance and highlights regions with significant stage characteristics.

[0093] (2) Within each anatomical region, the standard hemorrhage template and the interpretation diagram are compared as follows: Distribution similarity: Compare whether the voxel intensity distribution patterns of the two are consistent and whether the high response regions are concentrated in the same location.

[0094] Shape similarity: Under multiple intensity thresholds, two Figure 2 Value-based comparison compares the similarity of structural features such as regional morphology and connectivity.

[0095] Spatial location and orientation similarity: Calculate the difference in the centroid positions of the two images within the region, as well as the consistency of the principal orientation vectors (obtained from principal component analysis of voxel distribution).

[0096] Then, the three similarity results are fused according to a preset ratio to obtain the spatial similarity score of the anatomical region.

[0097] Optionally, the preset ratio in this embodiment can be, for example, 1 / 3, 1 / 3, and 1 / 3. Alternatively, the preset ratio can be set according to the actual situation, and there is no limitation here.

[0098] (3) Within each anatomical region, voxels are sorted by intensity from high to low, and the consistency of the two images in voxel importance order is evaluated. Furthermore, based on the region intensity distribution, an adaptive binarization threshold is determined, and the two images are calculated. Figure 2The degree of overlap in the value regions. Finally, the degree of consistency between the two graphs in the order of voxel importance and the two... Figure 2 The semantic consistency score of the anatomical region is obtained by weighted fusion of the overlap of the value regions.

[0099] (4) The spatial similarity score and semantic consistency score of each anatomical region are weighted and summed to obtain the final consistency score of the region.

[0100] (5) Finally, according to the anatomical region weights determined in step (1) above, the final consistency scores of each region are weighted and summed to obtain the global consistency score. The score is then mapped to the interval from 0 to 100 as a quantitative indicator of the overall consistency between the standard bleeding template and the interpretation diagram, that is, the consistency score between the standard bleeding template and the interpretation diagram.

[0101] Step 17: Calculate the initial brain hemorrhage volume based on the brain hemorrhage area image using a preset dual-path estimation model.

[0102] The dual-path estimation model uses a voxel count method to estimate the volume of cerebral hemorrhage in the first estimation path and an image feature regression neural network to estimate the volume of cerebral hemorrhage in the second estimation path.

[0103] Optionally, in this embodiment of the application, when calculating the initial brain hemorrhage volume using a preset dual-path estimation model, the first estimation path in the dual-path estimation model can be used to statistically analyze target voxels in the hemorrhage mask with a probability greater than or equal to a preset probability threshold based on the image of the brain hemorrhage region; the initial brain hemorrhage volume of the first estimation path can be calculated based on the target voxels and their unit voxel volumes.

[0104] Furthermore, based on the image of the brain hemorrhage region, the hemorrhage feature information is determined, including the maximum diameter, density peak, and texture directionality of the hemorrhage region; based on the hemorrhage feature information and the initial brain hemorrhage volume of the first estimation path, the initial brain hemorrhage volume of the second estimation path is obtained through the second estimation path in the dual-path estimation model.

[0105] Optionally, in the embodiments of this application, when the second estimation path uses an image feature regression neural network to estimate the brain hemorrhage volume, it can determine the feature vector based on the initial brain hemorrhage volume of the first estimation path and the hemorrhage feature information of the brain hemorrhage region image. The hemorrhage feature information includes the maximum diameter, density peak, and texture directionality of the hemorrhage region. Then, the feature vector is input into the image feature regression neural network, such as SVR or lightweight MLP, to predict the initial brain hemorrhage volume of the second estimation path.

[0106] Step 18: Adjust the fusion weights of the dual-path model based on the consistency score, and determine the final target cerebral hemorrhage volume based on the adjusted fusion weights and the initial cerebral hemorrhage volume.

[0107] In this embodiment of the application, before adjusting the fusion weights of the dual-path model based on the consistency score, the initial weights of the first and second estimated paths in the dual-path model can be determined as follows: (1) Obtain historical brain medical image data and the actual brain hemorrhage volume corresponding to the historical brain medical image data; (2) Based on historical brain medical image data, calculate the consistency score of historical brain medical image data, the initial brain hemorrhage volume of the first estimation path and the initial brain hemorrhage volume of the second estimation path according to steps 11 to 17 respectively. (3) The consistency score of historical brain medical image data is segmented according to the preset score segment threshold to obtain multiple consistency score intervals, such as 0–10, 10–20, 20–30, and so on.

[0108] (4) For each consistency score interval, the error level (e.g., deviation size) and stability (e.g., variance, fluctuation range) of the initial brain hemorrhage volume of the first estimation path, the initial brain hemorrhage volume of the second estimation path, and the actual brain hemorrhage volume are used.

[0109] (5) Based on the error level and stability, determine whether the initial brain hemorrhage volume of the first estimation path or the second estimation path in each consistency score interval is closer to the true value and more stable.

[0110] (6) Finally, based on the determination results of (5), the basic weights of the first estimation path and the second estimation path in each consistency score interval are assigned respectively.

[0111] For example, suppose the consistency score intervals are: Interval 1: [0, 10]; Interval 2: [10, 20]; Interval 3: [20, 30]; where Interval 2 contains 50 consistency score samples of historical brain medical image data, and among these samples, the first estimation path has a smaller average error but slightly larger volatility (some samples have large errors, some have small ones). The second estimation path has a slightly larger average error but smaller volatility (errors are more concentrated). In this case, we can consider that although the first estimation path has a smaller average error, its overall stability is not as good as the second estimation path due to its larger volatility. The second estimation path is slightly less accurate but more stable. Therefore, when assigning the base weights to the first and second estimation paths in Interval 2, we can consider that the two estimation paths are roughly equal in strength in this interval, without an absolute overwhelming advantage. Therefore, to avoid an excessively skewed weight distribution due to a slight difference, the base weight of the first estimation path can be set to 0.55, and the base weight of the second estimation path can be set to 0.45. This reflects a slight bias towards the first path while retaining a larger proportion of the second path, making good use of its stability.

[0112] For example, suppose there are very few consistency score samples of historical brain medical image data in interval 1, such as only 8. In this case, even if the first path is obviously more accurate, it will not be given a weight of 0.8 or 0.9 directly. Instead, it will be reduced, for example, set as follows: the basic weight of the first estimation path is 0.65; the basic weight of the second estimation path is 0.35.

[0113] In this embodiment of the application, the fusion weights of the dual-path model can be adjusted based on the consistency score in the following manner: If the local consistency score of a key anatomical region is significantly higher than that of the global consistency score, the base weight of the first estimation path is increased; or, if the consistency score of a non-key anatomical region is low and the characteristics of the second estimation path are stable, the base weight of the second estimation path is increased.

[0114] Key anatomical regions include the basal ganglia and brainstem. Non-key anatomical regions include other anatomical areas of the brain besides the basal ganglia and brainstem.

[0115] In this embodiment of the application, the target cerebral hemorrhage volume can be calculated according to the following calculation method:

[0116] in, Indicates the volume of the target brain hemorrhage; This represents the initial cerebral hemorrhage volume in the first estimation path; This represents the initial cerebral hemorrhage volume in the second estimation path; This represents the fusion weight of the first estimated path in the two-path model; This represents the fusion weight of the second estimated path in the two-path model.

[0117] It should be noted that the solution provided in this application is not simply to independently estimate multiple cerebral hemorrhage volumes using various methods and then sum them up using fixed or empirical weights. This is because such a method has significant shortcomings in practical applications: the calculated volumes are disconnected from the features of the case images, the weights cannot be adaptively adjusted for different hemorrhage stages, lesion locations, or imaging features, leading to deviations in the fusion results under complex or atypical lesions, and lacking explanatory support for the model's judgment criteria.

[0118] To overcome the technical problems associated with the simple approach of independently estimating multiple brain hemorrhage volumes using various methods and then summing them using fixed or empirical weights, this application introduces a standard hemorrhage template based on the stage of brain hemorrhage and the anatomical location of the lesion within the dual-path estimation model. Spatial similarity analysis is then performed using the gradient response interpretation map of a three-dimensional segmentation network to obtain a consistency score closely related to the current case. This score directly influences the generation of the dual-path fusion weights, tightly coupling the weight allocation process with the specific lesion distribution characteristics and the clinical rationality of the model's focus area, forming an interpretability-driven adaptive fusion mechanism. Therefore, this application can adaptively reduce the bias of a single path under different hemorrhage stages and anatomical locations, significantly improving the accuracy, generalization ability, and clinical reliability of volume estimation.

[0119] The method provided in this application introduces a three-dimensional segmentation network to accurately segment the brain hemorrhage region in brain medical image data. A standard hemorrhage template is constructed by combining the stage of brain hemorrhage and anatomical location. Then, an interpretation map is generated using gradient response analysis. Spatial similarity analysis is performed in combination with the standard template. The fusion weight of the dual-path estimation model is dynamically adjusted in the form of a score. This method can fully integrate the advantages of voxel statistics and image feature regression, which are two volume estimation paths. This can effectively improve the accuracy of hemorrhage volume estimation and solve the problems of existing technologies, such as reliance on manual work, large errors for irregular hemorrhage foci, weak ability to identify different hemorrhage stages, and estimation results that are too high or too low.

[0120] Example 2 To address the problems in existing technologies, such as large measurement errors and cumbersome operation of traditional geometric estimation methods for irregular hemorrhage foci, and inaccurate identification of hemorrhage morphology changes and CT value reduction in semi-automatic threshold segmentation methods, which result in poor volume measurement accuracy, reliance on manual labor, and low efficiency, making it difficult to meet the needs of rapid and accurate assessment in complex clinical scenarios, this application provides a brain hemorrhage volume measurement device. A schematic diagram of the specific structure of the device is shown below. Figure 2As shown, the system includes an acquisition module 21, a segmentation module 22, an identification module 23, a construction module 24, a generation module 25, an analysis module 26, an estimation module 27, and a determination module 28. The functions of each module are as follows: Acquisition module 21 is used to acquire brain medical image data to be measured; The segmentation module 22 is used to input brain medical image data into a three-dimensional segmentation network to segment the brain hemorrhage region and obtain an image of the brain hemorrhage region; The identification module 23 is used to identify the stage of brain hemorrhage in the brain hemorrhage area image based on the brain hemorrhage area image; Module 24 is used to construct a standard hemorrhage template based on the stage of cerebral hemorrhage and the anatomical location of the lesion in the image of the cerebral hemorrhage region; the standard hemorrhage template is used to characterize the distribution characteristics of different stages of cerebral hemorrhage in the corresponding anatomical region of the lesion. The generation module 25 is used to perform gradient response analysis based on the feature map of the last convolutional layer in the 3D segmentation network and generate an interpretation map. The interpretation map is used to characterize the contribution of each image voxel in the brain medical image data to the segmentation result of the brain hemorrhage region. Analysis module 25 is used to perform spatial similarity analysis based on the standard bleed template and the interpretation diagram to obtain a consistency score between the standard bleed template and the interpretation diagram; The estimation module 27 is used to calculate the initial brain hemorrhage volume based on the brain hemorrhage area image using a preset dual-path estimation model. The first estimation path of the dual-path estimation model estimates the brain hemorrhage volume using the voxel number statistics method, and the second estimation path estimates the brain hemorrhage volume using an image feature regression neural network. Module 28 is used to adjust the fusion weights of the dual-path model based on the consistency score, and to determine the final target cerebral hemorrhage volume based on the adjusted fusion weights and the initial cerebral hemorrhage volume.

[0121] Optional, build module 24, for: Obtain the standard CT hemorrhage density distribution characteristics corresponding to the different stages of cerebral hemorrhage; Identify the anatomical location of lesions in images of brain hemorrhage areas. The anatomical location of lesions is used to characterize the spatial location of lesions in the midline of the brain, basal ganglia, lobes, cerebellum and / or brainstem. Extract target 3D reference images corresponding to the stage and anatomical location of the cerebral hemorrhage from a preset template library; the preset template library contains a set of 3D reference images constructed by statistically analyzing multiple historical cerebral hemorrhage image samples with cerebral hemorrhage stage labels and anatomical region annotations; The standard CT hemorrhage density distribution features are fused and registered with the target three-dimensional reference image to obtain a standard hemorrhage template.

[0122] Optionally, module 25 is generated for: Gradient backpropagation is performed based on the feature map of the last convolutional layer in the 3D segmentation network to obtain the contribution of each image voxel to the segmentation result of the brain hemorrhage region. Generate a response weight map based on the degree of contribution; The response weight graph is normalized to generate an explanatory graph.

[0123] Optional, estimation module 27, used for: Using the first estimation path in the dual-path estimation model, target voxels with a probability greater than or equal to a preset probability threshold are statistically analyzed based on the brain hemorrhage region image in the hemorrhage mask. The initial cerebral hemorrhage volume of the first estimation path is calculated based on the target voxel and the unit voxel volume of the target voxel. Hemorrhage feature information is determined based on images of the brain hemorrhage region, including the maximum diameter, density peak, and texture directionality of the hemorrhage region. Based on the hemorrhage characteristics and the initial cerebral hemorrhage volume of the first estimation path, the initial cerebral hemorrhage volume of the second estimation path is obtained through the second estimation path in the dual-path estimation model.

[0124] Optionally, the 3D segmentation network is a dual-branch heterogeneous encoder structure, which includes two feature extraction branches: a structural path and a semantic path. The structural path adopts a convolutional structure based on 3D residual U-Net to extract low-level image features of the brain hemorrhage region in brain medical image data. The low-level image features include the density distribution, edge morphology and local texture of the brain hemorrhage region. The semantic path uses a multi-scale window attention module to perform global modeling of brain medical image data, which is used to extract the spatial context and global dependencies of the brain hemorrhage region; A cross-attention mechanism module is set between the structural path and the semantic path to combine the anatomical location of the lesion in the brain hemorrhage area image with the distribution features in the standard hemorrhage template constructed according to different brain hemorrhage stages, and to perform feature interaction fusion on the feature information extracted by the structural path and the semantic path.

[0125] Optionally, the 3D segmentation network includes a density-driven dynamic scaling module; The density-driven dynamic scaling module is used to adjust the size of the convolution window during the feature extraction process in the image segmentation process of the 3D segmentation network, based on the density changes and texture direction information of different lesion anatomical locations in the brain hemorrhage region image.

[0126] Optionally, segmentation module 22 is used for: Brain medical image data is input into a 3D segmentation network to extract low-level image features of the brain hemorrhage area from the brain medical image data through the structural path of the 3D segmentation network; Furthermore, the spatial context and global dependencies of the brain hemorrhage region are extracted through the semantic path of the 3D segmentation network; Based on low-level image features, contextual relationships, and global dependencies, feature interaction and fusion are performed through the cross-attention mechanism module of a 3D segmentation network to output an image of the brain hemorrhage region.

[0127] The device provided in this application introduces a three-dimensional segmentation network to accurately segment the brain hemorrhage region in brain medical image data. It then constructs a standard hemorrhage template by combining the stage of brain hemorrhage and anatomical location. Gradient response analysis is then used to generate an interpretation map. Spatial similarity analysis is performed using the standard template, and the fusion weight of the dual-path estimation model is dynamically adjusted in the form of a score. This fully integrates the advantages of both voxel statistics and image feature regression for volume estimation, thereby effectively improving the accuracy of hemorrhage volume estimation. It solves the problems of existing technologies, such as reliance on manual labor, large errors for irregular hemorrhage foci, weak ability to identify different hemorrhage stages, and estimation results that are too high or too low.

[0128] Example 3 Figure 3 To illustrate the hardware structure of an electronic device according to various embodiments of this application, the electronic device may include a processor 301 and a memory 302 storing computer program instructions. Specifically, the processor 301 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of this application.

[0129] Memory 302 may include mass storage for data or instructions. For example, and not limitingly, memory 302 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 302 may include removable or non-removable (or fixed) media. Where appropriate, memory 302 may be internal or external to an electronic device. In a particular embodiment, memory 302 may be a non-volatile solid-state memory.

[0130] In one embodiment, memory 302 may be read-only memory (ROM). In one embodiment, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0131] The processor 301 reads and executes computer program instructions stored in the memory 302 to implement any of the brain hemorrhage volume measurement methods in the above embodiments.

[0132] In one example, the electronic device may also include a communication interface 303 and a bus 310. For example, Figure 3 As shown, the processor 301, memory 302, and communication interface 303 are connected through bus 310 and complete communication with each other.

[0133] The communication interface 303 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0134] Bus 310 includes hardware, software, or both, that couples components of an electronic device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 310 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0135] Furthermore, in conjunction with the brain hemorrhage volume measurement method in the above embodiments, this application embodiment can provide a computer-readable storage medium for implementation. This computer-readable storage medium stores computer program instructions; when executed by a processor, these computer program instructions implement any of the brain hemorrhage volume measurement methods in the above embodiments.

[0136] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0137] The above description is merely a specific implementation example of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0138] Secondly, those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0139] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0140] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0141] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0142] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0143] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0144] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0145] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0146] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for measuring the volume of cerebral hemorrhage, characterized in that, include: Acquire medical image data of the brain to be measured; The brain medical image data is input into a three-dimensional segmentation network to segment the brain hemorrhage region, thereby obtaining an image of the brain hemorrhage region; Identify the stage of brain hemorrhage in the brain hemorrhage area image based on the brain hemorrhage area image; Based on the stage of cerebral hemorrhage and the anatomical location of the lesion in the image of the cerebral hemorrhage region, a standard hemorrhage template is constructed; the standard hemorrhage template is used to characterize the distribution characteristics of different stages of cerebral hemorrhage in the corresponding anatomical region of the lesion. Gradient response analysis is performed based on the feature map of the last convolutional layer in the three-dimensional segmentation network to generate an interpretation map. The interpretation map is used to characterize the contribution of each image voxel in the brain medical image data to the segmentation result of the brain hemorrhage region. Spatial similarity analysis is performed based on the standard bleed template and the interpretation diagram to obtain a consistency score between the standard bleed template and the interpretation diagram; Based on the image of the brain hemorrhage area, the initial brain hemorrhage volume is calculated using a preset dual-path estimation model. The first estimation path of the dual-path estimation model estimates the initial brain hemorrhage volume using the voxel number statistics method, and the second estimation path estimates the initial brain hemorrhage volume using an image feature regression neural network. The fusion weights of the dual-path model are adjusted based on the consistency score, and the final target cerebral hemorrhage volume is determined based on the adjusted fusion weights and the initial cerebral hemorrhage volume.

2. The method as described in claim 1, characterized in that, The standard hemorrhage template is constructed based on the stage of the cerebral hemorrhage and the anatomical location of the lesion in the image of the cerebral hemorrhage region, including: Obtain the standard CT hemorrhage density distribution characteristics corresponding to the stage of the cerebral hemorrhage; Identify the anatomical location of the lesion in the image of the brain hemorrhage region, the anatomical location of the lesion being used to characterize the spatial location of the lesion in the midline of the brain, basal ganglia, lobes, cerebellum and / or brainstem; Extract target three-dimensional reference images corresponding to the stage of cerebral hemorrhage and the anatomical location of the lesion from a preset template library; the preset template library contains a set of three-dimensional reference images constructed by statistically analyzing multiple historical cerebral hemorrhage image samples with cerebral hemorrhage stage labels and lesion anatomical region annotations; The standard CT hemorrhage density distribution features are fused and registered with the target three-dimensional reference image to obtain the standard hemorrhage template.

3. The method as described in claim 1, characterized in that, Gradient response analysis is performed based on the feature map of the last convolutional layer in the 3D segmentation network to generate an interpretation map, including: Gradient backpropagation is performed based on the feature map of the last convolutional layer in the 3D segmentation network to obtain the contribution of each image voxel to the segmentation result of the brain hemorrhage region; A response weight map is generated based on the degree of contribution. The interpretation graph is generated by normalizing the response weight graph.

4. The method as described in claim 1, characterized in that, Based on the image of the brain hemorrhage region, the initial brain hemorrhage volume is calculated using a preset dual-path estimation model, including: Using the first estimation path in the dual-path estimation model, target voxels with a probability greater than or equal to a preset probability threshold are statistically analyzed based on the brain hemorrhage region image; The initial cerebral hemorrhage volume of the first estimated path is calculated based on the target voxel and the unit voxel volume of the target voxel. Based on the image of the brain hemorrhage region, hemorrhage feature information is determined, including the maximum diameter, density peak value, and texture directionality of the hemorrhage region. Based on the hemorrhage feature information and the initial cerebral hemorrhage volume of the first estimation path, the initial cerebral hemorrhage volume of the second estimation path is obtained through the second estimation path in the dual-path estimation model.

5. The method as described in claim 1, characterized in that, The three-dimensional segmentation network is a dual-branch heterogeneous encoder structure, which includes two feature extraction branches: structural path and semantic path. The structural path adopts a convolutional structure based on three-dimensional residual U-Net to extract low-level image features of the brain hemorrhage region in the brain medical image data. The low-level image features include the density distribution, edge morphology, and local texture of the brain hemorrhage region. The semantic path performs global modeling of the brain medical image data by integrating a multi-scale window attention module, which is used to extract the spatial context and global dependencies of the brain hemorrhage region; A cross-attention mechanism module is provided between the structural path and the semantic path, which is used to combine the anatomical location of the lesion in the brain hemorrhage region image and the distribution characteristics in the standard hemorrhage template constructed according to different stages of brain hemorrhage to perform feature interaction fusion on the feature information extracted by the structural path and the semantic path.

6. The method as described in claim 1 or 5, characterized in that, The three-dimensional segmentation network includes a density-driven dynamic scale adjustment module; The density-driven dynamic scaling module is used to adjust the size of the convolution window during the feature extraction process based on the density changes and texture direction information of different lesion anatomical locations in the brain hemorrhage region image during the image segmentation process of the three-dimensional segmentation network.

7. The method as described in claim 1 or 5, characterized in that, The brain medical image data is input into a three-dimensional segmentation network to segment the brain hemorrhage region, resulting in an image of the brain hemorrhage region, including: The brain medical image data is input into the three-dimensional segmentation network to extract low-level image features of the brain hemorrhage region from the brain medical image data through the structural path of the three-dimensional segmentation network; Furthermore, the spatial context and global dependencies of the brain hemorrhage region are extracted through the semantic path of the three-dimensional segmentation network; Based on the low-level image features, contextual relationships, and global dependencies, the cross-attention mechanism module of the 3D segmentation network performs feature interaction fusion to output the image of the brain hemorrhage region.

8. A device for measuring the volume of cerebral hemorrhage, characterized in that, It includes an acquisition module, a segmentation module, an identification module, a construction module, a generation module, an analysis module, an estimation module, and a determination module, wherein: The acquisition module is used to acquire the brain medical image data to be measured; The segmentation module is used to input the brain medical image data into a three-dimensional segmentation network to segment the brain hemorrhage region and obtain an image of the brain hemorrhage region. The identification module is used to identify the stage of brain hemorrhage in the brain hemorrhage region image based on the brain hemorrhage region image; A construction module is used to construct a standard hemorrhage template based on the stage of the cerebral hemorrhage and the anatomical location of the lesion in the image of the cerebral hemorrhage region; the standard hemorrhage template is used to characterize the distribution characteristics of different stages of cerebral hemorrhage in the corresponding anatomical region of the lesion. The generation module is used to perform gradient response analysis based on the feature map of the last convolutional layer in the three-dimensional segmentation network and generate an interpretation map. The interpretation map is used to characterize the contribution of each image voxel in the brain medical image data to the segmentation result of the brain hemorrhage region. The analysis module is used to perform spatial similarity analysis based on the standard bleed template and the interpretation diagram to obtain a consistency score between the standard bleed template and the interpretation diagram; The estimation module is used to calculate the initial brain hemorrhage volume based on the brain hemorrhage area image using a preset dual-path estimation model. The first estimation path of the dual-path estimation model estimates the brain hemorrhage volume using the voxel number statistics method, and the second estimation path estimates the brain hemorrhage volume using an image feature regression neural network. The determination module is used to adjust the fusion weights of the dual-path model according to the consistency score, and determine the final target cerebral hemorrhage volume according to the adjusted fusion weights and the initial cerebral hemorrhage volume.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for measuring brain hemorrhage volume as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method for measuring cerebral hemorrhage volume as described in any one of claims 1 to 7.

11. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method for measuring brain hemorrhage volume as described in any one of claims 1 to 9.

Images