Automatic planning system and method for brain hemorrhage puncture path based on multi-modal images

Abstract

This application provides an automatic planning system and method for puncture path in cerebral hemorrhage based on multimodal imaging, including a planning host. The planning host is configured with an image spatial registration module, a structural mask generation module, a target reference generation module, a path set generation module, and a path evaluation and filtering module. The image spatial registration module is configured to perform image spatial registration processing on multimodal head image data to generate aligned image data. The structural mask generation module is configured to generate hematoma tissue masks and dangerous structure masks based on the aligned image data. The target reference generation module is configured to generate target puncture points and target puncture direction reference lines based on the hematoma tissue masks. The path set generation module is configured to generate a set of candidate puncture paths based on the aligned image data, preset skull quadrant constraints, and target puncture points. The path evaluation and filtering module is configured to generate path cost quantification values ​​based on the target puncture direction reference lines, dangerous structure masks, and candidate puncture path sets, and then filter to obtain the final target puncture path.

Description

Technical Field

[0001] This application relates to the field of medical digital imaging and assisted planning technology, and more specifically, to an automatic planning system and method for puncture path in cerebral hemorrhage based on multimodal imaging. Background Technology

[0002] Minimally invasive drainage surgery for cerebral hemorrhage requires precise puncture path planning to maximize hematoma removal while protecting normal brain tissue. Within the complex anatomical structure of the brain, accurately planning the internal hematoma puncture and drainage path to achieve effective decompression and avoid damage to vital brain tissue areas is a crucial prerequisite for safe surgical execution.

[0003] Existing puncture path planning schemes typically employ sampling-based path generation and step-by-step structure avoidance techniques. This scheme first performs simple spatial grid division based on a single image dataset, generating a set of basic random sampling points as the initial connection starting points. Then, a basic physical collision detection algorithm is used to eliminate paths that spatially overlap with obviously dangerous tissues. Finally, the system outputs a default usable connection that has not experienced a physical collision as a reference.

[0004] However, this planning scheme has obvious technical flaws. Due to the failure to fully integrate multimodal image features and the lack of anti-interference constraints for specific regional anatomical features, the generated random connections often fail to balance the comprehensive optimization of safe distance and angle. At the same time, relying solely on single collision detection lacks a deep global cost quantification evaluation mechanism, resulting in low avoidance accuracy of the final determined route in the face of complex and ever-changing dangerous structures. It is very easy to fall into local suboptimal routes and cannot meet the requirements of fully automatic and highly safe path planning in the target scenario. Summary of the Invention

[0005] This application provides an automatic planning system and method for puncture path in cerebral hemorrhage based on multimodal imaging, so as to at least alleviate the above-mentioned technical problems.

[0006] An automatic puncture path planning system for cerebral hemorrhage includes a planning host, which is configured with an image spatial registration module, a structural mask generation module, a target reference generation module, a path set generation module, and a path evaluation and filtering module. The image spatial registration module is configured to perform image spatial registration processing on multimodal head image data to generate aligned image data. The structural mask generation module is configured to generate hematoma tissue masks and dangerous structure masks based on the aligned image data. The target reference generation module is configured to generate target puncture points and target puncture direction reference lines based on the hematoma tissue masks. The path set generation module is configured to generate a set of candidate puncture paths based on the aligned image data, preset skull quadrant constraints, and target puncture points. The path evaluation and filtering module is configured to generate path cost quantification values ​​based on the target puncture direction reference lines, dangerous structure masks, and candidate puncture path sets, and filter the final target puncture path based on the path cost quantification values.

[0007] Optionally, when generating hematoma tissue masks and dangerous structure masks based on aligned image data, the structure mask generation module inputs the aligned image data into a pre-trained 3D image segmentation network, extracts and fuses image spatial features to obtain a fused feature map. The structure mask generation module is also configured to perform 3D pixel-level category division on the fused feature map using a preset category evaluation rule as the category division basis, generate a set of hematoma tissue feature pixels and a set of dangerous structure feature pixels, and map the set of hematoma tissue feature pixels to a hematoma tissue mask and the set of dangerous structure feature pixels to a dangerous structure mask, so that the dangerous structure mask includes a blood vessel distribution mask and a ventricular structure mask.

[0008] Optionally, when the structure mask generation module performs three-dimensional pixel-level category division on the fused feature map according to the preset category evaluation rules, the preset category evaluation rules include pre-configured category cluster centers and category overlap constraints. The structure mask generation module is configured to use the multi-dimensional feature vectors of the three-dimensional pixels in the fused feature map as evaluation objects, calculate the category distance between the multi-dimensional feature vectors and the cluster centers of the hematoma tissue category and the cluster centers of the dangerous structure category, and assign category labels to the three-dimensional pixels in combination with the category overlap constraints, so as to form a set of hematoma tissue feature pixels for mapping the hematoma tissue mask and a set of dangerous structure feature pixels for mapping the dangerous structure mask.

[0009] Optionally, when generating the target puncture target point and target puncture direction reference line based on the hematoma tissue mask, the target reference generation module performs centroid positioning and principal direction analysis on the three-dimensional pixel coordinates of the hematoma tissue mask to obtain the three-dimensional coordinate centroid of the hematoma tissue mask and the major axis features of the principal component analysis of the hematoma tissue mask; the target reference generation module is also configured to determine the target puncture target point based on the three-dimensional coordinate centroid of the hematoma tissue mask, and determine the target puncture direction reference line based on the target puncture target point and the major axis features of the principal component analysis of the hematoma tissue mask.

[0010] Optionally, when generating a candidate puncture path set based on aligned image data, preset skull quadrant constraints, and target puncture points, the path set generation module extracts skull surface contour distribution data from the aligned image data and determines allowed and prohibited sampling areas on the skull surface contour distribution data according to the preset skull quadrant constraints. The path set generation module is also configured to generate multiple initial puncture entry points within the allowed sampling areas and spatially connect each initial puncture entry point with the target puncture point to generate a candidate puncture path set.

[0011] Optionally, when generating a path cost quantification value based on the target puncture direction reference line, the hazardous structure mask, and the candidate puncture path set, the path evaluation and screening module uses the candidate puncture path in the candidate puncture path set as the evaluation object, determines the path length cost based on the candidate puncture path set, determines the hazardous structure avoidance cost based on the hazardous structure mask, and determines the direction deviation cost based on the target puncture direction reference line; the path evaluation and screening module is also configured to fuse the path length cost, hazardous structure avoidance cost, and direction deviation cost according to a preset path cost evaluation rule to generate a path cost quantification value.

[0012] Optionally, when the path evaluation and screening module integrates path length cost, dangerous structure avoidance cost, and directional deviation cost according to the preset path cost evaluation rules, the preset path cost evaluation rules include pre-configured safety distance constraints and cost weight relationships. The path evaluation and screening module is configured to use safety distance constraints to determine the path safety of dangerous structure avoidance cost, and when the safety distance constraints are met, to integrate path length cost, dangerous structure avoidance cost, and directional deviation cost according to the cost weight relationship to generate a path cost quantification value, and when the safety distance constraints are not met, to generate a path cost quantification value representing that the safety distance constraints are not met.

[0013] Optionally, when the path evaluation and filtering module obtains the final target puncture path based on the path cost quantification value, it constructs a path filtering index structure based on the candidate puncture path set and the path cost quantification value. The path filtering index structure organizes the candidate puncture paths and their corresponding path cost quantification values ​​according to the spatial adjacency relationship in the candidate puncture path set. The path evaluation and filtering module is also configured to determine the target candidate puncture path in the path filtering index structure using the path cost quantification value as the filtering basis, and confirm the target candidate puncture path as the final target puncture path.

[0014] Optionally, when performing image spatial registration processing on multimodal head image data, the image spatial registration module uses the acquired head computed tomography (CT) image sequence and head magnetic resonance imaging (MRI) image sequence together as multimodal head image data, and determines a fixed coordinate reference and a floating matching image from the multimodal head image data. The image spatial registration module is also configured to generate a spatial transformation relationship based on the correspondence between the anatomical features of the fixed coordinate reference and the floating matching image, and perform image spatial registration processing on the multimodal head image data according to the spatial transformation relationship to obtain aligned image data.

[0015] An automatic puncture path planning method for cerebral hemorrhage includes the following steps: performing image spatial registration processing on multimodal head image data to generate aligned image data; generating hematoma tissue masks and dangerous structure masks based on the aligned image data; generating target puncture points and target puncture direction reference lines based on the hematoma tissue masks; generating a set of candidate puncture paths based on the aligned image data, preset skull quadrant constraints, and target puncture points; generating path cost quantification values ​​based on the target puncture direction reference lines, dangerous structure masks, and candidate puncture path sets, and selecting the final target puncture path based on the path cost quantification values.

[0016] Optionally, generating a hematoma tissue mask and a dangerous structure mask based on the aligned image data includes: inputting the aligned image data into a pre-trained three-dimensional image segmentation network, extracting and fusing image spatial features to obtain a fused feature map; using a preset category evaluation rule as the category classification basis, performing three-dimensional pixel-level category classification on the fused feature map to generate a set of hematoma tissue feature pixels and a set of dangerous structure feature pixels, and mapping the set of hematoma tissue feature pixels to a hematoma tissue mask and the set of dangerous structure feature pixels to a dangerous structure mask, so that the dangerous structure mask includes a blood vessel distribution mask and a ventricular structure mask.

[0017] Optionally, a three-dimensional pixel-level classification is performed on the fused feature map using a preset category evaluation rule as the basis for category division. This includes: the preset category evaluation rule includes pre-configured category cluster centers and category overlap constraints; using the multi-dimensional feature vectors of the three-dimensional pixels in the fused feature map as evaluation objects, calculating the category distance between the multi-dimensional feature vectors and the cluster centers of the hematoma tissue category and the cluster centers of the dangerous structure category, and assigning category labels to the three-dimensional pixels in combination with the category overlap constraints, so as to form a set of hematoma tissue feature pixels for mapping the hematoma tissue mask and a set of dangerous structure feature pixels for mapping the dangerous structure mask.

[0018] Optionally, the target puncture point and target puncture direction reference line are generated based on the hematoma tissue mask, including: centroid positioning and principal direction analysis of the three-dimensional pixel coordinates of the hematoma tissue mask to obtain the three-dimensional coordinate centroid of the hematoma tissue mask and the major axis features of the principal component analysis of the hematoma tissue mask; determining the target puncture point based on the three-dimensional coordinate centroid of the hematoma tissue mask, and determining the target puncture direction reference line based on the target puncture point and the major axis features of the principal component analysis of the hematoma tissue mask.

[0019] Optionally, a set of candidate puncture paths is generated based on aligned image data, preset skull quadrant constraints, and target puncture points. This includes: extracting skull surface contour distribution data from aligned image data, and determining allowed and prohibited sampling areas on the skull surface contour distribution data according to preset skull quadrant constraints; generating multiple initial puncture entry points within the allowed sampling areas, and spatially connecting each initial puncture entry point with the target puncture point to generate a set of candidate puncture paths.

[0020] Optionally, a path cost quantification value is generated based on the target puncture direction reference line, the hazardous structure mask, and the candidate puncture path set. This includes: using the candidate puncture path in the candidate puncture path set as the evaluation object, determining the path length cost based on the candidate puncture path set, determining the hazardous structure avoidance cost based on the hazardous structure mask, and determining the direction deviation cost based on the target puncture direction reference line; and fusing the path length cost, hazardous structure avoidance cost, and direction deviation cost according to a preset path cost evaluation rule to generate a path cost quantification value.

[0021] Optionally, the path length cost, dangerous structure avoidance cost, and directional deviation cost are integrated according to a preset path cost evaluation rule, including: the preset path cost evaluation rule includes a pre-configured safety distance constraint and cost weight relationship; the path safety is determined by using the safety distance constraint to assess the dangerous structure avoidance cost, and when the safety distance constraint is met, the path length cost, dangerous structure avoidance cost, and directional deviation cost are integrated according to the cost weight relationship to generate a path cost quantification value; when the safety distance constraint is not met, a path cost quantification value representing the failure to meet the safety distance constraint is generated.

[0022] Optionally, the final target puncture path is obtained by filtering based on the path cost quantification value, including: constructing a path filtering index structure based on the candidate puncture path set and the path cost quantification value, the path filtering index structure organizing the candidate puncture paths and their corresponding path cost quantification values ​​according to the spatial adjacency relationship in the candidate puncture path set; determining the target candidate puncture path in the path filtering index structure using the path cost quantification value as the filtering basis, and confirming the target candidate puncture path as the final target puncture path.

[0023] Optionally, image spatial registration processing is performed on the multimodal head image data to generate aligned image data, including: taking the acquired head computed tomography image sequence and head magnetic resonance imaging image sequence together as multimodal head image data, and determining a fixed coordinate reference and a floating matching image from the multimodal head image data; generating a spatial transformation relationship based on the correspondence of anatomical features between the fixed coordinate reference and the floating matching image, and performing image spatial registration processing on the multimodal head image data according to the spatial transformation relationship to obtain aligned image data.

[0024] The technical advantages of the technical solution provided in this application are: This application presents an automatic planning system and method for cerebral hemorrhage puncture path based on multimodal imaging. Addressing the shortcomings of existing technologies, such as the failure to fully integrate image features and the lack of a deep evaluation mechanism leading to low avoidance accuracy and a high susceptibility to local suboptimal routes, this system utilizes an image spatial registration module to perform image spatial registration processing on multimodal head image data to generate aligned image data. This effectively solves the problem of limited perception from a single image. The fusion of multimodal image features enables the system to achieve high data completeness in perceiving the features of internal anatomical structures. Furthermore, a structure mask generation module generates a dangerous structure mask based on the aligned image data, and a target reference generation module generates the target puncture point and target puncture direction reference line based on the hematoma tissue mask, establishing a structured and refined spatial constraint benchmark.

[0025] Furthermore, the path set generation module generates a set of candidate puncture paths based on aligned image data, preset skull quadrant constraints, and target puncture points. This overcomes the blindness of traditional basic random sampling. The preset skull quadrant constraints can effectively exclude potential high-risk anatomical areas, ensuring that the generated candidate path set has high physical feasibility and directional accuracy from the initial stage.

[0026] Finally, the path evaluation and screening module generates path cost quantification values ​​based on the target puncture direction reference line, the hazardous structure mask, and the candidate puncture path set. The final target puncture path is then selected based on these quantified values, overcoming the limitations of traditional methods that rely solely on single physical interference detection. This closed-loop screening mechanism, based on multi-constraint collaborative processing and deep cost quantification, compared to traditional step-by-step approaches, enables the system to automatically make relatively optimal route decisions that consider the safety distance dimensions of all parties when facing complex and varied anatomical structures. This significantly improves puncture avoidance accuracy and the robustness of planning calculations. Attached Figure Description

[0027] Figure 1 This application provides an automatic puncture path planning system for cerebral hemorrhage.

[0028] Figure 2 This application provides an embodiment of an automatic puncture path planning method for cerebral hemorrhage.

[0029] Figure 3 This is an electronic device according to an embodiment of the present application.

[0030] Figure 4 This is a computer-readable storage medium according to an embodiment of the present application. Detailed Implementation

[0031] like Figure 1The image shown is an embodiment of an automatic puncture path planning system for cerebral hemorrhage according to this application. It includes a planning host, which is configured with an image spatial registration module, a structural mask generation module, a target reference generation module, a path set generation module, and a path evaluation and filtering module. The image spatial registration module is configured to perform image spatial registration processing on multimodal head image data to generate aligned image data. The structural mask generation module is configured to generate a hematoma tissue mask and a dangerous structure mask based on the aligned image data. The target reference generation module is configured to generate a target puncture point and a target puncture direction reference line based on the hematoma tissue mask. The path set generation module is configured to generate a candidate puncture path set based on the aligned image data, preset skull quadrant constraints, and the target puncture point. The path evaluation and filtering module is configured to generate a path cost quantification value based on the target puncture direction reference line, the dangerous structure mask, and the candidate puncture path set, and filter the path cost quantification value to obtain the final target puncture path.

[0032] Optionally, when generating hematoma tissue masks and dangerous structure masks based on the aligned image data, the structure mask generation module inputs the aligned image data into a pre-trained three-dimensional image segmentation network, extracts and fuses image spatial features to obtain a fused feature map; the structure mask generation module is further configured to perform three-dimensional pixel-level category division on the fused feature map using a preset category evaluation rule as the category division basis, generate a set of hematoma tissue feature pixels and a set of dangerous structure feature pixels, and map the set of hematoma tissue feature pixels to the hematoma tissue mask, and map the set of dangerous structure feature pixels to the dangerous structure mask, so that the dangerous structure mask includes a blood vessel distribution mask and a ventricular structure mask.

[0033] Preferably, the structure mask generation module is equipped with a pre-trained 3D image segmentation network during the generation of hematoma tissue masks and hazardous structure masks based on aligned image data. This 3D image segmentation network is not a general-purpose segmentation model, but is specifically designed for the application scenario of cerebral hemorrhage puncture path planning, and is designed to simultaneously and accurately segment multiple anatomical structures such as hematoma, blood vessels, and ventricles. The core architecture of this network consists of a 3D feature encoder, a feature fusion bridging structure, a 3D feature decoder, and a clustering-guided feature constraint module. These components support each other functionally and are tightly integrated in terms of data processing flow.

[0034] Preferably, the 3D feature encoder, serving as the front-end processing part of the network, consists of multiple cascaded 3D convolutional processing units. Each 3D convolutional processing unit includes a 3D convolutional layer, a 3D instance normalization layer, and a calibrated linear unit activation layer. The 3D convolutional layer receives multi-channel 3D volume data of aligned image data and performs synchronous sliding window feature scanning and extraction in the height, width, and depth dimensions of the image through multiple learnable 3D convolutional kernels. Unlike 2D convolution, this 3D operation can directly capture the continuous morphology and contextual relationships of brain tissue, hematoma, and dangerous structures in three-dimensional space, generating a set of multi-scale feature maps with spatial hierarchy. As the processing level deepens, the multi-scale feature maps extracted by the encoder gradually transition from low-level features representing local details such as hematoma edges and blood vessel orientation to high-level semantic features representing the overall layout of brain regions and the relative positions between structures.

[0035] Preferably, the feature fusion bridging structure connects the 3D feature encoder and the 3D feature decoder. This structure contains a multi-scale feature aggregation layer that receives multi-scale feature maps output from different depth levels of the 3D feature encoder as processing objects. Its processing logic is not a simple concatenation; instead, it first uses 3D adaptive mean pooling to uniformly adjust all feature maps with different spatial resolutions to a preset intermediate spatial scale, ensuring dimensionality consistency in subsequent fusion operations. Then, it performs element-wise weighted summation on the adjusted feature maps, where the weighting coefficients are dynamically optimized during network training using a backpropagation algorithm. The result of this fusion process is defined as the fused feature map, which integrates the feature information of different scale anatomical structures such as the hematoma body, small blood vessel branches, and ventricular boundaries in the same feature space, providing an information foundation for the subsequent decoder to simultaneously reconstruct multiple segmentation masks.

[0036] Preferably, the 3D feature decoder receives the fused feature map output from the feature fusion bridging structure and is responsible for restoring it to the spatial resolution of the original input image to complete pixel-level category prediction. Its structure consists of multiple 3D transposed convolutional processing units that are substantially symmetrical to the 3D feature encoder. Each 3D transposed convolutional processing unit performs a 3D transposed convolution operation. Mathematically, this operation upsamples and maps the spatial dimensions of the input feature map using learnable parameters, gradually restoring the spatial detail information compressed during downsampling in the encoding stage. Simultaneously, a skip connection mechanism is introduced within the decoder to directly transmit and stitch the uncompressed detail feature map from the corresponding level in the 3D feature encoder onto the corresponding feature map in the decoder. This approach ensures that during upsampling reconstruction, the decoder can accurately locate the approximate area of ​​the hematoma using high-level strong semantic information and precisely restore the clear boundary between the hematoma and adjacent dangerous structures (such as microvessels) using low-level detail information.

[0037] Preferably, to enhance the ability to distinguish different segmentation categories (especially hematoma and blood vessel, two small targets that may have similar gray levels in images) at the feature space level, a cluster-guided feature constraint module is introduced during the network training phase. This module does not directly operate on image pixels, but rather on the highest semantic layer feature map output by the 3D feature decoder. Internally, it pre-sets a set of trainable category cluster centers, specifically including a hematoma tissue category cluster center, a blood vessel category cluster center, and a ventricular structure category cluster center. The processing method involves first calculating the feature space distance between each feature vector point in the feature map and these three category cluster centers. This distance measures which anatomical structure a pixel is more inclined to in terms of high-level semantic features.

[0038] Preferably, the cluster-guided feature constraint module, based on the calculation of feature space distance, further combines a preset category overlap constraint to determine the category affiliation of pixels. This category overlap constraint is a distance margin value set during training, which mandates that the distance between a feature vector point and the cluster center of its own category must be at least smaller than its distance to the cluster centers of other categories by this margin value. Through this constraint mechanism, the model is driven to learn more discriminative and separating feature representations during training, causing the hematoma feature vectors to cluster more tightly around their cluster centers in their feature space, while maintaining a clear distance interval from the cluster centers of blood vessels and ventricles. This mechanism effectively alleviates the feature confusion problem caused by the high overlap of grayscale distribution between hematoma and adjacent blood vessels in the original CT image.

[0039] Preferably, after the network's forward propagation is complete, the structure mask generation module performs three-dimensional pixel-level category division based on the classification probability map output by the three-dimensional feature decoder and normalized by the Softmax function. Each pixel location in this probability map outputs a vector containing three values, representing the probability that the pixel belongs to hematoma, blood vessel, or ventricle, respectively. The module selects the category with the highest probability as the category label for that pixel according to a preset category evaluation rule. Accordingly, all pixels belonging to the hematoma label are grouped into one set, named the hematoma tissue feature pixel set; all pixels belonging to the blood vessel and ventricle labels are grouped into another set, named the dangerous structure feature pixel set.

[0040] Preferably, after obtaining the set of hematoma tissue feature pixels and the set of dangerous structure feature pixels, the structure mask generation module performs a mask mapping operation. Specifically, the set of hematoma tissue feature pixels is mapped into a binary three-dimensional matrix with the same spatial size as the original aligned image data. Pixels belonging to the set are assigned a first value representing the foreground, and the remaining positions are assigned a second value representing the background. This binary three-dimensional matrix is ​​the hematoma tissue mask. Similarly, the set of dangerous structure feature pixels is mapped into another binary three-dimensional matrix, which is the dangerous structure mask. Within this dangerous structure mask, based on the original category labels of the pixels, a three-dimensional sub-region composed of blood vessel category pixels and a three-dimensional sub-region composed of ventricular structure category pixels can be further distinguished. This ensures that the dangerous structure mask explicitly includes the blood vessel distribution mask and the ventricular structure mask, providing a clear and quantifiable avoidance space basis for the subsequent path set generation module and path evaluation and screening module.

[0041] Preferably, when constructing training samples for training the three-dimensional image segmentation network, the processing object is a large number of retrospectively collected multimodal head image data of cerebral hemorrhage cases. Unlike conventional single-modal data preparation, the training sample construction process in this application particularly emphasizes the alignment of multimodal information and the joint annotation of multiple anatomical structures. Specifically, for each case, the preoperative head computed tomography (CT) image sequence and CT angiography image sequence are first acquired. Using a rigid registration algorithm that uses mutual information as a similarity measure, the CT angiography image sequence is used as a floating matching image and rigidly registered to the spatial coordinate system of the head CT image sequence, which serves as a fixed coordinate reference, forming angiography aligned image data with voxels corresponding one-to-one with the head CT image sequence.

[0042] Preferably, after obtaining angiographic-aligned image data through spatial registration, experts with knowledge of neuroanatomical imaging perform multi-category three-dimensional pixel-by-pixel annotation on the head computed tomography (CT) image sequence. This annotation process does not simply delineate the region of interest in general; instead, it strictly follows a pre-defined annotation protocol to generate three mutually exclusive binary label matrices. First, high-density hemorrhage areas within the brain parenchyma are delineated layer by layer, forming a three-dimensional hematoma annotation mask. Second, guided by the angiographic-aligned image data, the course of major intracranial arteries, veins, and important perforating vessels is identified and delineated, forming a three-dimensional vascular annotation mask. Third, based on the anatomical location and morphology of the ventricular system, structures such as the lateral ventricle, third ventricle, and fourth ventricle are precisely marked, forming a three-dimensional ventricular structure annotation mask. For cases equipped with head magnetic resonance imaging (MRI) sequences, a standard brain functional area map is also mapped to the individual space using nonlinear registration technology, generating functional area annotation masks as additional supervisory signals.

[0043] Preferably, the head computed tomography (CT) image sequence, angiography-aligned image data, and the paired binary label matrix of each anatomical structure for a case with complete annotation are collectively defined as a training sample. To improve the network's adaptability to differences in image acquisition and changes in patient position, online data augmentation is applied to the image data and all corresponding annotation masks in each training sample before formal training. This data augmentation includes, but is not limited to: performing random rotations within a preset angle range in three-dimensional space, random scaling within a preset scale range, random translations within a preset voxel offset, and applying small random intensity perturbations to the image grayscale values. The image data of the training sample after data augmentation serves as the input to the three-dimensional image segmentation network, while its paired augmented annotation masks serve as the gold standard for the predicted output of the supervised network.

[0044] Preferably, the training process of the 3D image segmentation network, based on a large number of pre-built training samples, is a multi-task learning process that jointly optimizes the segmentation task loss and the clustering-guided feature constraint loss. The total training loss function is a weighted sum of two parts. The first part is the segmentation task loss used to constrain the final segmentation accuracy of the network. This segmentation task loss calculates the difference between the classification probability map output by the 3D feature decoder and the three gold standards in the training samples: hematoma annotation mask, blood vessel annotation mask, and ventricular structure annotation mask. Its calculation method uses a combination of Dice loss and cross-entropy loss. The Dice loss directly optimizes the spatial overlap between the predicted mask and the gold standard mask, and is insensitive to the volume ratio of small targets such as hematomas and small blood vessels. The cross-entropy loss, on the other hand, penalizes classification errors at the single voxel level. The two are complementary and jointly drive the network's pixel category label prediction to approach the gold standard.

[0045] Preferably, the second part is a cluster-guided feature constraint loss applied to the network feature space. This loss is calculated entirely within the cluster-guided feature constraint module. It processes the feature map output from the highest semantic layer of the 3D feature decoder, along with preset cluster centers for hematoma tissue, blood vessels, and ventricular structures. The design principle of this loss function is to calculate the average feature space distance between the feature vectors corresponding to all voxels marked as foreground by the hematoma label mask in the feature map and the hematoma tissue cluster centers, which is taken as the intra-cluster distance for hematoma tissue. Similarly, the intra-cluster distances for blood vessel and ventricular structure-related voxels are calculated respectively. Simultaneously, the pairwise feature space distances between all cluster centers are calculated as the inter-cluster distances. The goal of the cluster-guided feature constraint loss function is to reduce intra-cluster distances while increasing inter-cluster distances; its core is a contrastive loss that incorporates preset category overlap constraint margin values. The loss function imposes a mandatory constraint that the distance between any voxel and the cluster center of its class must not only be small, but also be at least smaller than its distance to the cluster center of any other class by a preset class overlap constraint margin value. If this condition is not met, a penalty will be incurred.

[0046] Preferably, the total loss function for network training is the sum of the segmentation task loss and the clustering-guided feature constraint loss multiplied by a preset balanced weight coefficient. In each iteration of the training process, a batch of training samples undergoes feature extraction by the 3D feature encoder, scale aggregation by the feature fusion bridging structure, and upsampling reconstruction by the 3D feature decoder, ultimately yielding a classification probability map and a high-level semantic feature map. Subsequently, the classification probability map is fed into the loss calculation stage to generate the segmentation task loss, and the high-level semantic feature map is fed into the clustering-guided feature constraint module to generate the clustering-guided feature constraint loss. After the total loss function value is calculated, the loss gradient is propagated backward from the network's end layer to the network's front end using the backpropagation algorithm. A stochastic gradient descent optimizer with a preset learning rate is then used to update all learnable parameters within the 3D feature encoder, feature fusion bridging structure, 3D feature decoder, and clustering-guided feature constraint module. This process is iterated across all training samples until the total loss function value converges to a relatively small stationary state, ultimately resulting in a 3D image segmentation network capable of performing multi-class joint segmentation on aligned image data.

[0047] Optionally, when the structure mask generation module performs three-dimensional pixel-level category division on the fused feature map according to a preset category evaluation rule, the preset category evaluation rule includes pre-configured category cluster centers and category overlap constraints; the structure mask generation module is configured to use the multi-dimensional feature vectors of the three-dimensional pixels in the fused feature map as evaluation objects, calculate the category distance between the multi-dimensional feature vectors and the hematoma tissue category cluster centers and the dangerous structure category cluster centers, and assign category labels to the three-dimensional pixels in combination with the category overlap constraints, so as to form the hematoma tissue feature pixel set for mapping the hematoma tissue mask and the dangerous structure feature pixel set for mapping the dangerous structure mask.

[0048] Preferably, when the structure mask generation module performs three-dimensional pixel-level category classification on the fused feature map according to a preset category evaluation rule, this preset category evaluation rule is designed as a nonlinear decision-making mechanism that combines prototype learning and margin constraints. Its processing object is the high-dimensional spatial vector corresponding to each three-dimensional pixel in the fused feature map. Specifically, the fused feature map is a high-dimensional data structure that is spatially aligned with the original input but assigns a set of lengths to feature vectors at each spatial location. The length, width, and height dimensions preserve the three-dimensional anatomical correspondence of the cranium, while each value in the feature channel dimension describes the multi-dimensional response intensity of the three-dimensional pixel to different anatomical structures such as hematoma, blood vessels, and ventricles after image spatial feature extraction and fusion. This module first reads the feature channel values ​​located at any set of spatial coordinates in the fused feature map as a numerical sequence with the same number of elements as the feature channel dimension, and defines this numerical sequence as the multi-dimensional feature vector of the three-dimensional pixel. This multi-dimensional feature vector constitutes the direct object of subsequent category distance calculation.

[0049] Preferably, as a core component of the preset category evaluation rules, the pre-configured category clustering centers are a set of feature prototype vectors obtained through iterative optimization by the clustering-guided feature constraint module under the supervision of a large number of training samples during the training phase of the 3D image segmentation network. Specifically, this application configures three category clustering centers corresponding one-to-one with anatomical structures: a hematoma tissue category clustering center, where each element value in its vector represents the most representative feature response pattern of hematoma tissue in the high-level semantic space learned by the 3D image segmentation network during training; a vascular category clustering center, which encodes the typical feature patterns of vascular structures; and a ventricular structure category clustering center, which encodes the typical feature patterns of ventricular boundaries. These category clustering centers do not exist in isolation, but are stored in the configuration parameters of the structure mask generation module in the form of multi-dimensional feature vectors, and are directly loaded into the feature space distance calculation graph when performing 3D pixel-level category division, and compared with each multi-dimensional feature vector to be classified.

[0050] Preferably, the process of calculating the category distance between the multidimensional feature vector and the cluster centers of each category in the structure mask generation module is essentially a distance metric performed in the high-dimensional feature space. Specifically, for the multidimensional feature vector of the fused feature map at any set of spatial coordinates, the module sequentially calculates the category distance between the multidimensional feature vector and the cluster centers of the hematoma tissue category, the blood vessel category, and the ventricular structure category. The calculation method of this category distance is defined as a normalized distance metric commonly used between two vectors in the feature space, such as the spatial distance obtained by summing the squares of the differences between corresponding elements of two vectors and then taking the square root. Its value directly reflects the degree of similarity between the multidimensional feature vector of a three-dimensional pixel and the prototype feature vector of a certain anatomical structure in high-level semantics. The smaller the distance value, the closer the feature response pattern of the pixel is to the typical features of the prototype of that category.

[0051] Preferably, after obtaining the calculation results of the distances for each category, the structure mask generation module does not simply assign category labels based on the nearest distance principle, but instead combines a preset category overlap constraint to perform the final determination. The category overlap constraint is a distance margin value predetermined through joint optimization during the network training phase. Its physical meaning is that it mandates that in the feature space, a pixel must be reliably classified into a certain category; its category distance to the cluster center of that category must not only be less than its category distance to all other category cluster centers, but numerically, this difference must be at least greater than or equal to the distance margin value. By introducing this constraint, the module actively avoids the label uncertainty caused by forcibly classifying pixels located in the blurred boundary region between two category cluster centers in the feature space into a certain category. Especially in scenarios where hematomas and adjacent blood vessels highly overlap in the original image grayscale, this mechanism can effectively suppress category confusion caused by feature mapping deviations.

[0052] Preferably, the structure mask generation module performs a logical determination of category label assignment for each 3D pixel in the fused feature map based on the calculated category distance and the configured distance margin value. Specifically, for a given 3D pixel, the module first determines the smallest distance value among the three category distances and assumes the category corresponding to the smallest distance value as the target candidate category. Next, the module calculates the difference between the smallest distance value of the target candidate category and the smaller of the other two category distances, and compares this difference with the distance margin value. If the difference is greater than or equal to the distance margin value, the 3D pixel is determined to satisfy the category overlap constraint, meaning that its association with the target candidate category in the feature space is clear and reliable. At this point, the anatomical type represented by the target candidate category is assigned as the category label of the 3D pixel. Conversely, if the difference is less than the distance boundary value, the three-dimensional pixel is determined not to meet the category overlap constraint, meaning that it is located in the fuzzy boundary region between two categories. In this case, the module will not assign it to any valid anatomical tissue category, but will instead mark it with a preset label that represents the background or uncertain region, thereby avoiding the introduction of unreliable fake structural information into subsequent mask mapping.

[0053] Preferably, based on the pixel-by-pixel category label assignment results with constraints, the structure mask generation module constructs two structured pixel index sets as direct data sources for mask mapping. The module traverses the entire fused feature map, aggregating the spatial coordinate indices of all 3D pixels assigned the hematoma category label to form a coordinate list, which is defined as the hematoma tissue feature pixel set. Simultaneously, the module merges the spatial coordinate indices of all 3D pixels assigned the blood vessel category label and all 3D pixels assigned the ventricular structure category label to form another coordinate list, which is defined as the dangerous structure feature pixel set. These two feature pixel sets precisely define the occupancy range of hematoma voxels and all dangerous structure voxels that need to be avoided in a unified 3D image space, providing a spatial localization foundation with solid feature differentiation for subsequent mask mapping operations.

[0054] Preferably, after obtaining the hematoma tissue feature pixel set and the dangerous structure feature pixel set, the structure mask generation module immediately enters the mask mapping stage, converting these two discrete coordinate index sets into binary three-dimensional matrices as standard inputs for the target reference generation module and the path evaluation and screening module. For the hematoma tissue feature pixel set, the module creates a three-dimensional matrix with the exact same height, width, and depth dimensions as the aligned image data, with all elements initially representing background values. Then, based on each coordinate index in the hematoma tissue feature pixel set, the module updates the element values ​​at the corresponding spatial positions in this matrix to values ​​representing the foreground. This updated three-dimensional matrix is ​​the hematoma tissue mask. For the dangerous structure feature pixel set, the module performs the same mapping operation, creating an initial background three-dimensional matrix of the same size, and updating the element values ​​at the corresponding positions of all coordinate indices in the dangerous structure feature pixel set to values ​​representing the foreground. This updated three-dimensional matrix is ​​the dangerous structure mask. In this mapping process, since the spatial index source difference between blood vessel category labels and ventricular structure category labels is naturally preserved within the set of dangerous structure feature pixels, the final generated dangerous structure mask can clearly contain the foreground voxel connected domains corresponding to the blood vessel distribution mask and the foreground voxel connected domains corresponding to the ventricular structure mask. This provides inherent class distinguishability for considering the avoidance of different dangerous structures in the subsequent path cost calculation.

[0055] Optionally, when generating the target puncture target point and the target puncture direction reference line based on the hematoma tissue mask, the target reference generation module performs centroid positioning and principal direction analysis on the three-dimensional pixel coordinates of the hematoma tissue mask to obtain the three-dimensional coordinate centroid of the hematoma tissue mask and the principal component analysis long axis feature of the hematoma tissue mask; the target reference generation module is further configured to determine the target puncture target point based on the three-dimensional coordinate centroid of the hematoma tissue mask, and determine the target puncture direction reference line based on the target puncture target point and the principal component analysis long axis feature of the hematoma tissue mask.

[0056] Preferably, after receiving the hematoma tissue mask from the structure mask generation module, the target reference generation module immediately initiates a spatial morphological analysis process for the hematoma region. This process aims to extract two key geometric features from the area occupied by the hematoma in the three-dimensional image space, which can objectively characterize its spatial distribution concentration trend and maximum extension direction. These features serve as the target location to be reached by the puncture needle tip and the ideal direction of travel that the puncture path should follow, respectively, thereby providing quantitative basis based on individualized anatomical morphology for subsequent candidate path generation and evaluation.

[0057] Preferably, the targeting reference generation module first performs centroid localization processing on the hematoma tissue mask to obtain the three-dimensional coordinate centroid of the hematoma tissue mask. The hematoma tissue mask is a binary three-dimensional matrix with the same spatial dimension as the aligned image data, where all foreground positions marked as hematoma voxels take the first value, and the remaining background positions take the second value. The targeting reference generation module traverses all spatial coordinates of this three-dimensional matrix, filters out the three-dimensional coordinate indices corresponding to all voxels with the first value, and collects these coordinate indices into a hematoma voxel coordinate list. Then, for all coordinate indices in this hematoma voxel coordinate list, the arithmetic mean is calculated on the left-right, front-back, and head-to-foot coordinate axes of the image, respectively, to obtain a set of values ​​containing the three coordinate means. The three-dimensional spatial point defined by these three coordinate means is the three-dimensional coordinate centroid of the hematoma tissue mask. In a geometric sense, this centroid corresponds to the center of gravity of all voxels in the hematoma in three-dimensional space. Its spatial position is stably located in the concentrated area of ​​the hematoma, effectively avoiding target selection deviation that may be caused by irregular local morphology or edge burrs of the hematoma.

[0058] Preferably, simultaneously with or immediately after completing the centroid localization, the targeting reference generation module performs principal direction analysis on the hematoma tissue mask to extract the long axis features of the hematoma tissue mask through principal component analysis. This analysis operation also directly processes the aforementioned hematoma voxel coordinate list. The targeting reference generation module first constructs a numerical matrix based on the hematoma voxel coordinate list. The number of rows in this matrix equals the number of hematoma voxels, and the number of columns equals the three spatial dimensions. Subsequently, each column of this matrix is ​​subjected to mean-reduction processing, that is, each element of each column is subtracted from the average value of all elements in that column, resulting in a mean-reduced coordinate data matrix, thereby shifting the spatial distribution center of the hematoma voxels to the origin of the coordinate system. Next, the covariance matrix of the mean-reduced coordinate data matrix is ​​calculated, resulting in a 3x3 real symmetric matrix. Each element in this covariance matrix describes the degree of covariance in the distribution of hematoma voxels across different spatial dimensions, essentially recording the extent of hematoma morphology extension and directional correlation characteristics in various spatial directions. Subsequently, eigenvalue decomposition was performed on the covariance matrix, yielding three non-negative eigenvalues ​​and their corresponding three mutually orthogonal three-dimensional eigenvectors. The target reference generation module arranged these three eigenvalues ​​in descending order of numerical value. The direction of the eigenvector corresponding to the largest eigenvalue represents the direction with the largest variance in the distribution of hematoma voxels in space, i.e., the longest extension axis of the overall hematoma morphology. This eigenvector with a unit length corresponding to the largest eigenvalue is defined as the long axis feature of the principal component analysis of the hematoma tissue mask. This feature clearly indicates the dominant extension direction of the hematoma in three-dimensional space, providing a directional prior derived from imaging morphology for selecting the needle insertion angle of the puncture path.

[0059] Preferably, after obtaining the three-dimensional coordinate centroid of the hematoma tissue mask, the target reference generation module directly establishes the three-dimensional spatial coordinates of the centroid as the target puncture point. This design is based on the clinical consensus of minimally invasive drainage surgery for cerebral hemorrhage, namely, the ideal placement of the drainage tube tip should be in the core area of ​​the hematoma to obtain a larger hematoma clearance volume and prevent the drainage tube tip from damaging normal brain tissue outside the hematoma wall. The centroid, as the spatial center of gravity of the hematoma voxel, is usually located in the relatively full central part of the hematoma. Using it as a target point can reduce the risk of perforating the hematoma wall while allowing the aspiration range of the drainage tube side hole to cover as much of the hematoma interior as possible, thereby improving the ability to control the amount of residual hematoma.

[0060] Preferably, based on the determined target puncture point and the principal component analysis (PCA) long axis features of the hematoma tissue mask, the target reference generation module further derives a target puncture direction reference line. Specifically, the target reference generation module uses the three-dimensional coordinates indicated by the target puncture point as a fixed spatial point and the unit vector indicated by the PCA long axis features of the hematoma tissue mask as a spatial direction. It then determines an infinitely extending straight line in the three-dimensional image space that passes through this fixed point and is parallel to this unit vector, and defines this line as the target puncture direction reference line. This reference line not only provides an ideal reference for the needle insertion direction, but its direction is also strictly aligned with the long axis of the hematoma's shape. This means that if the puncture path deviates minimally from the direction of this reference line, the drainage tube can naturally follow the long axis of the hematoma after entering it, helping to reduce resistance during aspiration and lowering the possibility of forming irregular residual cavities within the hematoma.

[0061] Preferably, the target puncture target points and target puncture direction reference lines generated by the target reference generation module serve as the core benchmarks for structured puncture planning and are organically output to the path set generation module and the path evaluation and screening module. Specifically, the target puncture target points provide the path set generation module with the necessary endpoints for connecting lines when generating candidate puncture paths, while the target puncture direction reference lines provide the path evaluation and screening module with a zero-cost direction for comparison when calculating directional deviation costs. This method, which integrates the location center and morphological orientation jointly analyzed from the hematoma 3D mask into the puncture planning process, differs from traditional methods that rely solely on a single geometric center or subjective manual marking, ignoring the directional characteristics of the hematoma. This ensures that the automatically planned path, in terms of both target point positioning and travel direction, achieves data-driven adaptation to the actual spatial occupancy characteristics of the patient's individual hematoma, creating objective prerequisites for improving hematoma clearance effectiveness.

[0062] Optionally, when generating a candidate puncture path set based on the aligned image data, preset skull quadrant constraints, and the target puncture target point, the path set generation module extracts skull surface contour distribution data from the aligned image data and determines allowed and prohibited sampling areas on the skull surface contour distribution data according to the preset skull quadrant constraints. The path set generation module is further configured to generate multiple initial puncture entry points within the allowed sampling areas and spatially connect each initial puncture entry point with the target puncture target point to generate the candidate puncture path set.

[0063] Preferably, after receiving the aligned image data from the image spatial registration module and the target puncture point from the target reference generation module, the path set generation module immediately initiates the process of constructing a candidate path set. This process does not blindly generate straight line connections throughout the entire cranial space; instead, it introduces pre-defined cranial quadrant constraints to spatially define the feasible area for puncture entry points. The technical objective is to pre-exclude, from a clinical anatomical perspective, areas of the skull surface that, if selected as puncture entry points, are highly likely to damage critical structures such as the optic canal, suboccipital foramen, important venous sinuses, or functional cortex. This establishes a spatial filter that conforms to neurosurgical safety standards at the source of candidate path generation, ensuring that all path options in subsequent evaluations possess basic anatomical feasibility.

[0064] Preferably, before initiating spatial constraints, the path set generation module first extracts the skull surface contour distribution data from the aligned image data. The aligned image data is a set of three-dimensional voxel matrices processed by image spatial registration, where each element represents the X-ray attenuation coefficient at the corresponding spatial location. The path set generation module uses a preset skull bone attenuation threshold range to threshold and classify the aligned image data, marking voxels whose voxel values ​​fall within this range as bone voxels and the remaining voxels as non-bone voxels, thus obtaining a binary three-dimensional bone mask of the same size as the aligned image data. Then, morphological boundary detection is performed on the binary three-dimensional bone mask, specifically by comparing the category labels of each bone voxel with other voxels in its six or twenty-six neighborhoods, identifying all voxels that are themselves marked as bones but have at least one neighboring voxel marked as non-bone as bone boundary voxels. These bone boundary voxels together constitute the inner and outer surface contours of the skull, and all their spatial coordinate indices are extracted and organized into a three-dimensional coordinate list, which is defined as the skull surface contour distribution data. The skull surface contour distribution data accurately describes the geometric shell position of the patient's skull in three-dimensional imaging space, providing a coordinate basis for the next step of applying quadrant constraints.

[0065] Preferably, the path set generation module, based on preset cranial quadrant constraints, further divides the skull geometry described by the cranial surface contour distribution data into permitted and prohibited sampling areas. The preset cranial quadrant constraints are a set of spatial partitioning rules pre-set within the planning host. These rules, using the cephalad-foot coordinate axes aligned with the image data as a reference, and combining the relative positions of cranial anatomical landmarks such as the coronal suture, sagittal suture, and external occipital protuberance in the three-dimensional image, divide the space covered by the cranial surface contour distribution data into several anatomical quadrants with different risk levels. Specifically, the rules explicitly mark the cranial surface coordinate ranges located in the sagittal sinus projection area, suboccipital region, and pterional protuberance—areas clinically identified as high-risk puncture channels—as prohibited sampling areas; while marking the relatively safer cranial surface coordinate ranges such as the frontal, parietal, and posterior temporal regions as permitted sampling areas. This pre-defined rule based on anatomical quadrants, unlike traditional online collision detection, encodes risk knowledge into spatial partitions in one go, ensuring initial anatomical safety during path generation without repeatedly performing complex intracranial structural interference checks at each entry point.

[0066] Preferably, the path set generation module performs the initial puncture entry point generation operation within the allowed sampling area. Based on a preset sampling density parameter, this module selects a set of discrete three-dimensional coordinate points from the coordinate index of the skull surface contour distribution data contained within the allowed sampling area using a uniform spatial sampling method. The sampling density parameter controls the spatial interval between adjacent initial puncture entry points; for example, this interval can be adjusted on the order of millimeters to achieve a balance between path diversity and computational overhead. Each selected three-dimensional coordinate point falls completely within the allowed sampling area, and since this area itself is defined based on skull boundary voxels, these coordinate points naturally lie on the surface of the skull shell. Each of these selected three-dimensional coordinate points is called an initial puncture entry point, and the set they collectively form provides diverse puncture starting points for candidate paths.

[0067] Preferably, after generating multiple initial puncture entry points, the path set generation module spatially connects each initial puncture entry point with the target puncture point to generate a candidate puncture path set. Specifically, the module uses the 3D coordinates of each initial puncture entry point as the starting endpoint of the connection and the 3D coordinates of the target puncture point as the ending endpoint. Then, using a 3D straight-line digitization algorithm, it generates a series of continuous voxel coordinates from the starting endpoint to the ending endpoint in the 3D image space with a preset step size. The straight-line path described by this sequence of voxel coordinates is defined as a candidate puncture path. After performing this connection operation on each initial puncture entry point, multiple candidate puncture paths are obtained, equal to the number of initial puncture entry points. The ordered organization of these candidate puncture paths constitutes the candidate puncture path set. Each path in this set naturally satisfies the hard spatial constraint of pointing from the permissible area on the skull surface to the hematoma target center, providing a space of alternative solutions that both comply with access safety specifications and offer diversity for subsequent refined evaluation of path costs.

[0068] Optionally, when generating a path cost quantification value based on the target puncture direction reference line, the hazardous structure mask, and the candidate puncture path set, the path evaluation and filtering module uses the candidate puncture path in the candidate puncture path set as the evaluation object, determines the path length cost based on the candidate puncture path set, determines the hazardous structure avoidance cost based on the hazardous structure mask, and determines the direction deviation cost based on the target puncture direction reference line; the path evaluation and filtering module is further configured to fuse the path length cost, the hazardous structure avoidance cost, and the direction deviation cost according to a preset path cost evaluation rule to generate the path cost quantification value.

[0069] Preferably, after obtaining the candidate puncture path set output by the path set generation module, the target puncture direction reference line output by the target reference generation module, and the hazardous structure mask output by the structure mask generation module, the path evaluation and screening module immediately initiates a multi-dimensional quantitative evaluation process for each candidate puncture path in the candidate puncture path set. This process treats each candidate puncture path as an independent evaluation object, extracting quantitative cost indicators from three dimensions: path length, degree of hazardous structure avoidance, and suitability of the puncture direction. Then, based on a preset fusion logic, these heterogeneous costs are integrated into a unified-dimensional path cost quantification value, replacing the single binary collision detection judgment in traditional schemes. This provides a continuous and comparable numerical benchmark for automatically selecting the globally optimal solution from numerous feasible paths.

[0070] Preferably, when determining the path length cost based on the set of candidate puncture paths, the path evaluation and screening module processes the three-dimensional spatial line segment represented by each candidate puncture path. This module defines the starting endpoint of a candidate puncture path as the initial puncture inlet point and the ending endpoint as the target puncture point, and calculates the straight-line distance in three-dimensional space between these two endpoints. This distance is obtained by squaring, summing, and then taking the square root of the differences in left-right, front-back, and head-to-foot coordinates between the starting and ending endpoints. The resulting value is the path length cost of the candidate puncture path. The physical meaning of the path length cost is that it quantifies the total distance of brain tissue that the drainage tube needs to traverse from the skull surface to the hematoma target. Its role in the cost system is to drive the optimization process to favor shorter puncture channels to reduce the potential interference range with normal brain tissue along the route.

[0071] Preferably, when determining the risk structure avoidance cost based on the risk structure mask, the path evaluation and screening module treats each candidate puncture path as a detection object and performs spatial overlap and distance analysis with the three-dimensional voxel space occupied by the risk structure mask. Specifically, the module first checks, voxel by voxel, along the voxel sequence of the candidate puncture path, within a preset safety distance threshold radius, whether there are voxels marked as foreground in the risk structure mask on the cross-section perpendicular to the path direction. If the shortest spatial distance between the risk structure foreground voxel and the path centerline at any detection location is less than the safety distance threshold, the path evaluation and screening module directly generates a risk structure avoidance cost representing the failure to meet the safety distance constraint for the candidate puncture path, and sets its value to an extremely large penalty value far exceeding the conventional path length cost. Conversely, if the shortest distance along the entire path is greater than or equal to the safety distance threshold, then the global minimum distance value among the shortest distances to the foreground voxels of the hazardous structure at all detection locations along the entire path is taken, and the reciprocal of this global minimum distance value is calculated. This reciprocal value is the cost of avoiding the hazardous structure for the candidate puncture path under the condition of safe passage. This approach ensures that the cost of paths within the safety distance boundary changes smoothly with the distance from the hazardous structure, with higher costs for closer paths. This guides the optimization process towards paths that provide a larger safety buffer zone for important anatomical structures.

[0072] Preferably, when determining the directional deviation cost based on the target puncture direction reference line, the path evaluation and screening module compares the spatial direction vector of the candidate puncture path with the direction vector of the target puncture direction reference line. The module extracts the path direction vector formed by the starting and ending points of the candidate puncture path, and simultaneously obtains the unit direction vector indicated by the target puncture direction reference line. Then, it calculates the absolute value of the angle between these two three-dimensional vectors. The cosine of this angle is used to measure the parallelism between the two directions, while the angle itself or its sine-based conversion value can serve as a direct source of the directional deviation cost. The physical meaning of this directional deviation cost is the degree of directional deviation between the puncture path and the natural long axis of the hematoma. Its driving optimization process tends to select directions that allow the drainage tube to be as parallel as possible to the long axis of the hematoma after entering it, thereby facilitating the attachment and aspiration of the drainage tube along the hematoma's extension direction, reducing lateral pressure on the hematoma wall and drainage obstruction caused by directional misalignment.

[0073] Preferably, after obtaining the path length cost, hazardous structure avoidance cost, and directional deviation cost for the same candidate puncture path, the path evaluation and screening module then performs a multi-cost fusion operation according to a preset path cost evaluation rule. This preset path cost evaluation rule includes a set of pre-configured cost weight relationships, specifically defining a first weight coefficient multiplied by the path length cost, a second weight coefficient multiplied by the hazardous structure avoidance cost, and a third weight coefficient multiplied by the directional deviation cost. During the fusion calculation, the module arithmetically sums the products of the path length cost and the first weight coefficient, the hazardous structure avoidance cost and the second weight coefficient, and the directional deviation cost and the third weight coefficient. The resulting single-valued result is the quantified path cost of the candidate puncture path. The magnitude of each weight coefficient reflects the relative emphasis on different dimensions of consideration in clinical practice, and this emphasis can be flexibly changed by adjusting the pre-configured file without modifying the underlying processing logic.

[0074] Preferably, after generating a corresponding path cost quantification value for each candidate puncture path in the candidate puncture path set according to the above processing, these paired values ​​are passed to the subsequent comparison stage of the path evaluation and screening module. This complete cost generation mechanism upgrades the isolated collision detection in the traditional scheme into a continuous function optimization problem that performs a weighted comprehensive evaluation of three dimensions: path length, safety distance, and directional compliance. For the automatic planning system, the continuous path cost quantification value overcomes the drawback of binary collision detection, which can only distinguish between pass and fail but cannot distinguish between excellent, average, and poor paths. This allows the system to identify the preferred scheme with a shorter diameter, more sufficient avoidance, and better coordination with the hematoma morphology based on the cost among a large number of candidate paths that meet basic safety requirements, providing a refined numerical driving basis for the final target puncture path selection.

[0075] Optionally, when the path evaluation and filtering module integrates the path length cost, the hazardous structure avoidance cost, and the direction deviation cost according to a preset path cost evaluation rule, the preset path cost evaluation rule includes a pre-configured safety distance constraint and a cost weight relationship. The path evaluation and filtering module is configured to use the safety distance constraint to determine the path safety of the hazardous structure avoidance cost, and when the safety distance constraint is met, to integrate the path length cost, the hazardous structure avoidance cost, and the direction deviation cost according to the cost weight relationship to generate the path cost quantification value, and when the safety distance constraint is not met, to generate the path cost quantification value representing that the safety distance constraint is not met.

[0076] Preferably, when the path evaluation and screening module enters the multi-cost fusion stage, it does not simply perform a weighted sum of the obtained path length cost, dangerous structure avoidance cost, and directional deviation cost. Instead, it first executes a gating decision step centered on a safety distance constraint. This safety distance constraint is a hard condition with veto power in the preset path cost evaluation rules, configured as a safety distance threshold defined in three-dimensional voxel space. The physical meaning of this safety distance threshold is the minimum tolerance distance that the puncture path must maintain between itself and anatomical structures that must be strictly avoided, such as blood vessels and ventricles. The path evaluation and screening module uses this safety distance constraint to compare the global minimum distance value associated with the dangerous structure avoidance cost, thereby determining whether the candidate puncture path is a spatially acceptable safe passage. This marks paths that are absolutely infeasible under clinical surgical standards at the source of cost fusion.

[0077] Preferably, the specific process of the path evaluation and screening module in performing path safety determination is as follows: it directly reads the original data source of the hazardous structure avoidance cost calculated for the candidate puncture path, namely, the global minimum distance value between the entire path and the foreground voxel of the hazardous structure mask. The module compares this global minimum distance value with a pre-configured safety distance threshold. If the global minimum distance value is less than the safety distance threshold, the candidate puncture path is determined not to meet the safety distance constraint. In this case, the path evaluation and screening module no longer reads the specific values ​​of the path length cost and direction deviation cost, nor does it call any weight coefficients in the cost weight relationship. Instead, it directly generates a dedicated path cost quantification value for the candidate puncture path, whose value is far greater than the normal fusion cost by an order of magnitude. The introduction of this extremely large penalty value ensures that in any subsequent path comparison and ranking, the unsafe path will undoubtedly be ranked after all safe paths, achieving hard exclusion based on the physiological safety baseline.

[0078] Preferably, when the path evaluation and screening module determines, through the above comparison, that the global minimum distance value is greater than or equal to the safe distance threshold, the candidate puncture path is deemed to meet the safe distance constraint. At this point, the path is confirmed to have basic physiological safety conditions for passage, and the module then enters the stage of constrained multi-cost fusion calculation based on the cost weight relationship. This branch logic ensures that only paths that pass the safety threshold need to undergo subsequent complex and refined weight adjustment calculations, thereby saving the computational overhead of performing ineffective weighted summations on a large number of unqualified paths, reflecting an engineering design that prioritizes safety and then selects the best efficiency.

[0079] Preferably, under the premise of meeting the safety distance constraint, the path assessment and screening module retrieves the cost weight relationship contained in the preset path cost assessment rules to initiate normalized cost weighted fusion. This cost weight relationship specifically consists of three predefined weight coefficients in the configuration file: a first weight coefficient weighted by the path length cost, a second weight coefficient weighted by the hazardous structure avoidance cost, and a third weight coefficient weighted by the direction deviation cost. These three weight coefficients form a weight vector with a fixed sum or a relative proportional relationship. The relative magnitude of each coefficient reflects the degree of bias in the system's trade-off between path length, risk buffer, and directional consistency when there is sufficient safety passage. The module first multiplies the path length cost by the first weight coefficient to obtain the weighted length component, multiplies the hazardous structure avoidance cost by the second weight coefficient to obtain the weighted avoidance component, and multiplies the direction deviation cost by the third weight coefficient to obtain the weighted direction component.

[0080] Preferably, after obtaining the weighted length component, weighted avoidance component, and weighted direction component, the path evaluation and screening module arithmetically sums these three weighted components. The single real value produced by this arithmetic summation is the quantified path cost of the candidate puncture path under the condition of satisfying the safety distance constraint. This final quantified path cost integrates information from three dimensions: the physical length attribute of the path, the degree of spatial avoidance of multiple hazardous structures, and the degree of conformity with the long axis of the hematoma, and is organically combined according to the relative importance ratio defined by the cost weight relationship. Unlike the traditional method of directly adding the cost items indiscriminately, this fusion method based on weight coefficients allows clinical parameters to be flexibly injected into the automated decision-making process. For example, in areas with dense functional areas, the second weight coefficient of the hazardous structure avoidance cost can be dynamically increased, making the path generated in this area more sensitive to changes in the safety distance in terms of overall cost.

[0081] Preferably, after completing the safety constraint determination and on-demand cost fusion for each candidate puncture path, the path cost quantification array obtained by the path evaluation and screening module naturally contains two distinct numerical groups: one part consists of eliminated paths assigned with maximum penalty values, and the other part consists of candidate safe paths with comparable continuous values ​​obtained based on weight fusion. This two-layer evaluation structure, which separates the hard safety constraints before the cost weight fusion calculation, distinguishes the entire screening mechanism from the traditional fuzzy comprehensive evaluation approach that conflates safety requirements with other objectives. It clearly transforms the clinical safety red line into an insurmountable decision boundary, while within this boundary, a Pareto optimal solution is searched in the multi-objective optimization space using a weight-guided cost function. This objective mechanism increases the probability that the final target puncture path is not only anatomically safe but also exhibits superior overall performance in terms of surgical efficiency and prognostic quality.

[0082] Optionally, when the path evaluation and filtering module obtains the final target puncture path based on the path cost quantification value, it constructs a path filtering index structure based on the candidate puncture path set and the path cost quantification value. The path filtering index structure organizes the candidate puncture paths and their corresponding path cost quantification values ​​according to the spatial adjacency relationship in the candidate puncture path set. The path evaluation and filtering module is further configured to determine the target candidate puncture path in the path filtering index structure using the path cost quantification value as the filtering criterion, and confirm the target candidate puncture path as the final target puncture path.

[0083] Preferably, after calculating the path cost quantification value for each candidate puncture path in the candidate puncture path set, the path evaluation and screening module immediately initiates a global optimization screening process based on a spatial index structure to determine the final target puncture path from among numerous candidate puncture paths. This process does not simply perform a one-time full sorting of all path cost quantification values, but first constructs a path screening index structure that reflects the relative positional relationships between candidate puncture paths in three-dimensional cranial space. The technical starting point of this design is that candidate puncture paths generated in automatically planned scenarios often exhibit spatial clustering in the skull surface entry area or intracranial pathways. If a hard sorting is performed solely based on cost values, the surgical entry point is likely to be concentrated in a very small surface area, ignoring the dispersed requirements of anatomical accessibility and clinical operability. By introducing spatial adjacency relationships as the organizational dimension of the index, this module can naturally maintain the rational distribution of the selected paths in three-dimensional space while selecting the best path based on cost.

[0084] Preferably, the specific process of constructing the path filtering index structure by the path evaluation and filtering module is as follows: each candidate puncture path in the candidate puncture path set is bound to its calculated path cost quantification value, forming a data pair. This module uses the spatial coordinates of the initial puncture entry point of the candidate puncture path on the skull surface contour distribution data as the basis for spatial addressing, dividing the entire allowable sampling area into a series of equally spaced spatial units in the image space. These spatial units constitute the partitioning units of the path filtering index structure. Based on the spatial unit into which the initial puncture entry point of each candidate puncture path falls, the module assigns the data pair of the candidate puncture path and its path cost quantification value to the corresponding spatial unit and organizes them in a list or tree structure. Thus, the path filtering index structure, according to the spatial adjacency relationship of each path entry point on the skull surface in the candidate puncture path set, weaves a large number of candidate puncture paths and their cost information into a data organization that can be quickly retrieved by region, replacing disordered linear listing with spatial proximity.

[0085] Preferably, after the path filtering index structure is constructed, the path evaluation and filtering module does not immediately search for the minimum value in the cost sequence globally. Instead, it first initiates a local selection operation within each spatial unit of the path filtering index structure. The module traverses each spatial unit, reads the path cost quantification value corresponding to each candidate puncture path within that unit, and determines the one with the smallest value through numerical comparison. The candidate puncture path corresponding to this value is marked as the locally optimal candidate path for that spatial unit. If there are no candidate puncture paths in a spatial unit, that unit is skipped directly. This unit-by-unit local filtering step ensures that among all candidate paths with spatially adjacent entry points, the one with the better overall cost performance represents that local area and enters the subsequent global comparison stage. This mechanism avoids a large number of redundant paths with similar entry points and costs entering the final selection pool simultaneously.

[0086] Preferably, after completing the unit-by-unit local optimization, the path evaluation and screening module summarizes the locally optimal candidate paths generated by all spatial units and forms a significantly reduced combination of refined candidate paths. At this refined candidate path combination level, the module performs the final global cost screening. Specifically, the module reads the path cost quantification value corresponding to each locally optimal candidate path in the refined candidate path combination and compares them to determine the locally optimal candidate path with the minimum path cost quantification value, and defines it as the target candidate puncture path. This target candidate puncture path, under the premise of satisfying various anatomical safety constraints and spatial distribution rationality, represents the overall better solution after weighted comprehensive evaluation of the three evaluation dimensions of path length, hazard avoidance, and directional compliance.

[0087] Preferably, after determining the target candidate puncture path, the path evaluation and screening module confirms the target candidate puncture path as the final target puncture path and uses it as the output of the entire automatic puncture path planning process. This final target puncture path contains a complete three-dimensional voxel sequence coordinate from the initial puncture entry point to the target puncture point, which can be directly mapped into the navigation coordinate system for use by intraoperative navigation equipment. This progressive screening mechanism, from coarse-grained spatial index construction to medium-grained local unit selection and then to fine-grained global cost comparison, differs from the single screening method of direct global sorting in traditional approaches. While ensuring superior overall path quality, it endows the final target puncture path with spatial regional representativeness, resulting in a more reasonable surface entry point distribution characteristic and less redundant approximate solution interference during clinical operation.

[0088] Optionally, when performing image spatial registration processing on multimodal head image data, the image spatial registration module uses the acquired head computed tomography (CT) image sequence and head magnetic resonance imaging (MRI) image sequence together as the multimodal head image data, and determines a fixed coordinate reference and a floating matching image from the multimodal head image data; the image spatial registration module is further configured to generate a spatial transformation relationship based on the anatomical feature correspondence between the fixed coordinate reference and the floating matching image, and perform image spatial registration processing on the multimodal head image data according to the spatial transformation relationship to obtain the aligned image data.

[0089] Preferably, after receiving the image data of the brain hemorrhage case to be processed, the image spatial registration module first acquires and confirms the composition of multimodal head image data. This module reads the same patient's head computed tomography (CT) image sequence and head magnetic resonance imaging (MRI) image sequence from the image archiving and communication system or local data storage, and combines these two sets of three-dimensional volume datasets with different imaging principles as multimodal head image data. The head CT image sequence provides high-contrast morphological information of the intracranial hemorrhage foci and skull structure, while the head MRI image sequence (which may contain other sequences) provides soft tissue anatomical details such as sulci and gyri, white matter fiber tract pathways, and functional area localization information. This operation of the image spatial registration module differs from the traditional approach of relying solely on a single image modality; it constructs a multi-channel data foundation containing complementary anatomical information from the source of data input, providing comprehensive input for subsequent structural segmentation and path planning.

[0090] Preferably, the image spatial registration module determines a fixed coordinate reference and a floating matching image from the constructed multimodal head image data. In the application scenario of cerebral hemorrhage puncture path planning, since the coordinates of the skull entry point and the measurement of the intracranial safe distance of the puncture path are naturally referenced to the computed tomography (CT) scan space that best reflects the skeletal anatomy, the image spatial registration module determines the three-dimensional spatial coordinate system of the head CT scan image sequence as the fixed coordinate reference and sets the head magnetic resonance imaging (MRI) image sequence as a floating matching image to be transformed to fit this fixed coordinate reference. This configuration ensures that all subsequently generated path parameters are unified within the CT scan space with a clear geometric scale and suitable for bony landmark localization, and can be directly read by navigation equipment without repeated conversion between cross-modal coordinates.

[0091] Preferably, the image spatial registration module generates a spatial transformation relationship based on the correspondence between anatomical features between a fixed coordinate reference and a floating matching image. Unlike a simple direct solution that maximizes pixel grayscale mutual information, this module designs a multi-level registration process that integrates anatomical landmark priors and voxel similarity measurements. First, the module extracts normalized coordinates of several clearly identifiable cranial anatomical landmarks, such as the anterior commissure, posterior commissure, nasal root, and external occipital protuberance, from both head computed tomography (CT) and magnetic resonance imaging (MRI) sequences. These anatomical landmarks form a sparse but high-confidence set of control point pairs. Based on the spatial deviation vector between corresponding point pairs in this control point pair set, the module uses the least squares method to solve for an initial rigid transformation matrix containing rotation, translation, and scaling components. This initial rigid transformation matrix serves as a coarse estimate of the spatial transformation relationship, quickly narrowing the spatial alignment of the two modalities at the global level.

[0092] Preferably, after applying the initial rigid transformation matrix to the floating matching image, the image spatial registration module further performs high-precision voxel-level mutual information registration to refine the spatial transformation relationship. This module uses the floating matching image transformed by the initial rigid transformation matrix as the current image to be evaluated and calculates the mutual information metric between it and the fixed-coordinate reference image. Mutual information, as a measure of statistical dependence, in image registration means that when the grayscale distributions representing different anatomical tissues in two images achieve optimal spatial correspondence, their mutual information reaches its maximum value. The image spatial registration module utilizes a preset learning rate parameter and a preset upper limit for the number of iterations, and continuously performs sub-voxel-level fine-tuning of the translation and rotation angle parameters in the current spatial transformation relationship (organized in the form of a three-row, four-column augmented transformation matrix) using a gradient ascent optimization algorithm. After each fine-tuning step, the mutual information metric is recalculated until the increase in the mutual information metric is less than a preset convergence threshold.

[0093] Preferably, the set of translation and rotation parameters that maximize the mutual information metric obtained at the end of the above iterative optimization process is combined with the scaling component contained in the initial rigid transformation matrix to form the final spatial transformation relationship. This spatial transformation relationship completely defines all the geometric operations required for accurate mapping from the coordinates of each voxel in the head MRI image sequence to the voxel coordinates in the head CT image sequence. The image spatial registration module performs three-dimensional spatial resampling on all voxels of the floating matching image, i.e., the head MRI image sequence, according to this finally determined spatial transformation relationship. Specifically, the module traverses the position of each output voxel on the three-dimensional grid defined by the fixed coordinate reference, uses the inverse transformation of the spatial transformation relationship to find the corresponding source position of the output voxel in the original floating matching image, performs interpolation calculation within the range of a preset interpolation kernel around the source position, and assigns the calculated intensity value to the output voxel. The new three-dimensional volume data obtained after completing this operation on all output voxels, which is in a completely unified spatial coordinate system with the head CT image sequence, is the aligned image data.

[0094] Preferably, the aligned image data generated by the image spatial registration module is not merely a simple geometric superposition of two images. Throughout the registration process, since the spatial transformation relationship is the product of optimization guided by anatomical landmarks and driven by global mutual information, the head MRI image sequence portion of the aligned image data has minimized positional and angular differences with the head CT image sequence, retaining only its inherent tissue contrast. This deep image spatial fusion allows the subsequent structural mask generation module, when processing the aligned image data using a 3D image segmentation network, to simultaneously observe the high-density features of hematoma, the angiographic enhancement features of blood vessels, and the MRI signal features of functional areas in the same spatial coordinates. This provides a multi-parameter information source for the joint and accurate segmentation of multiple anatomical structures, laying the data foundation for the robust operation of the entire automated planning pipeline in complex anatomical environments.

[0095] like Figure 2 The image shown is an embodiment of an automatic puncture path planning method for cerebral hemorrhage according to this application. The method includes the following steps: performing image spatial registration processing on multimodal head image data to generate aligned image data; generating a hematoma tissue mask and a hazardous structure mask based on the aligned image data; generating a target puncture point and a target puncture direction reference line based on the hematoma tissue mask; generating a candidate puncture path set based on the aligned image data, preset skull quadrant constraints, and the target puncture point; generating a path cost quantification value based on the target puncture direction reference line, the hazardous structure mask, and the candidate puncture path set; and selecting the final target puncture path based on the path cost quantification value.

[0096] like Figure 3 As shown in the figure, an electronic device according to an embodiment of this application includes a processor and a memory. The memory stores a computer program, and when the processor runs the computer program, it performs the steps of the automatic planning method for cerebral hemorrhage puncture path as described in the embodiment of this application.

[0097] like Figure 4 As shown, a computer-readable storage medium is provided in an embodiment of this application. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the automatic planning method for cerebral hemorrhage puncture path as described in the embodiment of this application.

[0098] Figures 2-4 For an exemplary explanation, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. An automatic puncture path planning system for cerebral hemorrhage, characterized in that, The system includes a planning host, which is equipped with an image spatial registration module, a structural mask generation module, a target reference generation module, a path set generation module, and a path evaluation and filtering module. The image spatial registration module is configured to perform image spatial registration processing on multimodal head image data to generate aligned image data. The structural mask generation module is configured to generate hematoma tissue masks and dangerous structure masks based on the aligned image data. The target reference generation module is configured to generate target puncture points and target puncture direction reference lines based on the hematoma tissue mask; the path set generation module is configured to generate a candidate puncture path set based on the aligned image data, preset skull quadrant constraints, and the target puncture points; the path evaluation and filtering module is configured to generate path cost quantification values ​​based on the target puncture direction reference lines, the hazardous structure mask, and the candidate puncture path set, and filter the final target puncture path based on the path cost quantification values.

2. The automatic puncture path planning system for cerebral hemorrhage as described in claim 1, characterized in that, When generating hematoma tissue masks and dangerous structure masks based on the aligned image data, the structure mask generation module inputs the aligned image data into a pre-trained three-dimensional image segmentation network, extracts and fuses image spatial features to obtain a fused feature map. The structure mask generation module is also configured to perform three-dimensional pixel-level category division on the fused feature map using a preset category evaluation rule as the category division basis, generate a set of hematoma tissue feature pixels and a set of dangerous structure feature pixels, and map the set of hematoma tissue feature pixels to the hematoma tissue mask and the set of dangerous structure feature pixels to the dangerous structure mask, so that the dangerous structure mask includes a vascular distribution mask and a ventricular structure mask.

3. The automatic puncture path planning system for cerebral hemorrhage as described in claim 2, characterized in that, When the structure mask generation module performs three-dimensional pixel-level category division on the fused feature map according to preset category evaluation rules, the preset category evaluation rules include pre-configured category cluster centers and category overlap constraints. The structure mask generation module is configured to use the multi-dimensional feature vectors of the three-dimensional pixels in the fused feature map as evaluation objects, calculate the category distance between the multi-dimensional feature vectors and the cluster centers of hematoma tissue and dangerous structures, and assign category labels to the three-dimensional pixels in combination with the category overlap constraints, so as to form the hematoma tissue feature pixel set for mapping the hematoma tissue mask and the dangerous structure feature pixel set for mapping the dangerous structure mask.

4. The automatic puncture path planning system for cerebral hemorrhage as described in claim 1, characterized in that, When the target reference generation module generates the target puncture target point and target puncture direction reference line according to the hematoma tissue mask, it performs centroid positioning and principal direction analysis on the three-dimensional pixel coordinates of the hematoma tissue mask to obtain the three-dimensional coordinate centroid of the hematoma tissue mask and the long axis feature of the principal component analysis of the hematoma tissue mask. The target reference generation module is further configured to determine the target puncture target point based on the three-dimensional coordinate centroid of the hematoma tissue mask, and to determine the target puncture direction reference line based on the target puncture target point and the principal component analysis long axis feature of the hematoma tissue mask.

5. The automatic puncture path planning system for cerebral hemorrhage as described in claim 1, characterized in that, When generating a candidate puncture path set based on the aligned image data, preset skull quadrant constraints, and the target puncture target point, the path set generation module extracts skull surface contour distribution data from the aligned image data and determines allowed and prohibited sampling areas on the skull surface contour distribution data according to the preset skull quadrant constraints. The path set generation module is also configured to generate multiple initial puncture entry points within the allowed sampling areas and spatially connect each initial puncture entry point with the target puncture target point to generate the candidate puncture path set.

6. The automatic puncture path planning system for cerebral hemorrhage as described in claim 1, characterized in that, When generating a path cost quantification value based on the target puncture direction reference line, the hazardous structure mask, and the candidate puncture path set, the path evaluation and screening module uses the candidate puncture path in the candidate puncture path set as the evaluation object, determines the path length cost based on the candidate puncture path set, determines the hazardous structure avoidance cost based on the hazardous structure mask, and determines the direction deviation cost based on the target puncture direction reference line. The path evaluation and screening module is also configured to fuse the path length cost, the hazardous structure avoidance cost, and the direction deviation cost according to a preset path cost evaluation rule to generate the path cost quantification value.

7. The automatic puncture path planning system for cerebral hemorrhage as described in claim 6, characterized in that, When the path evaluation and filtering module integrates the path length cost, the hazardous structure avoidance cost, and the direction deviation cost according to the preset path cost evaluation rules, the preset path cost evaluation rules include a pre-configured safety distance constraint and a cost weight relationship. The path evaluation and filtering module is configured to use the safety distance constraint to determine the path safety of the hazardous structure avoidance cost, and generate the path cost quantification value by integrating the path length cost, hazardous structure avoidance cost, and direction deviation cost according to the cost weight relationship when the safety distance constraint is satisfied, and generate the path cost quantification value representing that the safety distance constraint is not satisfied when the safety distance constraint is not satisfied.

8. The automatic puncture path planning system for cerebral hemorrhage as described in claim 1, characterized in that, When the path evaluation and filtering module obtains the final target puncture path based on the path cost quantification value, it constructs a path filtering index structure based on the candidate puncture path set and the path cost quantification value. The path filtering index structure organizes the candidate puncture paths and their corresponding path cost quantification values ​​according to the spatial adjacency relationship in the candidate puncture path set. The path evaluation and filtering module is also configured to determine the target candidate puncture path in the path filtering index structure using the path cost quantification value as the filtering criterion, and confirm the target candidate puncture path as the final target puncture path.

9. The automatic puncture path planning system for cerebral hemorrhage as described in claim 1, characterized in that, When performing image spatial registration processing on multimodal head image data, the image spatial registration module uses the acquired head computed tomography (CT) image sequence and head magnetic resonance imaging (MRI) image sequence together as the multimodal head image data, and determines a fixed coordinate reference and a floating matching image from the multimodal head image data. The image spatial registration module is also configured to generate a spatial transformation relationship based on the anatomical feature correspondence between the fixed coordinate reference and the floating matching image, and perform image spatial registration processing on the multimodal head image data according to the spatial transformation relationship to obtain the aligned image data.

10. A method for automatically planning puncture pathways in cases of cerebral hemorrhage, characterized in that, Includes the following steps: Image spatial registration is performed on multimodal head image data to generate aligned image data; Generate a hematoma tissue mask and a hazardous structure mask based on the aligned image data; The target puncture point and target puncture direction reference line are generated based on the hematoma tissue mask; a set of candidate puncture paths is generated based on the aligned image data, preset skull quadrant constraints, and the target puncture point; a path cost quantification value is generated based on the target puncture direction reference line, the dangerous structure mask, and the set of candidate puncture paths; and the final target puncture path is obtained by filtering based on the path cost quantification value.

Images