A real-time monitoring system for key tissues in spinal surgery based on endoscopic video stream

Abstract

The application discloses a kind of based on endoscope video stream's spinal column operation key tissue real-time monitoring method and system, it is related to computer vision target detection field.The method constructs standardization spinal endoscope dataset with anatomical structure annotation;Enhance the dataset using anatomical semantic constraint image enhancement strategy, obtain enhanced training dataset;Anatomical semantic constraint image enhancement strategy includes disabling the geometric transformation of destroying anatomical topology and limited HSV color space transformation;YOLOv11-Seg segmentation network and weighted loss function are constructed;Weighted loss function strengthens the loss weight of bounding box and classification;According to enhanced training dataset, train network, and apply trained segmentation network to identify key tissue of spinal column operation.The application can improve the recognition accuracy of anatomical structure, realize high-precision real-time operation navigation.

Description

Technical Field

[0001] This application relates to the field of computer vision target detection, and in particular to a method and system for real-time monitoring of key tissues in spinal surgery based on endoscopic video streams. Background Technology

[0002] In minimally invasive spinal surgeries (such as percutaneous endoscopic discectomy), surgeons observe the surgical area through an endoscopic screen. Currently, the identification of key anatomical structures during surgery (such as nerve roots, dural sac, and ligamentum flavum) mainly relies on the surgeon's personal experience and anatomical knowledge.

[0003] While general-purpose computer vision object detection models (such as the standard YOLO series) perform excellently on natural images, they have significant technical limitations when directly applied to spinal endoscopic surgery scenarios: General data augmentation disrupts anatomical semantics: General algorithms often use random rotations and flips (Flip / Rotate) to augment data. However, the anatomical structures of the human spine (such as the upper and lower lamina, and the left and right nerve roots) have strict spatial orientation. General geometric augmentation can disrupt this anatomical positional relationship, causing the model to learn incorrect semantics (e.g., misclassifying a flipped image as a contralateral structure), thus posing a serious risk of medical misdiagnosis.

[0004] Enhanced illumination leads to feature distortion: In endoscopic environments, the light source is singular and highly reflective. Aggressive color enhancement in general algorithms alters tissue color characteristics, making it difficult to distinguish between tissues with similar appearances (such as the ligamentum flavum and the dural sac).

[0005] Low accuracy in identifying fine tissues: Key neural tissue targets such as nerve roots are small and have blurred edges. The loss function weights of the standard model are not optimized for this, which can easily lead to missed detections or inaccurate segmentation edges. Summary of the Invention

[0006] The purpose of this application is to provide a method and system for real-time monitoring of key tissues in spinal surgery based on endoscopic video streams, which can improve the accuracy of anatomical structure identification and achieve high-precision real-time surgical navigation.

[0007] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method and system for real-time monitoring of key tissues in spinal surgery based on endoscopic video streams, including: Obtain historical video streams of spinal endoscopic surgeries; The anatomical structure categories in the historical video stream are labeled to obtain a standardized spinal endoscopy dataset with anatomical structure annotations; An image enhancement strategy with anatomical semantic constraints is used to enhance the standardized spinal endoscopy dataset to obtain an enhanced training dataset. The anatomical semantic constraint image enhancement strategy includes disabling geometric transformations that destroy anatomical topology and restricting HSV color space transformations. A YOLOv11-Seg segmentation network and a weighted loss function are constructed; the weighted loss function strengthens the loss weights for bounding boxes and classification. Based on the enhanced training dataset, train the YOLOv11-Seg segmentation network to obtain the trained segmentation network. The real-time video stream during spinal endoscopic surgery is input into a trained segmentation network to generate anatomical structure categories and binary masks of anatomical structures. Based on the anatomical structure category, the binary mask of the anatomical structure is mapped to different semi-transparent colors and superimposed on the original endoscopic image to obtain a real-time surgical navigation image with anatomical structure staining marks.

[0008] Secondly, this application provides a method and system for real-time monitoring of key tissues in spinal surgery based on endoscopic video streams, including: The historical data acquisition module is used to acquire historical video streams of spinal endoscopic surgeries; The annotation module is used to annotate the anatomical structure categories in the historical video stream to obtain a standardized spinal endoscopy dataset with anatomical structure annotations; An enhancement module is used to enhance the standardized spinal endoscopy dataset using an anatomically semantically constrained image enhancement strategy to obtain an enhanced training dataset; the anatomically semantically constrained image enhancement strategy includes disabling geometric transformations that disrupt anatomical topology and restricting HSV color space transformations; A network construction module is used to construct a YOLOv11-Seg segmentation network; a C2PSA spatial attention module is introduced into the neck network of the YOLOv11-Seg segmentation network. The training module is used to train the YOLOv11-Seg segmentation network based on the augmented training dataset to obtain the trained segmentation network. The application module is used to input the real-time video stream from spinal endoscopic surgery into the trained segmentation network to generate anatomical structure categories and binary masks of anatomical structures. The display module is used to map the binary mask of the anatomical structure into different semi-transparent colors according to the anatomical structure category, and overlay it on the original endoscopic image to obtain a real-time surgical navigation image with anatomical structure staining marks.

[0009] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method and system for real-time monitoring of key tissues in spinal surgery based on endoscopic video streams. It explicitly prohibits geometric transformations that disrupt anatomical topology, forcing the model to learn correct spatial topological relationships. Addressing the single cold light source characteristic of endoscopes, it employs restricted HSV color space transformation to prevent distortion of pathological features. A weighted loss function strengthens the loss weights for bounding boxes and classification, solving the segmentation problem of small and blurred tissues, thereby improving the accuracy of anatomical structure recognition and achieving high-precision real-time surgical navigation. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart illustrating a method for real-time monitoring of key tissues in spinal surgery based on endoscopic video stream, provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of the YOLOv11-Seg segmentation network provided in an embodiment of this application; Figure 3 This is a schematic diagram of the functional modules of a real-time monitoring system for key tissues in spinal surgery based on endoscopic video stream, provided in an embodiment of this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] In one exemplary embodiment, such as Figure 1 As shown, a method for real-time monitoring of key tissues in spinal surgery based on endoscopic video stream is provided. This method is executed by a computer device, specifically by a computer device such as a terminal or server alone, or by a terminal and server together. In the embodiments of this application, the method includes the following steps 101 to 107.

[0015] Step 101: Obtain historical video streams of spinal endoscopic surgeries.

[0016] Step 102: Label the anatomical structure categories in the historical video stream to obtain a standardized spinal endoscopy dataset with anatomical structure annotations.

[0017] Step 103: Employ an anatomically semantically constrained image enhancement strategy to enhance the standardized spinal endoscopy dataset and obtain an enhanced training dataset; the anatomically semantically constrained image enhancement strategy includes disabling geometric transformations that disrupt anatomical topology and restricting HSV color space transformations.

[0018] Step 104: Construct the YOLOv11-Seg segmentation network and a weighted loss function; the weighted loss function strengthens the loss weights for bounding boxes and classification.

[0019] Step 105: Train the YOLOv11-Seg segmentation network based on the augmented training dataset to obtain the trained segmentation network.

[0020] Step 106: Input the real-time video stream from the spinal endoscopic surgery into the trained segmentation network to generate anatomical structure categories and binary masks of the anatomical structures.

[0021] Step 107: Based on the anatomical structure category, the binary mask of the anatomical structure is mapped to different semi-transparent colors and superimposed on the original endoscopic image to obtain a real-time surgical navigation image with anatomical structure staining marks.

[0022] This application employs a deep learning method to process the endoscopic video stream in real time by implementing steps 101 to 107 as described above.

[0023] In another exemplary embodiment of this application, surgical video stream data acquisition includes: real-time video stream (RTSP / HDMI acquisition) or historical video files of spinal endoscopy.

[0024] In another exemplary embodiment of this application, step 102 described above may be replaced by steps 201 to 204.

[0025] Step 201: Extract RGB images from the historical video stream at a preset frequency, and remove the RGB images of invalid frames to obtain valid RGB images.

[0026] Frame extraction and cleaning: RGB images are extracted from the video stream at a frequency of 5-10 frames / second, and invalid frames that are blurry or have no surgical field of view are removed.

[0027] Step 202: Perform pixel-level polygon annotations on the anatomical structure categories in the valid RGB images to generate annotation data.

[0028] Seven key anatomical structures in the image (superior lamina, inferior lamina, facet joints, intervertebral discs, dural sacs, nerve roots, and ligamentum flavum) were annotated at the pixel level to generate a set of polygon coordinates. .

[0029] Step 203: Convert the labeled data into a YOLO format .txt tag file; the .txt tag file contains the category index, the normalized center point coordinates of each polygon, and the mask coefficients.

[0030] Step 204: Combine the valid RGB images and the .txt tag file to form a standardized spinal endoscopy dataset with anatomical structure annotations.

[0031] The standardized spinal endoscopy dataset can be divided into a training set, a validation set, and a test set.

[0032] In another exemplary embodiment of this application, an image enhancement strategy applying anatomical semantic constraints is described. ,in To enhance the parameter set, For the input image, Image enhancement strategies.

[0033] Because the spinal anatomy has a strict "up-down" and "left-right" spatial logic, geometric transformations that disrupt the anatomical topology are disabled, and the following geometric transformation parameters are forcibly set to 0: Rotation angle: ; Shear strength: ; Horizontal flip probability: ; Probability of vertical flip: .

[0034] In another exemplary embodiment of this application, endoscopic-specific lighting adaptation is implemented: a restricted HSV color space transformation is used to prevent distortion of pathological features, taking into account the single cold light source characteristic of the endoscope. The formula used for the restricted HSV color space transformation is: ; In the formula, This is the pixel value vector of the image after illumination enhancement in the HSV color space. This is the pixel value vector after converting the original input image to the HSV color space. This is the hue perturbation coefficient. This is the saturation perturbation coefficient. This represents the brightness perturbation coefficient.

[0035] The scope of the restriction is .

[0036] In another exemplary embodiment of this application, the image enhancement strategy with anatomical semantic constraints further includes: Mosaic stitching enhancement, which is used to randomly scale and stitch a preset number of images in the standardized spinal endoscopy dataset.

[0037] For example, enabling Mosaic enhancement (probability=1.0, scaling factor=0.7) randomly scales and stitches four training images to improve the model's ability to detect small targets (such as nerve roots).

[0038] In another exemplary embodiment of this application, the network architecture used in this application is designed as follows: YOLOv11-Seg is selected as the backbone network. For example... Figure 2 As shown, the YOLOv11-Seg segmentation network consists of a backbone network, a neck network, and a detection head.

[0039] Backbone: The C2f module is replaced with the C3k2 module (an improved convolutional block based on the Cross Stage Partial network) to reduce computation while maintaining feature extraction capabilities; and the SPPF (Spatial Pyramid Pooling - Fast) module is connected at the end to enhance multi-scale feature fusion.

[0040] Neck Network: Utilizing the PANet architecture, a C2PSA (Cross Stage Partial with Spatial Attention) spatial attention module is introduced into the neck network. This C2PSA module is located in the deepest layer (P5) of the PANet structure. This deepest layer corresponds to the large receptive field in endoscopic images. Although nerve roots are small, their positional relationship (topological structure) relative to the lamina and dural sac requires global context for judgment. The C2PSA spatial attention module, through its attention mechanism, allows the model to "focus" on key pixel regions within low-resolution, high-dimensional features, preventing the loss of edge information of small targets during subsequent upsampling and enhancing the model's ability to focus on small anatomical structures (such as nerve root edges).

[0041] Head (Detection Head): Employs a decoupled head structure and includes two parallel segmentation branches: the Proto Branch (for prototyping) and the Mask Coefficients Branch (for mask coefficient prediction). Proto Branch: Generates... A full-image prototype mask. Mask Coefficients Branch: Predicts for each detection box. Each mask coefficient.

[0042] In another exemplary embodiment of this application, the model is trained iteratively: Input: Augmented dataset, network structure.

[0043] Processing procedure: Transfer learning initialization: Load weight files (.pt) pre-trained on large-scale medical image datasets (such as lung-v8) or COCO datasets to accelerate convergence.

[0044] Optimizer configuration: Use SGD or AdamW optimizer, with momentum set to 0.937 and weight decay set to 0.0005.

[0045] Learning rate scheduling: The learning rate is adjusted using a cosine annealing strategy. : ; In the formula, , These represent the maximum and minimum values ​​of the learning rate, respectively. This represents the number of training epochs that have been performed so far. It is the independent variable in the formula and determines the current position on the cosine curve. This represents the maximum number of iterations or a period span of cosine annealing, when achieve When the learning rate drops to its minimum value .

[0046] Iteration and early stopping: Training is performed for a maximum of 20,000 epochs. If the validation set mAP50-95 metric does not improve within 100 consecutive epochs, the early stopping mechanism is triggered.

[0047] Output: The best model weight file (best.pt) for training convergence.

[0048] In another exemplary embodiment of this application, loss function reconstruction (Loss Function Engineering): total loss Composed of four parts, this application incorporates specific weight adjustments for medical scenarios. The weighted loss function is: ; In the formula, For the total loss, For bounding box loss, For classifying losses, For the distribution focus loss, To divide the loss, Weights for the bounding box loss. For classification loss weights, For the distribution focus loss weights, To divide the loss weights; (Enhance positioning accuracy) (Strengthen category differentiation) In this application, the bounding box loss weight is significantly increased (the default value is usually 7.5, but the importance of localization is emphasized here) to ensure that the model can achieve extremely high spatial localization of small tissues such as nerve roots in the complex dynamic context of endoscopy. Increasing the classification loss weight is to address the risk of misdiagnosis caused by similar tissue morphologies under endoscopy (such as the ligamentum flavum and dural sac having similar colors), forcing the model to learn more discriminative category features. The distribution focus loss weight is used to assist bounding box regression, especially when the target edges are blurred (such as the nerve root edge being mixed with surrounding fat), optimizing the boundary probability distribution. The segmentation loss weight is responsible for pixel-level mask generation, working in conjunction with the above weights to achieve collaborative optimization of detection and segmentation.

[0049] The bounding box loss uses CIoU Loss, taking into account the overlapping area, center point distance, and aspect ratio. .

[0050] The classification loss uses binary cross-entropy loss (BCE Loss): .

[0051] The segmentation loss employs a combination of binary cross-entropy loss and Dice Loss to address the issue of blurred tissue edges. .

[0052] In another exemplary embodiment of this application, step 106 above relates to TensorRT quantization acceleration and real-time inference.

[0053] TensorRT Quantization Acceleration: The trained segmentation network is accelerated using TensorRT quantization to obtain an accelerated segmentation network. The detailed acceleration process is as follows: The PyTorch model is exported to ONNX format, and then converted to TensorRTEngine. During this process, FP16 (half-precision floating-point) quantization is used to significantly reduce memory usage and improve inference speed while maintaining accuracy.

[0054] Real-time inference: Based on the real-time video stream from spinal endoscopic surgery, an accelerated segmentation network is used to output anatomical structure categories and binary masks of the anatomical structures. The detailed process of real-time inference is as follows: The input video stream undergoes normalization preprocessing (frame extraction and cleaning). The images are then input into the backbone network for feature extraction. The extracted features are fused at multiple scales through the neck network. The fused features are then fed into the detection head: the detection branch predicts the category and bounding box, while the segmentation branch includes mask coefficients and proto. Real-time mask generation: For each frame in the video stream, the model outputs a prototype mask. (size ) and target coefficient (size Final instance segmentation mask Generated via linear combination and Sigmoid activation: In the formula, This represents the Sigmoid activation function. Its technical function is to map the continuous values ​​resulting from a linear combination to the [0,1] interval, thereby generating a binary feature map representing the probability that each pixel belongs to the target tissue. Subsequently, the mask is cropped using the predicted bounding boxes to remove out-of-box noise. Post-processing (NMS): Non-maximum suppression (NMS) is applied, with an IoU threshold of 0.7, to remove overlapping and redundant detection boxes.

[0055] Visualization rendering: The generated binary mask is mapped to semi-transparent colors (e.g., nerve roots are displayed in yellow, and the dural sac in blue), overlaid on the original endoscopic image, and presented in real time on the surgical monitor. Output: Real-time surgical navigation screen with anatomical structure staining markers (latency <30ms).

[0056] The following comparative experiments verify that this application achieves significant beneficial effects compared to existing technologies: 1. Extremely high clinical real-time performance (accelerated validation with TensorRT) Experiments demonstrate that on the NVIDIA RTX 5090 platform, the inference time of the preferred model of this invention (YOLOv11) accelerated by TensorRT is only 14.71ms, with a frame rate as high as 67.99 FPS, fully meeting the stringent requirements of surgical navigation for low latency (<30ms). Even on a conventional medical graphics card (RTX 3060), near real-time performance of 41.28ms can be achieved.

[0057] 2. Excellent overall model performance In the model selection comparison experiment, YOLOv11 achieved the best balance between speed and accuracy while maintaining a low computational cost (295.9 GFLOPs), which is better than YOLOv8 (too cumbersome) and YOLOv9 (slower).

[0058] 3. Accurate identification of anatomical structures (robustness verification) By employing a training strategy that disables flipping / rotation as an anatomical semantic constraint, the model exhibits extremely high stability during training (supporting convergence for up to 20,000 epochs). This demonstrates that the strategy effectively avoids anatomical orientation confusion caused by erroneous data augmentation in general algorithms, significantly improving the reliability of identifying high-risk tissues such as nerve roots and dural sacs.

[0059] Additional Information: Experimental Data: This experiment constructed a dedicated dataset for spinal endoscopic surgery, containing 6539 images for training, 817 images for validation, and 818 images for testing. (Note: Some images may contain multiple tissues.) Upper lamina: 1241 images | 1266 masks; Inferior lamina: 1043 images | 1099 masks; Articular processes: 3284 images | 3358 masks; Intervertebral disc: 2314 images | 2489 masks; Dural sac: 1141 images | 1180 masks; Nerve root: 2001 images | 2240 masks; Ligamentum flavum: 2188 images | 2385 masks.

[0060] The accuracy of training different models is compared in Table 1.

[0061] Table 1. Accuracy of Model Training The Dice coefficient measures the pixel-level fit between the instance segmentation mask generated by the model and the ground truth (GT) annotated by the human. A high Dice coefficient demonstrates the extremely high reliability of this application in handling the edges of tiny tissues in spinal surgery.

[0062] The core inventive point of this application is: (1) A data augmentation strategy with anatomical semantic constraints was constructed, which forced the model to learn the correct spatial topological relationships by explicitly disabling operations such as flipping and rotating that disrupt anatomical orientation.

[0063] (2) A weighted loss function for the endoscopic environment was designed, which significantly improved the loss weights of bounding boxes (Box) and classification (Cls) to solve the segmentation problem of small and blurry tissues.

[0064] (3) Based on the preferred YOLOv11 / v12 architecture combined with TensorRT acceleration, high-precision real-time surgical navigation was achieved.

[0065] Based on the same inventive concept, this application also provides a real-time monitoring system for critical tissues in spinal surgery based on endoscopic video streams for implementing the methods described above. The solution provided by this system is similar to the implementation schemes described in the above methods. Therefore, the specific limitations in one or more embodiments of the real-time monitoring system for critical tissues in spinal surgery based on endoscopic video streams provided below can be found in the limitations of the above methods described above, and will not be repeated here.

[0066] In one exemplary embodiment, such as Figure 3 As shown, a real-time monitoring system for key tissues in spinal surgery based on endoscopic video stream is provided, including: a historical data acquisition module, an annotation module, an enhancement module, a network construction module, a training module, an application module, and a screen display module.

[0067] The system includes a historical data acquisition module for acquiring historical video streams of spinal endoscopy surgeries. An annotation module for labeling the anatomical structure categories in the historical video streams, obtaining a standardized spinal endoscopy dataset with anatomical structure annotations. An enhancement module for enhancing the standardized spinal endoscopy dataset using an image enhancement strategy constrained by anatomical semantics, obtaining an enhanced training dataset; the anatomical semantics-constrained image enhancement strategy includes disabling geometric transformations that disrupt anatomical topology and restricting HSV color space transformations.

[0068] The network construction module is used to construct the YOLOv11-Seg segmentation network; a C2PSA spatial attention module is introduced into the neck network of the YOLOv11-Seg segmentation network. The training module is used to train the YOLOv11-Seg segmentation network based on the augmented training dataset to obtain the trained segmentation network.

[0069] The application module is used to input the real-time video stream from spinal endoscopic surgery into the trained segmentation network, generating anatomical structure categories and binary masks of the anatomical structures. The display module is used to map the binary masks of the anatomical structures to different semi-transparent colors according to the anatomical structure categories, and overlay them on the original endoscopic image to obtain a real-time surgical navigation image with anatomical structure staining marks.

[0070] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0071] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A real-time monitoring system for key tissues in spinal surgery based on endoscopic video stream, characterized in that, include: The historical data acquisition module is used to acquire historical video streams of spinal endoscopic surgeries; The annotation module is used to annotate the anatomical structure categories in the historical video stream to obtain a standardized spinal endoscopy dataset with anatomical structure annotations; An enhancement module is used to enhance the standardized spinal endoscopy dataset using an anatomically semantically constrained image enhancement strategy to obtain an enhanced training dataset; the anatomically semantically constrained image enhancement strategy includes disabling geometric transformations that disrupt anatomical topology and using restricted HSV color space transformations; A network construction module is used to construct a YOLOv11-Seg segmentation network and a weighted loss function; a C2PSA spatial attention module is introduced into the neck network of the YOLOv11-Seg segmentation network; the weighted loss function strengthens the loss weights for bounding boxes and classification; The training module is used to train the YOLOv11-Seg segmentation network based on the augmented training dataset to obtain the trained segmentation network. The application module is used to input the real-time video stream from spinal endoscopic surgery into the trained segmentation network to generate anatomical structure categories and binary masks of anatomical structures. The screen display module is used to map the binary mask of the anatomical structure into different semi-transparent colors according to the anatomical structure category, and superimpose it on the original endoscope screen to obtain a real-time surgical navigation screen with anatomical structure staining marks; The geometric transformation parameters set for disabling the geometric transformation that disrupts the anatomical topology are as follows: ； ； ； ； In the formula, For rotation angle, Shear strength, The probability of horizontal flipping. This represents the probability of vertical flipping. The formula used for the restricted HSV color space transformation is: ； In the formula, This is the pixel value vector of the image after illumination enhancement in the HSV color space. This is the pixel value vector after converting the original input image to the HSV color space. This is the hue perturbation coefficient. This is the saturation perturbation coefficient. The value is the brightness perturbation coefficient; the limited range is... ; The image enhancement strategy with anatomical semantic constraints also includes: Mosaic stitching enhancement; Mosaic stitching enhancement is used to randomly scale and stitch together a preset number of images in the standardized spinal endoscopy dataset; The YOLOv11-Seg segmentation network includes: a backbone network, a neck network, and a detection head; In the backbone network, the C3k2 module is used to replace the C2f module, and an SPPF module is connected at the end; The neck network adopts the PANet structure, and a C2PSA spatial attention module is introduced into the neck network. The C2PSA spatial attention module is set in the deepest layer of the PANet structure. The detection head adopts a decoupled head structure and contains two parallel segmentation branches; the two parallel segmentation branches include a Proto prototype generation branch and a Mask coefficient prediction branch.

2. The real-time monitoring system for key tissues in spinal surgery based on endoscopic video stream according to claim 1, characterized in that, The anatomical structure categories in the historical video stream are labeled to obtain a standardized spinal endoscopy dataset with anatomical structure annotations, specifically including: RGB images are extracted from the historical video stream at a preset frequency, and invalid RGB images are removed to obtain valid RGB images; Pixel-level polygon annotations are performed on the anatomical structure categories in valid RGB images to generate annotation data; The labeled data is converted into a YOLO format .txt tag file; the .txt tag file contains the category index, the normalized center point coordinates of each polygon, and the mask coefficients; The valid RGB images and the .txt tag file are combined to form a standardized spinal endoscopy dataset with anatomical structure annotations.

3. The real-time monitoring system for key tissues in spinal surgery based on endoscopic video stream according to claim 1, characterized in that, When training the YOLOv11-Seg segmentation network: Load the weight file pre-trained on the medical image dataset or COCO dataset for transfer learning; Optimize using the SGD optimizer or the AdamW optimizer; The learning rate is adjusted using a cosine annealing strategy. If the validation set mAP50-95 metric does not improve within 100 consecutive rounds, the early stop mechanism is triggered.

4. The real-time monitoring system for key tissues in spinal surgery based on endoscopic video stream according to claim 1, characterized in that, The weighted loss function is: ； In the formula, For the total loss, For bounding box loss, For classifying losses, For the distribution focus loss, To divide the loss, Weights for the bounding box loss. For classification loss weights, For the distribution focus loss weights, To divide the loss weights; , , ; The bounding box loss uses CIoU Loss, the classification loss uses binary cross-entropy loss, and the segmentation loss uses a combination of binary cross-entropy loss and Dice Loss.

5. The real-time monitoring system for key tissues in spinal surgery based on endoscopic video stream according to claim 1, characterized in that, The real-time video stream from spinal endoscopic surgery is input into a trained segmentation network to generate anatomical structure categories and binary masks of the anatomical structures, specifically including: The trained segmentation network is accelerated by TensorRT quantization to obtain the accelerated segmentation network; Based on the real-time video stream during spinal endoscopic surgery, an accelerated segmentation network is used to output the anatomical structure category and the binary mask of the anatomical structure.

Images