Intraoperative neurovascular bundle identification and injury early warning method and system based on artificial intelligence

Abstract

The invention provides an intraoperative neurovascular bundle identification and injury early warning method and system based on artificial intelligence. According to the method, real-time image data are acquired through a multi-modal imaging device, multi-scale blood vessel features are extracted by adopting a three-stage cascade adaptive path convolution (APC) module, the long-range topology perception capability is enhanced in combination with a local-global interaction module (LGIM), global features are integrated by utilizing a cross-scale deformable convolutional network (CSDCN), and dimensions are compressed through a multi-branch feature distillation module (MFDM). And the multi-task classifier synchronously outputs the category, position and injury risk assessment of the blood vessel bundle, and real-time multi-modal early warning is realized through dynamic confidence threshold and mahalanobis distance anomaly detection. According to the invention, the accuracy and real-time performance of intraoperative nerve and blood vessel protection are obviously improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of medical artificial intelligence, and particularly relates to an intraoperative neural vascular bundle intelligent identification and real-time injury early warning method and system. BACKGROUND

[0002] In minimally invasive surgery and neurosurgery, accurately identifying neural vascular bundles and avoiding intraoperative injury are the key to ensuring surgical safety. However, the existing technology still has the following significant problems in practical application:

[0003] Insufficient recognition accuracy. The recognition rate of traditional methods based on threshold segmentation or morphological processing (such as region growing method, level set algorithm) for low-contrast blood vessels is generally lower than 90%, especially in the neural and vascular intersection area, the misjudgment rate is as high as 15%-20%. For example, the equal error rate (EER) of DenseNet-161 and other classic networks on the HKPU database can only reach 2.35%, which cannot meet the clinical accuracy requirements. Real-time defects. The existing 3DCNN or Transformer-based solutions (such as ViT-Cap) have a high computational complexity, with a single-frame processing delay exceeding 100ms (GPU platform), while the intraoperative operation requires a delay to be controlled within 30ms. In addition, the large number of parameters (such as ResNet+SE model parameter quantity > 50M) makes it difficult to deploy to edge devices.

[0004] Poor morphological adaptability. The fixed receptive field of traditional convolution kernels cannot adapt to complex geometric structures such as blood vessel bifurcation and bending. Although dynamic curved convolution extracts features through pre-set dimension directions, it still has a blind area in capturing oblique blood vessels, with an EER as high as 1.10% on the SDUMLA database. Lack of injury early warning. Current systems rely on postoperative offline analysis and lack real-time risk assessment mechanisms during surgery. A few studies (such as Triplet-Classifier GAN) attempt to detect abnormalities through generative adversarial networks, but the false positive rate exceeds 25%, and they cannot provide multi-modal early warning feedback. Inefficient use of multi-modal data. The heterogeneity of optical, ultrasonic, and other multi-modal imaging data makes it difficult to fuse features, and existing methods (such as PCA dimensionality reduction fusion) lose more than 30% of effective blood vessel information, affecting subsequent recognition accuracy.

[0005] In view of the above problems, there is an urgent need for an intraoperative neural vascular bundle identification technology with high accuracy, low delay, self-adaptability to blood vessel morphological variations, and support for real-time injury early warning. SUMMARY

[0006] The present application proposes an intraoperative neural vascular bundle identification and injury early warning method and system based on artificial intelligence, which systematically solves the problems of low recognition accuracy, insufficient real-time performance, poor blood vessel morphological adaptability, lack of intraoperative injury early warning, and inefficient multi-modal data fusion.

[0007] In a first aspect, the embodiments of the present application provide an intraoperative neurovascular bundle recognition and injury warning method based on artificial intelligence, which comprises:

[0008] S1, acquiring real-time multi-modal image data containing a neurovascular bundle through an intraoperative imaging device and performing standardization preprocessing;

[0009] S2, processing the preprocessed image by using a three-level cascaded adaptive path convolution (APC) module, adaptively capturing the geometric structure features of the neurovascular bundle such as bifurcation and trend through deformable convolution kernels in horizontal and vertical directions, and outputting multi-scale local feature maps;

[0010] S3, inputting the features extracted by the APC into a local-global interaction module (LGIM), calculating spatial and channel attention through block calculation, fusing local features and global context information, and enhancing the perception ability of long-distance topological relationship of the neurovascular bundle;

[0011] S4, globally integrating the multi-scale features by using a cross-scale deformable convolution network module (CSDCN), and compressing the feature dimension by a multi-branch feature distillation module (MFDM) to retain key discriminative features;

[0012] S5, inputting the reduced features into a multi-task classifier to output the category and position information of the neurovascular bundle; at the same time, judging the potential injury risk based on feature anomaly detection or confidence threshold, and triggering visual or sound warning signals in real time;

[0013] S6, training the network by using a cross-entropy loss function and an Adam optimizer, designing the model with a parameter amount ≤1.5M to support real-time processing; and deploying the model through an edge computing device to realize low-latency interaction.

[0014] Optionally, in an implementation manner of the first aspect of the present application, the data acquisition and preprocessing step in S1 further comprises:

[0015] acquiring multi-modal image data of the neurovascular bundle through an intraoperative optical imaging, ultrasound or near-infrared imaging device, and performing spatial alignment by using a deep learning-based registration algorithm to enhance the contrast of blood vessels and nerve tissues;

[0016] automatically detecting the potential region of the neurovascular bundle based on a lightweight segmentation network, generating an initial ROI mask, and optimizing the ROI boundary in combination with morphological operations;

[0017] performing brightness normalization by using histogram matching, enhancing the microvessel structure in combination with local contrast limited adaptive histogram equalization, and uniformly scaling the image to a preset size and normalizing the pixel value to a specified interval;

[0018] Deploying a preprocessing pipeline on an edge computing device, using GPU acceleration to achieve low-latency single-frame processing, and managing image sequences through a ring buffer to ensure real-time processing continuity;

[0019] The ROI extraction network uses adaptive path convolution (APC) as the backbone to adapt to the tortuous morphology of neural vascular bundles.

[0020] Optionally, in an implementation form of the first aspect of the present application, the structure of the adaptive path convolution (APC) module in S2 includes a multi-branch convolution design module, a feature fusion module, and a dynamic offset mechanism, wherein

[0021] The multi-branch convolution design module includes:

[0022] Deformable convolution branch: using a 3x3 deformable convolution kernel, learning to adapt to the bifurcation structure of blood vessels through an offset;

[0023] Horizontal direction perception branch: using a 1x9 horizontal direction convolution kernel, through a dynamic offset α x ∈[-1, 0, 1] to adapt to the horizontal bifurcation and bending structure of neural vascular bundles;

[0024] Vertical direction convolution branch: using a 9x1 vertical direction convolution kernel, through a dynamic offset α y ∈[-1, 0, 1] to capture the vertical direction features of neural vascular bundles;

[0025] The feature fusion module is: concatenating the feature maps output by the three branches, reducing the dimension through 1x1 convolution, and then generating the final output features through group normalization and LeakyReLU activation function, to output feature maps of three scales of 64x64, 32x32 and 16x16;

[0026] The dynamic offset mechanism is: the center coordinates of the horizontal and vertical direction convolution kernels are dynamically adjusted according to the vascular structure, and the offset is controlled through learnable parameters, so that the convolution kernel path fits the geometric morphology of neural vascular bundles;

[0027] The dynamic offset is predicted in real time through a lightweight subnetwork to adapt to the individualized variations of intraoperative neural vascular bundles.

[0028] Optionally, in an implementation form of the first aspect of the present application, the processing process of the local-global interaction module (LGIM) in S3 further includes:

[0029] S3.1, local feature extraction: dividing the feature map output by the adaptive path convolution (APC) into non-overlapping NxN local windows, independently calculating the attention for each window, and extracting the query vector Q, the key vector K and the value vector V, and compressing the dimension to 1 / 4 of the original channel number;

[0030] S3.2, dual-path attention calculation:

[0031] Spatial attention path Attention space : calculate the correlation weight of the spatial position in the window, the formula is:

[0032]

[0033] Channel attention path Attention channel : calculate the correlation weight of the feature across the channel, the formula is:

[0034]

[0035] where S(·) is a sparse function based on the significance sampling of the blood vessel structure, d k represents the dimension of the key vector K;

[0036] S3.3, feature fusion and enhancement: the outputs of spatial attention and channel attention are spliced in the channel dimension; feature fusion and dimension recovery are performed through 1×1 convolution; a jump connection is added to retain the original feature information;

[0037] S3.4, cross-window feature fusion: local and global features are dynamically fused by using a gated linear unit, and the formula is:

[0038] Output=Conv 1×1 (σ(Conv 1×1 (F local ))⊙)F global ,

[0039] where F local , F global are local and global features, respectively, and is an element-wise multiplication, Conv 1×1 (x) represents a 1×1 convolution operation on the input feature map x;

[0040] S3.5, dynamic receptive field adjustment: adjust the attention receptive field through deformable convolution to adapt to blood vessel structures of different sizes.

[0041] Optionally, in an implementation form of the first aspect of the present application, the S4 cross-scale deformable convolution network processing step further comprises:

[0042] S4.1, multi-scale feature alignment: dynamically align the blood vessel structures of different resolution feature maps through multi-level deformable convolution kernels, and the offset is predicted by a lightweight subnetwork;

[0043] S4.2, pyramid fusion: 32x32, 16x16, 8x8 feature maps are respectively upsampled to a unified scale and added, and then fused through 3x3 convolution;

[0044] S4.3, channel-space collaborative attention: CBAM attention module is applied to the fused features to simultaneously optimize channel and spatial weights.

[0045] Optionally, in an implementation form of the first aspect of the present application, the feature dimension compression in S4 through the multi-branch feature distillation module MFDM retains key discriminative features, comprising:

[0046] Multi-branch is adopted for feature extraction;

[0047] The three branch output features are spliced along the channel dimension and reduced to the target channel number through 1x1 convolution;

[0048] Channel attention is adopted to dynamically adjust the contribution weight of each branch, and the formula is:

[0049] F out =Sigmoid(W SE (F concat ))⊙F concat ,

[0050] Wherein, F concat represents the channel splicing result of the multi-branch output feature, W SE represents the channel attention weight matrix, Sigmoid(·) represents the activation strength of each channel, and represents element-wise multiplication;

[0051] In the cascade structure, the high-level features are added to the low-level features through the jump connection to avoid information loss.

[0052] Optionally, in an implementation form of the first aspect of the present application, the classification and damage warning step in S5 further comprises:

[0053] A fully connected layer and a softmax activation function are adopted to output the class probability distribution of the neural vascular bundle;

[0054] A coordinate regression branch is accessed in parallel to output the bounding box position (x, y, w, h) of the vascular bundle through a sigmoid function;

[0055] Confidence monitoring: when the highest class probability of the classifier is lower than the dynamic threshold, it is determined as an uncertain recognition area;

[0056] Feature anomaly detection: calculate the Mahalanobis distance between the test sample feature and the feature distribution of the training set, and trigger the corresponding multi-modal warning system and real-time guarantee measures when the distance exceeds 3σ.

[0057] Optionally, in an implementation form of the first aspect of the present application, the model optimization and deployment step in S6 further comprises:

[0058] An FP16 / FP32 mixed precision training strategy is adopted to accelerate the calculation through Tensor Core; dynamic loss scaling is implemented to prevent gradient underflow;

[0059] In the training phase, an L1 regularization term is added to the loss function to realize weight sparsification; in the deployment phase, feature channels with a contribution of less than 5% are removed through channel pruning;

[0060] The model is converted into a TensorRT engine optimized inference path and deployed on a Jetson AGX Xavier edge device for parallel processing using an ARM CPU and a Volta GPU;

[0061] The inference delay is continuously monitored, and when the delay exceeds the limit, the system automatically switches to a simplified model;

[0062] Through continuous learning technology, the system dynamically fine-tunes using intraoperative data, and sets a verification mechanism to ensure that the accuracy fluctuation does not exceed a preset value.

[0063] Optionally, in an implementation form of the first aspect of the present application, the loss function is:

[0064] L reg =αL cls +βL reg +γL ano ,

[0065] wherein L cls represents the classification loss, L reg represents the regression loss, and L ano represents the anomaly detection loss, and α, β, γ are the corresponding balance coefficients.

[0066] In a second aspect, the embodiments of the present application provide an intraoperative neural vascular bundle recognition and damage early warning system based on artificial intelligence, which is applied to the intraoperative neural vascular bundle recognition and damage early warning method based on artificial intelligence as described in the first aspect. The system comprises:

[0067] A data acquisition and preprocessing module is configured to acquire real-time multi-modal image data containing neural vascular bundles through an intraoperative imaging device and perform standardization preprocessing;

[0068] A dynamic structure feature extraction module is configured to process the preprocessed image using a three-level cascaded adaptive path convolution (APC) module, to adaptively capture geometric structure features of neural vascular bundles, including bifurcation and strike, through deformable convolution kernels in horizontal and vertical directions, and to output multi-scale local feature maps;

[0069] The global dependency modeling module is configured to input the features extracted by the APC into a local-global interaction module LGIM, and to enhance the perception ability of long-distance topological relations of the neurovascular bundle by calculating spatial and channel attention in blocks, and fusing local features and global context information;

[0070] The cross-region feature fusion and dimension reduction module is configured to perform global integration on multi-scale features by using a cross-scale deformable convolution network module CSDCN, and to compress the feature dimension by using a multi-branch feature distillation module MFDM, and to retain key discriminative features.

[0071] The classification and damage warning module is configured to input the features after dimension reduction into a multi-task classifier, and to output the category and position information of the neurovascular bundle; and to judge the potential damage risk based on feature anomaly detection or a confidence threshold, and to trigger a visual or sound warning signal in real time.

[0072] The model optimization and deployment module is configured to train the network by using a cross-entropy loss function and an Adam optimizer, and to support real-time processing with a lightweight design parameter amount of the model being less than or equal to 1.5M; and to realize low-delay interaction by deploying the model in an edge computing device during an operation.

[0073] In a third aspect, an electronic device is provided, including:

[0074] A processor;

[0075] A memory for storing processor-executable instructions;

[0076] When the processor is configured to execute the instructions, the processor is configured to implement the method for identifying and warning about damage of a neurovascular bundle in an operation based on artificial intelligence according to the first aspect.

[0077] In a fourth aspect, a computer-readable storage medium is provided, and the computer-readable storage medium stores a program, and the program instructs a device to execute the method for identifying and warning about damage of a neurovascular bundle in an operation based on artificial intelligence according to the first aspect.

[0078] The application discloses an intraoperative neurovascular bundle recognition and injury early warning method and system based on artificial intelligence, real-time image data is acquired through a multi-modal imaging device, after standardized preprocessing, a three-level cascaded adaptive path convolution (APC) module is used to extract multi-scale blood vessel features, and a local-global interaction module (LGIM) is combined to enhance long-range topological perception capability. Global features are integrated by using a cross-scale deformable convolution network (CSDCN), and dimensions are compressed and key features are retained through a multi-branch feature distillation module (MFDM). A multi-task classifier synchronously outputs blood vessel bundle categories, position information and injury risk assessment, and dynamic confidence threshold and Mahalanobis distance anomaly detection realize real-time visual / auditory early warning. The model is jointly optimized by using improved cross-entropy loss, CIoU loss (including aspect ratio penalty term) and Mahalanobis distance loss, and the parameter amount is controlled within 1.5M by combining L1 regularization (λ=0.001), and

[0079] TensorRT acceleration is realized on an edge device with a delay of <30ms. The accuracy and real-time performance of intraoperative neurovascular protection are significantly improved.

[0080] Advantages:

[0081] 1. High-precision recognition: multi-scale feature extraction and fusion of the adaptive path convolution (APC) module and the local-global interaction module (LGIM) improve the recognition accuracy.

[0082] 2. Strong real-time performance: lightweight design (parameter amount ≤1.5M) and model compression technology (L1 regularization, channel pruning) are adopted, combined with TensorRT acceleration optimization, single-frame processing delay <30ms is realized on an edge computing device, meeting the intraoperative real-time interaction demand.

[0083] 3. Multi-modal adaptability: optical, ultrasonic and near-infrared multi-modal imaging data are innovatively fused, and the blood vessel shape change under different imaging conditions is adaptively changed through dynamic standardization and deformable convolution kernel, improving the robustness in complex surgical field environment.

[0084] 4. Injury early warning sensitivity: based on the double mechanism of Mahalanobis distance anomaly detection and dynamic confidence threshold, early warning of intraoperative blood vessel injury risk is realized, and the false alarm rate is reduced by more than 40%.

[0085] 5. Good clinical practicability: visual, auditory and tactile multi-modal early warning feedback are adopted to directly guide the operation, and the lightweight design is convenient for integration into the existing surgical navigation system.

[0086] 6. Strong generalization ability: class weights and CIoU aspect ratio optimization are introduced in the loss function, effectively solving the problems of sample imbalance and blood vessel size difference, and adapting to neurovascular bundle recognition in different anatomical sites. BRIEF DESCRIPTION OF DRAWINGS

[0087] Figure 1 A flowchart of a method for intraoperative neurovascular bundle recognition and injury warning based on artificial intelligence is provided for an embodiment of the present application.

[0088] Figure 2 A single-stage APC module structure diagram is provided for an embodiment of the present application.

[0089] Figure 3 An APC feature extraction flowchart is provided for an embodiment of the present application.

[0090] Figure 4 An LGIM feature fusion flowchart is provided for an embodiment of the present application.

[0091] Figure 5 A multi-task classification warning flowchart is provided for an embodiment of the present application.

[0092] Figure 6 A system architecture diagram of an intraoperative neurovascular bundle recognition and injury warning system based on artificial intelligence is provided for an embodiment of the present application.

[0093] Figure 7 An electronic device schematic diagram is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0094] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, but not all the embodiments of the present application.

[0095] It should be noted that “at least one” in the embodiments of the present application means one or more, and more means two or more. Unless otherwise defined, all the technical and scientific terms used in the present application have the same meaning as understood by the person skilled in the art to which the present application belongs. The terms used in the specification of the present application are only for the purpose of describing the specific embodiments of the present application, and are not intended to limit the present application.

[0096] It should be noted that in the embodiments of the present application, the words "first", "second", etc. are used only for the purpose of distinguishing description, and cannot be understood as indicating or implying relative importance, nor can it be understood as indicating or implying sequence. The features limited by "first", "second" can be explicitly or implicitly included one or more of the features. In the description of the embodiments of the present application, the words "exemplary" or "for example" are used to represent as an example, illustration or description. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the use of "exemplary" or "for example" and the like is intended to present the relevant concept in a specific manner.

[0097] Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0098] Embodiment one

[0099] Figure 1 The flowchart of the intraoperative neural vascular bundle recognition and injury warning method based on artificial intelligence provided by an embodiment of the present application is shown in FIG. 1. As shown in FIG. 1, an intraoperative neural vascular bundle recognition and injury warning method based on artificial intelligence comprises the following steps. Figure 1

[0100] S1, acquiring real-time multi-modal image data containing neural vascular bundles through an intraoperative imaging device, and performing standardized preprocessing, including image denoising, size normalization and ROI extraction. This step provides high-quality input data for the subsequent deep learning model.

[0101] Specifically, in the present embodiment, the data acquisition and preprocessing step in S1 further comprises:

[0102] Multi-modal image fusion: acquiring multi-modal image data of neural vascular bundles through intraoperative optical imaging, ultrasound or near-infrared imaging devices; using a deep learning-based multi-modal image registration algorithm to spatially align the multi-modal images, and enhancing the contrast of blood vessels and neural tissue through feature-level fusion.

[0103] Adaptive ROI extraction: using a lightweight U-Net network to generate an initial ROI region, and combining morphological operations to optimize the boundary. Based on a pre-trained lightweight segmentation network (such as a U-Net variant), the potential region of the neural vascular bundle is automatically detected, and an initial ROI mask is generated; combining morphological operations (such as dilation and erosion) to optimize the ROI boundary and eliminate artifacts and irrelevant tissue interference.

[0104] ​Dynamic standardization: including CLAHE enhancement, 128x128 pixel size normalization and [0, 1] pixel value normalization. For intraoperative light changes, histogram matching-based brightness normalization is used; the microvessel structure is enhanced by local contrast limited adaptive histogram equalization (CLAHE); the image is uniformly scaled to 128x128 pixels, and the pixel value is normalized to the [0, 1] interval.

[0105] Real-time optimization: deploying the preprocessing pipeline on the edge computing device, using GPU acceleration to achieve a single-frame processing delay of <50ms; managing the image sequence through a ring buffer to ensure the continuity of real-time processing and early warning.

[0106] Preferably, the ROI extraction network uses adaptive path convolution (APC) as the backbone to adapt to the tortuous morphology of the neurovascular bundle.

[0107] S2, the pre-processed image is processed by using a three-level cascaded adaptive path convolution (APC) module, which adaptively captures the geometric structure features of the neurovascular bundle, including bifurcation and strike, through deformable convolution kernels in horizontal and vertical directions, and outputs multi-scale local feature maps. This module can effectively extract the complex structure information of the neurovascular bundle.

[0108] Specifically, in the present embodiment, the structure of the adaptive path convolution (APC) module in S2 includes:

[0109] Multi-branch convolution design. Figure 2 The single-level APC module structure diagram provided by an embodiment of the present application. Figure 3 The APC feature extraction flowchart provided by an embodiment of the present application. As shown in Figure 2 、 3 Deformable convolution branch: using a 3x3 deformable convolution kernel, adapt to the vascular bifurcation structure through offset learning;

[0110] Horizontal direction perception branch: using a 1x9 horizontal direction convolution kernel, adapt to the horizontal bifurcation and bending structure of the neurovascular bundle through dynamic offset (a x ∈[-1, 0, 1]).

[0111] Vertical strike convolution branch: using a 9x1 vertical direction convolution kernel, capture the vertical strike features of the neurovascular bundle through dynamic offset (a y ∈[-1, 0, 1]).

[0112] Feature fusion: concatenate the feature maps output by the above three branches (ConCat), reduce the dimension through 1x1 convolution, and then generate the final output features through group normalization (GroupNorm) and LeakyReLU activation function, outputting feature maps of three scales: 64x64, 32x32 and 16x16.

[0113] Dynamic shift mechanism: the center coordinates (x i ,y i ) of horizontal and vertical convolution kernels dynamically adjust according to the vascular structure, and the shift amount α c Through the control of learnable parameters, the convolution kernel path is fitted to the geometric morphology of the neurovascular bundle. Preferably, the dynamic shift amount α c is predicted in real time through a lightweight subnetwork to adapt to the individualized variations of the neurovascular bundle during surgery.

[0114] This module can effectively capture the complex geometric structure features of the neurovascular bundle through multi-branch convolution and dynamic shift mechanism, and provide high-quality feature representation for subsequent identification and damage warning.

[0115] Specifically, the processing process of the APC module further includes:

[0116] S2.1, multi-scale feature extraction architecture: equipped with deformable convolution unit (3x3 kernel), adapt to the vascular bifurcation topology through dynamic shift field. A three-level cascaded deformable convolution branch module group is adopted, and the output feature map size of each level module is sequentially decreased to 64x64, 32x32 and 16x16; each level deformable convolution branch module contains parallel horizontal direction convolution (1x9 kernel), vertical direction convolution (9x1 kernel) and standard convolution (3x3 kernel) branch.

[0117] S2.2, dynamic path planning mechanism: horizontal direction analysis branch: asymmetric horizontal convolution kernel (1x9) is adopted, and dynamic path adjustment mechanism is integrated, which is especially optimized for tracking the vascular horizontal bifurcation trajectory. Specifically, for horizontal direction convolution, the horizontal shift amount α i of the center point coordinates (x i ,y x ) is dynamically adjusted through learnable parameters, and the shift range is limited to [-1, 0, 1].

[0118] Vertical direction analysis branch: vertical direction convolution kernel (9x1) is configured, adaptive coordinate transformation unit is embedded, and is specially used for capturing the vertical direction features of blood vessels. Specifically, for vertical direction convolution, the vertical shift amount α i of the center point coordinates (x i ,y y ) is dynamically adjusted through independent learnable parameters, and the shift range is also limited to [-1, 0, 1].

[0119] S2.3, the feature fusion mode includes multi-modal feature integration and feature enhancement processing chain. Among them, the multi-modal feature integration includes: three-dimensional feature channel concatenation (ConCat), 1x1 convolution realizes cross-modal feature interaction. The feature enhancement processing chain includes: group normalization (GroupNorm) layer, LeakyReLU nonlinear activation unit, multi-scale feature output (64x64 to 16x16). Specifically, the feature maps output by the three branches are spliced in the channel dimension (ConCat operation); 1x1 convolution is used to reduce the dimension of the spliced features; and the final output features are generated through group normalization (GroupNorm) and LeakyReLU activation function.

[0120] S2.4, the adaptive enhancement strategy realizes a dynamic path planning mechanism, including a real-time geometric adaptation system and a lightweight control network. Among them, the real-time geometric adaptation system includes dynamic calibration of convolution kernel center coordinates and elastic offset strategy based on blood vessel morphology, and the lightweight control network includes a miniature parameter prediction subnetwork, intraoperative real-time blood vessel morphology adaptation and individualized anatomical structure response. Specifically, according to the clarity of the blood vessel structure in the feature map, the weight distribution of each branch is dynamically adjusted; the attention gate mechanism is introduced to strengthen the feature response to the bifurcation points and curved parts. According to the dynamic adjustment of the intraoperative image resolution, more cascade modules are used for high-resolution images to enhance the feature extraction capability.

[0121] S3, the features extracted by the APC are input into a local-global interaction module LGIM, which fuses local features and global context information by calculating spatial and channel attention in blocks, and enhances the perception ability of long-distance topological relationship of neurovascular bundles. This module helps to improve the model's understanding of the overall structure of neurovascular bundles.

[0122] Figure 4 The LGIM feature fusion flowchart provided by an embodiment of the present application is shown. Specifically, in the embodiment, as shown in Figure 4 the processing process of the local-global interaction module LGIM in S3 further includes:

[0123] Local feature extraction: the feature map output by the dynamic adaptive path convolution APC is divided into non-overlapping N*N local windows (N=4 or 8), and attention is calculated independently for each window. Query vector Q, key vector K and value vector V are extracted for each sub-block respectively, and the dimension is compressed to 1 / 4 of the original channel number.

[0124] In the double-path attention formula, the spatial attention path formula is:

[0125] where Q (query vector): query matrix extracted from input features, used to measure the relevance with other positions. K (key vector): key matrix interacting with Q, determines the attention weight distribution. V (value vector): matrix containing actual feature information, generates output after weighting.d k : feature dimension of key vector K, used to scale dot product results to prevent gradient vanishing. S (·): sparsification function, only keeps Top-K (e.g. K = 8) most relevant key-value pairs, reduces computational complexity. Byd k normalization factor, avoids the saturation of Softmax gradient caused by too large dot product results. Sparsification S (V) reduces the amount of calculation and focuses on key vascular structure features.

[0126] The formula of the channel attention path is: where Q T : transpose of query vector, used for cross-channel interaction. K: key vector, interacts with Q T to calculate the correlation between channels.d k : consistent with spatial attention, feature dimension of key vector. S (V): sparsification processing of value vector. Through channel attention, the weight of important vascular features (such as bifurcation points and main blood vessels) is enhanced.

[0127] Through deformable convolution, the attention receptive field is adjusted, and the outputs of spatial attention and channel attention are concatenated in the channel dimension; through 1 × 1 convolution, feature fusion and dimension recovery are performed; a skip connection is added to retain the original feature information. Through the dynamic sparse attention mechanism, the long-range dependency between windows is calculated, and only the Top-K significantly correlated window pairs are retained to reduce the computational complexity.

[0128] Cross-window feature fusion: adopts gated linear unit (GLU) to dynamically fuse local and global features, and the formula is:

[0129] Output = Conv 1×1 (σ (Conv 1×1 (F local )) ⊙) F global ,

[0130] where F local is the feature extracted by the local window, F global is the feature output by the global attention, and is the element-wise multiplication, which dynamically fuses local and global features. σ: Sigmoid activation function, generates weights between 0 and 1. Conv 1×1 (x) represents 1 × 1 convolution operation on input feature map x. Self-adaptive balance of local details (such as small blood vessels) and global structure (such as the direction of main blood vessels).

[0131] Dynamic receptive field adjustment:

[0132] Local processing replaces group attention: group computation is achieved through local window convolution, preserving spatial sensitivity. Sparse attention replaces multi-head mechanism: dynamically selects key window interaction, reduces computational complexity (complexity from O(N 2 ) to O(NlogN). Gated fusion replaces skip connection: enhances feature selection capability, equivalent to LGIM's concatenation + 1x1 convolution effect.

[0133] Global attention is introduced in the last LGIM module to capture the topological structure of the complete neurovascular bundle; deformable convolution is used to adjust the attention receptive field to adapt to different sizes of vascular structures. Preferably, the number of groups of the LGIM module is dynamically adjusted according to the feature map resolution, and more groups are used for high-resolution feature maps to reduce computational complexity.

[0134] This module effectively improves the performance of neurovascular bundle recognition and injury warning through the fusion of local and global features, dynamic sparse attention mechanism and deformable convolution.

[0135] S4, a cross-scale deformable convolution network module CSDCN is used to globally integrate multi-scale features, and a multi-branch feature distillation module MFDM is used to compress the feature dimension to retain key discriminative features. This step helps to reduce the amount of calculation while retaining key information.

[0136] Specifically, in the present embodiment, the cross-scale deformable convolution network processing step in S4 further comprises:

[0137] Multi-scale feature alignment: dynamically aligns the vascular structures of different resolution feature maps through a deformable convolution kernel (3x3), and the offset is predicted by a lightweight subnetwork. This method can adapt to different scales of vascular structures and improve the accuracy of feature extraction.

[0138] Pyramid fusion: upsample 32x32, 16x16, and 8x16 feature maps to a uniform scale, add them up, and then perform 3x3 convolution fusion. This method enhances the model's perception of different scale vascular structures through multi-scale feature fusion.

[0139] Channel-space collaborative attention: applies a CBAM attention module to the fused features to simultaneously optimize channel and spatial weights. This method combines channel attention and spatial attention to more effectively extract key features and improve model performance.

[0140] The multi-branch feature distillation module MFDM in S4 is used to compress the feature dimension to retain key discriminative features, including:

[0141] Multi-branch feature extraction: Branch one (local detail preservation): 1x1 convolution + depthwise separable convolution (3x3) is adopted, and the output channel number is 1 / 2 of the input feature. Branch two (global context modeling): non-local attention module is used to calculate the global feature correlation, and the output channel number is the same as the input. Branch three (structure enhancement): deformable convolution (3x3) is adopted to dynamically adapt to the blood vessel morphology, and the output channel number is 1 / 4 of the input feature.

[0142] Feature fusion and distillation: the three branch output features are spliced along the channel dimension (ConCat), and reduced to the target channel number through 1x1 convolution; channel attention (SE module) is used to dynamically adjust the contribution weight of each branch, and the formula is: F out =Sigmoid(W SE (F concat ))⊙F concat , where F concat represents the channel splicing result of multi-branch (such as local detail branch, global context branch, structure enhancement branch) output feature, with dimension BxC'HxW (C' is the total channel number after splicing), which is used to retain the complementary features (such as blood vessel details, long-range topology, etc.) extracted by different branches. W SE represents the channel attention weight matrix, which is generated through Squeeze-and-Excitation (SE) structure, with dimension C'xC', learning the importance weight of each channel, enhancing the response of key features (such as main blood vessels) and suppressing noise. Sigmoid(·) is used to normalize the weight to the interval [0, 1], which represents the activation intensity of each channel.

[0143] ⊙ represents element-wise multiplication, which dynamically scales the spliced features by channel weight. Through the channel attention mechanism, multi-branch features are adaptively fused, and the channel information most relevant to neural vascular bundle recognition is highlighted.

[0144] Cross-level information transmission: in the cascade structure, high-level features are added to low-level features through jump connection to avoid information loss.

[0145] MFDM realizes multi-view feature fusion of MPR through multi-branch structure (local + global + deformation); the max pooling branch of deformable convolution is more suitable for blood vessel morphology. Through multi-branch feature extraction, dynamic weight adjustment and cross-level information transmission, the feature representation ability is effectively improved, while the calculation efficiency and model performance are considered.

[0146] S5, input the reduced dimension feature into the multi-task classifier, output the category and position information of neural vascular bundle; at the same time, based on feature anomaly detection or confidence threshold, judge the potential damage risk, and trigger visual or sound warning signal in real time. This step realizes real-time warning of neural vascular bundle damage.

[0147] Figure 5 A multi-task classification and early warning process diagram is provided for an embodiment of the present application. Specifically, in the present embodiment, as shown in Figure 5 the classification and injury early warning step (S5) further includes:

[0148] Multi-task classifier design: adopt full connection layer and softmax activation function to output the category probability distribution of neural vascular bundle; access the coordinate regression branch in parallel, and output the boundary box position (x, y, w, h) of the vascular bundle through the sigmoid function.

[0149] Injury risk assessment mechanism: confidence monitoring and feature anomaly detection can be used. When confidence monitoring is used, the strategy is: when the highest category probability of the classifier is lower than the dynamic threshold (preset as 0.7), it is determined as an uncertain recognition area; when feature anomaly detection is used, the strategy is: calculate the Mahalanobis distance of the test sample feature and the feature distribution of the training set, and trigger the early warning when the distance exceeds 3σ.

[0150] The multi-modal early warning system can include visual prompts: marking the risk area with a flashing red outline on the intraoperative navigation interface, and displaying the confidence value; auditory prompts: playing different frequency beep sounds according to the risk level (low / medium / high); tactile feedback: transmitting the danger signal through the vibration motor of the handheld instrument (optional).

[0151] Real-time guarantee measures: adopt double buffering mechanism to process classification tasks, ensure that the single frame processing delay is less than 30ms; sliding window verification is performed on the danger signals detected in the continuous 3 frames to reduce the false alarm rate. Preferably, the dynamic threshold is dynamically adjusted according to the operation stage, and the threshold is automatically increased to 0.8 in the key dissection stage (such as nerve separation) to enhance the sensitivity.

[0152] The method realizes efficient and accurate neural vascular bundle recognition and injury early warning through multi-task classification, multi-modal early warning and real-time guarantee.

[0153] S6, the network is trained by using a cross-entropy loss function and an Adam optimizer, the model is lightweight and the design parameter amount is less than or equal to 1.5M, real-time processing is supported; in the operation, an edge computing device is deployed to realize low-delay interaction. The method ensures the performance of the model while taking into account the real-time performance and computing efficiency.

[0154] The loss function system adopts a multi-task joint optimization loss function. The main classification loss function is (improved cross-entropy):

[0155]

[0156] wherein w cClass weights (1.2:1.5:0.8 for vessel / neuron / background class weights) to address sample imbalance issues. i,c Indicator function (1 if sample i belongs to class c, 0 otherwise). i,c Probability that the model predicts sample i belongs to class c. L1 L1 regularization coefficient (0.001) to enforce model parameter sparsity, control parameter quantity <1.5M. L1 = 0.001 indicates control model sparsity. θ Model trainable parameters. Weighted cross-entropy to pay more attention to minority classes (e.g., neurons), L1 regularization to reduce redundant parameters.

[0157] Coordinate regression loss (CIoU Loss):

[0158]

[0159] Where IoU is the intersection over union of the predicted and true bounding boxes, measuring the degree of overlap. ρ represents the Euclidean distance between the center points of the predicted and true bounding boxes, d is the diagonal length of the minimum enclosing rectangle, and the center point error is normalized. v represents the aspect ratio penalty term, which is smoothed by the arctangent function to handle size differences. α represents the balance coefficient (default 1.0), which adjusts the weight of the aspect ratio term. The coordinate regression loss is used to optimize the position, overlap rate and aspect ratio of the bounding box at the same time, improving the positioning accuracy of the blood vessel bundle.

[0160] Anomaly detection loss (Mahalanobis distance):

[0161] L ano = max(0, D M (x) - τ),

[0162]

[0163] Where D M (x) represents the Mahalanobis distance between the test sample feature x and the distribution of the training set, measuring the degree of anomaly. μ represents the mean vector of the training set features. Σ: Covariance matrix of training set features. τ = 3σ is the dynamic threshold, σ is the standard deviation. The anomaly detection loss is used to detect the deviation of feature distribution, and triggers a warning when an abnormal blood vessel structure is encountered during surgery.

[0164] Classification loss (L cls ): Weighted cross-entropy loss is used, where the weight ratio of vessel, neuron and background classes is set to 1.2:1.5:0.8 to address class imbalance issues; Regression loss (L reg ): CIoU loss function is used to optimize the intersection over union (IoU), center point distance and aspect ratio similarity of the predicted bounding box; Anomaly detection loss (L ano ): Based on Mahalanobis distance (D MDynamically calculate the deviation of the feature distribution, when D M A loss penalty is triggered when the value is greater than 3σ.

[0165] Preferably, the aspect ratio penalty term v in the CIoU loss is calculated using the arctangent function to smooth out the geometric differences of vascular bundles of different sizes. Dynamic class weights are used to balance the sample imbalance problem. The CIoU Loss simultaneously optimizes the position, aspect ratio, and center point. Mahalanobis distance loss achieves end-to-end anomaly detection, and L1 regularization achieves model lightweighting (number of parameters ≤ 1.5M).

[0166] Regularization and sparsity constraints: Introduce an L1 regularization term (λ) into the total loss. L1 =0.001), forcing the model parameters to be sparsity, achieving lightweight design (number of parameters ≤1.5M).

[0167] The total loss function is the weighted sum of the losses in each sub-loss:

[0168] L reg =αL cls +βL reg +γL ano ,

[0169] Dynamic threshold adjustment mechanism: The classification confidence threshold is dynamically adjusted according to the importance of the intraoperative stage (default 0.7, increased to 0.8 in critical stages); the Mahalanobis distance threshold τ is updated online as the feature distribution changes to ensure the sensitivity of anomaly detection.

[0170] Specifically, in this embodiment, the model optimization and deployment step (S6) further includes:

[0171] Mixed-precision training mechanism: Employs a mixed-precision training strategy using FP16 / FP32, through...

[0172] NVIDIA Tensor Cores accelerate computation; Dynamic Loss Scaling is used to prevent gradient underflow. This method improves computational efficiency while maintaining model accuracy through mixed-precision training.

[0173] Progressive model compression: Training phase: An L1 regularization term (λ = 0.001) is added to the cross-entropy loss function to promote weight sparsity, which helps reduce the number of model parameters. Deployment phase: Channel pruning is used to remove feature channels with a contribution of less than 5%, further compressing the model size.

[0174] Edge computing optimization: The model is converted into the TensorRT engine, the computation graph is optimized and the inference path is solidified; it is deployed on the NVIDIA Jetson AGX Xavier edge device, utilizing an 8-core ARM CPU and a 512-core Volta GPU for parallel processing to improve inference efficiency.

[0175] Real-time monitoring system: Built-in performance analyzer continuously monitors inference latency (target <50ms / )

[0176] (frames); when the delay exceeds the threshold, automatically switch to the simplified model (parameter count ≤ 0.8M).

[0177] Online adaptive mechanism: Through continuous learning technology, the model parameters are dynamically fine-tuned using intraoperative data; a model update verification mechanism is set up to ensure that the accuracy fluctuation does not exceed ±2%. Preferably, the edge device is directly connected to the surgical field camera system through a 5G private network to establish an ultra-low latency transmission channel of <10ms to ensure real-time interaction.

[0178] This method achieves a balance between model performance and real-time performance through hybrid precision training, model compression, edge computing optimization, and real-time monitoring, making it suitable for real-time processing of intraoperative neurovascular bundle identification and damage warning.

[0179] Example 2

[0180] like Figure 6 As shown, this application provides an architecture diagram of an artificial intelligence-based intraoperative neurovascular bundle recognition and injury early warning system, which is applied to the artificial intelligence-based intraoperative neurovascular bundle recognition and injury early warning system as described in Embodiment 1. It includes a data acquisition and preprocessing module 11, a dynamic structural feature extraction module 12, a global dependency modeling module 13, a cross-regional feature fusion and dimensionality reduction module 14, a classification and injury early warning module 15, and a model optimization and deployment module 16.

[0181] Specifically, the data acquisition and preprocessing module 11 is used to acquire real-time multimodal image data containing neurovascular bundles through intraoperative imaging equipment and perform standardized preprocessing.

[0182] The dynamic structural feature extraction module 12 is used to process the preprocessed image using a three-level cascaded adaptive path convolution (APC) module. It adaptively captures the geometric structural features, including the bifurcation and direction of the neurovascular bundle, through deformable convolution kernels in the horizontal and vertical directions, and outputs multi-scale local feature maps.

[0183] The global dependency modeling module 13 is used to input the features extracted by the APC into the local-global interaction module LGIM, and enhance the perception of long-distance topological relationships of neurovascular bundles by merging local features and global contextual information through block-based computation of spatial and channel attention.

[0184] The cross-region feature fusion and dimensionality reduction module 14 is used to globally integrate multi-scale features using the cross-scale deformable convolutional network module CSDCN, and compress the feature dimension through the multi-branch feature distillation module MFDM to retain key discriminative features.

[0185] The classification and damage warning module 15 is used to input the dimensionality-reduced features into the multi-task classifier and output the category and location information of the neurovascular bundle; at the same time, it judges the potential damage risk based on feature anomaly detection or confidence threshold and triggers visual or sound warning signals in real time.

[0186] The model optimization and deployment module 16 is used to train the network using the cross-entropy loss function and the Adam optimizer. The model is designed to be lightweight with ≤1.5M parameters and supports real-time processing. During the operation, it is deployed through edge computing devices to achieve low-latency interaction.

[0187] Figure 7 This is an electronic device provided in one embodiment of this application. For example... Figure 7 As shown, the electronic device includes at least the following components: processor 101 and memory 100, communication interface 103, and bus 102.

[0188] In this embodiment of the application, memory 100 is used to store executable instructions of processor 101, which, when configured to execute instructions, implements the method as described in the first aspect.

[0189] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0190] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these devices is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0191] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0192] It should be noted that the term "computer" as used here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, "computer-readable recording media" refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage devices such as hard drives built into a computer.

[0193] Furthermore, a "computer-readable recording medium" can include: a medium that dynamically stores a program for a short period of time, such as a communication line used when transmitting a program via a network such as the Internet or a communication line such as a telephone line; or a medium that stores a program for a fixed period of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining with programs already recorded in the computer.

[0194] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (device group) composed of multiple devices. Each device constituting the device group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a device group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0195] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for intraoperative neurovascular bundle recognition and injury early warning based on artificial intelligence, characterized in that, The method includes: S1. Acquire real-time multimodal image data containing neurovascular bundles using intraoperative imaging equipment and perform standardized preprocessing; S2. The preprocessed image is processed by a three-level cascaded adaptive path convolution (APC) module. The geometric structural features, including the bifurcation and direction of the neurovascular bundle, are adaptively captured by deformable convolution kernels in the horizontal and vertical directions, and multi-scale local feature maps are output. S3. Input the features extracted by APC into the Local-Global Interaction Module (LGIM). By calculating spatial and channel attention in blocks, local features and global context information are fused to enhance the perception of long-distance topological relationships of neurovascular bundles. S4. The multi-scale deformable convolutional network module CSDCN is used to globally integrate multi-scale features, and the feature dimension is compressed by the multi-branch feature distillation module MFDM to retain key discriminative features. S5. Input the dimensionality-reduced features into the multi-task classifier and output the category and location information of the neurovascular bundle; at the same time, judge the potential damage risk based on feature anomaly detection or confidence threshold and trigger visual or sound warning signals in real time. S6. The network is trained using the cross-entropy loss function and the Adam optimizer. The model is designed with lightweight parameters ≤1.5M and supports real-time processing. It is deployed through edge computing devices during the operation to achieve low-latency interaction.

2. The method for intraoperative neurovascular bundle recognition and injury early warning based on artificial intelligence according to claim 1, characterized in that, The data acquisition and preprocessing steps in S1 further include: Multimodal image data of the neurovascular bundle are acquired through intraoperative optical imaging, ultrasound or near-infrared imaging equipment, and spatial alignment is performed using a deep learning-based registration algorithm to enhance the contrast between blood vessels and nerve tissue. The system automatically detects potential regions of neurovascular bundles based on a lightweight segmentation network, generates an initial ROI mask, and optimizes the ROI boundaries by combining morphological operations. Brightness normalization is achieved by using histogram matching, and microvascular structure is enhanced by adaptive histogram equalization with limited local contrast. The image is also scaled to a preset size and pixel values ​​are normalized to a specified range. Deploy a preprocessing pipeline on edge computing devices to achieve low-latency single-frame processing using GPU acceleration, and manage image sequences through a circular buffer to ensure continuous real-time processing. The ROI extraction network uses adaptive path convolution (APC) as its backbone to adapt to the tortuous shape of the neurovascular bundle.

3. The method for intraoperative neurovascular bundle recognition and injury early warning based on artificial intelligence according to claim 1, characterized in that, The adaptive path convolution (APC) module in S2 includes a multi-branch convolution design module, a feature fusion module, and a dynamic offset mechanism. The multi-branch convolution design module includes: Deformable convolutional branch: A 3×3 deformable convolutional kernel is used to learn and adapt to the vascular bifurcation structure through offset; Horizontal sensing branch: Uses a 1×9 horizontal convolution kernel, with dynamic offset α x ∈[-1,0,1] adapts to the horizontal bifurcation and tortuous structure of the neurovascular bundle; Vertical convolution branch: Uses a 9×1 vertical convolution kernel, with dynamic offset α y ∈[-1,0,1] captures the vertical orientation characteristics of neurovascular bundles; The feature fusion module is as follows: it concatenates the feature maps output from the three branches, reduces the dimensionality by 1×1 convolution, and generates the final output features by group normalization and LeakyReLU activation function, outputting feature maps of three scales: 64×64, 32×32 and 16×16. The dynamic offset mechanism is as follows: the center coordinates of the horizontal and vertical convolution kernels are dynamically adjusted according to the vascular structure, and the offset is controlled by learnable parameters so that the convolution kernel path conforms to the geometry of the neurovascular bundle. The dynamic offset is predicted in real time by a lightweight subnetwork to adapt to individual variations in the neurovascular bundle during surgery.

4. The method for intraoperative neurovascular bundle recognition and injury early warning based on artificial intelligence according to claim 1, characterized in that, The processing procedure of the local-global interaction module LGIM in S3 further includes: S3.1 Local Feature Extraction: The feature map output by Adaptive Path Convolution (APC) is divided into non-overlapping N×N local windows. Attention is calculated independently for each window, and query vector Q, key vector K, and value vector V are extracted. The dimension is compressed to 1 / 4 of the original number of channels. S3.2 Dual-path attention calculation: Spatial Attention Path space The formula for calculating the association weight of spatial positions within a window is as follows: Channel Attention Path channel The formula for calculating the feature association weights across channels is: Where S(·) is the sparsification function based on saliency sampling of vascular structures, d k This represents the dimension of the key vector K; S3.3 Feature Fusion and Enhancement: The outputs of spatial attention and channel attention are concatenated along the channel dimension; feature fusion and dimensionality restoration are performed through 1×1 convolution; skip connections are added to preserve the original feature information; S3.4 Cross-window feature fusion: Local and global features are dynamically fused using gated linear units, with the following formula: Output=Conv 1×1 (σ(Conv 1×1 (F local ))⊙)F global , Among them, F local F global These represent local and global features, respectively; ⊙ represents element-wise multiplication; Conv 1×1 (x) represents a 1×1 convolution operation on the input feature map x; S3.5 Dynamic receptive field adjustment: Adjust the attention receptive field through deformable convolution to adapt to vascular structures of different sizes.

5. The method for intraoperative neurovascular bundle recognition and injury early warning based on artificial intelligence according to claim 1, characterized in that, The multi-scale deformable convolutional network processing steps in S4 further include: S4.1 Multi-scale feature alignment: Vascular structures of feature maps at different resolutions are dynamically aligned through multi-level deformable convolutional kernels, and the offset is predicted by a lightweight sub-network. S4.2 Pyramid Fusion: The 32×32, 16×16, and 8×8 feature maps are upsampled to a uniform scale, then added together, and finally fused by 3×3 convolution; S4.3, Channel-Spatial Co-attention: Apply the CBAM attention module to the fused features to simultaneously optimize channel and spatial weights.

6. The method for intraoperative neurovascular bundle recognition and injury early warning based on artificial intelligence according to claim 5, characterized in that, In step S4, the feature dimension is compressed using the Multi-Branch Feature Distillation (MFDM) module, retaining key discriminative features, including: Feature extraction is performed using a multi-branch approach; The three output features are concatenated along the channel dimension and then reduced to the target number of channels using a 1×1 convolution. Channel attention is used to dynamically adjust the contribution weights of each branch, using the following formula: F out =Sigmoid(W SE (F concat ))⊙F concat , Among them, F concat W represents the channel concatenation result of multi-branch output features. SE This represents the channel attention weight matrix, Sigmoid(·) represents the activation intensity of each channel, and ⊙ represents element-wise multiplication; In a cascaded structure, high-level features are added to low-level features through skip connections to avoid information loss.

7. The method for intraoperative neurovascular bundle recognition and injury early warning based on artificial intelligence according to claim 6, characterized in that, The classification and damage warning steps in S5 further include: The class probability distribution of the neurovascular bundle is output using a fully connected layer and a softmax activation function; Parallel access to the coordinate regression branch, outputting the bounding box position (x,y,w,h) of the vascular bundle through the sigmoid function; Confidence monitoring: When the probability of the highest class in the classifier is lower than the dynamic threshold, it is determined to be an area of ​​uncertainty in identification; Feature anomaly detection: Calculate the Mahalanobis distance between the feature distribution of the test sample and the feature distribution of the training set. When the distance exceeds 3σ, trigger the corresponding multimodal early warning system and real-time protection measures.

8. The method for intraoperative neurovascular bundle recognition and injury early warning based on artificial intelligence according to claim 7, characterized in that, The model optimization and deployment steps in S6 further include: A mixed-precision FP16 / FP32 training strategy is adopted, and computation is accelerated through Tensor Cores; dynamic loss scaling is implemented to prevent gradient underflow. Training phase: Add an L1 regularization term to the loss function to achieve weight sparsity; Deployment phase: Remove feature channels with a contribution of less than 5% through channel pruning; The model was converted to the TensorRT engine to optimize the inference path and deployed on the Jetson AGX Xavier edge device, utilizing ARM CPU and Volta GPU for parallel processing; Continuously monitor inference latency, and automatically switch to a simplified model when latency exceeds the limit; By using continuous learning techniques and intraoperative data for dynamic fine-tuning, a verification mechanism is set up to ensure that the accuracy fluctuation does not exceed the preset value.

9. The method for intraoperative neurovascular bundle recognition and injury early warning based on artificial intelligence according to claim 1, characterized in that, The loss function is: L reg =αL cls +βL reg +γL ano , Among them, L cls L represents the classification loss. reg L represents the regression loss. ano This represents the anomaly detection loss, where α, β, and γ are the corresponding balance coefficients.

10. An artificial intelligence-based intraoperative neurovascular bundle recognition and injury early warning system, applied to the artificial intelligence-based intraoperative neurovascular bundle recognition and injury early warning method as described in any one of claims 1 to 9, characterized in that, The system includes: The data acquisition and preprocessing module is used to acquire real-time multimodal image data containing neurovascular bundles through intraoperative imaging equipment and perform standardized preprocessing. The dynamic structural feature extraction module is used to process the preprocessed image using a three-level cascaded adaptive path convolution (APC) module. It adaptively captures the geometric structural features, including the bifurcation and direction of the neurovascular bundle, through deformable convolution kernels in the horizontal and vertical directions, and outputs multi-scale local feature maps. The global dependency modeling module is used to input the features extracted by the APC into the local-global interaction module LGIM. By calculating spatial and channel attention in blocks, it fuses local features and global context information to enhance the perception of long-distance topological relationships of neurovascular bundles. The cross-region feature fusion and dimensionality reduction module is used to globally integrate multi-scale features using the cross-scale deformable convolutional network module CSDCN, and compress the feature dimension through the multi-branch feature distillation module MFDM to retain key discriminative features. The classification and damage warning module is used to input the dimensionality-reduced features into the multi-task classifier and output the category and location information of the neurovascular bundle; at the same time, it judges the potential damage risk based on feature anomaly detection or confidence threshold and triggers visual or sound warning signals in real time. The model optimization and deployment module is used to train the network using the cross-entropy loss function and the Adam optimizer. The model is designed to be lightweight with ≤1.5M parameters and supports real-time processing. During the procedure, it is deployed through edge computing devices to achieve low-latency interaction.

Images