Bone tumor CT image segmentation method and system based on image processing

Abstract

The invention relates to the technical field of medical image processing, and discloses a bone tumor CT image segmentation method and system based on image processing. The method comprises the following steps: establishing a bone tumor CT image feature expression structure containing a plurality of image region units, a region connection relationship and attribute description, and updating the structure based on a real-time image data stream containing a CT imaging sequence, pixel intensity distribution and the like; key segmentation boundaries, low-efficiency segmentation regions and difficult segmentation points are identified through an image analysis algorithm, and an initial segmentation strategy group covering region merging, boundary optimization and processing sequence adjustment schemes is generated; receiving a user interaction instruction, judging a correction operation category, and adjusting the initial segmentation strategy group in real time; the adjusted strategy group is implemented, and the feature expression structure is updated according to execution effect feedback. The method can adapt to complex features of the bone tumor CT image, and the segmentation flexibility and pertinence are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, in particular to a bone tumor CT image segmentation method and system based on image processing. BACKGROUND

[0002] In the clinical diagnosis and treatment planning of bone tumors, CT images have become the core means for obtaining key information such as tumor location, size and infiltration range due to their high-resolution imaging advantage of bone structure, and accurate CT image segmentation is an important prerequisite for subsequent medical analysis and clinical decision-making. With the rapid development of medical imaging technology, the efficiency of acquiring bone tumor CT images is continuously improving, but the imaging quality is easily affected by various factors, such as differences in bone morphology of different patients, density overlap between tumor tissue and normal bone tissue, fluctuations in imaging device parameters, and scanning environment interference, etc. These factors all bring many challenges to the accurate segmentation of CT images.

[0003] Currently, existing bone tumor CT image segmentation methods mostly rely on fixed feature extraction patterns, and the established image feature structures often cannot dynamically adapt to real-time changing image data streams. Such fixed structures usually only focus on single-dimensional information such as pixel intensity, ignoring the correlation between regional units and key factors such as imaging environment parameters, resulting in insufficient comprehensiveness and timeliness of feature expression, and it is difficult to accurately reflect the complex characteristics of bone tumor CT images. In the segmentation boundary recognition link, traditional algorithms often have problems such as fuzzy positioning of key boundaries, misjudgment of low-efficiency segmentation regions, and omission of segmentation difficulties, making the subsequent generated segmentation strategy lack of pertinence and unable to effectively deal with complex scenarios such as blurred boundaries between bone tumor tissue and surrounding normal tissue and uneven internal texture of the tumor.

[0004] The initial strategy design of existing segmentation methods is mostly single, usually only containing a single region processing or boundary optimization scheme, which is difficult to comprehensively cover various needs in the bone tumor CT image segmentation process. More critically, most methods lack effective interaction mechanisms with users, and cannot respond to personalized correction needs raised by doctors in actual operation in a timely manner, resulting in possible deviations between the segmentation results and the actual clinical needs. In addition, existing methods generally do not form a complete feedback loop, and the effect after segmentation execution cannot timely benefit the optimization of feature expression structure, making the adaptability of segmentation strategy difficult to continuously improve, and prone to problems such as decreased segmentation accuracy and redundant process in long-term application, which cannot meet the efficiency and accuracy requirements of bone tumor CT image segmentation in clinical scenarios. SUMMARY

[0005] The present application relates to the technical field of medical image processing, in particular to a bone tumor CT image segmentation method and system based on image processing.

[0006] To achieve the above object, the application provides a bone tumor CT image segmentation method based on image processing, which comprises the following steps: A feature expression structure of the bone tumor CT image is established, and the feature expression structure is updated based on real-time image data flow, wherein the feature expression structure comprises a plurality of image region units, connection relationships between the region units, and attribute descriptions of each region unit, and the real-time image data flow comprises CT imaging sequences, pixel intensity distribution, tissue texture features, and imaging environment parameters; Key segmentation boundaries, inefficient segmentation regions, and segmentation difficulties in the feature expression structure are identified through an image analysis algorithm; An initial segmentation strategy group is generated based on the key segmentation boundaries and the segmentation difficulties, wherein the initial segmentation strategy group comprises region merging schemes, boundary optimization schemes, and processing sequence adjustment schemes; User interaction instructions are received, and a correction operation category is determined based on the user interaction instructions, and the initial segmentation strategy group is adjusted in real time based on the correction operation category; the initial segmentation strategy group after the real-time adjustment is implemented, and the feature expression structure is updated according to the execution effect feedback.

[0007] Preferably, the process of establishing the feature expression structure of the bone tumor CT image comprises the following steps: A historical CT image data set is obtained, wherein the historical CT image data set relates to tumor shape changes, tissue structure distribution, and existing segmentation annotation information; multi-angle feature mining is performed on the historical CT image data set to generate attribute descriptions of the region units, wherein the attribute descriptions comprise region categories, gray scale statistical features, texture change degrees, historical segmentation accuracy rates, and dynamic importance measures, wherein the dynamic importance measures are obtained by weighted accumulation of the historical segmentation accuracy rate component, the current segmentation priority level component, and the region usage rate component; a network layout of the feature expression structure is constructed based on an image storage architecture, wherein the region units are connected to each other through connection edges, the connection edges represent region adjacency and connection tightness, each region unit is associated with an attribute description, and a complete unit representation is formed, which comprises region category labels, gray scale feature values, texture complexity indexes, historical segmentation accuracy rate values, and dynamic importance measure parameters.

[0008] Preferably, the connection tightness is determined in the following manner: Based on historical CT image data sets, the pixel interaction frequency and interaction scale between adjacent region units are counted; the connection tightness value is calculated by combining the pixel interaction frequency, interaction scale and attribute description of the region unit, and the calculation process includes dividing the pixel interaction frequency by the maximum frequency reference value to obtain the normalized frequency ratio, dividing the pixel interaction scale by the maximum scale reference value to obtain the normalized scale ratio, taking the average value of the gray feature values of the source region unit and the target region unit as the contribution measure, multiplying the frequency ratio, scale ratio and contribution measure by the preset weight coefficient respectively and summing them up, and the weight coefficient sum is one; the connection tightness value is used to assign a weight level to the connection edge, if the value reaches or exceeds the high threshold value, it is marked as high connection tightness, if the value is between the medium threshold value and the high threshold value, it is marked as medium connection tightness, and if the value is lower than the medium threshold value, it is marked as low connection tightness; the high threshold value is adjusted according to the dynamic importance measure, and only when the connection tightness value exceeds the adjusted high threshold value is it marked as a key connection path.

[0009] Preferably, the process of identifying key segmentation boundaries, inefficient segmentation regions and determining segmentation difficulty points in the feature expression structure by the image analysis algorithm comprises: A depth-first traversal algorithm is used to scan the feature expression structure, and all complete segmentation paths from the image edge to the tumor core are extracted; the image feature transmission efficiency of each segmentation path is calculated; the key segmentation path and the inefficient segmentation path are labeled based on the image feature transmission efficiency, specifically, the image feature transmission efficiency value is compared with the efficiency standard range, which includes the high efficiency standard value and the low efficiency standard value, if the efficiency value is greater than or equal to the high efficiency standard value, it is marked as a key segmentation path, if the efficiency value is less than the low efficiency standard value, it is marked as an inefficient segmentation path; difference analysis is performed on the key segmentation path and the inefficient segmentation path to confirm the segmentation difficulty point, the connection edge with the highest connection tightness value in the key segmentation path is marked as the core connection path, and additional feature redundancy is assigned to it, and the dynamic importance measure of all region units in the inefficient segmentation path is extracted, the unit with the lowest dynamic importance measure is selected, if the measures of multiple units are the same, the historical segmentation accuracy is compared, and the unit with the lowest accuracy is confirmed as the segmentation difficulty point; the core connection path information and the segmentation difficulty point information are stored in association.

[0010] Preferably, the calculation of the image feature transmission efficiency comprises: The dynamic importance metrics of all region units in the extraction path are extracted, and the geometric mean of the dynamic importance metrics of the units is calculated; the weight levels of all connection edges in the extraction path are extracted, and the arithmetic mean of the connection tightness values is calculated; the physical length of the path is counted, and the physical length of the path is defined as the sum of the spatial distances between adjacent units in the path; the image feature transmission efficiency is calculated through comprehensive evaluation logic, the geometric mean is multiplied by the arithmetic mean to obtain the comprehensive evaluation value, the physical length of the path is added to a minimum normal number to obtain the length adjustment value, and the image feature transmission efficiency value is obtained by dividing the comprehensive evaluation value by the length adjustment value; the higher the image feature transmission efficiency value, the better the feature transmission ability and stability of the path.

[0011] Preferably, the specific implementation of generating the initial segmentation strategy group comprises: A region merging scheme is generated for the inefficient segmentation path, the inefficient segmentation path and its segmentation difficulty point information are extracted, adjacent units of the segmentation difficulty point are traversed, adjacent units with feature redundancy satisfying a preset condition are screened, the screening condition is that the current feature utilization rate of the adjacent unit is lower than the utilization rate threshold and the gray feature value is greater than the current load, the feature gap of the segmentation difficulty point, i.e. the difference between the current demand feature amount and the current supply feature amount, is calculated, the feature demand is allocated based on the feature redundancy capacity of the adjacent unit and the connection tightness value, the adjacent unit with high connection tightness value is preferentially selected, the allocation amount does not exceed the upper limit of the percentage of the feature redundancy capacity of the adjacent unit, the feature allocation instruction is generated and added to the initial segmentation strategy group; A boundary optimization scheme is generated for the key segmentation path, the key segmentation path and its core connection path information are extracted, the current spatial length of the core connection path and the external imaging environment data are obtained, if the current spatial length exceeds the length threshold or the external imaging risk level is higher than the risk threshold, the path shortening strategy and the risk avoidance strategy are executed, the path shortening strategy searches for a connection path with a connection tightness value not lower than the original path and a shorter spatial length in the network layout, if no alternative path is found, a warning is triggered and an optimization suggestion report is generated, the risk avoidance strategy calculates the risk coefficient of the unit in the path based on the historical CT image data, screens the units with a risk coefficient lower than the threshold and adjacent in space as backup units, recalculates the connection tightness value and the spatial length after replacing the high-risk unit, generates the boundary adjustment instruction and adds it to the initial segmentation strategy group.

[0012] Preferably, the process of receiving the user interaction instruction and determining the correction operation category according to the user interaction instruction comprises: The system monitors user input devices in real time and captures user interaction commands. It parses the operation type and target parameters of these commands. The operation types include region selection, boundary correction, weight adjustment, and strategy confirmation. The target parameters include image coordinates, region unit identifiers, and segmentation strategy numbers. Based on a predefined command-operation mapping table, the parsed operation types are mapped to correction operation categories, including local optimization, global adjustment, strategy switching, and weight reallocation. The system then calls the corresponding strategy adjustment algorithm to adjust the initial segmentation strategy group in real time according to the correction operation category.

[0013] Preferably, the process of updating the feature representation structure based on execution effect feedback includes: After implementing the segmentation strategy, the segmentation result image is acquired, and the overlap between the actual segmented region and the expected segmented region is calculated as the segmentation accuracy. Resource consumption data during the segmentation process is monitored, including computation time and memory usage. An execution performance score is generated based on the segmentation accuracy and resource consumption data. The dynamic importance metric of relevant regional units and the connection tightness of connecting edges are adjusted according to the execution performance score. If the execution performance score is higher than the performance threshold, the dynamic importance metric of regional units in the involved path is increased, and the connection tightness level of relevant connecting edges is improved. If the execution performance score is lower than the performance threshold, a reconstruction check of the feature representation structure is triggered, and the attribute descriptions and connection relationships of regional units associated with paths with low scores are recalculated and updated.

[0014] Preferably, the generation of the initial segmentation strategy group based on key segmentation boundaries and segmentation difficulties further includes a processing order adjustment scheme, the generation process of which includes: Obtain the image feature transfer efficiency values ​​of all critical segmentation paths and inefficient segmentation paths; sort all paths according to the image feature transfer efficiency values ​​to generate a path processing sequence, with the sorting rule being that critical segmentation paths take precedence over inefficient segmentation paths, and paths of the same type with higher image feature transfer efficiency values ​​are processed first; assign a processing priority label to each path in the path processing sequence; generate a processing order adjustment scheme containing path processing order instructions based on the processing priority labels, and add it to the initial segmentation strategy group.

[0015] Preferably, the present invention also includes an image processing-based CT image segmentation system for bone tumors, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the image processing-based CT image segmentation method for bone tumors as described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By constructing a dynamically updated feature representation structure, comprehensive capture and real-time adaptation of multi-dimensional information from bone tumor CT images are achieved. This feature representation structure not only includes the basic attributes of image region units but also covers the connectivity relationships between regions. Furthermore, it is dynamically updated by combining multi-source real-time data streams such as CT imaging sequences, pixel intensity distribution, tissue texture features, and imaging environment parameters. This fully reflects the complex characteristics and dynamic changes of bone tumor CT images, making the description of image features more complete and timely, effectively avoiding the segmentation bias problems caused by information lag or missing dimensions in traditional fixed feature structures.

[0017] By leveraging image analysis algorithms to accurately identify key segmentation boundaries, inefficient segmentation regions, and segmentation difficulties, precise targeting is provided for the formulation of subsequent segmentation strategies. This targeted identification method can quickly pinpoint the core points and weak links in the segmentation process, avoiding the waste of resources and inefficiency caused by blindly formulating segmentation strategies. This allows the generated initial segmentation strategy set to directly address segmentation pain points, improving the rationality and effectiveness of the segmentation scheme from the source. The initial segmentation strategy set covers multiple schemes such as region merging, boundary optimization, and processing order adjustment, forming a comprehensive segmentation strategy system that can fully cope with various scenarios that may occur in the segmentation of bone tumor CT images, breaking the limitations of traditional single strategies that are difficult to adapt to complex segmentation needs.

[0018] The integration of user interaction commands makes the segmentation process more flexible and personalized, enabling timely responses to clinicians' professional judgments and correction needs during actual operations. By determining the type of correction operation based on user commands and adjusting the initial segmentation strategy group in real time, the segmentation strategy can closely align with actual application scenarios and professional needs. This effectively compensates for the shortcomings of automatic algorithm segmentation in adapting to personalized clinical needs, making the segmentation results more consistent with actual medical diagnostic standards.

[0019] By implementing the adjusted segmentation strategy group and updating the feature representation structure based on feedback from the execution effect, a complete closed-loop mechanism of "construction-identification-strategy-adjustment-feedback-update" is formed. This closed-loop design allows the feature representation structure and segmentation strategy to be continuously optimized during ongoing application, gradually improving its adaptability to different imaging conditions, different tumor types, and individual patient differences. In clinical applications, this method simplifies the segmentation process, reduces invalid segmentation operations, makes doctors' operations more efficient and convenient, and improves the reliability and consistency of segmentation results. It better adapts to the actual needs of CT image segmentation in bone tumor diagnosis and treatment planning, contributing to the precise conduct of clinical work. In addition, this method has good adaptability to bone tumor CT images of different qualities and features, and can maintain stable segmentation performance in complex imaging environments, expanding its applicability in practical clinical applications. Attached Figure Description

[0020] Figure 1 This is a diagram illustrating the structural analysis of the feature representation. Figure 2 A flowchart for establishing the feature representation structure of bone tumor CT images; Figure 3 A flowchart for calculating image feature transfer efficiency; Figure 4 The diagram shows the generation and optimization effect analysis of the initial segmentation strategy group. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0022] Please see Figure 1 This invention provides a bone tumor CT image segmentation method based on image processing. The method includes constructing a dynamic feature representation structure and combining it with real-time image data streams to achieve accurate segmentation. A feature representation structure for the bone tumor CT image is established, comprising multiple image region units, the connections between these units, and attribute descriptions for each unit. The real-time image data stream includes CT imaging sequences, pixel intensity distribution, tissue texture features, and imaging environment parameters. The feature representation structure is updated based on the real-time image data stream to ensure it reflects the latest image characteristics. Key segmentation boundaries, inefficient segmentation regions, and segmentation difficulties are identified using image analysis algorithms. The algorithms scan the feature representation structure to extract segmentation paths and evaluate their efficiency. An initial segmentation strategy group is generated based on the key segmentation boundaries and segmentation difficulties. This initial strategy group includes region merging schemes, boundary optimization schemes, and processing order adjustment schemes. User interaction commands are received from input devices such as mice or touchscreens. The type of correction operation is determined based on the user interaction commands, including local optimization or global adjustment. The initial segmentation strategy group is adjusted in real-time based on the correction operation category. This adjustment process involves modifying the region merging scheme or reallocating processing priorities. The initial segmentation strategy group, adjusted in real time, is implemented. The process involves performing segmentation operations and generating segmentation results. Based on feedback regarding the execution results, the feature representation structure is updated. This feedback includes segmentation accuracy and resource consumption data. The update operation adjusts the attribute descriptions and connectivity relationships of the region cells.

[0023] Example 1: See Figure 2Establishing the feature representation structure of bone tumor CT images begins with acquiring a historical CT image dataset. This dataset contains information on tumor shape changes, tissue structure distribution, and existing segmentation annotations. Data sources include hospital PACS systems and public medical imaging databases, and all data formats conform to the DICOM standard. Multi-dimensional feature mining is performed on the historical CT image dataset, covering spatial domain analysis, frequency domain analysis, and texture domain analysis. Spatial domain analysis primarily detects pixel gradient changes and region boundary features. Frequency domain analysis uses Fast Fourier Transform to identify periodic patterns in the images. Texture domain analysis calculates contrast and energy features using the gray-level co-occurrence matrix. The attribute description of the generated region units contains five key components. Region category labels adopt a three-level classification system, including bone tissue, soft tissue, and tumor regions. The gray-level statistical feature vector consists of three statistics: mean, variance, and skewness. The texture complexity coefficient is calculated based on entropy. Historical segmentation accuracy values ​​are derived from expert annotation verification results. The dynamic importance metric is fused using a linear weighted formula, incorporating components of historical segmentation accuracy, current segmentation priority, and region utilization. The preprocessing of historical CT image datasets comprises three key steps. Noise filtering utilizes anisotropic diffusion algorithms to effectively preserve edge information. Intensity normalization employs histogram matching to eliminate differences between different devices. Resolution unification is achieved through bicubic interpolation, scaling the images to standard pixel spacing. The implementation of multi-angle feature mining requires the construction of a complete feature extraction pipeline. This pipeline integrates two major technical approaches: traditional image processing algorithms and deep learning models. Traditional image processing algorithms include SIFT feature points and HOG feature descriptors. The deep learning model uses a pre-trained ResNet network to extract high-level semantic features.

[0024] The division of region units is based on a superpixel segmentation algorithm, which generates uniform image patches, each corresponding to a region unit. The calculation process for attribute descriptions follows strict mathematical definitions and standardized procedures. The network layout of the feature representation structure uses a graph structure for storage. Nodes in the graph represent region units, and edges represent spatial adjacency relationships. Node attributes store all feature data of the region unit, while edge attributes store connection density values ​​and weight levels. The image storage architecture design fully considers the needs of large-scale data processing. The distributed graph database Neo4j is used for efficient data management. The network layout construction algorithm includes three core steps: node creation, edge establishment, and attribute binding. The real-time image data stream update mechanism adopts an incremental update strategy. This strategy monitors newly arriving CT image sequences and automatically extracts feature vectors from the new images. The system inserts new nodes into the graph structure and dynamically adjusts the connection density values ​​of existing edges. Quality control for multi-angle feature mining includes two important steps: feature consistency verification evaluates feature stability by calculating intragroup correlation coefficients; and redundant feature removal uses principal component analysis for dimensionality reduction.

[0025] The dynamic importance metric for region units is updated every 10 new CT slices processed, using a sliding window mechanism to ensure the latest metric values ​​are always maintained. Network layout optimization includes a community detection algorithm to identify subgraph structures with low feature transfer efficiency. Compliance processing of the historical CT image dataset includes de-identification, effectively removing all patient privacy information and ensuring data security. The parallel computing architecture for multi-angle feature mining employs a MapReduce programming model, significantly improving processing efficiency. Version management of the feature representation structure uses a Git-based workflow. Each major update creates a new version branch, ensuring clarity in version management. The incremental update mechanism for the historical CT image dataset periodically scans newly acquired case data, automatically triggering feature recalculation. The entire process of establishing the feature representation structure forms a complete technical solution chain, involving multiple disciplines such as medical imaging, computer vision, and distributed computing. Quality control standards for the historical CT image dataset strictly adhere to ACR guidelines to ensure data quality. The establishment of the feature representation structure provides structured data support for subsequent segmentation algorithms. Through the synergistic application of multiple technologies, the system ultimately achieves accurate representation and efficient management of bone tumor CT image features. This system meets the dual requirements of real-time performance and accuracy for clinical applications.

[0026] Example 2: Connection density determination method is based on historical CT image datasets, which contain complete pixel-level interaction records and spatial topological information. The frequency of pixel interactions between adjacent regions is statistically analyzed using a sliding time window method, setting an observation window of 30 consecutive CT frames to calculate the number of pixel value changes. Interaction scale quantification combines morphological operations and region growing algorithms. Connected regions are detected through structuring elements, and the region growing algorithm expands the marked interaction region area from the seed point. The calculation is performed by comprehensively considering pixel interaction frequency, interaction scale, and region unit attribute descriptions (including grayscale feature values ​​and texture complexity indicators). The calculation process performs standardization processing: the pixel interaction frequency is divided by the dynamically updated maximum frequency benchmark value to obtain the normalized frequency ratio, and the pixel interaction scale is divided by the maximum scale benchmark value normalized according to the image resolution scale to obtain the normalized scale ratio. The grayscale feature values ​​of the source and target regions are calculated using a weighted average method considering region area weights to calculate the contribution metric. The frequency ratio, scale ratio, and contribution metric are weighted and summed using fixed weight coefficients optimized by a grid search method.

[0027] The constraint that the sum of weight coefficients is one is achieved through a normalization function, projecting any weight vector onto the unit hyperplane. A three-level classification system is used to assign weight levels based on connectivity density values: values ​​reaching or exceeding the high threshold of the 85th percentile in the historical dataset are marked as high connectivity density; values ​​between the 50th percentile threshold and the high threshold are marked as medium connectivity density; and values ​​below the medium threshold are marked as low connectivity density. The high threshold is adaptively adjusted based on the real-time dynamic importance metric of region units, and only when the connectivity density value exceeds the adjusted high threshold is it marked as a critical connection path. Connectivity density values ​​are updated synchronously with the image data stream, triggering recalculation with each new CT slice. The weight level allocation results affect the topological characteristics of the feature representation structure, thus determining the segmentation path selection strategy. Pixel interaction frequency statistics use time series analysis to detect periodic patterns, and interaction scale quantification includes pixel count area calculation and Fourier descriptor shape analysis. Version control is used to record change history for region unit attribute description updates, and the normalized weighting algorithm ensures numerical accuracy through double-precision floating-point operations. The maximum frequency baseline value is dynamically adjusted using an exponentially weighted moving average method to assign higher weights to recent data. The maximum scale baseline value is set to distinguish anatomical structures such as cortical bone, cancellous bone, and soft tissue. The contribution metric calculation incorporates spatial correlation analysis to detect the consistency of gray-level distribution between regions. The preset weight coefficients are optimized using a k-fold cross-validation method to divide the dataset into 10 mutually exclusive subsets.

[0028] The three-level classification system implementation includes a threshold learning phase, employing a supervised classification algorithm to train threshold parameters. Initial high threshold values ​​are calibrated using statistical hypothesis testing to verify significance, while the middle threshold is determined using clustering analysis to automatically identify natural boundary points. Dynamic importance measurement adjusts the high threshold using a linear transformation model to map threshold adjustment coefficients, and key connection path markings are stored in a graph database supporting complex queries. Connection density numerical storage uses a columnar database format to improve query efficiency, and the weighting logic includes edge case handling to ensure classification accuracy. Pixel interaction frequency statistical analysis includes a Poisson distribution fit test, and interaction size calculation introduces a 3D voxel analysis method to quantify the volume of interaction in three-dimensional space. Updates to region unit attribute descriptions trigger an event-driven mechanism to respond to data changes, and the computational efficiency of the normalized weighted algorithm is optimized through SIMD instruction vectorization. The data processing pipeline includes four stages: data acquisition, preprocessing, feature extraction, and numerical computation. An automatic annotation verification system ensures the quality of historical CT image datasets. Pixel interaction frequency statistical accuracy is improved through millisecond-level time sampling rates, and the accuracy of interaction size measurement is compared with expert annotation results using a multi-observer verification method. Consistency in attribute descriptions is ensured through data standardization protocols to guarantee process standardization, while the computation speed of the normalized weighted algorithm is optimized through JIT compilation technology.

[0029] The Paxos distributed consensus algorithm is used to coordinate the data views of multiple nodes to achieve global consistency of the maximum frequency baseline value. The adaptive adjustment of the maximum scale baseline value is optimized based on environmental feedback using a reinforcement learning strategy. The computational complexity of contribution metrics is reduced by feature hashing technology to decrease dimensionality, and the interpretability of weight coefficients is achieved by generating a decision model using the LIME local interpretation method. The performance evaluation of the three-level classification system uses PR curve analysis to analyze the balance between precision and recall. The initial setting of high threshold values ​​utilizes a multi-expert consensus mechanism to integrate clinical experience, while the dynamic adjustment of medium threshold values ​​uses a sliding window statistical method to track changes in data distribution. The persistence of key connection path markers uses WAL (Write-Ahead Log) technology to ensure data atomicity, and the connection tightness values ​​are normalized to the [0,1] interval using the sigmoid function. The weight level allocation decision adopts a multi-rule fusion strategy combining statistical rules and machine learning, and the spatiotemporal pattern analysis of pixel interaction frequency uses a 3D convolutional neural network to extract joint features.

[0030] Example 3: See Figure 3The process of identifying key segmentation boundaries, inefficient segmentation regions, and segmentation difficulties in the feature representation structure using image analysis algorithms employs a depth-first search algorithm to scan the feature representation structure. This algorithm systematically explores paths along connecting edges, starting from image edge nodes. All complete segmentation paths from image edges to the tumor core are extracted. A complete segmentation path must contain a continuous sequence of regional units and have a path length greater than a minimum threshold. The image feature transfer efficiency of each segmentation path is calculated using a comprehensive evaluation model that combines the dynamic characteristics and topological properties of the path. The formula for calculating the image feature transfer efficiency is:

[0031] in: This represents the image feature transfer efficiency value, which ranges from 0 to 1. This represents the dynamic importance measure of the i-th region unit in the path, and this value is directly obtained from the attribute description of the feature representation structure. This represents the total number of region units contained in the current path. This represents the connection tightness value of the j-th connecting edge. This represents the total number of connecting edges in the path. This represents the physical length of the path, which is the sum of the Euclidean distances between all adjacent cells in the path. This is the reference length. It is a very small positive number, set to 10. -6 To avoid division by zero errors. and The values ​​are 0.6 and 0.4, respectively, which are used to adjust the relative weights of dynamic importance measurement and connection tightness in efficiency calculation.

[0032] The A* algorithm for path shortening uses Euclidean distance as a heuristic function to estimate the cost of the remaining path. The search range is limited to the neighborhood of the original path, and the neighborhood size is optimized through experimental data. Warning message generation includes problem description and cause analysis, using natural language generation techniques. The optimization suggestion report structure covers current status analysis, improvement suggestions, and expected effect evaluation; the current status analysis is based on current path characteristics. The risk coefficient calculation for the risk avoidance strategy uses a logistic regression model, with training data from historical segmentation records. Backup unit selection criteria include spatial proximity and feature compatibility; spatial proximity requires unit distances to be within a specified range. Replacement operations update the topological relationships of the feature representation structure to ensure data integrity. Boundary adjustment command verification includes path connectivity testing and feature consistency checks; connectivity testing ensures the new path is unbroken. The event monitoring mechanism adopts the observer pattern to decouple event generation and processing logic. Interruption handler priority settings are based on the urgency of the operation to ensure real-time user interaction. Pattern recognition technology feature extraction includes gesture recognition and voice command parsing; gesture recognition supports multi-touch operations. Operation type classification uses a support vector machine model, with training samples from user operation history.

[0033] The instruction-operation mapping table maintenance provides a graphical configuration interface, allowing administrators to adjust mapping relationships. Local optimization category processing focuses on specific regional units, determined by coordinate positioning. Global adjustment category execution requires a full map traversal to update all relevant paths. Strategy switching category implementation uses conditional statements to evaluate the effect of the current strategy. The weight redistribution category algorithm uses gradient descent optimization to find the optimal weight combination. Feature allocation instruction execution monitors resource allocation effects, measured by feature utilization metrics. Region merging scheme effectiveness is evaluated using segmentation quality metrics, including the Dice coefficient and Hausdorff distance. Boundary optimization scheme performance testing includes processing time measurement using a high-precision timer. User interaction instruction response time is required to be less than 100 milliseconds, and response time testing simulates real-world usage scenarios. Initial segmentation strategy group persistent storage uses version control to record strategy group change history. Real-time adjustment conflict resolution uses an optimistic locking mechanism to detect concurrent modification conflicts. Feature representation structure updates are synchronized to ensure data consistency, implemented through transaction processing. System performance monitoring includes memory usage and CPU load metrics; memory usage monitoring prevents resource leaks.

[0034] The region merging scheme and boundary optimization scheme work together, taking into account the execution order, which affects the final segmentation result. The user interaction command processing flow includes an error recovery mechanism to handle abnormal input situations. The initial segmentation strategy group verification test includes boundary condition checks to test extreme input situations. System integration testing covers all correction operation categories to ensure that all modules work together normally. Feature allocation command table query optimization uses composite indexes to accelerate queries by time range and unit identifier. The path shortening strategy caching mechanism stores frequently used path results to reduce redundant calculation overhead. The risk avoidance strategy model update adopts an online learning approach to adapt to changes in data distribution. User interaction command security checks prevent malicious input, including parameter validation and access control. The initial segmentation strategy group export function supports multiple formats, including XML and JSON standard formats. System logs record detailed operation traces for auditing and troubleshooting. Performance optimization measures include database index optimization and algorithm complexity analysis; index optimization improves query speed. The user interface design follows ergonomic principles to improve operational efficiency. The specific implementation of generating the initial segmentation strategy group and the user interaction command processing form a complete closed loop, achieving a seamless connection from automatic generation to manual optimization. The region merging scheme addresses the problem of inefficient path resource allocation, while the boundary optimization scheme improves the accuracy of critical path segmentation. User interaction command processing ensures system flexibility and operability, allowing for real-time adjustments to meet individualized clinical needs.

[0035] Example 4: Specific Implementation of Initial Segmentation Strategy Group Generation. A region merging scheme is generated for inefficient segmentation paths. Information on inefficient segmentation paths is obtained from the feature representation structure analysis module, which outputs a path efficiency evaluation report. Extracting information about inefficient segmentation paths and their segmentation difficulties requires parsing the path topology data file, which is stored in Protocol Buffers serialization format. An improved breadth-first search algorithm is used to traverse the adjacent units of segmentation difficulties, with a maximum search depth of 5 neighboring units. A three-level filtering process is executed to filter adjacent units whose feature redundancy meets preset conditions. This process sequentially checks feature utilization status, grayscale feature value matching degree, and memory usage. The filtering condition is that the current feature utilization rate of the adjacent unit is lower than the dynamic utilization rate threshold and the difference in grayscale feature values ​​is less than the tolerance value. The dynamic utilization rate threshold is automatically adjusted between 30% and 70% based on system load. A supply-demand balance model is used to calculate the feature gap at segmentation difficulties, calculating the logarithmic difference between the target feature quantity and the currently held feature quantity. Based on the feature redundancy capacity and connectivity density values ​​of adjacent units, feature redundancy capacity is allocated according to the requirements. The feature redundancy capacity is updated every 200 milliseconds via a real-time monitoring system. The top 10% of adjacent units in terms of connectivity density are prioritized, and this ranking is obtained using a fast selection algorithm. The allocation amount is limited to 40% of the feature redundancy capacity of adjacent units, with this 40% upper limit enforced through a memory protection mechanism. Feature allocation instructions are generated and added to the initial segmentation strategy group. These instructions encapsulate semantic information in JSON-LD format.

[0036] For key segmentation paths, boundary optimization schemes are generated. Information about key segmentation paths is obtained from the message queue of the path analysis module. Extracting key segmentation paths and their core connection path information requires parsing graph database query results. The graph database query uses Cypher language to describe path patterns. The current spatial length of the core connection path and external imaging environment data are obtained. The current spatial length is calculated by accumulating the Euclidean distances of each segment. If the current spatial length exceeds the adaptive length threshold or the external imaging risk level is higher than the dynamic risk threshold (the adaptive length threshold is set to 20% of the image diagonal length), then path shortening and risk avoidance strategies are executed in parallel. The path shortening strategy initiates a multi-threaded search for optimized paths. The path shortening strategy searches for alternative paths in the feature representation structure using Algorithm A, whose heuristic function combines spatial distance and connection density weights. If the search times out and no alternative path is found, a three-level alarm mechanism is triggered, taking different measures based on the severity. The risk avoidance strategy calculates the risk probability of each unit in the path based on a Bayesian network. The Bayesian network training data includes historical segmentation failure cases. Candidate cells with a risk probability below 0.05 and a spatial distance within 5 mm are selected as backup cells. These candidate cells must pass topology consistency verification. After replacing a high-risk cell, the associated connectivity density value is updated immediately, using incremental calculations to reduce redundant computations. A boundary adjustment instruction is generated and added to the initial segmentation strategy group. This instruction includes a version number and digital signature.

[0037] Referring to Table 1, a multi-channel event monitoring system is established to receive user interaction commands and determine the correction operation category based on these commands. This system simultaneously handles mouse, keyboard, stylus, and voice input. User interaction command capture employs an interrupt-driven architecture, ensuring a response latency of less than 50 milliseconds. Parsing the operation type and target parameters of user interaction commands requires multimodal fusion analysis, which combines input device and interaction context. Operation types include region selection, boundary correction, weight adjustment, and policy confirmation, identified using a neural network classifier. Target parameters include coordinate points in the image coordinate system, region unit UUID identifiers, and segmentation policy hash values, parsed using regular expression matching. The parsed results are mapped to correction operation categories based on a predefined command-operation mapping table, which uses a red-black tree structure for fast lookup and updates. Correction operation categories include local optimization, global adjustment, policy switching, and weight redistribution, with combined operations achieved through bitmask encoding. The initial segmentation strategy group is adjusted in real time based on the corresponding strategy adjustment algorithm according to the correction operation category. The strategy adjustment algorithm adopts the differential evolution optimization method. The specific implementation of the region merging scheme is recorded in the feature allocation instruction table.

[0038] Table 1: Structure of Feature Allocation Instruction Table Field name Data type Explanation Constraint condition Instruction number BIGINT Instruction unique identifier Primary key, automatically incremented Source unit identifier UUID Feature supply unit identifier Foreign key, reference area unit table Target unit identifier UUID Feature receiving unit identifier Foreign key, reference area unit table Matching feature amount DECIMAL(5,4) Feature transfer amount Value range 0.0001-0.9999 Connection tightness requirement DECIMAL(3,2) Minimum connection tightness Must be greater than 0.50 Timestamp TIMESTAMP Instruction generation time Default current timestamp Execution status ENUM Instruction execution status pending, running, completed, failed The specific implementation of the boundary optimization scheme relies on a path reconstruction engine, which integrates multiple graph algorithm libraries. The current spatial length of the core connecting path is measured using Runge-Kutta integration, which improves the accuracy of curve length calculation. External imaging environment data includes noise distribution maps, contrast histograms, and resolution parameters; the noise distribution map is obtained through wavelet transform analysis. The calculation of the adaptive length threshold considers the image sampling theorem, which requires the minimum path length to be greater than twice the pixel spacing. The dynamic risk threshold is determined using a value-at-risk (VAT) model, which calculates the maximum possible loss at a 95% confidence level.

[0039] The A* algorithm for path shortening employs bidirectional search optimization, simultaneously searching from both the path's start and end points. Feasibility verification of alternative paths includes topology checks and smoothness tests; topology checks ensure path continuity without breaks. A three-tiered alarm mechanism includes logging, email notifications, and automatic system repair, with automatic repair attempting methods from a standard solution library. Optimization suggestion reports are generated using a template engine, which converts data structures into natural language descriptions. The Bayesian network training for risk avoidance strategies utilizes the EM algorithm, which handles missing values ​​in the training data. Candidate unit selection involves multi-metric comprehensive evaluation, weighted by calculating proximity and feature similarity. Transaction processing for replacement operations guarantees atomicity and consistency; atomicity is achieved through database transactions. Digital signatures for boundary adjustment instructions use the RSA algorithm, verifying instruction integrity and source reliability. The multi-channel event monitoring system employs the Reactor design pattern, handling high-concurrency input events. The interrupt-driven architecture prioritizes hardware and software interrupts; hardware interrupts respond to device inputs, while software interrupts handle logical events. The time window for multimodal fusion analysis is set to 500 milliseconds, and input events occurring within this 500-millisecond period are analyzed for correlation. The neural network classifier is trained using the cross-entropy loss function, which optimizes classification accuracy.

[0040] The instruction-operation mapping table is persistently stored using an SQLite database, which supports transaction operations and crash recovery. Local optimization is handled using gradient descent, which finds the optimal solution within the local neighborhood. Global adjustment requires a distributed computing framework, which distributes tasks across multiple computing nodes. Policy switching is implemented using a state machine model, which manages the transition logic between different policies. Weight redistribution uses Monte Carlo tree search, balancing exploration and utilization. The feature tuning instruction table provides a RESTful API for querying, supporting HTTP protocol calls. Cross-validation is used to evaluate the effectiveness of the region merging scheme, dividing the dataset into training and test sets. Flame graphs are used for performance analysis of the boundary optimization scheme, visualizing function call latency. Load testing is used to test the response time of user interaction instructions, simulating multi-user concurrent scenarios.

[0041] The initial segmentation strategy group uses the Git version control system for version management, which records the differences between each strategy modification. Real-time conflict detection employs a vector clock algorithm to detect the temporal order of concurrent operations. Feature representation structure updates are synchronized using the Paxos consensus algorithm, ensuring data consistency across multiple replicas. System monitoring metrics include memory usage, CPU load, and network I / O. Memory usage monitoring utilizes a garbage collection mechanism. The path shortening strategy's caching implementation uses the LRU eviction algorithm, retaining the most recently used path result. Risk mitigation strategies utilize an online learning framework for model updates, incrementally updating model parameters. Secure verification of user interaction commands uses a JWT token mechanism to verify user identity and permissions. The feature allocation command table backup strategy employs an off-site disaster recovery solution, ensuring data security and recoverability.

[0042] The complete process for generating the initial segmentation strategy group includes three main stages: parameter initialization, strategy generation, and effect evaluation. The parameter initialization stage loads system configurations and model parameters. The strategy generation stage processes optimization schemes for multiple paths in parallel. The effect evaluation stage employs a multi-index comprehensive evaluation system. The user interaction command processing pipeline includes four stages: input parsing, permission verification, operation execution, and result feedback. The input parsing stage decodes the input data format. The entire system adopts a microservice architecture to achieve high availability, deploying functional modules as independent services.

[0043] See Figure 4This chart demonstrates the generation process and optimization effect analysis of the initial segmentation strategy group. The bar chart shows the feature transfer efficiency values ​​of each segmentation path, with bars of different gray levels distinguishing between critical paths, inefficient paths, and ordinary paths. The markers in the chart indicate the application effects of the region merging scheme and the boundary optimization scheme, and the threshold line clarifies the criteria for path classification. The data in the chart covers the feature transfer efficiency of segmentation paths, path classification results, the application location of optimization strategies, and the expected improvement effect. This data reflects the system's ability to identify critical segmentation boundaries and inefficient segmentation regions, as well as the effectiveness of generating targeted optimization strategies. This chart allows for a comprehensive evaluation of the rationality of the segmentation strategy and the feasibility of the optimization scheme, providing data support for subsequent adjustments to user interaction.

[0044] Example 5: The complete implementation process of a bone tumor CT image segmentation system is demonstrated using a specific case, P2024CT015. The system first loads DICOM format CT image data, and the medical image processing library parses a 512×512 pixel matrix and a 1.25 mm slice thickness parameter. The feature representation structure contains 327 region units and 896 connecting edges. Each unit stores a 128-dimensional grayscale feature vector and a 256-dimensional texture descriptor. An improved depth-first search algorithm is used for path traversal, setting a recursive depth of 50 layers and employing memoization to avoid loops. The path analysis module identifies 3 key segmentation paths and 2 inefficient segmentation paths. In the inefficient PATH-7 path, the system locates the segmentation difficulty point U-115 unit with a dynamic importance metric of 0.23. By traversing adjacent units using a six-neighbor connectivity model, 3 qualified units with feature utilization rates below 65% and grayscale differences less than 0.15 are selected. The supply and demand balance algorithm calculates a feature gap of 0.31. The weighted round-robin scheduling is used to allocate feature amount according to connection density: 0.28 feature amount is allocated to unit U-119 (connection strength 0.87), and 0.03 feature amount is supplemented to unit U-121.

[0045] The PATH-3 critical path optimization employs a multi-objective engine to handle a path length of 47.3 mm and metal artifact interference. A bidirectional search A* algorithm, combining Euclidean distance and connectivity weights, generates an optimized path of 39.8 mm. Monte Carlo simulation runs for 1000 iterations, replacing high-risk unit U-087 (risk coefficient 0.91) with unit U-090 (risk coefficient 0.03). The user interface is implemented using HTML5 technology, supporting 10-point touch operation. A convolutional neural network recognizes the physician's selection operation, triggering an adaptive region growing algorithm. The conjugate gradient method adjusts the weights of 23 units within 10 iterations, and WebGL renders the segmentation results in real time. The system records parameters such as touch pressure and motion trajectory, and an incremental learning algorithm updates the segmentation strategy. A distributed architecture processes 14 feature allocation instructions, and multi-threaded computation shortens the critical path by 18.7%. A message queue ensures sequential execution of instructions, and an event-driven architecture handles 35 interaction requests. The final output is in DICOM-SEG format, with a directed acyclic graph managing 8 optimization steps.

[0046] Performance monitoring showed a processing time of 2 minutes and 17 seconds, with a peak memory usage of 1.2GB. The feature representation structure employs hierarchical storage, and a caching mechanism accelerates path lookup. A PID controller enables real-time strategy adjustment, and anomaly handling ensures data integrity. The system supports functional expansion through modular design, and verification tools ensure output conformity to the DICOM standard. The system automates the process from image loading to result generation. The feature representation structure supports precise path analysis, and the user interaction mechanism incorporates clinical experience, with multi-objective optimization balancing segmentation accuracy and efficiency. The processing results of case P2024CT015 validate the system's practicality and reliability in a clinical environment. The entire system adopts a microservice architecture to ensure high availability, with modules communicating through standard interfaces. The medical image processing library supports multiple CT equipment formats and features automatic calibration. The path analysis algorithm has been optimized to adapt to tumor lesions of different sizes. The feature allocation mechanism enables dynamic resource allocation, adjusting parameters according to real-time needs. The user interface has been optimized for human factors, reducing operator fatigue. The strategy generation module integrates multiple machine learning algorithms to continuously improve segmentation accuracy. The result output module provides multiple data formats for convenient subsequent statistical analysis. The system logs record the complete operation process, meeting medical audit requirements. Through continuous iterative optimization, the system has demonstrated stable performance in clinical applications across different hospitals.

[0047] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A bone tumor CT image segmentation method based on image processing, characterized in that, The method includes: A feature representation structure for bone tumor CT images is established, and the feature representation structure is updated based on real-time image data stream. The feature representation structure includes multiple image region units, the connection relationships between region units, and the attribute description of each region unit. The real-time image data stream includes CT imaging sequence, pixel intensity distribution, tissue texture features, and imaging environment parameters. Image analysis algorithms are used to identify key segmentation boundaries and inefficient segmentation regions in the feature representation structure and to determine segmentation difficulties. An initial segmentation strategy group is generated based on key segmentation boundaries and segmentation difficulties. The initial segmentation strategy group includes region merging scheme, boundary optimization scheme and processing order adjustment scheme. Receive user interaction instructions, determine the correction operation category based on the user interaction instructions, and adjust the initial segmentation strategy group in real time based on the correction operation category; implement the initial segmentation strategy group after real-time adjustment, and update the feature expression structure based on the execution effect feedback.

2. The bone tumor CT image segmentation method based on image processing according to claim 1, characterized in that, The process of establishing the feature representation structure of bone tumor CT images includes: A historical CT image dataset is acquired, which includes tumor shape changes, tissue structure distribution, and existing segmentation annotation information. Multi-angle feature mining is performed on the historical CT image dataset to generate attribute descriptions for region units. These attribute descriptions include region category, grayscale statistical features, texture variation degree, historical segmentation accuracy, and dynamic importance metric. The dynamic importance metric is obtained by weighted summation of historical segmentation accuracy, current segmentation priority, and region utilization rate. A network layout for feature representation is constructed based on the image storage architecture. In this network layout, region units are interconnected through connecting edges, which represent region adjacency and connection tightness. Each region unit is associated with an attribute description, forming a complete unit representation containing region category label, grayscale feature value, texture complexity index, historical segmentation accuracy value, and dynamic importance metric parameters.

3. The bone tumor CT image segmentation method based on image processing according to claim 2, characterized in that, The method for determining the connection tightness is as follows: Based on historical CT image datasets, the frequency and scale of pixel interactions between adjacent regions are statistically analyzed. Combining the pixel interaction frequency, interaction scale, and attribute descriptions of the regions, the connectivity density is calculated. The calculation process includes dividing the pixel interaction frequency by the maximum frequency benchmark to obtain the normalized frequency ratio, dividing the pixel interaction scale by the maximum scale benchmark to obtain the normalized scale ratio, taking the average grayscale feature values ​​of the source and target regions as the contribution measure, and multiplying the frequency ratio, scale ratio, and contribution measure by preset weight coefficients and summing them. The total sum of the weight coefficients is 1. The connection edges are assigned weight levels based on the connection density value. If the value reaches or exceeds the high threshold, it is marked as high connection density. If the value is between the medium and high thresholds, it is marked as medium connection density. If the value is below the medium threshold, it is marked as low connection density. The high threshold is adjusted according to the dynamic importance metric. Only when the connection density value exceeds the adjusted high threshold is it marked as a critical connection path.

4. The bone tumor CT image segmentation method based on image processing according to claim 2, characterized in that, The process of identifying key segmentation boundaries and inefficient segmentation regions in the feature representation structure and determining segmentation difficulties using image analysis algorithms includes: A depth-first search algorithm is used to scan the feature representation structure and extract all complete segmentation paths from the image edge to the tumor core. The image feature transfer efficiency of each segmentation path is calculated. Based on the image feature transfer efficiency, key segmentation paths and inefficient segmentation paths are marked. Specifically, the image feature transfer efficiency value is compared with an efficiency standard range, which includes both high-efficiency and low-efficiency standard values. If the efficiency value is greater than or equal to the high-efficiency standard value, it is marked as a key segmentation path; if the efficiency value is less than the low-efficiency standard value, it is marked as an inefficient segmentation path. Difference analysis is performed on key and inefficient segmentation paths to identify segmentation difficulties. The connection edges with the highest connection density in the key segmentation paths are marked as core connection paths, and additional feature redundancy is allocated to them. For inefficient segmentation paths, the dynamic importance metric of all regional units is extracted, and the unit with the lowest dynamic importance metric is selected. If multiple units have the same metric, the historical segmentation accuracy is compared, and the unit with the lowest accuracy is identified as a segmentation difficulty point. The core connection path information and the segmentation difficulty point information are stored together.

5. The bone tumor CT image segmentation method based on image processing according to claim 4, characterized in that, The calculation of the image feature transfer efficiency includes: Extract the dynamic importance metric of all regional units in the path and calculate the geometric mean of the dynamic importance metric of the units; extract the weight level of all connecting edges in the path and calculate the arithmetic mean of the connection density values; calculate the physical length of the path, which is defined as the sum of the spatial distances between adjacent units in the path; calculate the image feature transfer efficiency through a comprehensive evaluation logic, multiply the geometric mean and the arithmetic mean to obtain the comprehensive evaluation value, add the physical length of the path to the minimum normal number to obtain the length adjustment value, and divide the comprehensive evaluation value by the length adjustment value to obtain the image feature transfer efficiency value. The higher the image feature transfer efficiency value, the better the feature transfer capability and stability of the path.

6. The bone tumor CT image segmentation method based on image processing according to claim 5, characterized in that, The specific implementation of generating the initial segmentation strategy group includes: For the inefficient segmentation path generation region merging scheme, extract the inefficient segmentation path and its segmentation difficulty point information, traverse the adjacent units of the segmentation difficulty point, and filter the adjacent units whose feature redundancy meets the preset conditions. The filtering conditions are that the current feature utilization rate of the adjacent unit is lower than the utilization rate threshold and the grayscale feature value is greater than the current load. Calculate the feature gap of the segmentation difficulty point, that is, the difference between the current demand feature quantity and the current supply feature quantity. Based on the feature redundancy capacity and connection tightness value of the adjacent units, allocate feature demand, give priority to adjacent units with high connection tightness value, and the allocation amount does not exceed the percentage upper limit of the feature redundancy capacity of the adjacent units. Generate feature allocation instructions and add them to the initial segmentation strategy group. For the critical segmentation path boundary optimization scheme, the critical segmentation path and its core connection path information are extracted, and the current spatial length and external imaging environment data of the core connection path are obtained. If the current spatial length exceeds the length threshold or the external imaging risk level is higher than the risk threshold, the path shortening strategy and risk avoidance strategy are executed. The path shortening strategy searches for alternative paths with a connection density value not lower than the original path and a shorter spatial length through network layout. If no alternative path is found, a warning is triggered and an optimization suggestion report is generated. The risk avoidance strategy calculates the risk coefficient of the unit in the path based on historical CT image data, selects the unit with a risk coefficient lower than the threshold and spatially adjacent as the backup unit, replaces the high-risk unit, recalculates the connection density value and spatial length, generates a boundary adjustment instruction and adds it to the initial segmentation strategy group.

7. The bone tumor CT image segmentation method based on image processing according to claim 1, characterized in that, The process of receiving user interaction instructions and determining the correction operation category based on the user interaction instructions includes: The system monitors user input devices in real time and captures user interaction commands. It parses the operation type and target parameters of these commands. The operation types include region selection, boundary correction, weight adjustment, and strategy confirmation. The target parameters include image coordinates, region unit identifiers, and segmentation strategy numbers. Based on a predefined command-operation mapping table, the parsed operation types are mapped to correction operation categories, including local optimization, global adjustment, strategy switching, and weight reallocation. The system then calls the corresponding strategy adjustment algorithm to adjust the initial segmentation strategy group in real time according to the correction operation category.

8. The bone tumor CT image segmentation method based on image processing according to claim 1, characterized in that, The process of updating the feature representation structure based on execution effect feedback includes: After implementing the segmentation strategy, the segmentation result image is acquired, and the overlap between the actual segmented region and the expected segmented region is calculated as the segmentation accuracy. Resource consumption data during the segmentation process is monitored, including computation time and memory usage. An execution performance score is generated based on the segmentation accuracy and resource consumption data. The dynamic importance metric of relevant regional units and the connection tightness of connecting edges are adjusted according to the execution performance score. If the execution performance score is higher than the performance threshold, the dynamic importance metric of regional units in the involved path is increased, and the connection tightness level of relevant connecting edges is improved. If the execution performance score is lower than the performance threshold, a reconstruction check of the feature representation structure is triggered, and the attribute descriptions and connection relationships of regional units associated with paths with low scores are recalculated and updated.

9. The bone tumor CT image segmentation method based on image processing according to claim 4, characterized in that, The initial segmentation strategy group generated based on key segmentation boundaries and segmentation difficulties also includes a processing order adjustment scheme, the generation process of which includes: Obtain the image feature transfer efficiency values ​​of all critical segmentation paths and inefficient segmentation paths; sort all paths according to the image feature transfer efficiency values ​​to generate a path processing sequence, with the sorting rule being that critical segmentation paths take precedence over inefficient segmentation paths, and paths of the same type with higher image feature transfer efficiency values ​​are processed first; assign a processing priority label to each path in the path processing sequence; generate a processing order adjustment scheme containing path processing order instructions based on the processing priority labels, and add it to the initial segmentation strategy group.

10. A bone tumor CT image segmentation system based on image processing, 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 steps of the image processing-based CT image segmentation method for bone tumors as described in any one of claims 1 to 9.

Images