Endoscope video enhancement processing intelligent edge computing system

Abstract

The invention relates to the technical field of endoscope video processing and intelligent edge computing, in particular to an endoscope video enhancement processing intelligent edge computing system. Comprising a data acquisition module which is used for acquiring an original video frame sequence of edge endoscope equipment in real time; the feature extraction module is used for determining an instantaneous feature vector representing the dynamic change of the operation scene; the criticality quantification module is used for quantizing and generating a surgical event criticality score; the tuning logic module is used for generating a discrete and stable calculation normal form switching instruction; the assembly line switching module is used for responding to the calculation normal form switching instruction and executing an asynchronous weight preheating strategy; the utility evaluation module is used for constructing a dynamic utility function to evaluate system performance; and the threshold value correction module is used for performing closed-loop correction on the high-criticality threshold value by adopting a gradient rising strategy. According to the method, the stability of system decision making is enhanced, the smooth transition of the video processing flow among different calculation paradigms is ensured, and the robustness of the system is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of endoscopic video processing and intelligent edge computing, in particular to an endoscopic video enhancement processing intelligent edge computing system. BACKGROUND

[0002] In modern endoscopic video enhancement processing, edge computing units are increasingly widely used. However, the surgical scene is highly dynamic, which poses a severe dual challenge to the edge computing system: the system needs to ensure low latency and smoothness of the video stream during regular cruising, and must immediately switch to a high-precision video enhancement mode when key events such as bleeding and identification of key tissues occur. The current technical dilemma is that the computing power resources of edge computing units are usually limited. This inherent computing power limitation makes it difficult to simultaneously meet the above-mentioned low latency and high precision performance requirements, which are two conflicting and high-standard performance requirements, resulting in the system being unable to efficiently and smoothly switch between the two computing paradigms. Therefore, under the condition of limited edge computing power, how to design an intelligent resource allocation and decision-making mechanism to enable the system to adapt to the highly dynamic surgical scene, and to achieve intelligent, stable and seamless switching between low latency and high precision, is a technical problem that needs to be solved in the field. SUMMARY

[0003] To solve the above technical problems, the present application discloses an endoscopic video enhancement processing intelligent edge computing system, in particular, the technical scheme of the present application comprises: A data acquisition module for acquiring a sequence of original video frames of an edge endoscope device in real time; A feature extraction module for determining an instantaneous feature vector representing the dynamic changes of the surgical scene based on the sequence of original video frames; A key quantification module for quantifying a surgical event key degree score based on the instantaneous feature vector; A tuning logic module for executing a double-threshold hysteresis control logic to generate discrete and stable computing paradigm switching instructions based on the surgical event key degree score, a preset high key degree threshold and a preset low key degree threshold; A pipeline switching module for executing an asynchronous weight warm-up strategy using the hysteresis zone of the double-threshold hysteresis control logic in response to the computing paradigm switching instructions; An utility evaluation module for constructing a dynamic utility function to evaluate system performance; A threshold correction module for closed-loop correction of the high key degree threshold using a gradient ascent strategy based on the dynamic utility function.

[0004] Preferably, the feature extraction module is configured to determine the instantaneous feature vector, comprising: The field of view motion vector is obtained by analyzing the mean value of the optical flow field between consecutive frames; The proportion of the blood color gamut is obtained by statistically analyzing the pixel proportion of the current frame in a preset blood color space model. Tissue texture heterogeneity is obtained by calculating the Mahalanobis distance between the feature map extracted by a lightweight convolutional neural network and the preset conventional tissue baseline feature map.

[0005] Preferably, the key quantification module is used to quantify and generate a criticality score for a surgical event based on an instantaneous feature vector, including: A custom linear weighted fusion model is used to perform a weighted summation of the field of view motion vector, blood color gamut proportion, and tissue texture heterogeneity; The model's weight coefficients were obtained through offline calibration or regression training on a gold-standard visual dataset annotated by senior surgeons.

[0006] Preferably, the tuning logic module is used to execute dual-threshold hysteresis control logic, including: Define the states of high-precision computing paradigm and low-latency computing paradigm; When the criticality score of a surgical event exceeds the high criticality threshold, switch to the high-precision calculation paradigm. When the criticality score of a surgical event is less than the low criticality threshold, switch to the low-latency calculation paradigm. When the criticality score of a surgical event falls between the low criticality threshold and the high criticality threshold, the current calculation paradigm remains unchanged.

[0007] Preferably, the pipeline switching module is used to execute an asynchronous weight preheating strategy, including: When the system is in a low-latency computing paradigm, and the criticality score of the surgical event crosses the low criticality threshold and enters the hysteresis region; Immediately load the high-precision model weights required for the high-precision computing paradigm state asynchronously into the memory of the edge computing unit in the background computing thread.

[0008] Preferably, the pipeline switching module is further used for: When the criticality score of the surgical event continues to rise and crosses the high criticality threshold, the high-precision model weights are pre-warmed in the video memory, avoiding model loading delay and achieving low-latency switching. When the criticality score of a surgical event falls back within the hysteresis region and drops below the low criticality threshold, the preheating process is stopped or the loaded video memory is cleared to release computing resources.

[0009] Preferably, the preset high-criticality threshold and the preset low-criticality threshold are initially obtained through statistical calibration using the gold standard dataset. The low criticality threshold is set to a value that can filter out noise peaks in the criticality score during normal cruise scenarios; The high criticality threshold is set to a value that can capture the lowest criticality score in critical organization identification scenarios.

[0010] Preferably, the utility evaluation module is used to construct a dynamic utility function, including: Real-time acquisition of normalized instantaneous diagnostic accuracy and normalized instantaneous energy consumption; Among them, the normalized instantaneous energy consumption is obtained by normalizing the instantaneous power with the preset maximum and minimum operating power; Based on a preset energy consumption trade-off factor, a Pareto trade-off is performed between normalized instantaneous diagnostic accuracy and normalized instantaneous energy consumption to determine the cumulative utility within a preset time window, which is then used as the value of the dynamic utility function.

[0011] Preferably, the threshold correction module is used to employ a gradient ascent strategy, including: By applying a small perturbation to the high criticality threshold at runtime, the gradient of the dynamic utility function with respect to the high criticality threshold is approximately calculated. Based on the gradient and the preset learning rate, the high criticality threshold is iteratively adjusted to converge to the maximization of the dynamic utility function.

[0012] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention employs dual-threshold hysteresis control logic to generate discrete and stable computational paradigm switching instructions based on the criticality score of surgical events. By setting two criticality thresholds (high and low) and utilizing the hysteresis region between them, frequent and ineffective paradigm switching caused by small fluctuations in criticality scores around the thresholds is effectively avoided. This design enhances the stability of system decisions, ensures a smooth transition in the image processing flow between different computational paradigms, and improves the system's robustness.

[0013] 2. This invention utilizes the hysteresis zone of a dual-threshold hysteresis control logic to execute an asynchronous weight preheating strategy. When the criticality score enters the hysteresis zone from a low-latency state, the system immediately and asynchronously loads high-precision model weights into the video memory in the background. When the score finally crosses the high threshold, the weights have been preheated, thus avoiding model loading delays and achieving a high-speed, low-latency switch from a low-latency paradigm to a high-precision paradigm, ensuring immediate response in critical surgical scenarios.

[0014] 3. This invention constructs a dynamic utility function to evaluate the Pareto trade-off between instantaneous diagnostic accuracy and energy consumption in real time. The system employs a gradient ascent strategy, using this utility function to perform closed-loop correction on high-criticality thresholds. This design enables the system to adaptively adjust the sensitivity of the switching logic based on real-time performance, converging to maximize the dynamic utility function, thus achieving continuous performance optimization in complex edge computing environments.

[0015] 4. This invention, through a feature extraction module, comprehensively analyzes multi-dimensional information such as visual field motion vectors, blood color gamut proportions, and tissue texture heterogeneity to determine instantaneous feature vectors representing the dynamic changes of the surgical scene. The criticality quantification module further employs a fusion model calibrated based on a gold-standard dataset of senior surgeons to weightedly fuse the aforementioned features. This method ensures that the generated surgical event criticality scores accurately and reliably reflect the complexity and criticality of the real surgical scene. Attached Figure Description

[0016] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0018] Example 1 Please see Figure 1 A laparoscopic image enhancement processing intelligent edge computing system, comprising: The data acquisition module is used to acquire the raw video frame sequence of the edge endoscope device in real time; The feature extraction module is used to determine instantaneous feature vectors representing the dynamic changes of the surgical scene based on the original video frame sequence; The criticality quantification module is used to quantify and generate criticality scores for surgical events based on instantaneous feature vectors. The tuning logic module is used to execute dual-threshold hysteresis control logic based on the criticality score of the surgical event, the preset high criticality threshold, and the preset low criticality threshold, in order to generate discrete and stable computational paradigm switching instructions. The pipeline switching module is used to respond to computational paradigm switching instructions and execute asynchronous weight warm-up strategies by utilizing the hysteresis region of the dual-threshold hysteresis control logic. The utility evaluation module is used to construct dynamic utility functions to evaluate system performance. The threshold correction module is used to perform closed-loop correction of high criticality thresholds based on a gradient ascent strategy using a dynamic utility function. This invention provides an intelligent edge computing system for laparoscopic image enhancement processing, which can be applied in scenarios such as high-definition electronic laparoscopes or gastrointestinal endoscopes in operating rooms. It aims to resolve the contradiction between the limited computing power of edge computing units such as GPUs or NPUs and the high dynamic and high-precision image enhancement requirements in surgery. The system comprises seven core modules that work together to form a complete technology chain from scene perception and decision tuning to closed-loop optimization. The data acquisition module aims to provide stable and raw visual input for subsequent real-time analysis. The raw image frame sequence refers to the uncompressed or only lightly encoded image data stream directly output from the CMOS sensor of the endoscope device. In this embodiment, the module acquires the raw image frame sequence of the edge endoscope device in real time via a high-speed interface such as MIPI or SDI. This ensures that data enters the analysis pipeline with minimal delay; The feature extraction module aims to extract key information that can efficiently represent the dynamic changes of the surgical scene from the high-throughput raw video stream in a lightweight computational manner, providing a basis for subsequent criticality judgment. Based on the input raw video frame sequence, this module computes and determines an instantaneous feature vector that represents the dynamic changes of the surgical scene in parallel. This vector is multi-dimensional and contains a comprehensive description of scene motion, color, and texture. The key quantification module aims to fuse the multidimensional and heterogeneous features extracted by the preceding modules into a single, quantifiable scalar to drive subsequent tuning logic. This module quantifies and generates a criticality score for the surgical event based on the instantaneous feature vector. Criticality score of surgical events It refers to a dimensionless scalar value, whose function is to reflect the urgency or importance of clinical events such as bleeding and tissue identification contained in the current video stream in real time and quantitatively. Its source is the calculation result of the key quantification module. The tuning logic module is designed to adjust based on continuously changing... The module generates discrete, stable, and jitter-free computational paradigm switching decisions based on the criticality scores of surgical events to avoid frequent oscillations near critical states. Preset high criticality threshold and preset low criticality threshold Execute dual-threshold hysteresis control logic; and These are two core control parameters, derived from offline calibration or online correction, and they always meet the following requirements. The execution result of this logic is the generation of discrete, stable computational paradigm switching instructions, such as switching to a high-precision or low-latency state. The pipeline switching module aims to physically execute the switching of computational paradigms and proactively manage and optimize the significant latency caused by model loading during the switching process. This module responds to the computational paradigm switching instructions generated by the tuning logic module and innovatively utilizes the hysteresis region provided by the dual-threshold hysteresis control logic. lie in and In the interval between these intervals, an asynchronous weight warm-up strategy is executed to avoid model loading delays; The utility evaluation module aims to construct a top-level feedback mechanism for real-time assessment of the instantaneous balance between the conflicting objectives of diagnostic accuracy and computational energy consumption. This module achieves this by constructing a dynamic utility function. This function is used to quantitatively evaluate system performance; it comprehensively considers the system's instantaneous accuracy and energy consumption, and outputs a single utility score. The threshold correction module aims to adaptively optimize the core parameter of the control logic, namely the threshold, using the results of utility evaluation, so that the system can dynamically converge to the optimal operating point. This module is based on the dynamic utility function calculated by the utility evaluation module. A gradient ascent strategy is used to apply high criticality thresholds. Similarly, it can also be applied to Perform closed-loop correction; This embodiment constructs a complete closed-loop intelligent system through the close collaboration of the above seven modules, from scene perception feature extraction and key quantification to intelligent decision-making tuning logic, agile execution pipeline switching, adaptive optimization utility evaluation, and threshold correction; it solves the inherent contradiction that edge computing devices have limited computing power and cannot simultaneously meet the requirements of low-latency routine navigation and high-precision critical events throughout the entire surgical process. This system can autonomously and intelligently balance computing power allocation in highly dynamic surgical scenarios. It automatically switches to high-precision mode at critical moments such as bleeding and automatically switches back to low-latency mode at normal times. Throughout the process, it achieves smooth switching through asynchronous preheating and continuously optimizes the energy efficiency-precision ratio through closed-loop correction. Ultimately, it achieves intelligent, low-latency, and energy-efficient balanced image enhancement processing in highly dynamic scenarios.

[0019] Example 2 The feature extraction module, used to determine the instantaneous feature vector, includes: The field of view motion vector is obtained by analyzing the mean value of the optical flow field between consecutive frames; The proportion of the blood color gamut is obtained by statistically analyzing the pixel proportion of the current frame in a preset blood color space model. Tissue texture heterogeneity is obtained by calculating the Mahalanobis distance between the feature map extracted by a lightweight convolutional neural network and the preset conventional tissue baseline feature map. Based on Example 1, this embodiment specifies how the feature extraction module determines the instantaneous feature vector; the module calculates three independent, clinically decoupled feature components in parallel. Field of view motion vector This refers to the characteristics that characterize the field of view of the endoscope, i.e., the macroscopic motion amplitude and direction of the lens between consecutive frames; its function is to quantify the intensity of motion during routine cruising or large-amplitude scanning; in this embodiment, by analyzing consecutive frames, for example... and The optical flow is obtained by averaging the optical flow fields between different regions; after calculating the optical flow field, the optical flow vectors within the entire image or the region of interest (ROI) are averaged or moduloed to obtain a scalar or low-dimensional vector. ; Blood color gamut percentage This refers to the proportion of pixel colors in the current image frame that fall within a preset blood color space model; its function is to characterize the urgency of bleeding point localization or active bleeding events. In this embodiment, to determine this value, a preset blood color space model is established in a specific color space, such as CIE-Lab or HSV, defining a specific region in the Lab space; subsequently, the current frame is statistically analyzed during runtime. What percentage of the pixels fall within the region defined by the model, thus obtaining... ; Tissue texture heterogeneity This refers to the degree of difference between the tissue texture features in the current image frame and the baseline features of conventional healthy tissue; its function is to characterize the need for lesion identification or identification of key tissues such as nerves and blood vessels; in this embodiment, to determine this value, a shallow layer of a lightweight convolutional neural network (CNN) such as MobileNet is used to extract the current frame. The feature map is obtained by pre-compiling and storing a preset conventional tissue baseline feature map, such as averaging the feature maps of a large number of healthy tissue images; the Mahalanobis distance between the current feature map and the baseline feature map is calculated to obtain the feature map. Mahalanobis distance is used instead of Euclidean distance because it can take into account the correlation between features and is more robust to noise. This embodiment utilizes three specific and clinically decoupled feature dimensions. Represents sports, Representative color, Representing texture, the system can comprehensively, accurately, and with extremely low computational overhead perceive subtle changes in the surgical scene from three orthogonal clinical perspectives: motion, color, and texture. This lays the foundation for subsequent key quantification. The accuracy provides a solid and reliable multidimensional data foundation, ensuring... The score can sensitively reflect different types of surgical events.

[0020] Example 3 The criticality quantification module is used to quantify and generate criticality scores for surgical events based on instantaneous feature vectors, including: A custom linear weighted fusion model is used to perform a weighted summation of the field of view motion vector, blood color gamut proportion, and tissue texture heterogeneity; The model's weight coefficients were obtained through offline calibration or regression training on a gold-standard visual dataset annotated by senior surgeons. Based on Example 1, this embodiment focuses on how the key quantification module uses instantaneous feature vectors. To quantify and generate criticality scores for surgical events Specific limitations were imposed; Three heterogeneous features All values ​​are normalized to [0,1] and merged into a single scalar. This embodiment uses a custom linear weighted fusion model; its model is described as follows: ; in, The criticality score of the surgical event is dimensionless and calculated by this module, serving as the input to the tuning logic. The normalized field-of-view motion vector, blood color gamut proportion, and tissue texture heterogeneity, with a value range of [0,1], are input in real time by the feature extraction module; These are preset feature weight coefficients, dimensionless; To further clarify the weighting coefficients The source of these weights, serving as key adjustable parameters in this model, defines the degree to which different features contribute to the final criticality. Their values ​​are determined not by subjective setting or equivalence, but by obtaining them through offline calibration or multivariate logistic regression training on a gold-standard video dataset labeled by experienced surgeons. The operational details involve collecting calibration video data encompassing various scenarios such as routine navigation, bleeding point localization, and tissue identification, and calculating the features in this calibration dataset, denoted as . Each video segment was scored by several senior doctors on a scale of 0-10, indicating its criticality. Based on this, As the independent variable, As dependent variables, they constitute a set The optimal data set is fitted using, for example, multivariate logistic regression analysis. In this way, for example, when locating bleeding points, weight It will be trained to its maximum; during normal cruise, weight It will be lower; This embodiment employs a linear weighted model to achieve an extremely lightweight fusion method involving only a few multiplications and additions, with very low computational cost and high interpretability. The weight of each feature is directly linked to clinical significance; it can output results with extremely high real-time performance, such as milliseconds, and is highly correlated with the clinical judgment of experienced surgeons. The score provides accurate and latency-free decision input for the downstream tuning logic.

[0021] Example 4 The tuning logic module, used to execute dual-threshold hysteresis control logic, includes: Define the states of high-precision computing paradigm and low-latency computing paradigm; When the criticality score of a surgical event exceeds the high criticality threshold, switch to the high-precision calculation paradigm. When the criticality score of a surgical event is less than the low criticality threshold, switch to the low-latency calculation paradigm. When the criticality score of a surgical event falls between the low criticality threshold and the high criticality threshold, the current calculation paradigm remains unchanged. Based on Example 1, this embodiment provides specific limitations on how the tuning logic module executes the dual-threshold hysteresis control logic, also commonly referred to as Schmitt trigger logic. To clarify the two computational paradigm states of system operation, they are defined as follows: High-precision computing paradigm state The system calls complex, computationally expensive enhancement algorithms such as Generative Adversarial Networks (GANs) and deep learning super-resolution models to maximize the diagnostic accuracy and clarity of the video, but this results in high computational latency and power consumption. Low-latency computing paradigm state The system only runs basic, low-computing-power algorithms such as basic denoising and sharpening, with the aim of ensuring absolute smoothness and minimum power consumption of the video stream, but the enhancement effect is limited. Its switching logic is as follows: In order to... and Achieve stable switching between them and avoid The score is in or Small fluctuations near the critical point cause system state jitter; the tuning logic module executes the following state machine logic in each calculation cycle, for example, every 100 milliseconds: when the criticality score of the surgical event... Greater than the high criticality threshold For example, if a sudden massive hemorrhage is detected, The number of cases surged, and the system switched to a high-precision computing paradigm. When the criticality score of the surgical event Less than the low criticality threshold For example, resuming regular cruise. The system has switched to a low-latency computing paradigm after the initial drop. When the criticality score of the surgical event Between low criticality threshold With high criticality threshold In between The system maintains the current computational paradigm state unchanged; This embodiment introduces... and During the hysteresis region, the system achieves extremely high state stability; the system will not be affected by... The score fluctuates slightly around the threshold, which is normal, such as due to slight camera movement. It avoids frequent and invalid switching of computing paradigms caused by value fluctuations; it enhances the consistency of system state and avoids additional computing power overhead and screen lag or flicker that can be perceived by users due to frequent switching.

[0022] Example 5 The pipeline switching module is used to execute asynchronous weight warm-up strategies, including: When the system is in a low-latency computing paradigm, and the criticality score of the surgical event crosses the low criticality threshold and enters the hysteresis region; Immediately and asynchronously load the high-precision model weights required for the high-precision computing paradigm state into the GPU memory of the edge computing unit in the background computing thread; The pipeline switching module is further used for: When the criticality score of the surgical event continues to rise and crosses the high criticality threshold, the high-precision model weights are pre-warmed in the video memory, avoiding model loading delay and achieving low-latency switching. When the criticality score of the surgical event falls back in the hysteresis region and drops below the low criticality threshold, the preheating is stopped or the loaded video memory is cleared to release computing resources. Based on Example 4, this example elaborates on how the pipeline switching module uses the hysteresis region of the tuning logic to execute the asynchronous weight warm-up strategy. The triggering conditions for warm-up and the two different results of the warm-up strategy, namely hit and abort, together constitute a complete low-latency switching scheme. The technical problem with this embodiment is that, from State switched to State, its total switching delay Mainly due to decision delay For millisecond-level and model loading latency High-precision model weights can be loaded from storage to VRAM in a timeframe of hundreds of milliseconds. This is the main cause of lag; the purpose of this strategy is to avoid it before the switch occurs by warming up. ; The preheating trigger logic occurs when the system is in a low-latency computing paradigm state. And the criticality score of the surgical event The upward value increases as it crosses the low criticality threshold. And enter the lag zone When this occurs, the system controller interprets it as a warning signal that a critical event may be imminent and immediately loads the necessary data asynchronously in a background computing thread. The high-precision model weights required for the state, such as .pth or .onnx files, are stored in the VRAM of edge computing units, such as GPUs; the technical consideration is that, since the system is still in a state of flux, this is necessary for the system to achieve its intended purpose. The current state indicates a low computational load, thus providing sufficient background resources for asynchronous loading with minimal impact on the foreground. The smoothness of the production line; The preheating result logic is as follows: When preheating is successful: after preheating begins, if... The score continued to rise and crossed the high criticality threshold. The tuning logic module switches to the input based on the output. The system receives the instruction; at this point, because the high-precision model weights have already been pre-warmed in the video memory, the system avoids model loading delays. Total switching delay That is, only the decision delay is reduced, thus achieving, for example, low latency and smooth switching from hundreds of milliseconds to milliseconds; when warm-up is aborted: after warm-up begins, if The score fluctuated within the hysteresis zone before falling back and dropping below the low criticality threshold. The system determines that the warning has been lifted, and the tuning logic module either holds or switches back based on the issued signal. The command indicates the status; at this point, the background thread will stop warming up if it is not yet complete or clear the loaded video memory to release valuable video memory and computing resources and avoid unnecessary occupation. This embodiment cleverly utilizes the hysteresis region of the control logic as a warning signal and buffer zone, thus mitigating high latency. The model loading action is hidden in the background and executed asynchronously; this enables the system to achieve near-instantaneous response and warm-up hit when it really needs to switch to high-precision mode, and to release resources and stop warm-up in time when there is a false alarm, thus optimizing the user experience and perfectly solving the problem of stuttering and latency spikes when switching models on edge devices under resource-constrained conditions.

[0023] Example 6 The preset high-keyness threshold and preset low-keyness threshold were initially obtained through statistical calibration using the gold standard dataset. The low criticality threshold is set to a value that can filter out noise peaks in the criticality score during normal cruise scenarios; The high criticality threshold is set to a value that can capture the lowest criticality score in criticality identification scenarios; Based on Example 4, this embodiment uses a preset high criticality threshold. Compared with the preset low criticality threshold The source and basis for the initial values ​​are explained; this provides the system with reasonable and evidence-based cold start parameters. and The initial values ​​were obtained through statistical calibration on the gold standard video dataset; against The calibration logic: low criticality threshold It is configured to filter calibration criticality scores in regular cruise scenarios. The value of the noise peak; its operable implementation involves extracting all video segments in the dataset that are labeled by doctors as routine cruises or non-critical events, and calculating the criticality scores of these segments. distributed; It is set as the high quantile of this distribution, such as the 95th or 99th quantile; the technical consideration is that... It must be high enough to ensure that the vast majority, such as 99%, of the noise generated by normal operation, such as normal lens movement and washing, is eliminated. Fluctuations, with their peaks still... Therefore, the state transition will not be triggered. Asynchronous preheating ensures low power consumption and stability of the system; against Calibration logic: High criticality threshold It is configured to capture key tissue identification or use criticality scores for labeling in bleeding scenarios. The valley value; its operable implementation involves extracting all video segments from the dataset that have been labeled as key events by doctors, such as bleeding or tissue identification, and calculating the labeling criticality score of these segments. distributed; Set as the lower quantile of the distribution, such as the 1st or 5th quantile, or even the minimum value, while ensuring that the value is strictly greater than... The technical considerations are as follows: Must be low enough but higher than To ensure that all or nearly 100% of events deemed critical by doctors are [critical / critical]. Even the lowest point of the score can be higher than This ensures high sensitivity and guarantees that critical events will be triggered. state; This embodiment employs a threshold setting method based on statistical calibration. Its gain effect lies in providing initial parameters with clear clinical significance and statistical basis for the tuning logic. To ensure stability and specificity, Ensuring sensitivity ensures that the system has high sensitivity and stability from the initial deployment stage, avoiding the uncertainty caused by manually setting parameters, and providing a reasonable and efficient starting point for subsequent online adaptive correction.

[0024] Example 7 The utility evaluation module, used to construct dynamic utility functions, includes: Real-time acquisition of normalized instantaneous diagnostic accuracy and normalized instantaneous energy consumption; Among them, the normalized instantaneous energy consumption is obtained by normalizing the instantaneous power with the preset maximum and minimum operating power; Based on a preset energy consumption trade-off factor, a Pareto trade-off is made between normalized instantaneous diagnostic accuracy and normalized instantaneous energy consumption to determine the cumulative utility within a preset time window, which is then used as the value of the dynamic utility function. Based on Example 1, this embodiment describes how the utility evaluation module constructs a dynamic utility function. This was explained in detail; this function is the mathematical foundation for achieving closed-loop optimization. To improve diagnostic accuracy For example, units are confidence level and system energy consumption. For example, to find a Pareto optimal balance between two conflicting performance objectives, each measured in watts, this embodiment constructs a unified dynamic utility function. Its technological motivation lies in solving... and The conflict in the dimensions between them necessitates normalization of both. Sub-features that have undergone normalization; Normalized instantaneous diagnostic accuracy : It is a dimensionless value in the interval [0,1]; its source is, for example, Enhance the diagnostic confidence of the model output under certain conditions, or A baseline value, such as 0.1, in a given state is normalized to the [0,1] interval using a mapping function such as Sigmoid or a linear mapping. Normalized instantaneous energy consumption : It is also a dimensionless value in the interval [0,1]; it is calculated by... With the preset maximum operating power and minimum operating power The formula is obtained by normalization calculation; Regarding the source of parameters: Instantaneous power, measured in watts, is collected in real time by the power consumption monitor of the edge computing unit; The unit watt is a physical parameter obtained through benchmark testing of hardware. Construction of the utility function: Based on the above normalized values, it is constructed within a preset time window. For example Cumulative utility per second Its formula is expressed as: ; in, In order to be in arrive The cumulative utility within the time window is dimensionless and is calculated by this module, serving as input to the threshold correction module. This is a dimensionless energy consumption trade-off factor; to ensure parameters Feasibility, As a key adjustable hyperparameter, it is preset by the system designer or user according to the power consumption constraints of the application scenario; for example, for battery-powered portable endoscopes, The value should be set relatively high, for example... This indicates that the system is more energy-efficient; for surgical towers with fixed power supply, The value can be set to a lower value, for example... This indicates that the system tends to pursue high precision. This embodiment constructs this dynamic utility function. It uses an interpretable trade-off factor to balance the two core performance indicators of the system: diagnostic accuracy and real-time energy consumption. This was unified into a single, quantifiable optimization objective. This makes system performance no longer a vague, multi-objective concept, but rather something that can be maximized. The values ​​are used for quantitative evaluation and iterative optimization, providing a solid mathematical foundation for achieving closed-loop adaptive correction.

[0025] Example 8 The threshold correction module, used for employing a gradient ascent strategy, includes: By applying a small perturbation to the high criticality threshold at runtime, the gradient of the dynamic utility function with respect to the high criticality threshold is approximately calculated. Based on the gradient and the preset learning rate, the high criticality threshold is iteratively adjusted to converge to the maximization of the dynamic utility function. Based on Example 7, this embodiment describes how the threshold correction module uses a gradient ascent strategy to adjust the threshold for high criticality. The process of closed-loop correction was explained. The goal of this module is to adjust control parameters. Similarly, adjustments can also be made. To maximize the dynamic utility function The technical motivation lies in employing the gradient ascent, or hill-climbing algorithm, to... Able to follow The increasing direction is iteratively corrected and eventually converges to The local or global optimum; The formula for the gradient ascent strategy is expressed as follows: ; in, For the next optimization cycle, such as the next time window of value; For the current period value; The learning rate is a small, dimensionless positive number, such as 0.001; it is derived from a pre-defined hyperparameter used to control the correction speed and stability. For utility For threshold The gradient is dimensionless. To ensure the feasibility of gradients, in real-world complex systems, utility With threshold analytic derivatives between It is difficult to calculate directly; therefore, this embodiment approximates the gradient by applying a small perturbation to the high criticality threshold during runtime; the operable steps involve, for example, using the finite difference method for approximation: in the current cycle, using Run the system and calculate In the next cycle, use in It is a preset small perturbation value, such as 0.01, used to calculate the system. Based on this, the gradient can be approximately calculated. Finally, using this approximate gradient and substituting it into the gradient ascent formula above, we can calculate... The value is used to complete one iteration of correction; This embodiment introduces a gradient-ascending closed-loop correction strategy, the gain effect of which is to enable the system to have online, runtime adaptive learning capabilities; the system no longer permanently depends on a static initial threshold, but can adapt to the energy efficiency-accuracy performance during actual operation. The value of the core control parameters is autonomously and continuously optimized. This allows the system to dynamically adapt to different surgical styles, such as aggressive or conservative surgeons, hardware aging, such as increased GPU power consumption, or specific clinical environments, such as smoke interference, and always automatically converge to the Pareto optimal operating state.

[0026] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A smart edge computing system for endoscopic image enhancement processing, characterized in that, include: The data acquisition module is used to acquire the raw video frame sequence of the edge endoscope device in real time; The feature extraction module is used to determine instantaneous feature vectors representing the dynamic changes of the surgical scene based on the original video frame sequence; The criticality quantification module is used to quantify and generate criticality scores for surgical events based on instantaneous feature vectors. The tuning logic module is used to execute dual-threshold hysteresis control logic based on the criticality score of the surgical event, the preset high criticality threshold, and the preset low criticality threshold, in order to generate discrete and stable computational paradigm switching instructions. The pipeline switching module is used to respond to computational paradigm switching instructions and execute asynchronous weight warm-up strategies by utilizing the hysteresis region of the dual-threshold hysteresis control logic. The utility evaluation module is used to construct dynamic utility functions to evaluate system performance. The threshold correction module is used to perform closed-loop correction of high criticality thresholds based on a gradient ascent strategy using a dynamic utility function.

2. The intelligent edge computing system for endoscopic image enhancement processing according to claim 1, characterized in that, The feature extraction module is used to determine the instantaneous feature vector, including: The field of view motion vector is obtained by analyzing the mean value of the optical flow field between consecutive frames; The proportion of the blood color gamut is obtained by statistically analyzing the pixel proportion of the current frame in a preset blood color space model. Tissue texture heterogeneity is obtained by calculating the Mahalanobis distance between the feature map extracted by a lightweight convolutional neural network and the preset conventional tissue baseline feature map.

3. The intelligent edge computing system for endoscopic image enhancement processing according to claim 1, characterized in that, The criticality quantification module is used to quantify and generate criticality scores for surgical events based on instantaneous feature vectors, including: A custom linear weighted fusion model is used to perform a weighted summation of the field of view motion vector, blood color gamut proportion, and tissue texture heterogeneity; The model's weight coefficients were obtained through offline calibration or regression training on a gold-standard visual dataset annotated by senior surgeons.

4. The intelligent edge computing system for endoscopic image enhancement processing according to claim 1, characterized in that, The tuning logic module is used to execute dual-threshold hysteresis control logic, including: Define the states of high-precision computing paradigm and low-latency computing paradigm; When the criticality score of a surgical event exceeds the high criticality threshold, switch to the high-precision calculation paradigm. When the criticality score of a surgical event is less than the low criticality threshold, switch to the low-latency calculation paradigm. When the criticality score of a surgical event falls between the low criticality threshold and the high criticality threshold, the current calculation paradigm remains unchanged.

5. The intelligent edge computing system for endoscopic image enhancement processing according to claim 1, characterized in that, The pipeline switching module is used to execute an asynchronous weight preheating strategy, including: When the system is in a low-latency computing paradigm, and the criticality score of the surgical event crosses the low criticality threshold and enters the hysteresis region; Immediately load the high-precision model weights required for the high-precision computing paradigm state asynchronously into the memory of the edge computing unit in the background computing thread.

6. The intelligent edge computing system for endoscopic image enhancement processing according to claim 5, characterized in that, The pipeline switching module is further used for: When the criticality score of the surgical event continues to rise and crosses the high criticality threshold, the high-precision model weights are pre-warmed in the video memory, avoiding model loading delay and achieving low-latency switching. When the criticality score of a surgical event falls back within the hysteresis region and drops below the low criticality threshold, the preheating process is stopped or the loaded video memory is cleared to release computing resources.

7. The intelligent edge computing system for endoscopic image enhancement processing according to claim 1, characterized in that, The preset high-keyness threshold and preset low-keyness threshold were initially obtained through statistical calibration using the gold standard dataset. The low criticality threshold is set to a value that can filter out noise peaks in the criticality score during normal cruise scenarios; The high criticality threshold is set to a value that can capture the lowest criticality score in critical organization identification scenarios.

8. The intelligent edge computing system for endoscopic image enhancement processing according to claim 1, characterized in that, The utility evaluation module is used to construct a dynamic utility function, including: Real-time acquisition of normalized instantaneous diagnostic accuracy and normalized instantaneous energy consumption; Among them, the normalized instantaneous energy consumption is obtained by normalizing the instantaneous power with the preset maximum and minimum operating power; Based on a preset energy consumption trade-off factor, a Pareto trade-off is performed between normalized instantaneous diagnostic accuracy and normalized instantaneous energy consumption to determine the cumulative utility within a preset time window, which is then used as the value of the dynamic utility function.

9. The intelligent edge computing system for endoscopic image enhancement processing according to claim 1, characterized in that, The threshold correction module is used to employ a gradient ascent strategy, including: By applying a small perturbation to the high criticality threshold at runtime, the gradient of the dynamic utility function with respect to the high criticality threshold is approximately calculated. Based on the gradient and the preset learning rate, the high criticality threshold is iteratively adjusted to converge to the maximization of the dynamic utility function.

Images