A deep learning-based real-time positioning monitoring system for recurrent laryngeal nerve surgery

Abstract

This invention relates to the field of object recognition technology for nerves, and more particularly to a deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery. The system includes a preoperative analysis module to acquire the expected nerve path sequence to determine a first variation risk region; an intraoperative acquisition module to acquire intraoperative video, electromyographic signals, and sequences corresponding to instrument positions; a priori constraint recognition module to determine the intraoperative nerve recognition path sequence and visual recognition confidence level; a path deviation correction module to correct the intraoperative nerve recognition path sequence and visual recognition confidence level; a distance state analysis module to determine whether the real-time distance between the instrument and the nerve is within a distance boundary state; and a warning monitoring module to determine the existence of hidden warning regions based on spontaneous electromyographic activity indicators combined with visual recognition correction confidence level, and to monitor the rate of change of spontaneous electromyographic activity indicators to indicate the nerve stress response state. This invention improves the accuracy of real-time intraoperative localization and monitoring of recurrent laryngeal nerve surgery.

Description

Technical Field

[0001] This invention relates to the field of object recognition technology for nerves, and more particularly to a real-time intraoperative localization and monitoring system for the recurrent laryngeal nerve based on deep learning. Background Technology

[0002] With the increasing prevalence of cervical and thoracic surgeries such as thyroidectomy, parathyroidectomy, and neck lymph node dissection, recurrent laryngeal nerve injury, as one of the most serious intraoperative complications, has attracted widespread attention. To reduce the risk of nerve injury, current techniques utilize intraoperative nerve monitoring. This involves recording laryngeal muscle electromyography (EMG) signals using electrodes on the endotracheal intubation tube, combined with probes to induce EMG responses in the tissue, to achieve nerve localization and functional assessment. This has become an important means of assisting in the identification and protection of the recurrent laryngeal nerve during thyroid surgery.

[0003] However, the aforementioned existing technologies have significant limitations in clinical applications. Traditional monitoring methods are intermittent, and nerve damage cannot be detected during the time intervals between explorations, creating numerous monitoring blind spots. When nerves are subjected to continuous traction, heat conduction, or compression damage during surgery, the system cannot provide timely warnings. Visual recognition struggles to accurately identify nerves in complex anatomical variations and cannot provide warnings when nerves are covered by tissue. Furthermore, when intraoperative tissue traction and displacement cause significant deviations between preoperative images and actual positions, preoperative images cannot guide real-time intraoperative procedures. More critically, existing technologies lack a real-time fusion verification mechanism between visual localization and electrophysiological function; visual recognition and electrophysiological monitoring are disconnected, failing to form a complementary verification loop.

[0004] Chinese Patent Publication No. CN120526290A discloses a method and system for real-time identification and dynamic tracking of the recurrent laryngeal nerve based on a neural network. The method includes selecting a predetermined number of frame-annotated images of the recurrent laryngeal nerve region, aligning edge features and semantic features with specific features of the recurrent laryngeal nerve, and adjusting the parameters of the neural network to obtain initial neural network parameters. Based on the semantic similarity between the predetermined number of frame-annotated images of the recurrent laryngeal nerve region and real-time acquired images of the recurrent laryngeal nerve region, the initial neural network parameters are dynamically adjusted to obtain final neural network parameters. The final neural network parameters are then used to perform real-time identification and dynamic tracking of the recurrent laryngeal nerve in real-time acquired intraoperative videos of the recurrent laryngeal nerve, resulting in real-time identification and dynamic tracking results of the recurrent laryngeal nerve.

[0005] Therefore, the following problems exist in the existing technology: This invention aims to overcome the problems in the prior art that do not consider the real-time cross-validation relationship between visual position and electrophysiological function during recurrent laryngeal nerve surgery, and do not consider the dynamic collaborative early warning relationship between intraoperative spatial distance information and functional status information. These problems result in the inability to verify visual segmentation when there are misjudgments, and the inability to perform linkage analysis on the dual abnormalities of instrument spatial proximity and nerve function deterioration, leading to low accuracy of real-time positioning monitoring during recurrent laryngeal nerve surgery. Summary of the Invention

[0006] To address this, the present invention provides a deep learning-based real-time positioning and monitoring system for recurrent laryngeal nerve surgery. This system overcomes the problems in existing technologies, such as the failure to consider the real-time cross-validation relationship between visual position and electrophysiological function during recurrent laryngeal nerve surgery, the failure to consider the dynamic collaborative early warning relationship between intraoperative spatial distance information and functional status information, the inability to verify visual segmentation by combining it with electrophysiological monitoring when there are misjudgments, and the inability to perform linkage analysis on the dual abnormalities of instrument spatial proximity and nerve function deterioration, which in turn lead to low accuracy in real-time positioning and monitoring during recurrent laryngeal nerve surgery.

[0007] To achieve the above objectives, the present invention provides a deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery, comprising: The preoperative analysis module obtains the expected neural pathway sequence based on the preoperative neck images of the target user, analyzes the path uncertainty index based on the path points corresponding to the expected neural pathway sequence, and compares the path uncertainty index with a preset path uncertainty index threshold to determine the first variation risk area. The intraoperative acquisition module is used to acquire in real time the intraoperative continuous video frame sequence of the target user's neck, the intraoperative electromyographic signal sequence, and the intraoperative instrument position sequence. The prior constraint recognition module performs prior constraints on the intraoperative neck continuous video frame sequence based on the neural expected path sequence and the first variation risk region combined with preset constraint coefficients, so as to determine the intraoperative neural recognition path sequence and the corresponding visual recognition confidence. The path deviation correction module performs a real-time deviation comparison between the intraoperative neurological identification path sequence and the expected neurological path sequence to determine whether to dynamically adjust the constraint coefficient, and performs real-time correction on the intraoperative neurological identification path sequence based on the adjusted constraint coefficient to determine the intraoperative neurological identification path correction sequence and the corresponding visual recognition correction confidence level. The distance state analysis module compares the real-time distance between the neurosurgical identification path correction sequence and the intraoperative instrument position sequence with a preset distance threshold to determine whether the real-time distance is in a distance boundary state. The warning monitoring module is used to respond to the real-time distance being at the distance boundary state, analyze spontaneous electromyographic activity indicators based on the intraoperative electromyographic signal sequence, and perform consistency verification based on the visual recognition correction confidence and the spontaneous electromyographic activity indicators to determine whether there is a hidden warning area, and in response to the existence of a hidden warning area, monitor the rate of change of the spontaneous electromyographic activity indicators to indicate the state of neural stress response.

[0008] Furthermore, the preoperative analysis module calculates the local curvature based on adjacent path points in the expected neural path sequence to determine the path uncertainty index.

[0009] Furthermore, the preoperative analysis module determines the first variation risk region based on the comparison results of the path uncertainty index being greater than a preset path uncertainty index threshold.

[0010] Furthermore, the prior constraint recognition module determines the intraoperative neurosurgical recognition path sequence based on the prior projection region sequence and the enhanced region sequence combined with the corresponding constraint coefficients, and determines the visual recognition confidence level based on the probability value corresponding to the intraoperative neurosurgical recognition path sequence. Specifically, the prior projection region sequence corresponding to the intraoperative continuous video frame sequence of the neck is determined based on the neural expected path sequence, and the enhancement region sequence of the prior projection region sequence is determined based on the first variation risk region.

[0011] Furthermore, the path deviation correction module calculates the real-time path deviation based on the intraoperative neurosurgical path sequence and the expected neurosurgical path sequence, and determines to dynamically adjust the constraint coefficient based on the comparison result of the real-time path deviation being greater than a preset path deviation threshold.

[0012] Furthermore, the path deviation correction module performs real-time correction on the neurosurgical identification path sequence based on the first constraint correction coefficient and the second constraint correction coefficient to determine the neurosurgical identification path correction sequence and the corresponding visual recognition correction confidence level. Wherein, the first constraint correction coefficient is determined after reducing and adjusting the first constraint coefficient according to the first preset amplitude; The second constraint correction coefficient is determined by amplifying and adjusting the second constraint coefficient according to the second preset amplitude.

[0013] Furthermore, the distance state analysis module determines that the real-time distance is in a distance boundary state based on the comparison result of the real-time distance between the neurosurgical identification path correction sequence and the intraoperative instrument position sequence being less than a preset distance threshold.

[0014] Furthermore, the warning monitoring module determines spontaneous electromyographic activity indicators based on the ratio of real-time electromyographic activity intensity to a preset real-time electromyographic activity benchmark intensity. The real-time electromyographic activity intensity is determined based on the root mean square amplitude of the electromyographic signal within a first preset time window calculated from the intraoperative electromyographic signal sequence.

[0015] Furthermore, the warning monitoring module determines the existence of a hidden warning area based on the fact that both the visual recognition correction confidence level and the spontaneous electromyography activity index meet the first preset condition; The first preset condition is that the visual recognition correction confidence level is less than a preset visual recognition confidence level threshold and the spontaneous electromyographic activity index is greater than a preset spontaneous electromyographic activity index threshold.

[0016] Furthermore, the warning monitoring module is used to calculate the rate of change of neural electrical activity of the spontaneous electromyographic activity index within a second preset time window, and to compare the rate of change of neural electrical activity with preset first and second thresholds for the rate of change of neural electrical activity, so as to indicate the state of neural stress response.

[0017] Compared with existing technologies, the advantages of this invention lie in its complete prior transfer system, which integrates preoperative image anatomical analysis with intraoperative visual recognition guidance. This system applies prior constraints to the deep learning segmentation process using the expected neural path sequence and the first variation risk region, enabling visual recognition to obtain the expected location of the nerve pathway and the distribution information of highly variable regions at the beginning of surgery, reducing the probability of segmentation misjudgment due to intraoperative tissue deformation. The path deviation correction module calculates the deviation between the intraoperative recognition path and the preoperative expected path in real time and dynamically adjusts the constraint coefficients, enabling adaptive recognition of anatomical variations in non-recurrent laryngeal nerves. When there are significant differences between the intraoperative anatomical structure and the preoperative model, it automatically reduces dependence on the prior path and enhances visual sensitivity to variable regions, solving the problem of recognition failure caused by mismatch between preoperative images and intraoperative anatomy. By employing a dual-condition triggering mechanism based on both instrument spatial distance information and electromyographic function information, high-intensity electrophysiological analysis is initiated when the instrument approaches the nerve. Simultaneously, when the distance between the instrument and the nerve is normal but electromyographic activity is abnormally elevated, a hidden warning area determination is independently triggered. This forms a bidirectional complementary early warning capability based on spatial proximity and functional abnormality, thereby improving the accuracy of real-time positioning monitoring during recurrent laryngeal nerve surgery.

[0018] Furthermore, this invention achieves preoperative neural course variation risk analysis based on geometric quantification by calculating the local curvature of adjacent path points in the expected neural path sequence and using this as a path uncertainty index. A greater curvature indicates a more drastic change in the expected direction of the nerve at that location, meaning a significantly increased probability of abnormal bending, sharp turns, or anatomical variations in the nerve at that point. This provides a clear quantitative basis for enhancing sensitivity during subsequent intraoperative visual segmentation. By introducing the aforementioned path uncertainty index into the system's preoperative analysis module, this invention enables the deep learning visual segmentation network to identify in advance which areas within the central cervical region are high-risk variation areas and automatically improve the sensitivity of nerve detection within these areas, thereby improving the real-time performance and accuracy of variation identification and nerve localization during dissection.

[0019] Furthermore, this invention identifies path points with path uncertainty indices exceeding a threshold and their spatial neighborhoods as first variation risk regions, transforming high variation risk locations identified in preoperative image analysis into spatial region labels directly usable for intraoperative visual segmentation. The first variation risk region marks which areas require higher visual detection sensitivity, enabling the deep learning segmentation network to preferentially identify neural structures within these regions. By proactively labeling variation risk regions preoperatively, preoperative prediction and proactive intraoperative enhancement of detection are achieved.

[0020] Furthermore, this invention corrects the original segmentation probability map by combining prior projection region sequences and enhanced region sequences with constraint coefficients. This achieves a deep fusion of anatomical prior information and deep learning segmentation results. This design allows the visual segmentation network to focus on the expected nerve pathway region using preoperative images when performing nerve recognition, reducing false positives such as misidentifying fascial cords, blood vessels, and other nerve-like structures as the recurrent laryngeal nerve. By combining the enhanced region sequence corresponding to the first variation risk region with the corresponding constraint coefficients, the sensitivity of nerve detection can be automatically improved in areas with high incidence of anatomical variations, ensuring that nerves can still be reliably identified in variation scenarios. The calculation of visual recognition confidence is based on the average probability of all pixels within the corrected probability map contour. This confidence reflects the model's certainty regarding the recognition result and is indirectly regulated by the prior constraint coefficients.

[0021] Furthermore, this invention establishes a self-checking mechanism for intraoperative and preoperative consistency by calculating the deviation between the intraoperative neurological path sequence and the preoperative expected neurological path sequence in real time and comparing it with a preset threshold. This mechanism can continuously monitor the degree of deviation between the actual intraoperative anatomical structure and the preoperative neurological identification during the operation. If the deviation exceeds the threshold, the system automatically determines that the reliability of the prior information of the preoperative neurological identification has decreased, thereby triggering a dynamic adjustment process of the constraint coefficient. Compared with static reliance on preoperative image guidance, the deviation comparison mechanism of this invention enables the system to perceive intraoperative anatomical changes and actively adjust the guidance strategy. When encountering rare variations such as non-recurrent laryngeal nerve, the system can autonomously reduce its reliance on preoperative erroneous information, avoiding the risk of preoperative models misleading intraoperative operations.

[0022] Furthermore, this invention achieves refined dynamic correction of the constraint coefficients in intraoperative scenarios by adjusting the first constraint correction coefficient and the second constraint correction coefficient separately. The reduction adjustment of the first constraint coefficient enables the system to reduce its spatial dependence on the expected path region when it detects a deviation between the preoperative model and the actual anatomy, avoiding the neural network from mistaking non-neural structures for neurons due to erroneous prior information. The amplification adjustment of the second constraint coefficient enhances the system's visual detection sensitivity to areas of variation risk under the same conditions, ensuring that even when the actual neural pathway deviates completely from the expected path, the system can still capture neural image features through enhanced visual sensitivity.

[0023] Furthermore, this invention compares the real-time distance with a distance threshold to determine if the real-time distance is at a distance boundary, providing a precise and quantifiable basis for judging the distance boundary state between the device and the nerve. Continuous calculation of the real-time minimum spatial distance enables the system to capture every minute change in distance between the device and the nerve, providing temporally aligned spatial conditions for the precise triggering of subsequent electrophysiological analysis, and enabling joint analysis of spatial proximity risk and electrophysiological monitoring response.

[0024] Furthermore, this invention uses the root mean square amplitude (RMS amplitude) as the real-time electromyographic activity intensity and the average value five minutes before the start of surgery as the baseline intensity. It then calculates the ratio of the real-time EMG intensity to the baseline intensity to determine the spontaneous EMG activity index, thus constructing a standardized spontaneous EMG activity assessment index that eliminates the influence of individual differences, electrode contact quality, and initial muscle tension. The RMS amplitude is the most commonly used time-domain quantization parameter for EMG signal energy, effectively reflecting the combined effect of motor unit recruitment and firing frequency, and exhibits better stability and noise resistance compared to the original amplitude value.

[0025] Furthermore, this invention establishes a dual-modal complementary verification system of visual and electrophysiological mechanisms by simultaneously satisfying a visual recognition confidence level less than a preset visual recognition confidence threshold and a spontaneous electromyographic activity index greater than a preset spontaneous electromyographic activity index threshold, thereby determining the hidden warning region. Low visual confidence indicates that the deep learning segmentation network has not detected a clear neural structure in the current field of vision, possibly because the nerve is covered by fascia, located in a visual blind spot, or in adhesion tissue. Conversely, high spontaneous electromyographic activity indicates that even if the visual system does not recognize the nerve, the spontaneous electrical activity of the laryngeal muscles is significantly higher than the baseline level, suggesting that the nerve is being stimulated in some form. The conclusions of the two modalities should normally be consistent; when a contradictory combination occurs, it highly likely indicates the presence of an occluded nerve. This hidden warning region determination logic based on dual-modal contradictions solves the blind spot problem of pure visual methods in occluded scenarios, enabling the system to indirectly locate nerves through abnormal electrophysiological signals even when vision is completely ineffective.

[0026] Furthermore, this invention constructs a hierarchical prompting system for neural stress response based on the dynamic rate of change of electrical activity by calculating the rate of change of spontaneous electromyographic activity indicators within a second preset time window and comparing it stepwise with the first and second thresholds for the rate of change of electrical activity. This hierarchical prompting is entirely based on quantifiable electrophysiological amplitude change parameters, providing trend-based auxiliary information for intraoperative procedures. Attached Figure Description

[0027] Figure 1 This is a module connection diagram of the deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery, according to an embodiment of the present invention. Figure 2 This is a logic decision diagram for determining the first mutation risk region in an embodiment of the present invention; Figure 3 A flowchart illustrating the workflow for determining the intraoperative neurosurgical identification path sequence and visual identification confidence level in embodiments of the present invention; Figure 4 This invention provides a logic decision diagram for dynamically adjusting constraint coefficients in an embodiment of the invention. Detailed Implementation

[0028] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0029] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0030] It should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0031] As an explanation, the deep learning-based real-time intraoperative localization and monitoring system for the recurrent laryngeal nerve in this embodiment of the invention is mainly applied to cervical and thoracic surgical scenarios requiring exposure and protection of the recurrent laryngeal nerve, providing real-time auxiliary monitoring for difficulties in nerve identification and abnormal nerve activity. Specifically, this includes, but is not limited to, nerve localization assistance in thyroidectomy, parathyroidectomy, neck lymph node dissection, substernal goiter surgery, recurrent thyroid surgery, and recurrent laryngeal nerve reconstruction. In these scenarios, bleeding and tissue exudation around the inferior thyroid artery may obscure the area where the recurrent laryngeal nerve is located, and anatomical variations of the recurrent laryngeal nerve may frequently lead to difficulties in intraoperative nerve identification. The system described in this embodiment of the invention is particularly suitable for real-time auxiliary monitoring of difficulties in intraoperative recurrent laryngeal nerve identification and abnormal nerve activity.

[0032] Please see Figure 1 The diagram shown is a module connection diagram of the deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery according to an embodiment of the present invention. The deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery according to an embodiment of the present invention includes: The preoperative analysis module obtains the expected neural pathway sequence based on the preoperative neck images of the target user, analyzes the path uncertainty index based on the path points corresponding to the expected neural pathway sequence, and compares the path uncertainty index with a preset path uncertainty index threshold to determine the first variation risk area. The intraoperative acquisition module is used to acquire in real time the intraoperative continuous video frame sequence of the target user's neck, the intraoperative electromyographic signal sequence, and the intraoperative instrument position sequence. The prior constraint recognition module performs prior constraints on the intraoperative neck continuous video frame sequence based on the neural expected path sequence and the first variation risk region combined with preset constraint coefficients, so as to determine the intraoperative neural recognition path sequence and the corresponding visual recognition confidence. The path deviation correction module performs a real-time deviation comparison between the intraoperative neurological identification path sequence and the expected neurological path sequence to determine whether to dynamically adjust the constraint coefficient, and performs real-time correction on the intraoperative neurological identification path sequence based on the adjusted constraint coefficient to determine the intraoperative neurological identification path correction sequence and the corresponding visual recognition correction confidence level. The distance state analysis module compares the real-time distance between the neurosurgical identification path correction sequence and the intraoperative instrument position sequence with a preset distance threshold to determine whether the real-time distance is in a distance boundary state. The warning monitoring module is used to respond to the real-time distance being at the distance boundary state, analyze spontaneous electromyographic activity indicators based on the intraoperative electromyographic signal sequence, and perform consistency verification based on the visual recognition correction confidence and the spontaneous electromyographic activity indicators to determine whether there is a hidden warning area, and in response to the existence of a hidden warning area, monitor the rate of change of the spontaneous electromyographic activity indicators to indicate the state of neural stress response.

[0033] Specifically, the preoperative analysis module obtains the expected neural pathway sequence based on the target user's preoperative neck images.

[0034] In this embodiment, the expected neural pathway sequence, which is based on the preoperative neck images of the target user, is used to provide preoperative anatomical prior information for intraoperative neural localization and characterize the conventional course of the recurrent laryngeal nerve.

[0035] The preoperative neck image refers to the central cervical region image from the preoperative enhanced computed tomography (CT) scan of the neck of the target user. The central cervical region extends from the lower edge of the cricoid cartilage to the upper edge of the clavicle. In this embodiment, automated semantic segmentation of key anatomical structures within the central cervical region is performed. A three-dimensional convolutional neural network can be used to determine binary masks for each anatomical structure, providing anatomical boundary constraints for generating the expected neural pathways. As a preferred embodiment, a three-dimensional convolutional neural network, such as a three-dimensional U-shaped network architecture, is used to automatically perform semantic segmentation of key anatomical structures within the central cervical region, including the tracheoesophageal groove, thyroid gland, and carotid sheath, to determine the binary masks for each anatomical structure. Based on the conventional anatomical rules of the prior art, which correspond to the recurrent laryngeal nerve ascending along the tracheoesophageal groove and crossing the inferior thyroid artery at the lower pole of the thyroid gland, an equidistant point set is generated on the central curve of the region determined by the above segmentation results. The distance between adjacent points is set to one millimeter, thereby determining the expected neural pathway sequence.

[0036] Specifically, the preoperative analysis module calculates local curvature based on adjacent path points in the expected neural path sequence to determine the path uncertainty index.

[0037] In this embodiment, the path uncertainty index, determined based on the local curvature corresponding to the path point in the expected neural path sequence, is used to quantify the degree of local directional change of the recurrent laryngeal nerve in the expected course path, in order to identify potential locations of variation risk and to distinguish between normal physiological curvature and abnormal course deviation. When the path point is located in the natural transition segment of the physiological curvature of the tracheoesophageal groove, its curvature is small; when the path point is located in the bypass crossing segment of the lower pole of the thyroid gland, its curvature will increase; when the actual course of the nerve at the location of the path point deviates from the normal curvature, the calculated value of the local curvature will deviate significantly from the normal range, thereby being used to identify potential locations of variation risk.

[0038] The local curvature, determined by the ratio of the angle between adjacent path points in the expected neural path sequence to a preset reference angle, reflects the degree of local directional change of the recurrent laryngeal nerve along its expected path. A greater local curvature indicates a steeper nerve trajectory within that segment, increasing the likelihood of abnormal curvature or anatomical variations. In this embodiment, the determination of the angle between the (i-1), i, and i+1 points in the path point sequence is as follows: using the i-th point as the vertex, the angle between the vector pointing from the (i-1)-th point to the i-th point and the vector pointing from the i-th point to the i+1-th point is calculated. This angle ranges from 0° to 180°. Preferably, the preset reference angle is set to 90°. An angle less than 90° indicates a gentler directional change, while an angle greater than 90° indicates a more drastic directional change. Choosing 90° as the preset reference angle effectively distinguishes between normal physiological curvature and abnormal curvature or anatomical variations requiring attention.

[0039] Please see Figure 2 As shown, it is a logic decision diagram for determining the first mutation risk region in an embodiment of the present invention.

[0040] Specifically, the preoperative analysis module determines the first variation risk region based on the comparison results of the path uncertainty index being greater than a preset path uncertainty index threshold.

[0041] In this embodiment, the preoperative analysis module includes a first variation risk region determination unit. This unit generates a spherical spatial region centered on the spatial coordinates of high-variation-risk path points to determine the first variation risk region. This region is used to indicate anatomical areas where the geometry of the preoperative expected path exhibits abnormal distortion. The first variation risk region characterizes a set of locations within the central neck region where the local curvature significantly deviates from the typical range of its own path. It assigns higher sensitivity to intraoperative visual segmentation, indicating abnormal distortion in the geometry of the preoperative expected path at these locations, requiring greater attention during intraoperative visual segmentation. By using the first variation risk region as input for the enhanced region sequence in the subsequent prior constraint recognition module, the system can identify anatomical areas requiring special attention before surgery begins, assigning higher visual segmentation sensitivity to these areas.

[0042] As a preferred implementation, the spatial coordinates of the high-variance-risk path point are used as the center point, and a spherical spatial region with a radius of three millimeters is generated. This spherical region is the first variation-risk region. When the spherical regions generated by multiple adjacent high-variance-risk path points overlap, these overlapping spherical regions are merged into a connected first variation-risk region.

[0043] In this embodiment, high-risk path points are identified by comparing each path point's path uncertainty index with a preset path uncertainty index threshold point by point. This method is used to identify path points where local directional changes significantly deviate from the typical range. When the path uncertainty index of a path point exceeds the path uncertainty index threshold, that path point is determined to be a high-risk path point.

[0044] In this embodiment, the path uncertainty index threshold is determined based on the path uncertainty index corresponding to all path points in the expected neural path sequence, and is used to define the upper limit of the typical fluctuation range of the path. As a preferred implementation, the process of determining the path uncertainty index threshold specifically involves: obtaining the path uncertainty index corresponding to each path point in the expected neural path sequence, determining the median and interquartile range of the path uncertainty index, and the path uncertainty threshold being the sum of the median and twice the interquartile range. For explanation, if the path uncertainty index of the recurrent laryngeal nerve exceeds the sum of its own median and twice the interquartile range, it means that the local directional change at that point has significantly exceeded the typical fluctuation range of the target user's recurrent laryngeal nerve path, which may correspond to anatomical variations such as non-recurrent nerves causing abnormal curvature of the expected path or image segmentation errors.

[0045] Specifically, the intraoperative acquisition module is used to acquire in real time the intraoperative continuous video frame sequence of the target user's neck, the intraoperative electromyographic signal sequence, and the intraoperative instrument position sequence.

[0046] In this embodiment, the intraoperative continuous neck video frame sequence, which is a sequence of continuously acquired video frames of the target user's neck during surgery, is used to provide a visual data source for intraoperative nerve localization. As a preferred implementation, the intraoperative acquisition module uses a surgical microscope or endoscopic camera to continuously acquire video images of the central area of ​​the neck in the surgeon's operating region at a frequency of no less than 30 frames per second, thereby obtaining the intraoperative continuous neck video frame sequence.

[0047] In this embodiment, the intraoperative electromyography (EMG) signal sequence is a sequence composed of real-time acquired EMG signals of the target user during surgery. It is used to record the spontaneous electrical activity of the intralarynx muscles under nerve innervation and the electrical activity responses induced by stimulation probes. As a preferred embodiment, a modified neuromonitoring endotracheal cannula is used, with a four-channel annular surface electrode embedded in its wall. Each electrode has a contact area of ​​25 square millimeters and an electrode spacing of 8 millimeters. The electrical activity signals of the intralarynx muscle region are continuously acquired at a sampling rate of 10,000 sampling points per channel per second to obtain the intraoperative EMG signal sequence. This EMG signal sequence records the spontaneous electrical activity of the intralarynx muscles under nerve innervation and the electrical activity responses induced by stimulation probes.

[0048] In this embodiment, the intraoperative instrument position sequence, which is a sequence of instrument positions acquired in real time by the target user during the operation, is used to track the three-dimensional spatial position of the surgical instrument tip within the operating range of the central cervical region. As a preferred implementation, an existing electromagnetic tracking method is used to acquire the three-dimensional spatial coordinates of the surgical instrument tip at a frequency of 60 times per second. Specifically, the instrument refers to surgical tools that may be close to the recurrent laryngeal nerve within the operating range of the central cervical region, including but not limited to dissecting forceps, ultrasonic scalpels, and suction devices, to acquire the intraoperative instrument position sequence.

[0049] Please see Figure 3 As shown, it is a flowchart of the process for determining the intraoperative neurosurgical identification path sequence and visual identification confidence level in an embodiment of the present invention.

[0050] Specifically, the prior constraint recognition module determines the intraoperative neurosurgical recognition path sequence based on the prior projection region sequence and the enhanced region sequence combined with the corresponding constraint coefficients, and determines the visual recognition confidence based on the probability value corresponding to the intraoperative neurosurgical recognition path sequence. Specifically, the prior projection region sequence corresponding to the intraoperative continuous video frame sequence of the neck is determined based on the neural expected path sequence, and the enhancement region sequence of the prior projection region sequence is determined based on the first variation risk region.

[0051] In this embodiment, intraoperative video frame preprocessing includes acquiring and preprocessing the current frame image from a continuous intraoperative neck video frame sequence to provide standardized input for the deep learning segmentation network. As a preferred implementation, the prior constraint recognition module acquires the current frame image from the continuous intraoperative neck video frame sequence and preprocesses it, including performing white balance correction, brightness normalization, and scaling to 512×512 pixels using any method in the prior art. The preprocessed image is then input into an improved U-Net++ deep learning segmentation network in the prior art, outputting a three-channel original segmentation probability map.

[0052] In this embodiment, the prior projection region sequence, generated based on the projection of the neural expected path sequence into the current video frame coordinate system, is used to define the approximate spatial range of the neural expected path. As a preferred implementation, a strip-shaped region sequence extending 2.5 mm to each side along the projection centerline is used to generate the prior projection region sequence.

[0053] In this embodiment, the enhanced region sequence, determined based on the spatial relationship between the projection of the first mutation risk region and the prior projection region sequence, is used to identify regions of special interest requiring higher visual segmentation sensitivity. As a preferred implementation, after projecting the first mutation risk region onto the current video frame coordinate system, the overlapping portion between this projection and the prior projection region sequence is taken as the enhanced region sequence.

[0054] In this embodiment, the prior constraint recognition module includes a constraint coefficient determination unit, which includes a first constraint coefficient determined based on the difference between a first probability and a second probability, and a second constraint coefficient determined based on the difference between a third probability and the second probability. These constraints are used to quantify the degree of difference in neural classification probabilities within and outside the prior projection region and within the enhancement region. The constraint coefficients are used to compensate for the visual segmentation sensitivity of neural signals in different regions by differentiating the constraint strength. A larger constraint coefficient pulls the original probability towards the upper probability limit; a smaller constraint coefficient pulls it closer. The second constraint coefficient is greater than the first constraint coefficient, thus providing a stronger enhancement effect to compensate for the visual segmentation sensitivity of neural signals within the enhancement region. The constraint coefficients are used to amplify the probability value of a pixel belonging to the recurrent laryngeal nerve in the probability map output by deep learning, based on whether the pixel coordinates are located within the prior projection region and whether they are located within the first mutation risk region.

[0055] Specifically, the first probability is determined based on the median of the original neural classification probabilities of all pixels within the prior projection region in historical data where no abnormalities were found during or after surgery. The second probability is determined based on the median of the original neural classification probabilities of all pixels outside the prior projection region in historical data where no abnormalities were found during or after surgery. The third probability is determined based on the median of the original neural classification probabilities of all pixels within the first variation risk region in historical data where no abnormalities were found during or after surgery.

[0056] In this embodiment, the prior constraint probability enhancement includes linearly interpolating the original neural classification probability based on the pixel coordinate position and corresponding constraint coefficients to linearly bring the probability closer to the upper probability limit, thereby improving the neural recognition sensitivity. As a preferred embodiment, if the pixel coordinates are located within the prior projection region sequence but not within the enhancement region sequence, a first constraint coefficient is used to linearly interpolate between the original probability and the upper probability limit, correcting the probability to the product of the original probability and the first constraint coefficient, thus linearly bringing the probability closer to its upper limit and determining the enhanced neural classification probability. If the pixel coordinates are located within both the prior projection region sequence and the enhancement region sequence, a second constraint coefficient is used to linearly interpolate between the original neural classification probability and the upper probability limit for that pixel, correcting the probability to the product of the original probability and the second constraint coefficient, thus linearly bringing the probability closer to the upper probability limit and determining the enhanced neural classification probability.

[0057] Understandably, prior constraints are used to apply anatomical prior guidance to the deep learning visual segmentation process by utilizing neural expected path information and first variant risk region information obtained from preoperative image analysis, limiting the neural search space to the expected anatomical channel, thereby reducing the segmentation false alarm rate caused by image noise, hemorrhage or tissue deformation.

[0058] In this embodiment, the intraoperative neurosurgical identification path sequence is determined by extracting the centerline point sequence of the nerve contour in the main recurrent laryngeal nerve region based on the probability map constructed after nerve classification probability enhancement. This sequence is used to characterize the spatial trajectory of the recurrent laryngeal nerve identified in real time during surgery. Any method in the prior art can be used to extract the nerve contour and centerline point sequence from the enhanced probability map to determine the intraoperative neurosurgical identification path sequence. It is understood that the centerline of the nerve contour represents the trajectory of the recurrent laryngeal nerve in space, i.e., a continuous curve along the nerve extension direction, composed of points located on the geometric central axis among all points within the contour, used to quantitatively describe the nerve's position, direction, and spatial morphology. As a preferred implementation, for the segmented binary mask region of the recurrent laryngeal nerve, a set of connected points with a single pixel width is extracted using a skeletonization algorithm in the prior art. These points are then sorted according to the natural extension direction of the nerve and sampled at equal intervals to form an ordered point sequence, which is the nerve centerline. Each point on this centerline represents the geometric center position of the nerve within that section; connecting all points sequentially presents the complete three-dimensional path of the nerve.

[0059] Specifically, the neural contour of the main recurrent laryngeal nerve region is determined by connected component analysis based on a binary neural mask, used to eliminate noise areas and retain the effective neural region. As a preferred implementation, connected component analysis is performed on the binary neural mask to extract all connected regions. The connected region with the largest area is retained as the main recurrent laryngeal nerve region, while noise areas with too small an area are eliminated. Morphological operations are then used to optimize the contour of the retained main neural region to obtain a smooth and complete neural contour. Morphological closing operations (using circular structural elements with a radius of three pixels to fill small holes and gaps within the contour) and morphological opening operations (also using circular structural elements with a radius of three pixels to eliminate burrs and isolated protrusions) are performed sequentially on the retained main neural region to obtain a smooth and complete neural contour.

[0060] Specifically, the extraction and sampling of the main recurrent laryngeal nerve skeleton includes extracting a single-pixel-width neural skeleton based on the smoothed neural contour and performing equidistant sampling to generate an ordered two-dimensional centerline point sequence. As a preferred embodiment, based on the smoothed neural contour, a Zhang-Suen parallel thinning algorithm is used to extract a single-pixel-width neural skeleton. The skeleton points are arranged in a connected order to form a one-dimensional point sequence, and then this point sequence is sampled equidistantly. The sampling interval is set to the image pixel distance corresponding to one millimeter, resulting in an ordered two-dimensional centerline point sequence to determine the intraoperative neural recognition path sequence.

[0061] In this embodiment, the visual recognition confidence score is determined based on the arithmetic mean of the neural probability values ​​of all pixels within the corrected probability map contour region. It reflects the degree of certainty the deep learning model has regarding the current recognition result, and its value ranges from 0 to 1. The closer the confidence score is to 1, the clearer the neural contour and the more certain the model's recognition. The neural probability values ​​are applied using the neural classification probability values ​​of each pixel, corrected based on constraint coefficients.

[0062] Please see Figure 4 As shown, it is a logic decision diagram for determining the dynamic adjustment of constraint coefficients in an embodiment of the present invention.

[0063] Specifically, the path deviation correction module calculates the real-time path deviation based on the intraoperative neurosurgical path sequence and the expected neurosurgical path sequence, and determines to dynamically adjust the constraint coefficient based on the comparison result that the real-time path deviation is greater than a preset path deviation threshold.

[0064] In this embodiment, the path deviation correction module compares the current real-time path deviation with a path deviation threshold to determine whether dynamic adjustment of the constraint coefficients is necessary. When the real-time path deviation is greater than the path deviation threshold, it is determined that there is a significant difference between the actual intraoperative anatomical structure and the conventional anatomical assumptions based on the preoperative images, and it is determined that dynamic adjustment of the first and second constraint coefficients is necessary. When the real-time path deviation is less than or equal to the path deviation threshold, no adjustment of the constraint coefficients is required, i.e., no correction is needed for the intraoperative neurosurgical identification path sequence and the corresponding visual recognition confidence. The intraoperative neurosurgical identification path sequence determined in the prior constraint identification module is transmitted to the distance state analysis module, and the visual recognition confidence is transmitted to the warning monitoring module to execute the subsequent workflow of the system in this embodiment of the invention.

[0065] In this embodiment, the real-time path deviation is determined by acquiring the intraoperative neural path identification sequence and the preoperative expected neural path sequence at a preset frequency, and is used to quantify the spatial deviation between the actual intraoperative neural course and the preoperative expected path. As a preferred implementation, the path deviation correction module acquires the intraoperative neural path identification sequence and the preoperative expected neural path sequence from the intraoperative acquisition module at a preset frequency, such as once per second. The Euclidean distance between each coordinate point in the intraoperative neural path identification sequence and the nearest coordinate point in the preoperative expected neural path sequence is calculated, and the arithmetic mean of these nearest distances for all coordinate points is taken to obtain the real-time path deviation.

[0066] In this embodiment, a path deviation threshold is used to distinguish between normal intraoperative tissue displacement and significant anatomical deviation. As a preferred implementation, historical data with no abnormalities during or after the operation are acquired. The path deviation threshold is determined based on the 95th percentile of the real-time path deviation in the historical data. This threshold can distinguish between normal intraoperative tissue displacement and significant anatomical deviation. When the real-time path deviation exceeds this threshold, it indicates that the current deviation level exceeds the deviation level of 95% of the abnormal data in the historical data.

[0067] Specifically, the path deviation correction module performs real-time correction on the neurosurgical identification path sequence based on the first constraint correction coefficient and the second constraint correction coefficient to determine the neurosurgical identification path correction sequence and the corresponding visual recognition correction confidence level. Wherein, the first constraint correction coefficient is determined after reducing and adjusting the first constraint coefficient according to the first preset amplitude; The second constraint correction coefficient is determined by amplifying and adjusting the second constraint coefficient according to the second preset amplitude.

[0068] In this embodiment, the path deviation correction module initiates a real-time correction process for the constraint coefficient after determining that dynamic adjustment of the constraint coefficient is required, in order to adapt to the significant differences between the actual anatomical structure during surgery and the preoperative expectations.

[0069] Specifically, the path deviation correction module applies the adjusted constraint correction coefficients to the prior constraint recognition process of the current and subsequent video frames to recalculate the corrected neural recognition results. Specifically, the adjusted first and second constraint correction coefficients are applied to the prior constraint recognition process of the current and subsequent video frames, and the probability map, the corrected neural recognition path sequence, and the corrected visual recognition confidence are recalculated using the same determination process as described above for the intraoperative neural recognition path sequence and visual recognition confidence.

[0070] In this embodiment, the constraint correction coefficient is determined based on the product of the original constraint coefficient and the corresponding preset amplitude, and is used to generate the adjusted prior constraint strength parameter. The first constraint correction coefficient is determined based on the product of the first constraint coefficient and the first preset amplitude, and the second constraint correction coefficient is determined based on the product of the second constraint coefficient and the second preset amplitude.

[0071] In this embodiment, the first preset amplitude, which is determined based on the reciprocal of the deviation ratio, is used to control the degree of reduction in dependence on unreliable expected path regions. As the deviation ratio increases, the first preset amplitude decreases in a reciprocal relationship. The more the real-time path deviation exceeds the threshold, the greater the reduction in the first constraint coefficient by the system, thereby reducing dependence on unreliable expected path regions.

[0072] In this embodiment, the second preset amplitude, which is determined based on the deviation ratio, is used to control the degree of enhancement of the visual detection sensitivity of the first variation risk area. As the deviation ratio increases, the second preset amplitude increases proportionally. The more the real-time path deviation exceeds the threshold, the greater the amplification of the second constraint coefficient by the system, thereby enhancing the visual detection sensitivity of the first variation risk area.

[0073] In this embodiment, the deviation ratio is determined based on the ratio of the current real-time path deviation to the path deviation threshold. It is used to quantify the degree to which the deviation exceeds the threshold and to provide a basis for subsequent preset amplitude calculation.

[0074] Specifically, the distance state analysis module determines that the real-time distance is in a distance boundary state based on the comparison result of the real-time distance between the neurosurgical identification path correction sequence and the intraoperative instrument position sequence being less than a preset distance threshold.

[0075] In this embodiment, the distance state analysis module is used to obtain the current intraoperative neurosurgical identification path correction sequence from the path deviation correction module or, in the absence of correction, the current intraoperative neurosurgical identification path sequence from the prior constraint identification module, and to obtain the intraoperative instrument position sequence from the intraoperative acquisition module.

[0076] In this embodiment, the minimum real-time distance is compared with a preset distance threshold to determine whether the surgical instrument has entered the nerve proximity area. When the minimum real-time distance is less than the distance threshold, it is determined that the surgical instrument has entered the nerve proximity area, and the real-time distance is determined to be at the distance boundary.

[0077] In this embodiment, the minimum real-time distance is determined based on the minimum Euclidean distance between the instrument end coordinates and each nerve path point in the intraoperative nerve identification path sequence, and is used to quantify the closest spatial distance between the surgical instrument and the nerve.

[0078] In this embodiment, a distance threshold is used to distinguish between the typical distance range between the instrument and nerve during routine operations and the less frequent, closer distance range. As a preferred implementation, the process of determining the distance threshold includes: acquiring historical data showing no abnormalities during and after the operation; determining the distance threshold based on the 10th percentile of the real-time distance in the historical data; this threshold can distinguish between the typical distance range between the instrument and nerve during routine operations and the less frequent, closer distance range; when the real-time distance is less than this threshold, it indicates that the current distance between the instrument and nerve has reached a level that is only achieved in a few cases in historical situations where there are no abnormalities during and after the operation, requiring increased monitoring sensitivity and further monitoring of electromyographic activity.

[0079] Specifically, the warning monitoring module determines spontaneous electromyographic activity indicators based on the ratio of real-time electromyographic activity intensity to a preset real-time electromyographic activity benchmark intensity. The real-time electromyographic activity intensity is determined based on the root mean square amplitude of the electromyographic signal within a first preset time window calculated from the intraoperative electromyographic signal sequence.

[0080] In this embodiment, the warning monitoring module responds to the real-time distance being at the distance boundary state and executes the analysis process of spontaneous electromyographic activity indicators to initiate electrophysiological monitoring when the device is in a high-risk state of approaching the nerve.

[0081] In this embodiment, the spontaneous electromyographic activity index, which is determined based on the ratio of real-time electromyographic activity intensity to real-time electromyographic activity baseline intensity, is used to quantify the degree of deviation of the spontaneous contraction level of the laryngeal muscles from the normal state. A spontaneous electromyographic activity index greater than 1 indicates that the current spontaneous contraction level of the laryngeal muscles is higher than the expected normal level after surgery. The larger the value, the higher the intensity of electrical activity of the laryngeal muscles innervated by the nerve, and the greater the possibility that the nerve is subjected to traction, compression or thermal stimulation.

[0082] In this embodiment, the real-time electromyographic activity intensity is characterized by root mean square (RMS) calculation based on a time window of the intraoperative EMG signal sequence. As a preferred implementation, the signal segment of the first 100 milliseconds of the intraoperative EMG signal sequence at the current moment is acquired, and the signal segments corresponding to the subsequent 100 milliseconds of the intraoperative EMG signal sequence are continuously monitored. Within the aforementioned 200-millisecond time window, the EMG signal is squared, the arithmetic mean of all squared values ​​is calculated, and the square root of this mean is taken to obtain the RMS amplitude within the window. This value is then determined as the real-time EMG activity intensity. It is understood that the RMS amplitude can effectively characterize the real-time EMG activity intensity, and the RMS calculation preserves the overall energy of the signal, avoiding the influence of the cancellation of positive and negative potentials.

[0083] In this embodiment, the real-time electromyographic activity baseline intensity is determined based on the arithmetic mean of all real-time electromyographic activity intensities calculated within a preset time period before the start of surgery, and is used to establish an individualized background level of electromyographic activity. As a preferred implementation, the arithmetic mean of all real-time electromyographic activity intensities calculated within 5 minutes before the start of surgery is used as the real-time electromyographic activity baseline intensity.

[0084] Specifically, the warning monitoring module determines the existence of a hidden warning area based on the fact that both the visual recognition correction confidence level and the spontaneous electromyography activity index meet the first preset condition. The first preset condition is that the visual recognition confidence level is less than a preset visual recognition correction confidence level threshold and the spontaneous electromyographic activity index is greater than a preset spontaneous electromyographic activity index threshold.

[0085] In this embodiment, the warning monitoring module obtains the visual recognition correction confidence level from the path deviation correction module or, in the absence of correction, the visual recognition confidence level from the prior constraint recognition module, and obtains spontaneous electromyographic activity indicators.

[0086] In this embodiment, a hidden warning region is defined as follows: under the condition that a first preset condition is met, a spherical spatial coordinate set is generated centered on the projection of the electrode side with the strongest intraoperative electromyographic signal. This set is used to identify areas where nerves are not detected by visual segmentation but where spontaneous electrical activity of the laryngeal muscles is significantly enhanced. As a preferred embodiment, a spherical spatial coordinate set with a radius of five millimeters, centered on the projection of the electrode side with the strongest intraoperative electromyographic signal, is defined as the hidden warning region. It is understood that the significance of the hidden warning region lies in the fact that, when no nerves are detected by visual segmentation but spontaneous electrical activity of the laryngeal muscles is significantly enhanced, the system determines that there may be nerves covered by tissue and invisible within this region, and defines this region's spatial coordinate set as the hidden warning region.

[0087] In this embodiment, the first preset condition is determined by jointly judging the results of comparing the visual recognition confidence level with the visual confidence level threshold and the spontaneous electromyography (EMG) index with the spontaneous EMG index threshold, and is used to trigger the recognition of the hidden warning area. In this embodiment, the analysis based on the first preset condition includes: determining whether the visual recognition correction confidence level is less than the visual confidence level threshold, and simultaneously determining whether the spontaneous EMG index is greater than the spontaneous EMG index threshold. If and only if both of the above conditions are met, it is determined that the first preset condition is satisfied at the current time.

[0088] In this embodiment, the visual confidence threshold is used to distinguish between the confidence level when the nerve is clearly visible during routine surgery and the lower confidence level when the nerve is obscured or at the edge of the visual field. As a preferred implementation, the visual confidence threshold is determined by acquiring historical data showing no abnormalities during or after surgery, statistically analyzing the distribution of visual recognition confidence with or without correction in the historical data, and taking the 10th percentile as the visual confidence threshold.

[0089] Understandably, the visual confidence threshold can distinguish between the confidence level when nerves are clearly visible during routine surgery and the lower confidence level when nerves are occluded or located at the edge of the visual field. When the real-time visual recognition confidence is less than this threshold, it indicates that no clear neural visual features could be extracted from the current frame, possibly due to tissue occlusion or the nerve being located in a blind spot. By using the visual confidence threshold, the system can objectively determine the reliability of the visual segmentation results, providing a quantitative basis for subsequent consistency verification with electrophysiological information, and avoiding the mistaken identification of low-confidence areas as having no nerves when visual recognition is inherently unreliable.

[0090] In this embodiment, the spontaneous electromyography (EMG) activity threshold is used to distinguish between the normal fluctuation range of EMG signals during routine surgical procedures and the range of significantly increased signals caused by the proximity of instruments to nerves. As a preferred implementation, the process for determining the spontaneous EMG activity threshold involves acquiring historical data showing no abnormalities during or after surgery, statistically analyzing the distribution of spontaneous EMG activity in the historical data, and determining the 90th percentile as the spontaneous EMG activity threshold.

[0091] Understandably, the spontaneous electromyography (EMG) activity threshold can distinguish between the normal fluctuation range of EMG signals during routine surgical procedures and the range of significantly elevated signals caused by instrument proximity, traction, or nerve compression. When the real-time spontaneous EMG activity index exceeds the threshold, it indicates that the current electrical activity intensity of the laryngeal muscles has exceeded the level of most times in historically normal cases, potentially indicating mechanical stimulation of the nerve. By using the spontaneous EMG activity threshold, the system can identify events of significant increases in EMG signals, triggering visual-electrophysiological consistency checks, thereby discovering potentially obscured nerves or early nerve stress responses.

[0092] Specifically, the warning monitoring module is used to calculate the rate of change of spontaneous electromyographic activity index within a second preset time window, and to compare the rate of change of neural activity with preset first and second thresholds for the rate of change of neural activity, so as to indicate the state of neural stress response.

[0093] In this embodiment, after the warning monitoring module responds to the existence of a hidden warning area, it executes the calculation of the rate of change of neural electrical activity and the stress response status prompt process to dynamically monitor and grade the potential neural stimulation.

[0094] In this embodiment, the rate of change of neural electrical activity is determined based on the relative change between the current average value and the previous average value within the window, and is used to quantify the trend of change in electromyographic activity between adjacent time windows. In this embodiment, the difference between the current average value and the previous average value is calculated, and the ratio of the difference to the previous average value is determined as the rate of change of neural electrical activity.

[0095] In this embodiment, the current average value of the window is calculated by taking the arithmetic mean of the spontaneous electromyographic activity index sequence within the length of the preceding second preset time window, with the current time as the window endpoint. This average value represents the average level of electromyographic activity within the current time window. The preceding average value of the window is calculated by taking the arithmetic mean of the spontaneous electromyographic activity index sequence within the length of the preceding second preset time window, with the preceding window endpoint as the window endpoint. This average value represents the average level of electromyographic activity within the preceding time window. As a preferred implementation, the second preset time window is set to 10 seconds to define the time range for calculating the rate of change of neural electrical activity.

[0096] In this embodiment, the threshold for the rate of change of neural electrical activity includes a first threshold for the rate of change of neural electrical activity and a second threshold for the rate of change of neural electrical activity, which are used to distinguish different stages of change in electromyographic signals. As a preferred implementation, the process of determining the threshold for the rate of change of neural electrical activity includes: acquiring historical data showing no abnormalities during and after surgery; calculating the relative rate of change of spontaneous electromyographic activity indicators between consecutive windows using a ten-second time window for each data point; statistically analyzing the distribution of the rate of change across all windows in all data points; and determining the 33rd and 67th percentiles of this distribution as the first and second thresholds for the rate of change of neural electrical activity, respectively.

[0097] In this embodiment, the neural stress response state is a category of neural stress level distinguished by the rate of change of laryngeal muscle electrical activity intensity, divided into first-level stress response state and second-level stress response state. This state is only used as auxiliary monitoring information for dynamic changes in electrophysiological amplitude and does not constitute a medical diagnosis conclusion.

[0098] In this embodiment, the Level 1 stress response state is determined based on the comparison between the rate of change of neural electrical activity and the first and second thresholds of neural electrical activity change. This determines whether to output a Level 1 stress response state, indicating that the nerve may be receiving continuous but not yet intense stimulation. It is understood that the first threshold of neural electrical activity change is used to distinguish between the stable fluctuation phase and the slight upward trend phase of the electromyographic signal. When the real-time rate of change is greater than or equal to the first threshold and less than the second threshold, it indicates that the electromyographic signal shows a slight upward trend, suggesting that the nerve may be receiving continuous but not yet intense stimulation. The system outputs a Level 1 stress response state, suggesting that the operator slow down the operation and pay attention to the instrument position.

[0099] In this embodiment, the secondary stress response state is determined based on the comparison between the rate of change of neural electrical activity and a second threshold for the rate of change of neural electrical activity. This determines whether to output a secondary stress response state to indicate that the nerve is experiencing clear, continuous mechanical compression or thermal stimulation. It is understood that the second threshold for the rate of change of neural electrical activity is used to distinguish between a slight increase and a significantly rapid increase. When the real-time rate of change is greater than or equal to the second threshold, it indicates that the electromyographic signal has increased significantly and rapidly within ten seconds, suggesting that the nerve is experiencing clear, continuous mechanical compression or thermal stimulation. The system then outputs a secondary stress response state to alert attention to the hidden warning area.

[0100] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery, characterized in that, include: The preoperative analysis module obtains the expected neural pathway sequence based on the preoperative neck images of the target user, analyzes the path uncertainty index based on the path points corresponding to the expected neural pathway sequence, and compares the path uncertainty index with a preset path uncertainty index threshold to determine the first variation risk area. The intraoperative acquisition module is used to acquire in real time the intraoperative continuous video frame sequence of the target user's neck, the intraoperative electromyographic signal sequence, and the intraoperative instrument position sequence. The prior constraint recognition module performs prior constraints on the intraoperative neck continuous video frame sequence based on the neural expected path sequence and the first variation risk region combined with preset constraint coefficients, so as to determine the intraoperative neural recognition path sequence and the corresponding visual recognition confidence. The path deviation correction module performs a real-time deviation comparison between the intraoperative neurological identification path sequence and the expected neurological path sequence to determine whether to dynamically adjust the constraint coefficient, and performs real-time correction on the intraoperative neurological identification path sequence based on the adjusted constraint coefficient to determine the intraoperative neurological identification path correction sequence and the corresponding visual recognition correction confidence level. The distance state analysis module compares the real-time distance between the neurosurgical identification path correction sequence and the intraoperative instrument position sequence with a preset distance threshold to determine whether the real-time distance is in a distance boundary state. The warning monitoring module is used to respond to the real-time distance being at the distance boundary state, analyze spontaneous electromyographic activity indicators based on the intraoperative electromyographic signal sequence, and perform consistency verification based on the visual recognition correction confidence and the spontaneous electromyographic activity indicators to determine whether there is a hidden warning area, and in response to the existence of a hidden warning area, monitor the rate of change of the spontaneous electromyographic activity indicators to indicate the state of neural stress response.

2. The deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery according to claim 1, characterized in that, The preoperative analysis module calculates local curvature based on adjacent path points in the expected neural path sequence to determine the path uncertainty index.

3. The deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery according to claim 2, characterized in that, The preoperative analysis module determines the first variation risk region based on the comparison results of the path uncertainty index being greater than the preset path uncertainty index threshold.

4. The deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery according to claim 3, characterized in that, The prior constraint recognition module determines the intraoperative neurosurgical recognition path sequence based on the prior projection region sequence and the enhanced region sequence combined with the corresponding constraint coefficients, and determines the visual recognition confidence based on the probability value corresponding to the intraoperative neurosurgical recognition path sequence. Specifically, the prior projection region sequence corresponding to the intraoperative continuous video frame sequence of the neck is determined based on the neural expected path sequence, and the enhancement region sequence of the prior projection region sequence is determined based on the first variation risk region.

5. The deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery according to claim 4, characterized in that, The path deviation correction module calculates the real-time path deviation based on the intraoperative neurosurgical path sequence and the expected neurosurgical path sequence. Based on the comparison result that the real-time path deviation is greater than a preset path deviation threshold, it determines to dynamically adjust the constraint coefficient.

6. The deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery according to claim 5, characterized in that, The path deviation correction module performs real-time correction on the neurosurgical recognition path sequence based on the first constraint correction coefficient and the second constraint correction coefficient, so as to determine the neurosurgical recognition path correction sequence and the corresponding visual recognition correction confidence. Wherein, the first constraint correction coefficient is determined after reducing and adjusting the first constraint coefficient according to the first preset amplitude; The second constraint correction coefficient is determined by amplifying and adjusting the second constraint coefficient according to the second preset amplitude.

7. The deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery according to claim 6, characterized in that, The distance state analysis module determines that the real-time distance is in a distance boundary state based on the comparison result of the real-time distance between the neurosurgical identification path correction sequence and the intraoperative instrument position sequence being less than a preset distance threshold.

8. The deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery according to claim 7, characterized in that, The warning monitoring module determines spontaneous electromyographic activity indicators based on the ratio of real-time electromyographic activity intensity to a preset real-time electromyographic activity benchmark intensity. The real-time electromyographic activity intensity is determined based on the root mean square amplitude of the electromyographic signal within a first preset time window calculated from the intraoperative electromyographic signal sequence.

9. The deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery according to claim 8, characterized in that, The warning monitoring module determines the existence of a hidden warning area based on the fact that both the visual recognition correction confidence level and the spontaneous electromyography activity index meet the first preset condition. The first preset condition is that the visual recognition correction confidence level is less than a preset visual recognition confidence level threshold and the spontaneous electromyographic activity index is greater than a preset spontaneous electromyographic activity index threshold.

10. The deep learning-based real-time intraoperative localization and monitoring system for recurrent laryngeal nerve surgery according to claim 9, characterized in that, The warning monitoring module is used to calculate the rate of change of spontaneous electromyographic activity index within a second preset time window, and to compare the rate of change of neural activity with preset first and second thresholds for the rate of change of neural activity, so as to indicate the state of neural stress response.

Images