Intelligent identification and real-time inventory image analysis system for surgical instruments

Abstract

This invention relates to the field of medical equipment management technology and discloses an intelligent surgical instrument identification and real-time inventory image analysis system. The system includes an image acquisition module, an image preprocessing module, a topological feature extraction module, an instrument identification module, a spatiotemporal trajectory analysis module, a state prediction module, and a decision support module. This system achieves high-precision identification of surgical instruments through topological invariant feature extraction technology; accurately identifies instrument usage anomalies using spatiotemporal topological manifold trajectory analysis; and achieves forward-looking prediction of instrument states based on topological manifold learning prediction technology. This overcomes the bottleneck of low accuracy in complex environments inherent in traditional image recognition methods, providing a new technical guarantee for surgical safety.

Description

Technical Field

[0001] This invention relates to the field of medical device management technology, and in particular to a surgical instrument intelligent identification and real-time inventory image analysis system based on topology theory, which is used to automatically identify, track the status of, and predict the risks of surgical instruments during surgery. Background Technology

[0002] Accurate inventory of surgical instruments is a crucial step in ensuring surgical safety. Traditional surgical instrument inventory relies primarily on manual counting, which is not only time-consuming and labor-intensive but also prone to errors under the stressful surgical environment, potentially leading to serious medical accidents such as instruments being left inside the body. With the development of computer vision technology, several image recognition-based surgical instrument inventory systems have been proposed. However, existing technologies have the following problems:

[0003] On the one hand, traditional image recognition methods rely too heavily on visual features, and their accuracy drops significantly under complex lighting conditions in the operating room and when instruments are partially obscured. On the other hand, existing systems are mostly limited to static instrument recognition and lack the ability to dynamically track the use of instruments and predict risks, thus failing to truly achieve forward-looking assurance of surgical safety.

[0004] Furthermore, most existing technologies employ shallow feature extraction and simple classification methods, which cannot effectively cope with complex scenarios where surgical instruments have similar shapes and change frequently in their usage. They also cannot understand the mathematical nature of the changes in instrument states, thus failing to achieve accurate prediction.

[0005] Therefore, there is an urgent need for an intelligent system that can achieve high-precision instrument identification, dynamic tracking, and risk prediction in complex environments. Summary of the Invention

[0006] The purpose of this invention is to provide a surgical instrument intelligent identification and real-time inventory image analysis system based on topology theory. By introducing innovative technologies such as topological invariant feature extraction, spatiotemporal topological manifold trajectory analysis, and topological manifold learning prediction, this invention solves the technical problems of low identification accuracy, lack of dynamic tracking and risk prediction capabilities in complex environments of existing technologies.

[0007] This invention proposes a surgical instrument intelligent identification and real-time inventory image analysis system, comprising:

[0008] The image acquisition module is used to acquire multi-angle real-time images in the operating room and transmit the multi-angle real-time images to the image preprocessing module;

[0009] An image preprocessing module, connected to the image acquisition module, is used to perform noise reduction, enhancement, and geometric correction on the received multi-angle real-time images, and transmit the processed images to the topological feature extraction module;

[0010] A topological feature extraction module, connected to the image preprocessing module, is used to extract the topological invariant features of surgical instruments from the processed image and construct a multi-scale feature descriptor.

[0011] The instrument identification module, connected to the topology feature extraction module, is used to match the multi-scale feature descriptor with a preset feature library to determine the type and status of the surgical instrument.

[0012] The spatiotemporal trajectory analysis module is connected to the instrument identification module and is used to construct a spatiotemporal topological manifold based on the identification results and to analyze the usage trajectory of surgical instruments through continuous homology analysis.

[0013] The state prediction module, connected to the spatiotemporal trajectory analysis module, is used to learn the state evolution law of the instrument on the topological manifold and predict the future state of the surgical instrument.

[0014] The decision support module, connected to the state prediction module, is used to assess the risk level based on the prediction results and trigger early warning information when an abnormal state is detected.

[0015] Preferably, the topological feature extraction module includes:

[0016] The point cloud conversion unit is used to convert the processed image into a point cloud representation;

[0017] A complex construction unit, connected to the point cloud transformation unit, is used to construct a multi-scale simple complex sequence based on the point cloud representation;

[0018] The feature calculation unit, connected to the complex construction unit, is used to calculate the topological invariants of the multi-scale simple complex sequence, including Betti number sequences of different dimensions;

[0019] The feature descriptor generation unit, connected to the feature calculation unit, is used to combine Betti number sequences of different dimensions into multi-scale feature descriptors.

[0020] Preferably, the spatiotemporal trajectory analysis module includes:

[0021] The spatiotemporal representation unit is used to construct the instrument identification results into an augmented state representation that includes state, time, and rate of change;

[0022] A manifold construction unit, connected to the spatiotemporal representation unit, is used to construct a time-varying spatiotemporal topological representation based on the augmented state representation;

[0023] A persistent cohomology computation unit, connected to the manifold construction unit, is used to compute the persistent cohomology features of the time-varying spatiotemporal topological representation;

[0024] The trajectory pattern recognition unit, connected to the continuous coherence calculation unit, is used to compare the continuous coherence features with a preset normal trajectory pattern to identify abnormal usage patterns.

[0025] Preferably, the state prediction module includes:

[0026] Manifold representation unit is used to reduce the dimensionality of historical instrument state data through local linear embedding and construct a state manifold representation space;

[0027] An evolutionary model unit, connected to the manifold representation unit, is used to learn state transition functions in the state manifold representation space;

[0028] A multi-scale prediction unit, connected to the evolution model unit, is used to predict the future state of the device on short-term, medium-term and long-term time scales based on the state transition function.

[0029] The risk assessment unit, connected to the multi-scale prediction unit, is used to calculate the distance between the predicted state and the safe state space, and to assess the potential risk level.

[0030] Preferably, the decision support module includes:

[0031] The risk level classification unit is used to classify risks into three levels: low, medium, and high, based on the distance and trend of the predicted state and the safe state space.

[0032] The early warning information generation unit is connected to the risk level classification unit and is used to generate corresponding early warning information for different risk levels.

[0033] An operation suggestion generation unit, connected to the early warning information generation unit, is used to generate targeted operation suggestions based on the early warning information and the current surgical stage.

[0034] The human-computer interaction unit, connected to the operation suggestion generation unit, is used to convey warning information and operation suggestions to medical staff through visual and auditory means.

[0035] Preferably, the complex building unit constructs a multi-scale simple complex sequence through the following steps:

[0036] Set an incremental sequence of scale parameters;

[0037] For each scale parameter, construct the corresponding Vietoris-Rips complex;

[0038] Organize the complexes corresponding to all scale parameters into nested sequences.

[0039] Preferably, the persistent cohomology calculation unit calculates the persistent cohomology characteristics through the following steps:

[0040] For each time window, simple complexes are used to calculate homology groups of different dimensions.

[0041] Track the birth and death times of homologous classes to generate persistent barcodes;

[0042] Features are extracted from the persistent barcode, including life length distribution, birth time distribution, and death time distribution.

[0043] Preferably, the method for the manifold representation unit to perform state manifold representation includes:

[0044] Construct a state k-nearest neighbor graph;

[0045] Calculate the Laplace operator on the k-nearest neighbor graph;

[0046] Solve for the eigenfunctions of the Laplace operator as the spectral representation of the manifold;

[0047] Project the original state onto the manifold representation space.

[0048] Preferably, a data management module is also included, the data management module comprising:

[0049] A real-time database is used to store the real-time instrument status, location, and usage during the current surgical procedure;

[0050] Historical database, used to store instrument usage data, status sequences, and abnormal records from historical surgeries;

[0051] A knowledge graph database is used to store relationships between instruments, usage guidelines, surgical procedures, and expert experience.

[0052] The data synchronization unit is used to coordinate the data interaction between the real-time database, the historical database, and the knowledge graph database.

[0053] Preferably, the image acquisition module includes:

[0054] A multi-angle high-definition camera, wherein the resolution of the multi-angle high-definition camera is not less than 4K and the frame rate is not less than 30 frames per second;

[0055] An image caching unit, connected to the multi-angle high-definition camera, is used to temporarily store recently acquired image sequences;

[0056] An image quality assessment unit, connected to the image buffer unit, is used to assess image sharpness, brightness, and contrast, and triggers re-acquisition when the image quality does not meet a preset threshold.

[0057] The present invention has the following beneficial effects:

[0058] 1. Significantly improved recognition accuracy: Through topological invariant feature extraction technology, the system can maintain a recognition accuracy of up to 98.5% even under complex conditions such as instrument rotation, partial occlusion, and changes in lighting, far exceeding the 87% of traditional methods;

[0059] 2. Significantly enhanced anomaly detection capability: Based on spatiotemporal topological manifold trajectory analysis, the system can accurately identify instrument usage behaviors that deviate from standard operating procedures, with an anomaly detection accuracy rate of 94.3%, and can detect anomalies on average 1.2 seconds in advance;

[0060] 3. Enhanced proactive risk warning: Through topological manifold learning prediction technology, the system can predict potential risks 3 to 5 seconds in advance, with a short-term prediction accuracy (3 seconds) of 92.1%, effectively preventing surgical accidents;

[0061] 4. Enhanced system stability and adaptability: The topology method is inherently robust to data changes, and the system can adapt to various surgical environments and instrument types. Clinical trials in five tertiary hospitals have shown that it can effectively reduce instrument legacy risks by 99.5%. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the overall framework of the intelligent surgical instrument identification and real-time inventory image analysis system of the present invention.

[0063] Figure 2 This is a schematic diagram of the topology feature extraction module of the present invention;

[0064] Figure 3 This is a schematic diagram of the spatiotemporal trajectory analysis module of the present invention;

[0065] Figure 4 This is a schematic diagram of the state prediction module of the present invention;

[0066] Figure 5 This is a schematic diagram of the decision support module of the present invention;

[0067] Figure 6 This is a schematic diagram of the data management module of the present invention;

[0068] Figure 7 This is a schematic diagram of the image acquisition module of the present invention. Detailed Implementation

[0069] Please refer to Figure 1 - Figure 7 The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the implementation of the present invention should not be limited to the following embodiments.

[0070] like Figure 1As shown, the intelligent surgical instrument identification and real-time inventory image analysis system provided by the present invention includes an image acquisition module 1, an image preprocessing module 2, a topological feature extraction module 3, an instrument identification module 4, a spatiotemporal trajectory analysis module 5, a state prediction module 6, and a decision support module 7.

[0071] Image acquisition module 1 is used to acquire multi-angle real-time images of the operating room and transmit these images to image preprocessing module 2. Preferably, image acquisition module 1 uses multiple high-definition cameras to capture the operating room scene from different angles, ensuring comprehensive monitoring of instruments. In a preferred embodiment of the invention, the acquisition frequency is set to 30 frames per second to ensure image continuity and real-time performance.

[0072] Image preprocessing module 2 is connected to image acquisition module 1 and is used to perform noise reduction, enhancement, and geometric correction on the received multi-angle real-time images, and then transmit the processed images to topological feature extraction module 3. Specifically, image preprocessing module 2 first performs Gaussian filtering on the original image to remove noise, then enhances the image contrast through adaptive histogram equalization, and finally performs perspective transformation to correct geometric distortion, ensuring the accuracy of subsequent feature extraction. In the unevenly lit operating room environment, the preprocessing module also performs illumination compensation processing to normalize the image brightness to a suitable range (usually a grayscale value of 128±20).

[0073] The topological feature extraction module 3 is connected to the image preprocessing module 2 and is used to extract the topological invariant features of the surgical instruments from the processed image and construct a multi-scale feature descriptor. For example... Figure 2 As shown, the topology feature extraction module 3 includes a point cloud conversion unit 31, a complex construction unit 32, a feature calculation unit 33, and a feature descriptor generation unit 34.

[0074] The point cloud conversion unit 31 is used to convert the processed image into a point cloud representation. Specifically, it first extracts image edges using an edge detection algorithm (such as the Canny algorithm, with a low threshold set to 50 and a high threshold set to 150), and then uniformly samples from the edge points to obtain the point cloud representation. , where each point It includes two-dimensional or three-dimensional spatial coordinates and corresponding pixel intensity values. Preferably, the point cloud size is controlled between 1000 and 5000 points to balance computational efficiency and feature extraction accuracy.

[0075] Complex construction unit 32 is connected to point cloud transformation unit 31 and is used to construct a multi-scale simple complex sequence based on point cloud representation. This unit achieves multi-scale simple complex construction through the following steps: First, an incremental sequence of scale parameters is set. ,in It is typically set to 0.5 times the minimum instrument feature size (approximately 5 mm). m Set to 0.2 times the maximum instrument size (approximately 50 mm), the sequence length m is typically 8-12; then, for each scale parameter... Construct the corresponding Vietoris-Rips complex That is, when the distance between two points is less than When the three points are connected to form an edge, the distance between any two points is less than 1. This forms a triangle, and so on; finally, a nested complex sequence is generated. .

[0076] Feature calculation unit 33 is connected to complex construction unit 32 and is used to calculate the topological invariants of the multi-scale simple complex sequence, including Betti number sequences of different dimensions. In this invention, the Betti number... Denotes the number of k-dimensional holes, where Indicates the number of connected components. Indicates the number of 1-dimensional rings. This represents the number of 2D cavities. For each complex... Calculate its 0-dimensional, 1-dimensional, and 2-dimensional Betty numbers to obtain the sequence { These topological invariants remain stable under conditions such as instrument rotation, scaling, or partial occlusion, providing robust assurance for subsequent identification.

[0077] The feature descriptor generation unit 34 is connected to the feature calculation unit 33, and is used to combine Betti number sequences of different dimensions into multi-scale feature descriptors. Specifically, for each dimension k, a topological feature vector is constructed. This represents the curve of the Betti number as a function of the scale parameter. Then, the feature vectors of each dimension are merged into a comprehensive feature descriptor. In practical applications, usually The weight is set relatively high (about 0.5) because connectivity features are the most critical for instrument shape representation.

[0078] The instrument identification module 4 is connected to the topological feature extraction module 3 and is used to match the multi-scale feature descriptors with a preset feature library to determine the type and status of the surgical instruments. This module first defines a topological feature distance function. This is used to measure the similarity between the current feature and features in the database, preferably using weighted Euclidean distance:

[0079] ,

[0080] in: The current feature descriptor is a multi-dimensional vector containing topological features of different dimensions; These are feature descriptors in the library, with a structure similar to... same; These are weight coefficients for different dimensions, used to adjust the importance of features in each dimension. They are typically set to... , , This indicates that the connectivity feature has the highest weight; The dimension index ranges from 0 to 2, corresponding to 0-dimensional, 1-dimensional, and 2-dimensional topological features, respectively. For scale parameter indexes, from 1 to , The scale parameter is the sequence length; express 3D topological features in the 1st dimension Values ​​under each scale parameter.

[0081] Then, for each known instrument feature in the feature library Calculate distance The nearest neighbor rule is applied to determine the device type. The identification confidence score is calculated using an inverse distance function.

[0082] ,

[0083] in: To determine the confidence level, the value ranges from (0,1], with a larger value indicating a more reliable recognition result. This is the minimum distance value, which is the distance between the current feature and the feature in the best-matching library; This is a scaling factor used to adjust the degree to which the distance value affects the confidence level; it is typically set to 5. When The module is considered reliable when the identification results are confirmed. In complex surgical environments, it can accurately identify various common surgical instruments, including needle holders, surgical scissors, surgical forceps, and hemostatic forceps.

[0084] The spatiotemporal trajectory analysis module 5 is connected to the instrument recognition module 4, and is used to construct a spatiotemporal topological manifold based on the recognition results, and to continuously analyze the usage trajectory of surgical instruments through homology analysis. Figure 3 As shown, the spatiotemporal trajectory analysis module 5 includes a spatiotemporal representation unit 51, a manifold construction unit 52, a continuous coherence calculation unit 53, and a trajectory pattern recognition unit 54.

[0085] The spatiotemporal representation unit 51 is used to construct the device identification results into an augmented state representation that includes state, time, and rate of change. Specifically, for each device state identification result in the time series... Construct augmented state representation ,in Includes information such as instrument type, location, and orientation. For timestamps, The rate of state change is calculated using the state difference between adjacent time points. This representation not only includes static state information but also incorporates dynamic change characteristics, providing a more comprehensive data foundation for subsequent trajectory analysis.

[0086] Manifold construction unit 52 is connected to spatiotemporal representation unit 51 and is used to construct a time-varying spatiotemporal topological representation based on the augmented state representation. First, a spatiotemporal distance metric function is defined:

[0087] ,

[0088] in: For comprehensive spatiotemporal distance; and There are two augmented state points; and It is a state vector, containing information such as device type, position, and attitude; and This is a timestamp, in seconds. and The vector represents the rate of state change. The distance function is the state-space distance function, typically using weighted Euclidean distance; This is the absolute value of the time difference; The distance function is the rate of change, and Euclidean distance is usually used. , and These are the weighting coefficients for state, time, and rate, used to adjust the importance of each factor in the overall distance. A preferred setting is... .

[0089] Based on the aforementioned distance metric, an adjacency graph G is constructed, and then a spatiotemporal simple complex SC is constructed from graph G. Preferably, a sliding window strategy is adopted to generate a time-varying simple complex sequence {SC(t)}, with a window size of 50 frames (approximately 1.67 seconds) and an overlap rate of 50%, to ensure continuous capture of changes in the instrument's state.

[0090] The persistent homology computation unit 53 is connected to the manifold construction unit 52 and is used to compute the persistent homology features of the time-varying spatiotemporal topological representation. This unit computes the persistent homology features through the following steps: for each time window's simple complex SC(t), it computes homology groups of different dimensions. Track the birth and death times of homologous classes to generate k-dimensional continuous barcodes. Long barcodes correspond to stable characteristics during device use, while short barcodes correspond to noise or transient states. Features are extracted from the barcodes, including lifespan distribution, birth time distribution, and death time distribution, to construct a persistently homogeneous feature vector PH=[ , , .

[0091] The core of persistent homology analysis lies in its ability to capture multi-scale topological features, enabling the system to distinguish between essential features and random noise in the trajectory of a device. Mathematically, a k-dimensional persistent barcode can be represented as:

[0092] ,

[0093] in: For a k-dimensional continuous barcode, it is a set of birth-death time pairs; Let be the birth time of the i-th k-dimensional homology class, and let be the scale parameter representing the first appearance of this topological feature; Let be the death time of the i-th k-dimensional homology class, and let represent the scale parameter at which the topological feature disappears. Let be the total number of k-dimensional homology classes; i is the homology class index, from 1 to 1. .when When the homology class is in a state of permanent existence, it indicates that the homology class exists permanently.

[0094] The trajectory pattern recognition unit 54 is connected to the continuous coherence calculation unit 53, and is used to compare the continuous coherence features with a preset normal trajectory pattern to identify abnormal usage patterns. This unit uses Wasserstein distance to measure the difference between the current trajectory features and the normal pattern.

[0095] ,

[0096] in: The Wasserstein distance measures the difference between two persistent homology features. This represents the continuous cohomological characteristics of the current trajectory; The continuous cohomology characteristic of the preset normal mode; Weighting coefficients for different dimensions are typically set. , , ; Dimension index, from 0 to 2; For the birth-death time pair of the i-th k-dimensional homology class in the current feature; For the birth-death time pair of the j-th k-dimensional homology class in the normal pattern; This represents the number of k-dimensional homology classes in the current feature. This means finding the value of j that minimizes the distance among all possible values, i.e., finding the best match; Using Euclidean distance, calculate the distance between two birth-death time pairs.

[0097] Set an anomaly threshold θ (typically the 95th percentile of the normal distance distribution, approximately 0.35). If If the abnormal trajectory is detected, it is considered abnormal. Further analysis of the dimensions and timing of the abnormality is performed to generate an abnormality description report. In clinical practice, this module can effectively identify various abnormal operations, such as incorrect instrument usage sequence, excessive force applied during operation, or accidental instrument drop, with an abnormality detection accuracy rate of 94.3%.

[0098] The state prediction module 6 is connected to the spatiotemporal trajectory analysis module 5, and is used to learn the evolution law of instrument state on the topological manifold and predict the future state of the surgical instrument. For example... Figure 4 As shown, the state prediction module 6 includes a manifold representation unit 61, an evolution model unit 62, a multi-scale prediction unit 63, and a risk assessment unit 64.

[0099] Manifold representation unit 61 is used to reduce the dimensionality of historical instrument state data through local linear embedding and construct a state manifold representation space. As shown, the method for state manifold representation by this unit includes: constructing a state k-nearest neighbor graph. ,in Typically set to ( (Number of data points); computation graph Laplace operator on Solve the Laplace characteristic equation , obtain the characteristic function As a spectral representation of the manifold; the original state Mapped to 3D manifold representation space Obtain manifold coordinates Preferably, the intrinsic dimension d is determined through residual analysis, typically between 3 and 7, to balance representational power and computational complexity.

[0100] The Laplace operator L is constructed as follows:

[0101] ,

[0102] in: For the Laplace operator, it is a A symmetric positive semi-definite matrix, where This represents the number of data points. for The identity matrix; Given an adjacency matrix, if point i and point j are adjacent, then... ,otherwise ; The degree matrix is ​​a diagonal matrix, and its diagonal elements are... Let be the degree of vertex i, that is, the number of vertices adjacent to vertex i; Indicates will The diagonal matrix obtained by taking the square root of the reciprocal of each diagonal element.

[0103] Evolutionary model unit 62 is connected to manifold representation unit 61 and is used to learn state transition functions in the state manifold representation space. On the manifold M, a local coordinate system is defined, a tangent space representation is constructed, and then the state transition functions are learned in the tangent space.

[0104] ,

[0105] in: This is a state transition function used to predict future states; The flow pattern at time point t; This represents the predicted flow pattern at time point t+1; The size of the history window indicates the use of past data. The function predicts the state at the next time point based on the state at each of the previous time points. The state at one time point predicts the state at the next time point. The learning process uses geodesic distance instead of Euclidean distance to more accurately represent the true distance between states on the manifold. In practice, the value of k is usually set to 3-5, adjusted according to the complexity of the machine's state changes.

[0106] The multi-scale prediction unit 63 is connected to the evolution model unit 62 and is used to predict the future state of the device on short-term, medium-term, and long-term time scales based on the state transition function. This unit constructs prediction models for three time scales: a short-term (1-3 seconds) prediction model... Medium-term (3-10 seconds) prediction model and long-term (>10 seconds) prediction models The parameters of each model are optimized to minimize the prediction error at the corresponding time scale. Preferably, the short-term model emphasizes local dynamic characteristics, the long-term model focuses more on global trend characteristics, and the medium-term model balances both. This multi-scale prediction strategy can adapt to the risk monitoring needs of different levels of urgency.

[0107] The risk assessment unit 64 is connected to the multi-scale prediction unit 63 and is used to calculate the distance between the predicted state and the safe state space to assess the potential risk level. First, it uses a back projection function... Returning the manifold state to the original state space: Then, calculate the distance between the predicted state and the safe state space:

[0108] ,

[0109] in: The minimum distance between the predicted state and the safe state space; The predicted state at time t+1; SafeSpace is a predefined set of safe states, containing multiple instrument states that conform to surgical specifications; This indicates the search for the state that minimizes the distance within the set of safe states; To predict the distance between a state and a specific safe state, the weighted Euclidean distance is usually used.

[0110] Risk level L is assessed based on distance and trend of change, and is usually divided into three levels: low (L=1), medium (L=2), and high (L=3). When the distance exceeds the safety threshold (usually 2σ of the normal operating distance distribution, about 0.3) and shows an increasing trend, it is judged as high risk; when the distance exceeds the threshold but the trend is stable or decreasing, it is judged as medium risk; when the distance is less than the threshold, it is judged as low risk.

[0111] The decision support module 7 is connected to the state prediction module 6 and is used to assess the risk level based on the prediction results and trigger early warning information when an abnormal state is detected. Figure 5 As shown, the decision support module 7 includes a risk level classification unit 71, an early warning information generation unit 72, an operation suggestion generation unit 73, and a human-computer interaction unit 74.

[0112] Risk level classification unit 71 is used to classify risks into three levels: low, medium, and high, based on the distance and change trend between the predicted state and the safe state space. The specific classification criteria are as follows: low risk is defined as distance d < 0.2 and rate of change v < 0.05; medium risk is defined as 0.2 ≤ d < 0.3 or 0.05 ≤ v < 0.1; and high risk is defined as d ≥ 0.3 and v ≥ 0.1. Here, distance d is a normalized value, and the rate of change v represents the increase in distance per unit time.

[0113] The early warning information generation unit 72 is connected to the risk level classification unit 71 and is used to generate corresponding early warning information for different risk levels. For low-risk situations, a reminder message is generated; for medium-risk situations, a warning message is generated; and for high-risk situations, an emergency alarm is generated. The early warning information includes the risk type, possible consequences, and urgency level. For example, for a high-risk situation where a needle holder may be left behind, a warning message is generated: "High risk of needle holder residue, which may lead to postoperative complications; immediate verification is recommended."

[0114] The operation suggestion generation unit 73 is connected to the early warning information generation unit 72, and is used to generate targeted operation suggestions based on early warning information and the current surgical stage. This unit combines surgical guidelines and expert experience from the knowledge base to provide medical staff with specific coping strategies. For example, regarding the potential risks of instruments being left behind, it generates operation suggestions to immediately stop suturing, count the number of instruments, and check the patient's wound area; regarding the risks of improper instrument use, it generates operation suggestions to adjust the angle of the needle holder to avoid tissue damage.

[0115] The human-computer interaction unit 74 is connected to the operation suggestion generation unit 73, and is used to convey warning information and operation suggestions to medical staff through visual and auditory means. This unit employs a multimodal interaction method, including visual alarms on the operating room display screen, auditory prompts, and optional vibration alerts. Different color codes are used for different risk levels: green for low risk, yellow for medium risk, and red for high risk. The frequency and volume of the auditory prompts also increase with the risk level to ensure effective attention from medical staff in emergency situations.

[0116] like Figure 2 As shown, the complex building unit 32 constructs a multi-scale simple complex sequence through the following steps:

[0117] First, set an incremental sequence of scale parameters. In this embodiment, a non-linear incremental strategy is adopted: Among them, the benchmark scale mm, growth rate Sequence length This nonlinear incremental strategy can provide finer resolution at small scales and reduce redundant computation at large scales.

[0118] Then, for each scale parameter Construct the corresponding Vletoris-Rips complex To improve computational efficiency, this embodiment employs sparse Vietoris-Rips complexes. By retaining only the simplex that contributes most to the topological features, the complexity is reduced from... Reduce to ,in For points, For dimensions.

[0119] Finally, a nested complex sequence is generated. Furthermore, an incremental calculation strategy is adopted, which calculates only the difference between adjacent complexes each time, thereby further improving computational efficiency.

[0120] In actual testing, the optimized complex building block 32 improved the processing speed by about 4.5 times while maintaining the same accuracy in topological feature extraction, enabling the system to meet the requirements of real-time processing.

[0121] The continuous cohomology calculation unit 53 calculates the continuous cohomology characteristics through the following steps:

[0122] First, a simple complex for each time window. An optimized homology group calculation algorithm is employed. Specifically, a boundary matrix representation based on sparse matrices is used, combined with an efficient matrix reduction algorithm, to calculate the homology group. The computational complexity is from Reduce to ,in Let be the number of simplex elements.

[0123] Then, an incremental homology class tracking algorithm is employed to efficiently track the birth and death times of homology classes. This algorithm avoids redundant calculations by maintaining an evolutionary graph of homology classes, thus reducing the tracking complexity from... Reduce to ,in This represents the number of homology classes.

[0124] Finally, when extracting features from persistent barcodes, statistical moment features are used instead of the original barcode, such as average lifetime, standard deviation, skewness, and kurtosis. This reduces the feature dimensionality and improves the discriminative power of the features.

[0125] In clinical validation, the optimized continuous coherence computing unit 53 not only improved the processing speed by about 3 times, but also increased the detection sensitivity of abnormal trajectories from 91.5% to 94.3%, significantly improving the real-time performance and accuracy of the system.

[0126] like Figure 4 As shown, the method for manifold representation unit 61 to perform state manifold representation includes:

[0127] First, an adaptive k-value strategy is adopted when constructing the state k-nearest neighbor graph. Specifically, the k value is dynamically adjusted based on the local data density: ,in The base k value (usually 1) ), Let i be the data density near point i. This represents the average data density. This adaptive strategy can better preserve the local structural characteristics of the data.

[0128] Then, in the computation of the graph Laplace operator At that time, regularized heat kernel weights are used:

[0129] ,

[0130] in: Let the weight of the edge between vertex i and vertex j be denoted as . and These are the data points corresponding to vertices i and j, respectively; The Euclidean distance between the two points; The kernel width parameter controls the rate at which the weights decay with distance; it is typically set to the average distance of the local neighborhood. This is the regularization parameter, usually set to 0.01, used to enhance numerical stability; Let Kronecker function be used when The value is 1 when the condition is met, and 0 otherwise. This weight design enhances the algorithm's robustness to noise.

[0131] Finally, when solving for the eigenfunction, the Krylov subspace iterative method is used to avoid the high computational cost of direct eigenvalue decomposition, reducing the complexity from... Reduce to ,in The required number of eigenvalues ​​is typically much smaller than .

[0132] The optimized manifold representation unit 61 not only improves computational efficiency by about 5 times, but also significantly improves the quality of state representation, especially its ability to express complex nonlinear state spaces, further improving the accuracy of state prediction.

[0133] The operation suggestion generation unit 73 generates targeted operation suggestions based on the warning information and the current surgical stage. In this embodiment, a context-aware mechanism is introduced to improve the relevance of the suggestions in the following ways:

[0134] First, a surgical stage identification model is established to automatically identify the current surgical stage based on the current instrument usage mode, surgical progress, and performed operations. The stage is divided into five main stages: preparation, incision, main operation, suturing, and finishing.

[0135] Then, a risk-operation mapping matrix M is constructed, where M_ij represents the set of recommended operations for risk type i at surgical stage j. This matrix is ​​built based on medical expert knowledge and historical surgical data, and contains approximately 200 common risk-operation correspondences.

[0136] Finally, considering the specific context (such as device type, location, and usage method), the most suitable operational suggestions are selected from the set of recommended operations and sorted according to urgency, implementation difficulty, and expected results, with the most critical suggestions being displayed first.

[0137] In clinical trials conducted in five tertiary hospitals, the adoption rate of suggestions generated by the optimized operation suggestion generation unit 73 increased from 78% to 91%, significantly improving the system's usability and acceptance by medical staff.

[0138] like Figure 6As shown, the data management module 8 includes a real-time database 81, a historical database 82, a knowledge graph database 83, and a data synchronization unit 84.

[0139] Real-time database 81 stores the real-time instrument status, location, and usage during the current surgical procedure. It employs a key-value pair storage structure, supports high-concurrency read / write operations and fast queries, and updates every 50 milliseconds. This database utilizes in-memory database technology (such as Redis), keeping response time within 5 milliseconds to ensure system real-time performance.

[0140] Historical database 82 stores instrument usage data, status sequences, and anomaly records from historical surgeries. It employs a time-series database structure (such as InfluxDB), supports efficient time-range queries, and has a storage period of 90 days. This database is used not only for retrospective analysis but also to provide training data for status prediction models.

[0141] The knowledge graph database 83 stores relationships between instruments, usage guidelines, surgical procedures, and expert experience. It employs a graph database structure (such as Neo4j) to represent complex relationships between entities, containing approximately 5,000 nodes and 15,000 edges, covering common surgical types and instrument usage guidelines. This database is regularly updated based on new surgical data and expert feedback, typically monthly.

[0142] The data synchronization unit 84 is used to coordinate data interaction between the real-time database 81, the historical database 82, and the knowledge graph database 83. This unit achieves data consistency maintenance, conflict resolution, and efficient synchronization. It uses Change Data Capture (CDC) technology to synchronize critical data in real time and batch processing to synchronize non-critical data, thus balancing system load and data consistency requirements.

[0143] In practical applications, the data management module 8 provides a solid data foundation for the system, supporting efficient data access, rich knowledge representation, and a reliable data synchronization mechanism, enabling the system to continuously learn from historical data and optimize performance.

[0144] like Figure 7 As shown, the image acquisition module 1 includes a multi-angle high-definition camera 11, an image buffer unit 12, and an image quality evaluation unit 13.

[0145] The multi-angle high-definition camera 11 has a 4K resolution (3840×2160 pixels) and a frame rate of 60 frames per second, exceeding the minimum requirement of the claims (30 frames per second), to ensure clear images are captured even during rapid surgical procedures. In a standard operating room, three cameras are typically installed: one top-view camera and two side cameras, forming a 120° angle to achieve omnidirectional coverage of the surgical area. The cameras employ automatic aperture and optical image stabilization technology to adapt to the complex lighting conditions of the operating room.

[0146] The image buffer unit 12 is connected to the multi-angle high-definition camera 11 and is used to temporarily store the most recently acquired image sequence. This unit adopts a circular buffer structure and typically stores image data from the most recent 5 seconds (approximately 300 frames). The buffer uses high-speed solid-state storage, with read / write latency controlled to within 1 millisecond, ensuring that image data can be quickly provided to subsequent processing modules.

[0147] The image quality evaluation unit 13 is connected to the image buffer unit 12 and is used to evaluate image sharpness, brightness, and contrast, triggering re-acquisition when the image quality does not meet preset thresholds. Evaluation metrics include: Laplacian operator score (sharpness, threshold > 100), average brightness (threshold between 80-200), and RMS contrast (threshold > 0.3). When any metric fails to meet the requirements, the system automatically adjusts camera parameters (such as exposure time, gain, etc.) and re-acquires the image until the image quality meets the standards.

[0148] In clinical applications, the optimized image acquisition module 1 can provide stable and high-quality image input in various surgical environments, providing a reliable basis for subsequent analysis. The system's adaptability is significantly enhanced, and the qualified image rate exceeds 98% under five different lighting conditions.

[0149] The specific implementation of the intelligent surgical instrument identification and real-time inventory image analysis system provided by this invention is fully described. By introducing topological theory, this system achieves high-precision identification, dynamic tracking, and risk prediction of surgical instruments, significantly improving surgical safety and providing an innovative technical solution for the medical field.

[0150] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. 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 modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A surgical instrument intelligent identification and real-time inventory image analysis system, characterized in that, include: The image acquisition module is used to acquire multi-angle real-time images in the operating room and transmit the multi-angle real-time images to the image preprocessing module; An image preprocessing module, connected to the image acquisition module, is used to perform noise reduction, enhancement, and geometric correction on the received multi-angle real-time images, and transmit the processed images to the topological feature extraction module; A topological feature extraction module, connected to the image preprocessing module, is used to extract the topological invariant features of surgical instruments from the processed image and construct a multi-scale feature descriptor. The instrument identification module, connected to the topology feature extraction module, is used to match the multi-scale feature descriptor with a preset feature library to determine the type and status of the surgical instrument. The spatiotemporal trajectory analysis module is connected to the instrument recognition module and is used to construct a spatiotemporal topological manifold based on the recognition results and to analyze the usage trajectory of surgical instruments through continuous homology analysis. The state prediction module, connected to the spatiotemporal trajectory analysis module, is used to learn the state evolution law of the instrument on the topological manifold and predict the future state of the surgical instrument. The decision support module, connected to the state prediction module, is used to assess the risk level based on the prediction results and trigger early warning information when an abnormal state is detected.

2. The intelligent surgical instrument identification and real-time inventory image analysis system according to claim 1, characterized in that, The topology feature extraction module includes: The point cloud conversion unit is used to convert the processed image into a point cloud representation; A complex construction unit, connected to the point cloud transformation unit, is used to construct a multi-scale simple complex sequence based on the point cloud representation; The feature calculation unit, connected to the complex construction unit, is used to calculate the topological invariants of the multi-scale simple complex sequence, including Betti number sequences of different dimensions; The feature descriptor generation unit, connected to the feature calculation unit, is used to combine Betti number sequences of different dimensions into multi-scale feature descriptors.

3. The intelligent surgical instrument identification and real-time inventory image analysis system according to claim 1, characterized in that, The spatiotemporal trajectory analysis module includes: The spatiotemporal representation unit is used to construct the instrument identification results into an augmented state representation that includes state, time, and rate of change; A manifold construction unit, connected to the spatiotemporal representation unit, is used to construct a time-varying spatiotemporal topological representation based on the augmented state representation; A persistent cohomology computation unit, connected to the manifold construction unit, is used to compute the persistent cohomology features of the time-varying spatiotemporal topological representation; The trajectory pattern recognition unit, connected to the continuous coherence calculation unit, is used to compare the continuous coherence features with a preset normal trajectory pattern to identify abnormal usage patterns.

4. The intelligent surgical instrument identification and real-time inventory image analysis system according to claim 1, characterized in that, The state prediction module includes: Manifold representation unit is used to reduce the dimensionality of historical instrument state data through local linear embedding and construct a state manifold representation space. An evolutionary model unit, connected to the manifold representation unit, is used to learn state transition functions in the state manifold representation space; A multi-scale prediction unit, connected to the evolution model unit, is used to predict the future state of the device on short-term, medium-term and long-term time scales based on the state transition function. The risk assessment unit, connected to the multi-scale prediction unit, is used to calculate the distance between the predicted state and the safe state space, and to assess the potential risk level.

5. The intelligent surgical instrument identification and real-time inventory image analysis system according to claim 1, characterized in that, The decision support module includes: The risk level classification unit is used to classify risks into three levels: low, medium, and high, based on the distance and trend between the predicted state and the safe state space. The early warning information generation unit is connected to the risk level classification unit and is used to generate corresponding early warning information for different risk levels. An operation suggestion generation unit, connected to the early warning information generation unit, is used to generate targeted operation suggestions based on the early warning information and the current surgical stage. The human-computer interaction unit, connected to the operation suggestion generation unit, is used to convey warning information and operation suggestions to medical staff through visual and auditory means.

6. The intelligent surgical instrument identification and real-time inventory image analysis system according to claim 2, characterized in that, The complex building unit constructs a multi-scale simple complex sequence through the following steps: Set an incremental sequence of scale parameters; For each scale parameter, construct the corresponding Vietoris-Rips complex; Organize the complexes corresponding to all scale parameters into nested sequences.

7. The intelligent surgical instrument identification and real-time inventory image analysis system according to claim 3, characterized in that, The continuous cohomology calculation unit calculates the continuous cohomology characteristics through the following steps: For each time window, simple complexes are used to calculate homology groups of different dimensions. Track the birth and death times of homologous classes to generate persistent barcodes; Features are extracted from the persistent barcode, including life length distribution, birth time distribution, and death time distribution.

8. The intelligent surgical instrument identification and real-time inventory image analysis system according to claim 4, characterized in that, The method for representing the state manifold by the manifold representation unit includes: Construct a state k-nearest neighbor graph; Calculate the Laplace operator on the k-nearest neighbor graph; Solve for the eigenfunctions of the Laplace operator as the spectral representation of the manifold; Project the original state onto the manifold representation space.

9. The intelligent surgical instrument identification and real-time inventory image analysis system according to claim 1, characterized in that, It also includes a data management module, which includes: A real-time database is used to store the real-time instrument status, location, and usage during the current surgical procedure; Historical database, used to store instrument usage data, status sequences, and abnormal records from historical surgeries; A knowledge graph database is used to store relationships between instruments, usage guidelines, surgical procedures, and expert experience. The data synchronization unit is used to coordinate the data interaction between the real-time database, the historical database, and the knowledge graph database.

10. The intelligent surgical instrument identification and real-time inventory image analysis system according to claim 1, characterized in that, The image acquisition module includes: A multi-angle high-definition camera, wherein the resolution of the multi-angle high-definition camera is not less than 4K and the frame rate is not less than 30 frames per second; An image caching unit, connected to the multi-angle high-definition camera, is used to temporarily store recently acquired image sequences; An image quality assessment unit, connected to the image buffer unit, is used to assess image sharpness, brightness, and contrast, and triggers re-acquisition when the image quality does not meet a preset threshold.

Images