A brain surgery image data intelligent storage method, device and equipment

Abstract

The application discloses a brain surgery image data intelligent storage method, device and equipment, and relates to the technical field of data storage. The method comprises the following steps: acquiring a brain surgery image sequence and performing spatial standardization to generate three-dimensional body data; acquiring a lesion, a key anatomical structure and uncertainty information based on a three-dimensional semantic analysis model, and constructing a region of interest; determining a hierarchical storage strategy and encoding storage according to the spatial relationship between an image block and the lesion and the key anatomical structure; constructing a semantic anchor object graph comprising an image frame, the region of interest, a measurement result and an analysis conclusion, and establishing version association; realizing image data retrieval based on a key slice index, and performing association consistency checking through content hashing and version information.

Description

Technical Field

[0001] This invention relates to the field of data storage technology, and more specifically to a method, apparatus, and device for intelligent storage of neurosurgical image data. Background Technology

[0002] Neurosurgical imaging typically includes various three-dimensional image data such as computed tomography (CT), magnetic resonance imaging (MRI), and their enhanced sequences. In scenarios such as preoperative assessment, postoperative follow-up, and emergency screening, it is usually necessary to simultaneously extract key anatomical structures such as lesion areas, blood vessels, and functional areas, abnormal sections, and measurement results and analytical conclusions obtained from intelligent analysis from the imaging data. With the increase in image resolution and the number of examination sequences, the data volume of neurosurgical imaging data continues to grow. The storage and retrieval of this image data requires simultaneously addressing issues such as information fidelity across different regions, correlation of image analysis results, and rapid localization of key image content.

[0003] In existing technologies, neurosurgical imaging data is typically stored uniformly according to patient information, examination time, or image sequence, and loaded as a complete image sequence when retrieved. Since different regions within the same image sequence vary in importance for subsequent diagnosis, measurement, and review, a uniform storage method makes it difficult to differentiate between lesion areas, areas with uncertain lesion boundaries, areas adjacent to key anatomical structures, and ordinary background areas. Storing the entire image sequence in a lossless manner results in significant storage space consumption; while compressing the entire image sequence using conventional compression methods may lead to the loss of image information on lesion boundaries and key anatomical structures such as blood vessels, thus affecting subsequent measurements and image review.

[0004] On the other hand, the correspondence between raw image data and the lesion annotations, measurement results, and analysis conclusions generated by intelligent analysis usually relies on file names, sequence order, or manual records, lacking a stable association mechanism that can simultaneously represent image instances, image frames, spatial locations, regions of interest, and data versions. When image data is reanalyzed, updated, or multiple versions are accessed, inaccurate correlations between image data and corresponding annotations, measurement results, and analysis conclusions can easily occur.

[0005] In addition, when doctors review neurosurgical images, they usually need to load the complete image sequence first, and then manually locate the key slices related to the lesion or key anatomical structure from a large number of image slices. This makes the process of reviewing image data lack a priority location mechanism for abnormal image content.

[0006] Therefore, existing methods for storing and accessing neurosurgical images still require an intelligent image data storage method that can differentiate the storage of different image regions based on the results of image semantic analysis, establish a stable correlation between image data and analysis data, and determine priority access objects based on abnormal image content. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent storage method, device, and equipment for neurosurgical image data. By performing three-dimensional registration and semantic analysis on neurosurgical image sequences, information such as lesions, key anatomical structures, and boundary uncertainties is obtained. Based on this, a semantic anchoring object map and a hierarchical storage strategy are constructed, allowing image regions of different importance to adopt appropriate storage methods. This reduces the storage space occupied by background images while protecting image information related to lesions and key anatomical structures. Simultaneously, image blocks, masks, measurement results, and analysis conclusions are associated through anchor keys, and a key slice index is established using abnormal slice scoring. This enables priority location of abnormal slices and their associated data during retrieval, thereby improving the accuracy of association and retrieval efficiency of neurosurgical image data.

[0008] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides an intelligent storage method for neurosurgical image data, the method comprising: Acquire neurosurgical image sequences, perform instance analysis and 3D registration on the neurosurgical image sequences, and generate a set of image frames containing instance identifiers, frame numbers and spatial coordinate information; The image frame set is input into the trained three-dimensional neurosurgical image semantic analysis model to obtain lesion mask, key anatomical structure mask, boundary uncertainty map, abnormal slice score, measurement results and analysis conclusions; Regions of interest are determined based on lesion masks and key anatomical structure masks. Image frames, regions of interest, measurement results, and analysis conclusions are organized into a semantic anchoring object graph. Anchor keys containing instance identifiers, frame numbers, spatial coordinates, region of interest numbers, content hashes, and version numbers are generated for objects in the semantic anchoring object graph. The fidelity level of the image block is determined based on the lesion mask distance field, segmentation probability entropy, distance to key anatomical structures, and abnormal slice score. Based on the fidelity level, the image block is divided into lesion core non-destructive block, near-non-destructive block with uncertain boundaries, key structure adjacent protection block, and background compression block, generating a hierarchical storage strategy. Each image block is encoded according to a hierarchical storage strategy, and the encoded image blocks, along with their corresponding masks, measurement results, and analysis conclusions, are stored together using anchor keys. A key slice index is established based on the abnormal slice score. When an image retrieval request is received, the abnormal slice is determined based on the key slice index, and the image block corresponding to the abnormal slice and its associated data are loaded through the anchor key. After receiving a complete sequence loading request, the remaining image frames are loaded.

[0009] Secondly, the present invention also provides an intelligent storage device for neurosurgical image data, the device comprising: The sequence acquisition module is used to acquire neurosurgical image sequences, perform instance parsing and 3D registration on the neurosurgical image sequences, and generate a set of image frames containing instance identifiers, frame numbers and spatial coordinate information. The model training module is used to input the set of image frames into the trained three-dimensional neurosurgical image semantic analysis model to obtain lesion masks, key anatomical structure masks, boundary uncertainty maps, abnormal slice scores, measurement results, and analysis conclusions. The region determination module is used to determine the region of interest based on the lesion mask and the key anatomical structure mask, and organize the image frame, region of interest, measurement results and analysis conclusions into a semantic anchoring object graph. It generates anchor keys for objects in the semantic anchoring object graph, which include instance identifier, frame number, spatial coordinates, region of interest number, content hash and version number. The grading module is used to determine the fidelity level of an image block based on the lesion mask distance field, segmentation probability entropy, distance to key anatomical structures, and abnormal slice scores. Based on the fidelity level, the image block is divided into a lesion core non-destructive block, a near-non-destructive block with uncertain boundaries, a key structure adjacent protection block, and a background compression block, generating a hierarchical storage strategy. The associated storage module is used to encode each image block according to the hierarchical storage strategy, and to associate and store the encoded image blocks, along with the corresponding masks, measurement results and analysis conclusions, according to the anchor key. The index building module is used to build a key slice index based on the abnormal slice score. When an image retrieval request is received, the abnormal slice is determined according to the key slice index, and the image block corresponding to the abnormal slice and its associated data are loaded through the anchor key. After receiving a complete sequence loading request, the remaining image frames are loaded.

[0010] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the intelligent storage method for neurosurgical image data provided in the first aspect.

[0011] Fourthly, the present invention also provides a computer-readable storage medium on which a computer program is stored, wherein when the computer program is executed by a processor, it implements the intelligent storage method for neurosurgical image data provided in the first aspect.

[0012] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the intelligent storage method for neurosurgical image data provided in the first aspect.

[0013] The beneficial effects of this invention are as follows: The method of this invention obtains information such as lesions, key anatomical structures, and boundary uncertainties by performing three-dimensional registration and semantic analysis on neurosurgical image sequences. Based on this, a semantic anchoring object map and a hierarchical storage strategy are constructed, so that image regions of different importance adopt appropriate storage methods, reducing the storage occupation of background images while protecting the image information related to lesions and key anatomical structures. At the same time, by linking image blocks, masks, measurement results, and analysis conclusions through anchoring keys, and establishing a key slice index using abnormal slice scoring, abnormal slices and their associated data can be prioritized during retrieval, thereby improving the accuracy of association and retrieval efficiency of neurosurgical image data.

[0014] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating an intelligent storage method for neurosurgical image data according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an intelligent storage device for neurosurgical image data according to an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device structure provided in an embodiment of this application. Detailed Implementation

[0016] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] It should be noted that references to "an embodiment," "embodiment," "example embodiment," etc., in this specification refer to the described embodiment including specific features, structures, or characteristics; however, not every embodiment must include these specific features, structures, or characteristics. Furthermore, such expressions do not refer to the same embodiment. Moreover, when describing specific features, structures, or characteristics in conjunction with embodiments, whether or not explicitly described, it is indicated that incorporating such features, structures, or characteristics into other embodiments is within the knowledge of those skilled in the art.

[0018] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0019] like Figure 1 As shown in this embodiment, an intelligent storage method for neurosurgical image data is provided for parsing, semantic analysis, tissue association, hierarchical encoding, storage, and retrieval of three-dimensional image data generated during neurosurgical examination. Neurosurgical image data can be three-dimensional image data formed from computed tomography (CT) images, magnetic resonance imaging (MRI) images, and their enhanced sequences. An image sequence refers to a set of image data formed from multiple image frames with a common sequence identifier, generated from the same neurosurgical examination; an image frame is a single frame of image data that constitutes a three-dimensional image sequence, containing not only pixel data but also header information characterizing the frame's source, spatial location, and acquisition parameters. The entire processing first converts the raw image data into a set of image frames with a unified spatial reference and stable positioning information. Then, a three-dimensional neurosurgical image semantic analysis model is used to identify lesions and key anatomical structures from the images, further forming regions of interest, semantic anchoring object maps, and hierarchical storage strategies. Finally, priority retrieval is completed based on the abnormal slice index, thereby enabling the spatial information obtained from the front-end image parsing to permeate the subsequent semantic analysis, storage, and retrieval processes.

[0020] S101, acquire neurosurgical image sequences, perform instance analysis and 3D registration on the neurosurgical image sequences, and generate a set of image frames containing instance identifiers, frame numbers and spatial coordinate information.

[0021] The system receives all image frame files of the same inspection sequence via the DICOM communication interface and reads the instance identifier from the header information of each image frame. (Retrieve SOPInstanceUID field), sequence identifier (Retrieve SeriesInstanceUID field), layer thickness (Retrieve SliceThickness field), pixel spacing (Retrieve PixelSpacing field), patient coordinate direction (Retrieve the direction cosine vectors corresponding to the row and column directions from the ImageOrientationPatient field) and the patient's coordinate origin. (Retrieve the 3D coordinates of the top-left pixel of the corresponding frame from the ImagePositionPatient field), and fix the patient coordinate system to the DICOM patient coordinate system LPS, where the coordinate axes are defined as follows: The axis points to the patient's left side. The axis points to the back of the patient. The axis points towards the patient's head; according to and Calculate the slice normal unit vector The symbol × represents the vector cross product, used to determine the slice stacking direction, and to arrange all image frames along the edge. The frame number is obtained by sorting the slice positions in ascending order. , where frame number It is an integer starting from 1 and incrementing, corresponding one-to-one with the sorted image frames; The sorted image frames are stacked to form the initial three-dimensional volume data, and the first frame is used as the basis for the data. As a unified directional benchmark, geometric consistency correction and 3D registration are performed on all frames under this benchmark. 3D registration employs rigid body transformation and uses mutual information as a similarity metric. The optimizer uses gradient descent, with a maximum number of iterations set to 100 and a learning rate set to [value missing]. When the mutual information gain between two adjacent iterations is less than The system determines if the alignment has converged and outputs the registration transform. The registered 3D volume data is resampled to a uniform voxel spacing. The resampling interpolation method uses trilinear interpolation and uses the resampling result as the standard representation of the three-dimensional volume data. At the same time, it generates a spatial index consistent with the standard representation for each image frame, which is used to achieve stable positioning in subsequent object maps and hierarchical storage. Spatial indexes are obtained by mapping voxel indices to patient coordinates, and the mapping relationship is as follows: ; in This represents a three-dimensional spatial coordinate vector in the patient's coordinate system. Voxel index in standard representation The corresponding patient coordinate origin vector, These represent the voxel indices in the standard representation, corresponding to the row, column, and layer directions, respectively. These represent the standard voxel spacing values. These represent the unit direction vectors in the row, column, and layer directions of the standard representation, respectively, and are consistent with the reference direction; A frame index record is written for each image frame. The frame index record is generated by... , Frame number and the slice center coordinates of the frame in standard representation Composition, where the center coordinates of the slice The image frame is calculated by substituting the central voxel index in the slice plane corresponding to the frame into the mapping relationship, and the "image pixel data, header information field, registration transformation parameters, and frame index record" are encapsulated into an image frame object and written into the image frame set, thereby obtaining an image frame set containing instance identifier, frame number, and spatial coordinate system information.

[0022] S102 inputs the set of image frames into the trained three-dimensional neurosurgical image semantic analysis model to obtain lesion masks, key anatomical structure masks, boundary uncertainty maps, abnormal slice scores, measurement results, and analysis conclusions.

[0023] Resampled 3D volumetric data is extracted from the image frame set, and its voxel gray levels are preprocessed by quantile cropping and linear normalization, where the lower limit of cropping is taken as the th quantile of all pixel gray levels. The quantile and the upper limit of cropping are taken as the first of all gray values. Quantiles, after cropping, linearly map the grayscale to... And maintain voxel spacing The data remains unchanged, while the 3D volume data is adjusted using a center-clipping method. The fixed input size of the voxel is used; if the original size is insufficient, it is symmetrically filled with a grayscale value of 0. A trained 3D neurosurgical image semantic analysis model is used to perform forward inference on a fixed-size input. The 3D neurosurgical image semantic analysis model is a 3D multi-scale encoder-dual segmenter decoder structure with abnormal slice supervision output. The multi-scale encoder contains 4 scale levels with the number of channels in each level being 32, 64, 128, and 256 respectively. Each level consists of two... The system employs 3D convolution, instance normalization, and ReLU activation, and scale transformation is achieved through downsampling convolution with a stride of 2. Both the lesion segmentation decoding branch and the key anatomical structure segmentation decoding branch utilize progressively upsampling deconvolution and skip connections to fuse with features from the corresponding encoding levels. The output layers of both segmentation branches are... 3D convolution followed by Sigmoid to obtain the lesion probability map With critical structure probability map ,in This represents an integer voxel index in the row direction that is consistent with the row direction unit vector r. Represents the column direction voxel index integer and the column direction unit vector Consistent, k represents an integer voxel index in the layer direction that is consistent with the unit vector s in the layer direction. Voxel representation The probability that it belongs to a lesion and the range of values ​​is: Voxel representation The probability of belonging to a key anatomical structure and the range of values ​​are: ; With a fixed threshold right Binarization is performed to obtain the lesion mask. With a fixed threshold right Binarization is performed to obtain a mask of key anatomical structures. Furthermore, a three-dimensional connected component sieving process is performed on the binarization result to remove connected components with a voxel count less than 100, thereby ensuring mask stability. Calculate the boundary uncertainty map based on the lesion probability map. Its calculation uses binary probabilistic information entropy and performs normalization, specifically as follows: ; in Voxel representation The boundary uncertainty at the given location and the range of values ​​is Voxel representation The probability of lesions at the location and satisfying Represents the natural logarithm function. The constant, expressed as the natural logarithm, is used to normalize the second type of entropy to... And to avoid Triggering numerical anomalies will Truncation to closed interval Then substitute the values ​​into the calculation; By frame number As a slice index, the three-dimensional volume data is decomposed into a sequence of two-dimensional slices according to a fixed layer orientation. For each slice... Calculate the score of the abnormal slice The score for an abnormal slice, used to characterize its degree of abnormality, is a weighted sum of three features: the proportion of lesion volume, the mean of boundary uncertainty, and the overlap ratio of key structures. ,in Indicates frame number is The abnormal slice score of the slice and the value range is Represent the fixed weights of the three features and take... Slice The proportion of lesion volume within the slice is defined as the percentage of the slice that meets the following criteria. The ratio of the number of voxels to the total number of pixels in the slice. Slice The inner boundary has an uncertain mean and is defined as the mean of all pixels in the slice. Arithmetic mean Slice The overlap ratio of internal key structures is defined as the percentage of structures that simultaneously satisfy the following conditions in the slice. and The number of voxels and the content of the slice The ratio of voxel counts, and when the voxel count of the lesion is 0. Set to 0; After obtaining the lesion mask, key anatomical structure mask, boundary uncertainty map, and abnormal slice score, the measurement results are calculated and the analysis conclusions are generated. The measurement results include at least the lesion volume. Maximum diameter of lesion Minimum distance from lesion to key anatomical structures The volume of the lesion Multiply the number of voxels in the lesion mask by the voxel volume The maximum diameter of the lesion was obtained by converting the value to mL. The minimum distance from the lesion to the key anatomical structure is obtained by finding the two-dimensional principal axis of the maximum connected component of the lesion mask and calculating the Euclidean distance of its maximum endpoint in the patient coordinate system. The result is obtained by calculating the three-dimensional Euclidean distance transformation of the key anatomical structure mask and taking the minimum value at the voxel position of the lesion mask, with the unit being mm; The analysis conclusion text is generated based on a fixed analysis conclusion template. The template includes fields such as "whether a lesion was detected," "lesion volume," "maximum diameter of the lesion," "minimum distance to key anatomical structures," and "set of abnormal slice frame numbers." Among these, "whether a lesion was detected" is determined by the lesion volume. It is determined that the "abnormal slice frame number set" is composed of those that satisfy... Frame number It is composed and consistent with subsequent key slice indexes; The 3D neurosurgical image semantic analysis model was trained using a fixed process before deployment. The training sample set consisted of 500 3D neurosurgical image samples, each containing voxel-level annotations of lesions, key anatomical structures, and abnormal slice frame numbers. The training input adopted the same preprocessing and size specifications as the inference input. The training objective was to update the model parameters using a total loss function, where the total loss was obtained by summing the lesion segmentation loss, key structure segmentation loss, abnormal slice classification loss, and boundary entropy constraint loss according to their weights, with the weights of the four losses fixed at 1.0, 1.0, 0.5, and 0.1, respectively. The optimizer used was Adam with a learning rate of [missing information]. The batch size is 1, the maximum number of training rounds is 200, and the decrease in total loss on the validation set over 10 consecutive rounds is less than 1. Training is stopped and the model weights are fixed, thus obtaining a trained three-dimensional neurosurgical image semantic analysis model for inference.

[0024] S103. Determine the region of interest based on the lesion mask and the key anatomical structure mask, and organize the image frame, region of interest, measurement results and analysis conclusions into a semantic anchoring object graph. Generate anchoring keys for objects in the semantic anchoring object graph that include instance identifier, frame number, spatial coordinates, region of interest number, content hash and version number.

[0025] Based on lesion masking Masking of key anatomical structures Construct a set of regions of interest, specifically for... and The three-dimensional 26-neighborhood connected component labeling is performed to obtain a set of connected components, and each connected component is mapped to a candidate region of interest; Using the patient coordinate system as a unified spatial reference, all voxels within each candidate region of interest are indexed. The voxel-to-patient coordinate mapping relationship is converted into three-dimensional spatial coordinates, thereby calculating the minimum bounding box and centroid coordinates of the candidate region of interest. Based on this, the two types of candidate regions of interest are fused. The fusion rule is that when the bounding boxes of the lesion candidate region of interest and the key anatomical structure candidate region of interest in the patient coordinate system overlap at the voxel level or the closest distance between the two bounding boxes does not exceed 2mm, they are merged into the same region of interest; otherwise, they remain as different regions of interest. Each region of interest (ROI) after fusion is assigned a unique ROI ID, and the 3D volume data is processed according to a fixed image block size. The voxels are divided into non-overlapping blocks. For each image block, the overlap relationship between it and the voxels of each region of interest is determined. The image blocks containing voxels of any region of interest are aggregated into the corresponding region of interest and the spatial coordinates of the center of the image block are recorded for subsequent retrieval and positioning. Furthermore, the image frames, regions of interest, measurement items, and analysis conclusions are organized into a semantic anchoring object graph. The object graph is stored using a deterministic data structure of "node table + edge table," where the node table contains the UID written for each image frame node. Frame number and the spatial coordinates of the slice center Write for each region of interest node Reference frame number With the coordinates of the centroid space The reference frame number The frame number with the largest voxel count across all frames is selected and used to establish a stable anchor between the region of interest and a specific slice. An anchor key is generated for each node and used as the primary index of the node. The anchor key is defined as follows: ; in Indicates anchor key, The instance identifier is consistent with SOPInstanceUID, f represents the frame number, and the frame number is taken for image frame nodes. For region of interest nodes, their associated measurement item nodes, and analysis conclusion nodes, the reference frame number is taken. , Represents a spatial coordinate vector and takes the values ​​of image frame nodes. For nodes in the region of interest, their associated measurement item nodes and analysis conclusion nodes, take... RID represents the region of interest number, with 0 for image frame nodes and the corresponding number for the region of interest node and its associated nodes. Represents the content hash. Indicates the version number; Content hash The SHA-256 algorithm is used to calculate and the serialization method is fixed to ensure consistency. The content hash of the image frame node is calculated using the byte sequence of the pixel matrix after resampling of the frame. The content hash of the region of interest node is calculated using the UTF-8 byte sequence of the index list of image blocks contained in the region of interest and the bounding box parameters. The content hash of the measurement item node is calculated using the UTF-8 byte sequence of the JSON text consisting of the measurement item name, value and unit after sorting by key name. The content hash of the analysis conclusion node is calculated using the UTF-8 byte sequence of the complete text after filling the analysis conclusion template. In the edge table, establish referencing edges and containing edges. The containing edges point from image frame nodes to the region of interest nodes covered by that frame, and the anchor key K of the pointed-to node is written into the edge attribute. The referencing edges point from measurement item nodes and analysis conclusion nodes to the region of interest nodes they reference, and the anchor key of the referenced region of interest node is written into the edge attribute. This allows the reference relationship between measurement results and analysis conclusions to the region of interest to be consistently resolved at the levels of instance identifier, frame number, and spatial coordinates. When a retrieval request triggers the loading of associated data, it is based on the anchor key carried in the request. After locating the target node and its external reference edges, loading the corresponding image blocks, masks, measurement results, and analysis conclusions, recalculate its content hash and compare it with the content hash in the anchor key. and version number A consistency check is performed. If the content hash is inconsistent or the version number does not match, the process backtracks along the previous version reference edge recorded in the semantic anchoring object graph to the version number. The historical nodes are loaded and their associated data is loaded. At the same time, associated verification records are generated and written with verification timestamp, request anchor key, comparison failure field, backtracking version number and backtracking path. When a new image sequence is written or a new analysis is performed on the same image sequence, a new anchor key is generated for the new result and used to determine the region of interest (ROI) with existing anchor keys. The determination rule is the instance identifier. Same and reference frame number Same and centroid space coordinates If the distance in the three-dimensional Euclidean distance does not exceed 1mm, and the judgment rule is met, the version number will be... Increment the value by 1 based on the original version and connect the new node to the original region of interest node through a "version evolution edge" while retaining the old version node without overwriting it. If the judgment rule is not met, assign a new region of interest number. And establish new nodes for regions of interest; During retrieval, the latest version of the node set before the corresponding time point is selected based on the retrieval time parameter, or the node set with the version number equal to the specified version number is directly located based on the specified version number, and the corresponding version of the image block, mask, measurement results and analysis conclusions are loaded from the object graph edge table accordingly.

[0026] S104. Based on the lesion mask distance field, segmentation probability entropy, distance to key anatomical structures, and abnormal slice score, the fidelity level of the image block is determined, and the image block is divided into lesion core non-destructive block, boundary uncertain near non-destructive block, key structure adjacent protection block and background compression block according to the fidelity level, and a hierarchical storage strategy is generated.

[0027] The image blocks are used as input objects, and the 3D volume data is processed according to a fixed image block size. Voxels are divided into non-overlapping blocks and each image block is assigned a block number. An image block refers to a cubic data unit composed of continuous voxels on a three-dimensional volumetric data system consistent with the patient's coordinate system. The block number is... It is an integer starting from 0 and incrementing, and it corresponds one-to-one with the image block; Determine the block center voxel index for each image block. ,in This represents the center voxel index triplet of image block b and is related to the voxel index. Within the same index space, the central voxel index is taken as the integer midpoint between the start and end indices of the image block in the row, column, and layer directions; Based on lesion mask Constructing the lesion voxel set Ω les And based on key anatomical structures mask Construct the key structural voxel set Ω str ; in , And further, using the 6-neighborhood rule from Ω les Extracting the voxel set of lesion boundaries Ω les Among them, the voxel at the lesion boundary satisfies that it belongs to Ω. les Furthermore, at least one voxel within its 6-neighborhood does not belong to Ω. les voxels; The distances used for stratification are calculated, and a lesion mask distance field and distances to key anatomical structures are generated. The distance calculation adopts a unified definition of the shortest voxel distance. ; in Indicates from voxel index The shortest voxel distance to the voxel set Ω, where q represents the voxel index triplet of the position to be calculated, and Ω represents the set consisting of voxel index triplets. Represents a set For any voxel index triple in the set Ω, min represents the operation of finding the minimum value on the set Ω. This represents the Euclidean norm, and the result is in units corresponding to the voxel spacing. Since the voxel spacing has been fixed as... Therefore, the distance unit is mm; for each image block Calculate the distance from the center of the lesion to the boundary of the lesion. And As a block-level representation of the lesion mask distance field, while when When Forced to be 0 to satisfy the definition that "the shortest distance is 0 when the center is located inside the voxel of the lesion"; For each image block Calculate the distance from the center of the block to the critical structure And As a block-level representation of key anatomical distances; Calculate the block-level statistics of the segmentation probability entropy. The segmentation probability entropy refers to the probability map of lesions. With critical structure probability map The voxel-level uncertainties were calculated separately using type II information entropy and normalized to [the appropriate value]. Block-level statistics Image block The arithmetic mean of the lesion uncertainty and the critical structure uncertainty of all voxels; Calculate the coverage ratio of the lesion mask The proportion of lesion coverage by masking Defined as an image block Internal satisfaction voxel count and image block Total number of elements The ratio; Image block type determination and hierarchical storage strategy are completed using a fixed threshold, with the first coverage threshold set to [value missing]. The entropy threshold is set to The first distance threshold is set to The second distance threshold is set to And judge them sequentially according to the rules of mutual exclusion and priority, satisfying... The image block was determined to be a lesion core without any defects, and was not determined to be a lesion core without defects and met the following conditions. and The image block was determined to be a near-lossless block with uncertain boundaries, and was not classified into the first two categories and met the following conditions. Image blocks that are identified as critical structure adjacent protection blocks are identified as background compression blocks; Image blocks that are not hit by the first three categories are identified as background compression blocks. Number each image block Block center voxel index lesion mask coverage ratio Block-level segmentation probability entropy lesion boundary distance Key structural distance The final determined image block type is written into the hierarchical storage strategy list, and this list is used as direct input data for differentiated encoding and storage.

[0028] S105, each image block is encoded according to the hierarchical storage strategy, and the encoded image blocks, along with the corresponding masks, measurement results, and analysis conclusions, are stored together according to the anchor key.

[0029] Read the list of tiered storage policies and assign them to image block numbers. To traverse the index, encoding and associated data writing are performed on each image block sequentially. Each image block is derived from resampled 3D volume data and retains its 16-bit signed integer grayscale value. 16-bit signed integer grayscale value This refers to the integerized result obtained by restoring the grayscale value of the DICOM pixel value according to RescaleSlope and RescaleIntercept and then interpolating it using trilinear interpolation during resampling. When an image block is determined to be a lesion core non-destructive block, the image block is encoded in JPEG-LS lossless mode and the encoding mode identifier is written into the block file header. At the same time, the lesion mask block and the key anatomical structure mask block corresponding to the image block are generated into a mask file using slice run-length encoding and compressed using Zstandard before being written. When an image block is determined to be a near-lossless block with uncertain boundaries, the image block is encoded using JPEG-LS near-lossless mode with the maximum absolute error threshold set to [value missing]. The maximum absolute error threshold refers to the absolute difference between the reconstructed grayscale and the original grayscale at each voxel, where the difference does not exceed [a certain threshold]. And the boundary uncertainty map corresponding to the image block. Cut off to Press back The uncertainty is quantized into an 8-bit unsigned integer block and written to the uncertainty file using Zstandard compression. At the same time, the corresponding mask file is written in a lossless manner. When an image block is identified as a protected block adjacent to a critical structure, the image block is encoded using JPEG-LS near-lossless mode with a maximum absolute error threshold set to [value missing]. Contour preservation storage is performed on key anatomical structure mask blocks. Contour preservation storage refers to performing contour preservation storage on each frame numbered... Two-dimensional slice mask Perform 8-neighbor boundary tracing to generate a closed contour point sequence and index it using voxel indexes. The integer sequence is written into the contour segment, and then all contour segments are Z-standard compressed to form a contour file, thus ensuring that the mask contours of key anatomical structures can be accurately reconstructed during decoding and version backtracking; when an image block is determined to be a background compression block, the background encoding parameters of the block are determined according to the storage budget mapping table and background encoding is performed. The storage budget mapping table is based on the available storage budget for the background area. The total prime number of the background region The number of bits available per voxel calculated Input and output the maximum absolute error threshold of the background. ,in In the formula This indicates the number of bits available per voxel, expressed in bits per voxel. This indicates the available storage budget for the background area, expressed in bytes. The background region represents a total number of prime numbers and is an integer. The mapping table is fixed as follows: Time to take ,when Time to take ,when Time to take And using JPEG-LS near-lossless mode to Encode the background compression block; After encoding each image block, a compressed block file is generated and its content hash is calculated. Then, based on the anchor key structure, the anchor key of the compressed block file is generated: ; in For instance identification, Image block center voxel index The corresponding frame number, Image block center voxel index The patient coordinate system three-dimensional spatial coordinates are obtained through mapping relationships. RID is the region of interest number, and when the image patch does not belong to any region of interest... To obtain the content hash by performing SHA-256 on the byte sequence of the compressed block file, Version number; Compressed block files, mask files, uncertainty files, or profile files are written to the storage medium in object storage mode and anchored. As the primary key, it is written into the key-value index table to realize the binding of "anchor key-file location". At the same time, the measurement result file and the analysis conclusion file are written into the same key-value index table with the anchor key of their corresponding nodes respectively, and the bound anchor key is recorded in the reference edge attribute of the semantic anchoring object graph. Scoring of abnormal slices With frame number Write to the key slice index file, which records all elements that meet the criteria. A set of frame numbers and a set of Regions of Interest (RIDs) covered by these frame numbers are provided to support retrieval and location.

[0030] S106. Establish a key slice index based on the abnormal slice score. When an image retrieval request is received, determine the abnormal slice based on the key slice index, and load the image block corresponding to the abnormal slice and its associated data through the anchor key. After receiving a complete sequence loading request, load the remaining image frames.

[0031] The key slice index is a data structure used to quickly locate abnormal image slices. It uses the frame number of the abnormal slice that meets the abnormality judgment criteria as the index content, and simultaneously records the set of Regions of Interest (RIDs) covered by these frame numbers. In this way, when doctors need to review the examination results, they can first determine the image slices and their corresponding regions of interest that need to be retrieved first based on the key slice index, without having to traverse the entire 3D image sequence first.

[0032] Upon receiving an image retrieval request, the system receives the target information and target version number carried in the request. First, it reads the key slice index file and parses the semantic anchoring object graph according to the frame number order in the key slice index to obtain a set of compressed block anchor keys associated with the key slices. Then, it prioritizes loading the compressed block files corresponding to these key slices, along with their corresponding mask files, uncertainty files, or contour files, through a key-value index table. Simultaneously, it loads the corresponding measurement result files and analysis conclusion files based on the RID reference edges in the semantic anchoring object graph, and decodes and renders the loaded data.

[0033] For example, when a neurosurgical MRI scan generates multiple consecutive image frames, if some slices have high abnormality scores, the system first reads the frame numbers of these slices from the key slice index. Then, based on the frame number and RID, it locates the corresponding compressed blocks, lesion masks, key anatomical structure masks, measurement results, and analysis conclusions. The physician first obtains these image contents directly related to the abnormal area. When further observation of the complete three-dimensional brain image is needed, a complete sequence loading request is sent to the system. The system then loads the compressed block files corresponding to the remaining image frames in batches according to frame number order, thus completing the retrieval of the entire three-dimensional volumetric data.

[0034] In version retrieval scenarios, if the retrieval request specifies a particular version, the system directly locates the corresponding version's node set from the semantic anchoring object graph based on the specified version number, and parses the image blocks, masks, measurement results, and analysis conclusions corresponding to that version through the object graph edge table. If the retrieval request includes a retrieval time parameter, the system selects the latest version's node set prior to that time point for loading. If the recalculated content hash after loading data is inconsistent with the anchor key record, or if the data version number does not match, the system traces back to the historical version according to the aforementioned version evolution relationship and generates an associated verification record. In this way, the storage, updating, reanalysis, and retrieval of image data can all be completed under a unified version relationship.

[0035] Thus, this embodiment forms a complete data processing closed loop: First, an image frame set with a unified spatial reference is formed through DICOM instance parsing, frame sorting, 3D rigid body registration, and unified resampling; then, lesion masks, key anatomical structure masks, boundary uncertainty maps, abnormal slice scores, measurement results, and analysis conclusions are obtained through a 3D neurosurgical image semantic analysis model; next, regions of interest are constructed using the semantic results of lesions and key structures, and a semantic anchoring object graph is established using anchor keys containing instance identifiers, frame numbers, spatial coordinates, RIDs, content hashes, and version numbers; based on this, image blocks are layered according to lesion distance, segmentation probability entropy, key structure distance, and lesion coverage ratio, and encoded using lossless, near-lossless, protective, and background compression methods respectively; finally, the encoding results are uniformly associated and stored with the semantic analysis data, and a key slice index is established based on the abnormal slice score. During retrieval, abnormal slices and their associated data are loaded first, and then the complete image sequence is loaded as needed. The data generated in the previous step serves as the basis for the processing in the next step, forming a continuous data processing chain for image spatial localization, semantic analysis, data association, hierarchical storage, and retrieval.

[0036] Through the aforementioned image analysis and 3D registration process, different image frames can be unified to the same patient coordinate system and standard voxel space. A stable image localization basis is established through instance identification, frame number, and spatial coordinates, providing a unified benchmark for the spatial correspondence of lesions, key anatomical structures, and image blocks. Through the 3D semantic analysis process, lesion regions, key anatomical structure regions, boundary uncertainty, and abnormal slice degree can be obtained simultaneously. Furthermore, measurement results such as lesion volume, maximum lesion diameter, and minimum distance from the lesion to key anatomical structures are generated, providing semantic basis for subsequent storage strategies derived from the image content itself. Through regions of interest and semantic anchoring object maps, image frames, image blocks, lesion and key structural regions, and measurement results can be integrated. The results and analysis conclusions are organized in a unified manner, and the relationship between data content and version is continuously verified using content hashing and version numbers. Through a hierarchical storage strategy, image blocks of different semantic importance can be entered into corresponding encoding paths. The core lesion area is encoded using lossless encoding, the area with uncertain boundaries is encoded using near-lossless encoding, the area adjacent to the key structure is encoded using protective encoding, and the background area is compressed and encoded according to the storage budget. Through key slice indexing and priority retrieval mechanism, abnormal slices and their associated images and analysis data can be obtained first when doctors retrieve images, and then the complete sequence can be loaded according to the actual request. This allows the image analysis results to participate in the determination of storage fidelity level and retrieval order, forming a complete technical closed loop from image analysis to storage to retrieval.

[0037] Based on the same inventive concept, this application also provides a neurosurgical image data intelligent storage device for implementing the above-mentioned intelligent storage method for neurosurgical image data. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the neurosurgical image data intelligent storage device provided below can be found in the limitations of the neurosurgical image data intelligent storage method described above, and will not be repeated here.

[0038] In one embodiment, such as Figure 2 As shown, a smart storage device for neurosurgical image data is provided, the device comprising: The sequence acquisition module 30 is used to acquire neurosurgical image sequences, perform instance parsing and three-dimensional registration on the neurosurgical image sequences, and generate a set of image frames containing instance identifiers, frame numbers and spatial coordinate information. The model training module 31 is used to input the set of image frames into the trained three-dimensional neurosurgical image semantic analysis model to obtain lesion masks, key anatomical structure masks, boundary uncertainty maps, abnormal slice scores, measurement results and analysis conclusions. The region determination module 32 is used to determine the region of interest based on the lesion mask and the key anatomical structure mask, and organize the image frame, region of interest, measurement results and analysis conclusions into a semantic anchoring object graph, and generate anchor keys for objects in the semantic anchoring object graph that include instance identifier, frame number, spatial coordinates, region of interest number, content hash and version number; The grade determination module 33 is used to determine the fidelity level of the image block based on the lesion mask distance field, segmentation probability entropy, key anatomical structure distance and abnormal slice score, and divide the image block into lesion core non-destructive block, boundary uncertain near non-destructive block, key structure adjacent protection block and background compression block according to the fidelity level, and generate a hierarchical storage strategy. The associated storage module 34 is used to encode each image block according to the hierarchical storage strategy, and to associate and store the encoded image blocks, as well as the corresponding masks, measurement results and analysis conclusions according to the anchor key. The index building module 35 is used to build a key slice index based on the abnormal slice score. When an image retrieval request is received, the abnormal slice is determined according to the key slice index, and the image block corresponding to the abnormal slice and its associated data are loaded through the anchor key. After receiving a complete sequence loading request, the remaining image frames are loaded.

[0039] This application also provides an electronic device, in some embodiments, referring to... Figure 3 As shown, the electronic device 700 includes an input unit 710, a memory 720, a processor 730, and an output unit 740. The memory 720 stores program instructions that can be executed on the processor 730. The processor 730 can execute the intelligent storage method and / or technical solution for neurosurgical image data based on the foregoing embodiments by calling the program instructions. The electronic device 700 can be a mobile terminal device such as a mobile phone or a computer.

[0040] Furthermore, embodiments of this application also provide a computer-readable storage medium for storing a computer program that executes a method for intelligent storage of neurosurgical image data. For example, computer program instructions, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. The program instructions that invoke the methods of this application may be stored in a fixed or removable storage medium, and / or transmitted via data streams in broadcast or other signal carrying media, and / or stored in a storage medium that operates according to the program instructions.

[0041] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0042] The technical features of the above embodiments can be arbitrarily integrated. For the sake of brevity, not all possible integrations of the technical features in the above embodiments are described. However, as long as the integration of these technical features does not contradict each other, they should be considered to be within the scope of this specification.

[0043] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for intelligent storage of neurosurgical image data, characterized in that, The method includes: Obtain neurosurgical image sequences, perform instance parsing and three-dimensional registration on the neurosurgical image sequences, and generate a set of image frames containing instance identifiers, frame numbers and spatial coordinate information; The image frame set is input into the trained three-dimensional neurosurgical image semantic analysis model to obtain lesion mask, key anatomical structure mask, boundary uncertainty map, abnormal slice score, measurement results and analysis conclusions; The region of interest is determined based on the lesion mask and the key anatomical structure mask, and the image frame, the region of interest, the measurement results and the analysis conclusion are organized into a semantic anchoring object graph. An anchor key containing instance identifier, frame number, spatial coordinates, region of interest number, content hash and version number is generated for the objects in the semantic anchoring object graph. The fidelity level of the image block is determined based on the lesion mask distance field, segmentation probability entropy, distance to key anatomical structures, and abnormal slice score. Based on the fidelity level, the image block is divided into lesion core non-destructive block, boundary uncertain near non-destructive block, key structure adjacent protection block, and background compression block, generating a hierarchical storage strategy. Each image block is encoded according to the hierarchical storage strategy, and the encoded image blocks, along with their corresponding masks, measurement results, and analysis conclusions, are stored together according to the anchor key. A key slice index is established based on the abnormal slice score. When an image retrieval request is received, the abnormal slice is determined based on the key slice index, and the image block corresponding to the abnormal slice and its associated data are loaded through the anchor key. After receiving a complete sequence loading request, the remaining image frames are loaded.

2. The intelligent storage method for neurosurgical image data as described in claim 1, characterized in that, The step of performing instance analysis and 3D registration on the neurosurgical image sequence to generate an image frame set includes: The instance identifier, sequence identifier, slice thickness, pixel spacing, patient coordinate direction, and patient coordinate origin of each image frame in the neurosurgical image sequence are read. The slice stacking direction is determined according to the patient coordinate direction, and the image frames are sorted according to their positions in the slice stacking direction to generate frame numbers. The sorted image frames are stacked to form initial three-dimensional volume data. The spatial direction of the first image frame is used as a unified direction reference. Three-dimensional rigid body registration is performed on the initial three-dimensional volume data, and the registered three-dimensional volume data is resampled to a unified voxel spacing. Based on the mapping relationship between voxel indexes and patient coordinates under a uniform voxel spacing, a frame index record is generated for each image frame. The frame index record includes at least the instance identifier, the sequence identifier, the frame number, and the corresponding slice center spatial coordinates. The frame index record is then written into the image frame set.

3. The intelligent storage method for neurosurgical image data as described in claim 1 or 2, characterized in that, The three-dimensional neurosurgical image semantic analysis model, trained by inputting the set of image frames, obtains lesion masks, key anatomical structure masks, boundary uncertainty maps, abnormal slice scores, measurement results, and analysis conclusions, including: The grayscale texture features, spatial location features and context features of the three-dimensional volume data are extracted by the multi-scale encoder of the three-dimensional neurosurgical image semantic analysis model, and the lesion probability map and the key anatomical structure probability map are output by the lesion segmentation decoding branch and the key anatomical structure segmentation decoding branch, respectively. Thresholding is performed on the lesion probability map and the key anatomical structure probability map respectively to obtain the lesion mask and the key anatomical structure mask; The probability entropy is calculated based on the category probability of each voxel in the lesion probability map, and the probability entropy is normalized to obtain the boundary uncertainty map. The three-dimensional volume data is divided into two-dimensional slices according to the frame number. The corresponding abnormal slice score is obtained by weighting the lesion volume ratio, the mean of boundary uncertainty, and the overlap ratio of key structures in each two-dimensional slice. The measurement results are calculated based on the lesion mask and the key anatomical structure mask, and the analysis conclusion is generated based on the measurement results and the abnormal slice scoring matching analysis conclusion template.

4. The intelligent storage method for neurosurgical image data as described in claim 1, characterized in that, The step of determining the region of interest based on the lesion mask and the key anatomical structure mask, and constructing the semantic anchoring object map, includes: Three-dimensional connected component labeling is performed on the lesion mask and the key anatomical structure mask respectively, and the obtained connected components are determined as candidate regions of interest; The voxel indices in each candidate region of interest are converted into patient coordinates. The three-dimensional bounding box and centroid of the corresponding candidate region of interest are determined based on the converted spatial coordinates. The candidate regions of interest are fused according to the spatial overlap or spatial distance between different candidate regions of interest to obtain a set of regions of interest. A unique region of interest number is assigned to each region of interest. The three-dimensional volume data is divided into multiple non-overlapping image blocks. Based on the voxel overlap relationship between each image block and the region of interest, the image block containing the voxel of the region of interest is associated with the corresponding region of interest, and the center spatial coordinates of the image block are recorded. Image frames, regions of interest, measurement items, and analysis conclusions are respectively used as nodes in the semantic anchoring object graph. An inclusion edge is established between the image frame and the region of interest, and a reference edge is established between the measurement item and the analysis conclusion and the region of interest it references. An anchor key is generated for each node, and the anchor key is used as the index of the corresponding node. The anchor key includes instance identifier, frame number, spatial coordinates, region of interest number, content hash and version number.

5. The intelligent storage method for neurosurgical image data as described in claim 1, characterized in that, The process of determining the fidelity level of an image block based on the lesion mask distance field, segmentation probability entropy, distance to key anatomical structures, and abnormal slice scores, and generating a hierarchical storage strategy, includes: The lesion mask distance field is determined based on the shortest voxel distance between the center of the image block and the boundary of the lesion mask. The segmentation probability entropy corresponding to the image block is calculated based on the category probabilities of the lesion probability map and the key anatomical structure probability map. The key anatomical structure distance is determined based on the shortest voxel distance between the center of the image block and the key anatomical structure mask. Calculate the lesion mask coverage ratio in each image block and classify each image block according to a preset priority; Image blocks whose lesion mask coverage reaches the first coverage threshold are identified as the core non-destructive blocks of the lesion. Image blocks that are not identified as the core non-destructive block of the lesion, whose segmentation probability entropy reaches the entropy threshold, and whose distance from the boundary of the lesion mask does not exceed the first distance threshold are identified as near-non-destructive blocks with uncertain boundaries. Image blocks that are not identified as the core non-invasive block of the lesion or the near non-invasive block with indeterminate boundaries, and whose distance from the mask of the key anatomical structure does not exceed the second distance threshold, are identified as the key structure adjacent protection blocks. Image blocks not identified by the above image block types are identified as background compression blocks, and the block number, block center position, and image block type of each image block are written into the hierarchical storage strategy list.

6. The intelligent storage method for neurosurgical image data as described in claim 5, characterized in that, The step of encoding each image block according to the hierarchical storage strategy and storing the encoded image blocks in association includes: The core non-destructive block of the lesion is encoded using a non-destructive encoding method, and the corresponding original grayscale value is saved; The near-lossless block with boundary uncertainty is encoded using a near-lossless encoding method with a first maximum absolute error constraint, and the corresponding boundary uncertainty map is saved. The critical structure adjacent protection block is encoded using a near-lossless encoding method with a second maximum absolute error constraint, and the corresponding critical anatomical structure mask is stored with contour preservation. The encoding parameters of the background compression block are determined based on the storage budget of the background region, and the background compression block is compressed and encoded according to the encoding parameters. After encoding each image block, a corresponding compressed block file is generated and the content hash of the compressed block file is calculated. The compressed block file, the corresponding mask file, and the content hash are then associated and stored with the anchor key in the semantic anchoring object graph.

7. An intelligent storage device for neurosurgical image data, characterized in that, The device includes: The sequence acquisition module is used to acquire neurosurgical image sequences, perform instance parsing and three-dimensional registration on the neurosurgical image sequences, and generate a set of image frames containing instance identifiers, frame numbers and spatial coordinate information. The model training module is used to input the set of image frames into the trained three-dimensional neurosurgical image semantic analysis model to obtain lesion masks, key anatomical structure masks, boundary uncertainty maps, abnormal slice scores, measurement results, and analysis conclusions. The region determination module is used to determine the region of interest based on the lesion mask and the key anatomical structure mask, and organize the image frame, the region of interest, the measurement result and the analysis conclusion into a semantic anchoring object graph, and generate anchor keys for the objects in the semantic anchoring object graph, including instance identifier, frame number, spatial coordinates, region of interest number, content hash and version number; The grading module is used to determine the fidelity level of an image block based on the lesion mask distance field, segmentation probability entropy, distance to key anatomical structures, and abnormal slice score. Based on the fidelity level, the image block is divided into a lesion core non-destructive block, a near-non-destructive block with uncertain boundaries, a key structure adjacent protection block, and a background compression block, and a hierarchical storage strategy is generated. The associated storage module is used to encode each image block according to the hierarchical storage strategy, and to associate and store the encoded image blocks, as well as the corresponding masks, measurement results and analysis conclusions according to the anchor key. The index building module is used to build a key slice index based on the abnormal slice score. When an image retrieval request is received, the abnormal slice is determined based on the key slice index, and the image block corresponding to the abnormal slice and its associated data are loaded through the anchor key. After receiving a complete sequence loading request, the remaining image frames are loaded.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the intelligent storage method for neurosurgical image data as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the intelligent storage method for neurosurgical image data as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the intelligent storage method for neurosurgical image data as described in any one of claims 1 to 6.

Images