Spine bone minimally invasive surgery target real-time detection method and device based on SCI

Abstract

The invention discloses a spine bone minimally invasive surgery target real-time detection method and device based on SCI, and relates to the field of image processing, and the method comprises the steps: S1, obtaining a spine bone minimally invasive surgery target real-time gray level image, carrying out the preprocessing of the gray level image, and marking the gray level image to make a data set; s2, constructing a target detection model comprising a Backbone part, a self-distillation part, a Neck part and a Head part; the Backbone part comprises a self-calibration illumination enhancement operation; constructing a loss function of the model based on GIoU loss; and S3, inputting the data set into a target detection model, and obtaining a self-calibration illumination enhanced spine bone minimally invasive surgery target real-time detection image. According to the method, self-calibration illumination enhancement and knowledge distillation technologies are combined, so that the quality of a real-time detection image of a spine bone minimally invasive surgery target is remarkably improved; an SE attention mechanism and a GIoU loss function are introduced to enhance the feature extraction capability and bounding box adjustment precision of the model, and the recognition rate and accuracy of target detection are improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a real-time detection method and device for a spinal minimally invasive surgery target based on SCI. Background Art

[0002] Minimally invasive spinal surgery is an important means of treating spinal degenerative diseases, spinal tumors, spinal fractures and other diseases. It is widely used because of its small trauma and fast recovery. However, this type of surgery has extremely high requirements for real-time detection and precise positioning of the lesion area during surgery. Due to factors such as the complex anatomical structure of the spine and the limited surgical field of view, traditional image-guided methods are difficult to fully meet the needs. Although preoperative reconstructed images and intraoperative real-time images (such as X-rays, CT or MRI) provide assistance for surgical navigation, these image data are often affected by noise interference and insufficient contrast due to uneven lighting, instrument occlusion and dynamic changes in tissues, which reduces the accuracy and safety of the surgery.

[0003] To address the above problems, image processing methods based on self-calibrated illumination enhancement (SCI) have gradually gained attention in recent years. The SCI module can dynamically adjust the brightness and contrast of the image according to the actual lighting conditions in the surgical scene, thereby effectively reducing the image quality degradation caused by uneven lighting or overexposure. With the help of this image enhancement technology, the clarity and contrast of surgical images can be significantly improved, thereby improving the detection accuracy of target lesions during surgery and reducing misjudgments and operational errors caused by image quality problems.

[0004] The SCI-based real-time detection method for minimally invasive spinal surgery targets can provide clearer and more accurate image guidance in complex and changing surgical environments by integrating advanced image enhancement technology and efficient real-time detection algorithms. This can not only significantly improve the safety and success rate of surgery, but also effectively reduce the risk of misjudgment and missed detection during surgery, further guarantee the patient's prognosis, and at the same time demonstrate important clinical application value and broad development potential. Summary of the invention

[0005] In view of the above problems, the present invention proposes a real-time detection method and device for spinal minimally invasive surgery targets based on SCI, the purpose of which is to significantly improve the quality of real-time detection images of spinal minimally invasive surgery targets by combining self-calibrated illumination enhancement (SCI) with knowledge distillation technology, and optimize the problem of false detection and missed detection. At the same time, the SE attention mechanism and GIoU loss function are introduced to further enhance the feature extraction ability and bounding box adjustment accuracy of the model, thereby improving the recognition rate and accuracy of target detection, making the present invention more robust and adaptable in complex surgical scenarios.

[0006] On the one hand, the real-time detection method of the target of minimally invasive spine surgery based on SCI has the following specific steps:

[0007] S1, obtain the real-time grayscale image of the target of minimally invasive spine surgery, pre-process the grayscale image and annotate it to make a data set;

[0008] S2, construct an object detection model including Backbone part, self-distillation part, Neck part and Head part;

[0009] S3, uses the target detection model to process the dataset and obtain images of real-time detection of targets in minimally invasive spine surgery enhanced by self-calibration illumination;

[0010] The Backbone part includes an SCI layer, an SE layer, and several alternating modules; the Backbone part obtains a data set, which is processed by the SCI layer and the SE layer in sequence after low-level feature processing, and then is processed by each alternating module in sequence for feature extraction; the feature representation output by each alternating module is output to the self-distillation part and the Neck part respectively; the SCI layer performs self-calibration illumination enhancement; the SE layer uses the SE attention mechanism for feature enhancement;

[0011] The Neck part obtains feature representation and outputs feature maps of several dimensions;

[0012] The Head part obtains a dimensional feature map and outputs an image of real-time detection of spinal minimally invasive surgery targets enhanced by self-calibration illumination;

[0013] The self-distillation part obtains the feature representation from the alternating module to construct a self-distillation loss; the self-distillation loss combined with the GIoU loss constitutes the total loss function of the target detection model.

[0014] Preferably, in S1, the preprocessing of the grayscale image includes: removing the annotation information around the real-time detection image of the spinal minimally invasive surgery target and enhancing the image contrast;

[0015] The calculation formula for enhancing the image contrast is: ;in, Represents the value of the original image pixel, Represents the pixel value after corresponding enhancement processing, represents the logarithmic function, and C represents a constant.

[0016] Preferably, the SCI layer performs self-calibration illumination enhancement, specifically comprising:

[0017] First, the illumination model of the intraoperative images in the dataset is estimated as follows:

[0018] ;

[0019] in, Represents the original image at pixel position The brightness value of Represents pixels Real scene reflection characteristics at Represents pixels The light intensity at

[0020] Secondly, the illumination components of the intraoperative images are estimated using Gaussian filtering. , the formula is as follows:

[0021] ;

[0022] in, represents a Gaussian filter with a standard deviation of ; represents the original estimated illumination component, represents the enhancement adjustment amount provided by the self-calibration enhancement module;

[0023] Next, the reflection component of the intraoperative image is separated using the following formula:

[0024] ;

[0025] in, Represents a constant, used to avoid the denominator being zero;

[0026] Finally, the intraoperative image is subjected to a self-calibrated illumination enhancement step: the reflectance component Enhance the reflected component With light component Recombining to get the final enhanced image, the formula is as follows:

[0027] ;

[0028] in, Represents the image at pixel position after self-calibration illumination enhancement The brightness value at ; Represents the contrast adjustment parameter, which is used to control the degree of image contrast enhancement. The value range is .

[0029] Preferably, after obtaining the final enhanced image, the method further comprises: Boundary processing is performed; the boundary processing includes edge-preserving filtering.

[0030] Preferably, the SE attention mechanism is used for feature enhancement, specifically as follows:

[0031] First, the feature map of each channel is compressed into a scalar through global average pooling to obtain a global feature description. The specific calculation formula is as follows:

[0032] ;

[0033] in, Represents the global description scalar of the cth channel; H represents the height of the feature map, and W represents the width of the feature map; Indicates the input feature map in the cth channel pixel value;

[0034] Secondly, after obtaining the global description scalar of each channel, the global description is processed through two fully connected layers and activation functions to generate the weight of each channel. The specific calculation formula is as follows:

[0035] ;

[0036] in, represents the recalibrated channel weight vector; and Represent the weight matrices of the fully connected layers respectively; Represents the Sigmoid activation function; represents the SiLU activation function;

[0037] Finally, the obtained channel weights are remapped back to the feature map of each channel, and the features of each channel are reweighted by element-by-element multiplication. The specific calculation formula is as follows:

[0038] ;

[0039] in, represents the feature map of the cth channel after recalibration by the SE module; Represents the feature map of the cth channel of the input feature map.

[0040] Preferably, the number of the alternating modules is three, which are an upper alternating module, a middle alternating module and a lower alternating module in the order in which the data sets flow in the model.

[0041] Preferably, the feature representation obtained from the alternating module is constructed from the distillation loss, as follows:

[0042] First, get the output of the upper alternating module , and the adaptive convolution adjuster is used for adjustment; the calculation formula is as follows:

[0043] ;

[0044] in, represents the convolution regulator applied to the upper alternating module, represents the convolution parameters; Represents the output of the upper alternating module after convolution adjustment;

[0045] Secondly, get the output of the middle-level alternating module , using adaptive convolution adjuster to adjust; the calculation formula is as follows:

[0046] ;

[0047] in, represents the convolution regulator applied to the middle layer alternating module, represents the convolution parameters; Represents the output of the middle-layer alternating module after convolution adjustment;

[0048] Finally, the output of the lower alternating module is obtained ,use Construct the self-distillation loss; the total self-distillation loss is calculated as:

[0049] ;

[0050] in, represents the loss from distillation; express Norm.

[0051] Preferably, the total loss function of the target detection model is expressed as:

[0052] ;

[0053] ;

[0054] in, represents the total loss function; represents the loss from distillation; represents the confidence loss; represents GIoU loss; represents the classification loss; , , , They represent weights respectively; A represents the predicted box; B represents the real box; C represents the minimum closed area containing A and B; Indicates area; Represents the overlapping area of ​​two boxes; Represents the union area of ​​two boxes.

[0055] Preferably, the Head part includes several SE attention modules; each SE attention module performs feature enhancement on a dimensional feature map.

[0056] On the other hand, the real-time detection device for minimally invasive spine surgery targets based on SCI includes the following:

[0057] Image acquisition and preprocessing module, which acquires real-time grayscale images of the spinal minimally invasive surgery target, preprocesses the grayscale images and annotates them to produce a data set;

[0058] Model building module, used to build the target detection model including Backbone part, self-distillation part, Neck part and Head part;

[0059] A target real-time detection image acquisition module is used to process the data set using a target detection model to obtain images of real-time detection of targets in minimally invasive spine surgery enhanced by self-calibration illumination;

[0060] The Backbone part includes an SCI layer, an SE layer, and several alternating modules; the Backbone part obtains a data set, which is processed by the SCI layer and the SE layer in sequence after low-level feature processing, and then is processed by each alternating module in sequence for feature extraction; the feature representation output by each alternating module is output to the self-distillation part and the Neck part respectively; the SCI layer performs self-calibration illumination enhancement; the SE layer uses the SE attention mechanism for feature enhancement;

[0061] The Neck part obtains feature representation and outputs feature maps of several dimensions;

[0062] The Head part obtains a dimensional feature map and outputs an image of real-time detection of spinal minimally invasive surgery targets enhanced by self-calibration illumination;

[0063] The self-distillation part obtains the feature representation from the alternating module to construct the self-distillation loss; the total loss function of the target detection model is based on the GIoU loss and the self-distillation loss.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] (1) The present invention introduces a self-calibrated illumination enhancement module (SCI layer) after initially extracting low-level features, which significantly improves the quality of the input image and makes the model more stable and accurate when processing low-quality images in minimally invasive spinal surgery. This improvement is particularly effective in complex surgical environments and helps to significantly improve the reliability of target detection.

[0066] (2) After initially completing the self-calibrated illumination enhancement (SCI) processing, the present invention further introduces an edge-preserving filtering step to perform boundary processing on the enhanced image. This subsequent processing significantly reduces the artifacts and over-enhancement problems that may occur during the image enhancement process, especially in the edge areas of the image. Through this optimization step, the quality of the final output image is further improved, making the target detection of the model in minimally invasive spinal surgery smoother and more natural, and significantly improving the detection reliability and accuracy in complex surgical environments.

[0067] (3) The present invention introduces the SE attention mechanism into the image processed by the SCI module, realizes the adaptive calibration of channel features, significantly suppresses the interference of irrelevant features, and optimizes the allocation of computing resources. This improvement effectively improves the model's feature expression ability and target detection accuracy, significantly improves its performance in real-time detection of targets in minimally invasive spine surgery, enhances the model's robustness and operating efficiency, and thus provides more reliable support for clinical diagnosis.

[0068] (4) The Backbone part of the present invention introduces a self-distillation module, which aims to enhance the consistency and relevance of image feature information during model transmission. Through the self-distillation mechanism, the model can self-optimize during the training process, making the extraction of key features more accurate and robust. This mechanism particularly improves the model's ability to identify and focus on spinal minimally invasive surgery targets, ensuring that the model has higher robustness and sensitivity when processing complex real-time detection images of spinal minimally invasive surgery targets, thereby significantly improving the detection accuracy and overall diagnostic effect of spinal minimally invasive surgery targets.

[0069] (5) The present invention introduces knowledge distillation technology into the target detection model. Through the multi-branch self-distillation loss mechanism, each module is guided to obtain performance similar to that of the lower-level module at the feature level, thereby enhancing the consistency of image feature information during model transmission. This method effectively reduces the impact of image quality on model performance and optimizes the false detection and missed detection problems in real-time detection images of minimally invasive spinal surgery targets.

[0070] (6) The present invention introduces the SE attention mechanism in the first, second, and third layers of the Head part, respectively, and performs fine-grained channel weighting and calibration on the output feature map. After SE processing, the feature expression ability of each channel is further enhanced, so that the model can more effectively focus on the surgical target area and suppress the interference of irrelevant background features.

[0071] (7) The present invention introduces the GIoU loss function when constructing the overall loss function, replacing the traditional rectangular box loss function, thereby providing the model with a more comprehensive overlap measurement, faster convergence speed, and better bounding box adjustment capabilities. This improvement significantly improves the robustness of the model, making it more adaptable to a variety of complex scenarios, and effectively reduces the problems of false detection and missed detection in the real-time detection images of spinal minimally invasive surgery targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The present invention is further described in detail below in conjunction with the accompanying drawings;

[0073] Figure 1 It is a flow chart of a method for real-time detection of spinal minimally invasive surgery targets based on SCI according to an embodiment of the present invention;

[0074] Figure 2 It is a schematic diagram of the preprocessing steps of the real-time detection method of the spinal minimally invasive surgery target based on SCI according to an embodiment of the present invention;

[0075] Figure 3 A schematic diagram of a model structure of a real-time detection method for a spinal minimally invasive surgery target based on SCI according to an embodiment of the present invention;

[0076] Figure 4 It is a structural block diagram of a real-time detection device for minimally invasive spinal surgery targets based on SCI according to an embodiment of the present invention. DETAILED DESCRIPTION

[0077] The present invention is further described below through specific implementation modes.

[0078] See also Figure 1 As shown, the real-time detection method of the target of minimally invasive spine surgery based on SCI is as follows:

[0079] S1, obtain the real-time grayscale image of the target of minimally invasive spine surgery, preprocess the grayscale image and annotate it to make a data set.

[0080] Preprocessing includes removing the annotation information around the real-time detection image of the spinal minimally invasive surgery target and enhancing the image contrast. The preprocessing effect is as follows: Figure 2 As shown in Figure 2, the calculation formula for enhancing image contrast is:

[0081] ;

[0082] in, is the value of the original image pixel, is the pixel value after enhancement processing. Represents a logarithmic function, and the constant C is used to make the grayscale dynamic range of the transformed image meet the requirements.

[0083] S2, builds a target detection model including Backbone part, self-distillation part, Neck part and Head part.

[0084] This embodiment uses the YOLOv5 target detection model, which has the characteristics of real-time detection, high precision, and high efficiency. It also has excellent background separation capabilities and robust detection performance, and can provide reliable target detection results in the complex environment of minimally invasive spine surgery. At the same time, the modular design of YOLOv5 is highly flexible and can adapt to a variety of application requirements, providing strong support for real-time target detection in minimally invasive spine surgery.

[0085] like Figure 3 As shown in the figure, the Backbone part includes an SCI layer, a SE layer and three alternating modules. The data set obtained by the Backbone part is processed by the SCI layer (self-calibration enhancement module) and the SE layer (using the SE attention mechanism) after low-level feature processing, and then extracted by three alternating modules. The feature representation output by each alternating module is output to the self-distillation part and the Neck part respectively.

[0086] The backbone part adopts feature extraction modules arranged from top to bottom, and the self-distillation module constructs a multi-branch distillation path to assist the model in building the loss function. The feature extraction module includes three alternating modules arranged from top to bottom. The upper alternating module at the upper layer includes C3 structure, CBS structure, and C3 structure. The middle alternating module at the middle layer is CBS structure. The lower alternating modules at the lower layer include C3 structure, CBS structure, C3 structure and fast spatial pyramid pooling structure (SPPF, Spatial Pyramid Pooling-Fast). Three sets of feature representations generated by the feature extraction module are obtained respectively, and the three sets of feature representations are input into the self-distillation module to construct a multi-branch distillation path to guide the alternating modules of the upper and middle layers to obtain performance similar to that of the alternating modules of the lower layer from the feature level, thereby improving the correlation of image feature information in model transmission, making the model more robust and reducing the impact of image quality on model performance.

[0087] The self-distillation block obtains the feature representation from the alternating module to construct the self-distillation loss; the total loss function of the target detection model is constructed based on the GIoU loss (instead of the rectangular box loss) and the self-distillation loss.

[0088] The Neck part adopts the FPN network and PANET network that are fused with each other. The Neck part obtains the feature representation from the alternating module and outputs several dimensional feature maps.

[0089] The Head part adopts the SE attention mechanism and CBS structure to obtain the dimensional feature map of the Neck part and output an image for real-time detection of spinal minimally invasive surgery targets with self-calibration illumination enhancement.

[0090] Specifically, in the Backbone part, the two CBS structures of the first unit extract low-level features from the real-time detection images of spinal minimally invasive surgery targets in the dataset, and introduce the self-calibration enhancement (SCI) module. The process formula is as follows:

[0091] First, the illumination model of the intraoperative image is estimated:

[0092] ;

[0093] in, Represents the original image at pixel position The brightness value, Represents pixels The real scene reflection characteristics at Represents pixels The light intensity at .

[0094] Secondly, the illumination components of the intraoperative images are estimated using filtering methods, such as Gaussian filtering. , the process formula is as follows:

[0095] ;

[0096] in, represents a Gaussian filter with a standard deviation of , Used to represent the original estimated lighting component, The amount of boost adjustment provided for the self-calibration enhancement module (SCI).

[0097] Next, the reflection component of the intraoperative image is separated by Divide by the estimated lighting component , separate the reflection component , the process formula is as follows:

[0098] ;

[0099] in, is a small constant used to avoid the denominator being zero and ensure the stability of the calculation.

[0100] Finally, the intraoperative images were subjected to a self-calibration illumination enhancement step; the reflectance component Enhance the reflected component With light component Recombined to get the final enhanced image , the image size is (320×320×64), the process formula is as follows:

[0101] ;

[0102] in, Refers to the image at the pixel position after self-calibration illumination enhancement The brightness value at ; is a contrast adjustment parameter used to control the degree of image contrast enhancement. Its value range is usually .

[0103] This embodiment introduces a self-calibrated illumination enhancement (SCI) module after initially extracting low-level features, which significantly improves the quality of the input image, making the model more stable and accurate when processing low-quality images in minimally invasive spine surgery. This improvement is particularly effective in complex surgical environments and helps to significantly improve the reliability of target detection.

[0104] In addition, in order to avoid excessive enhancement or edge artifacts in the intraoperative image enhancement process, the final image A boundary processing mechanism is introduced in the image processing, such as edge-preserving filtering, to further improve the image visual effect. The image size after processing is (320×320×64). The process formula is as follows:

[0105] ;

[0106] in, Indicates the final image after boundary processing at pixel position The brightness value, Represents an edge-preserving filter that reduces artifacts and preserves sharpness along image edges.

[0107] After initially completing the self-calibration illumination enhancement (SCI) processing, this embodiment further introduces an edge-preserving filtering step to perform boundary processing on the enhanced image. This subsequent processing significantly reduces the artifacts and over-enhancement problems that may occur during the image enhancement process, especially in the edge areas of the image. Through this optimization step, the quality of the final output image is further improved, making the target detection of the model in minimally invasive spinal surgery smoother and more natural, and significantly improving the detection reliability and accuracy in complex surgical environments.

[0108] Specifically, in the Backbone part, the SE attention mechanism in the SE layer is as follows:

[0109] First, the feature map of each channel is compressed into a scalar through global average pooling to obtain a global feature description. The specific calculation formula is as follows:

[0110] ;

[0111] in, represents the global description scalar of the cth channel, H and W represent the height and width of the feature map respectively, Indicates the input feature map in the cth channel pixel values.

[0112] Secondly, after obtaining the global description of each channel, the global description is processed through two fully connected layers and activation functions to generate the weight of each channel. The specific calculation formula is as follows:

[0113] ;

[0114] in, represents the recalibrated channel weight vector, and represents the weight matrix of the fully connected layer, represents the Sigmoid activation function, Represents the SiLU activation function.

[0115] Furthermore, the obtained channel weights are remapped back to the feature map of each channel, and the features of each channel are reweighted by element-by-element multiplication. The resulting image size is (320×320×64), and the specific calculation formula is as follows:

[0116] ;

[0117] in, represents the feature map of the cth channel after recalibration by the SE module, Represents the feature map of the cth channel of the input feature map.

[0118] This embodiment introduces the SE attention mechanism into the image processed by the SCI module, realizes the adaptive calibration of channel features, significantly suppresses the interference of irrelevant features, and optimizes the allocation of computing resources. This improvement effectively improves the model's feature expression ability and target detection accuracy, significantly improves its performance in real-time detection of spinal minimally invasive surgery targets, enhances the model's robustness and operating efficiency, and thus provides more reliable support for clinical diagnosis.

[0119] Specifically, in the self-distillation part, the construction of the self-distillation loss is as follows:

[0120] First, get the output of the upper alternating module ,at this time The size of is (80×80×256), and it is adjusted using a 1×1 adaptive convolution adjuster, and the calculation formula is as follows:

[0121] ;

[0122] in, represents the convolution regulator applied to the upper alternating module, Represents the convolution parameters, adjusted The size is (20×20×1024).

[0123] Secondly, get the output of the middle-level alternating module ,at this time The size of is (40×40×512), and it is adjusted using a 1×1 adaptive convolution adjuster. The calculation formula is as follows:

[0124] ;

[0125] in, represents the convolution regulator applied to the middle layer alternating module, Represents the convolution parameters, adjusted The size is (20×20×1024).

[0126] Furthermore, the output of the lower alternating module is obtained (20×20×1024), respectively calculated and and Between The norm is then used to construct the self-distillation loss, which is calculated as follows:

[0127] ;

[0128] in, express Norm.

[0129] Finally, the total self-distillation loss is calculated as:

[0130] .

[0131] The Backbone part of this embodiment introduces a self-distillation module, which aims to enhance the consistency and relevance of image feature information during model transmission. Through the self-distillation mechanism, the model can self-optimize during the training process, making the extraction of key features more accurate and robust. This mechanism particularly improves the model's ability to identify and focus on minimally invasive spinal surgery targets, ensuring that the model has higher robustness and sensitivity when processing complex real-time detection images of minimally invasive spinal surgery targets, thereby significantly improving the detection accuracy and overall diagnostic effect of minimally invasive spinal surgery targets.

[0132] This embodiment introduces knowledge distillation technology into the target detection model, and guides each module to obtain performance similar to that of the lower-level module at the feature level through a multi-branch self-distillation loss mechanism, thereby enhancing the consistency of image feature information during model transmission. This method effectively reduces the impact of image quality on model performance and optimizes the false detection and missed detection problems in real-time detection images of minimally invasive spinal surgery targets.

[0133] Specifically, in the Neck part and the Head part, the specific connection relationship is as follows:

[0134] Step S21, the lower layer alternating module is used as the input of the FPN network. The input is sequentially passed through the CBS structure and up-sampled, and then matched with the output of the middle layer alternating module for dimension splicing. The splicing result is then sequentially passed through the C3 structure and the CBS structure to complete feature fusion;

[0135] Step S22: The feature map outputted from step S21 is further upsampled and matched with the output of the upper alternating module, and then dimensionally spliced. The spliced ​​result is then feature fused through the C3 structure. The fusion result is input into the Head part, and is used as the first-layer output feature map after passing through the SE attention mechanism and the CBS structure.

[0136] Step S23: After the feature map outputted from step S22 passes through the CBS structure, it is matched with the output of step S21 and then dimensionally spliced. The spliced ​​result is then fused through the C3 structure. The fusion result is input into the Head part and passes through the SE attention mechanism and the CBS structure as the second-layer output feature map.

[0137] Step S24, the feature map outputted from step S23 is passed through the CBS structure and matched with the feature map outputted from the lower alternating module in step S21 through the output of the CBS structure, and then dimensionally spliced. The spliced ​​result is then fused through the C3 structure. The fusion result is input into the Head part, and is output as the third-layer output feature map through the SE attention mechanism and the CBS structure.

[0138] This embodiment introduces the SE attention mechanism in the first, second, and third layers of the Head part, and performs fine-grained channel weighting and calibration on the output feature map. After SE processing, the feature expression ability of each channel is further enhanced, allowing the model to more effectively focus on the surgical target area and suppress the interference of irrelevant background features.

[0139] Specifically, the total loss function of the target detection model is expressed as follows:

[0140] Overall loss of real-time detection method for minimally invasive spine surgery based on SCI , and its calculation formula is as follows:

[0141] ;

[0142] in, represents the confidence loss, represents the rectangular box loss, represents the classification loss, represents the self-distillation loss, , , , Represent weights respectively.

[0143] Furthermore, in the overall loss function GIoU loss is introduced in the calculation stage, and the original rectangular box loss The calculation formula becomes:

[0144] ;

[0145] Secondly, the GIoU loss is calculated according to the formula:

[0146] ;

[0147] Among them, A represents the predicted box, B represents the real box, and C represents the minimum closed area containing A and B. Indicates area; Represents the overlapping area of ​​two boxes (i.e., the part covered by both boxes at the same time, calculating the intersection area); Represents the union area of ​​two boxes (i.e., all parts covered by the two boxes, calculating the union area).

[0148] Constructing an overall loss of real-time detection method for spinal minimally invasive surgery based on SCI , and its calculation formula is as follows:

[0149] ;

[0150] in, represents the confidence loss, represents the GIoU loss, represents the classification loss, represents the self-distillation loss, , , , Represents weight.

[0151] This embodiment introduces the GIoU loss function when constructing the overall loss function, replacing the traditional rectangular box loss function, thereby providing the model with a more comprehensive overlap measurement, faster convergence speed, and better bounding box adjustment capabilities. This improvement significantly improves the robustness of the model, making it more adaptable to a variety of complex scenarios, and effectively reduces the problems of false detection and missed detection in the real-time detection images of spinal minimally invasive surgery targets.

[0152] S3, uses the target detection model to process the dataset and obtain images of real-time detection of targets in minimally invasive spine surgery enhanced by self-calibration illumination.

[0153] like Figure 4 As shown, the present invention also discloses a real-time detection device for spinal minimally invasive surgery targets based on SCI, comprising:

[0154] The image acquisition and preprocessing module 401 acquires a real-time grayscale image of the target of minimally invasive spine surgery, preprocesses the grayscale image and annotates it to produce a data set;

[0155] A model building module 402 is used to build a target detection model including a Backbone part, a self-distillation part, a Neck part and a Head part;

[0156] The target real-time detection image acquisition module 403 is used to process the data set using the target detection model to acquire the image of the real-time detection of the target of the spinal minimally invasive surgery enhanced by the self-calibration illumination.

[0157] Among them, the Backbone part includes an SCI layer, a SE layer and several alternating modules; the Backbone part obtains a data set, and after the data set is processed by low-level features, it is processed by the SCI layer and the SE layer in sequence, and then the features are extracted in sequence by each alternating module; the feature representation output by each alternating module is output to the self-distillation part and the Neck part respectively; the SCI layer performs self-calibrated illumination enhancement; the SE layer uses the SE attention mechanism for feature enhancement.

[0158] The Neck part obtains feature representation and outputs feature maps of several dimensions.

[0159] The Head part obtains a dimensional feature map and outputs a self-calibrated illumination-enhanced image for real-time detection of minimally invasive spinal surgery targets.

[0160] The self-distillation part obtains the feature representation from the alternating module to construct a self-distillation loss; the self-distillation loss combined with the GIoU loss constitutes the total loss function of the target detection model.

[0161] The specific implementation of the SCI-based real-time detection device for minimally invasive spine surgery targets is the same as the SCI-based real-time detection method for minimally invasive spine surgery targets, and will not be repeated in this embodiment.

[0162] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A real-time detection method for spinal minimally invasive surgery targets based on SCI, characterized in that: The steps include: S1, obtain the real-time grayscale image of the target of minimally invasive spine surgery, pre-process the grayscale image and annotate it to make a data set; S2, construct an object detection model including Backbone part, self-distillation part, Neck part and Head part; S3, uses the target detection model to process the dataset and obtain images of real-time detection of targets in minimally invasive spine surgery enhanced by self-calibration illumination; The Backbone part includes an SCI layer, an SE layer, and several alternating modules; the Backbone part obtains a data set, which is processed by the SCI layer and the SE layer in sequence after low-level feature processing, and then is processed by each alternating module in sequence for feature extraction; the feature representation output by each alternating module is output to the self-distillation part and the Neck part respectively; the SCI layer performs self-calibration illumination enhancement; the SE layer uses the SE attention mechanism for feature enhancement; The Neck part obtains feature representation and outputs feature maps of several dimensions; The Head part obtains a dimensional feature map and outputs an image of real-time detection of spinal minimally invasive surgery targets enhanced by self-calibration illumination; The self-distillation part obtains the feature representation from the alternating module to construct a self-distillation loss; the self-distillation loss combined with the GIoU loss constitutes the total loss function of the target detection model.

2. The real-time detection method for minimally invasive spine surgery targets based on SCI according to claim 1 is characterized in that: In S1, the preprocessing of the grayscale image includes: removing the annotation information around the real-time detection image of the spinal minimally invasive surgery target and enhancing the image contrast; The calculation formula for enhancing the image contrast is: ;in, Represents the value of the original image pixel, Represents the pixel value after corresponding enhancement processing, represents the logarithmic function, and C represents a constant.

3. The real-time detection method for minimally invasive spine surgery targets based on SCI according to claim 1, characterized in that: The SCI layer performs self-calibrated illumination enhancement, specifically including: First, the illumination model of the intraoperative images in the dataset is estimated as follows: ； in, Represents the original image at pixel position The brightness value of Represents pixels Real scene reflection characteristics at Represents pixels The light intensity at Secondly, the illumination components of the intraoperative images are estimated using Gaussian filtering. , the formula is as follows: ； in, represents a Gaussian filter with a standard deviation of ; represents the original estimated illumination component, represents the enhancement adjustment amount provided by the self-calibration enhancement module; Next, the reflection component of the intraoperative image is separated using the following formula: ； in, Represents a constant, used to avoid the denominator being zero; Finally, the intraoperative image is subjected to a self-calibrated illumination enhancement step: the reflectance component Enhance the reflected component With light component Recombining to get the final enhanced image, the formula is as follows: ； in, Represents the image at pixel position after self-calibration illumination enhancement The brightness value at ; Represents the contrast adjustment parameter, which is used to control the degree of image contrast enhancement. The value range is .

4. The real-time detection method for minimally invasive spine surgery targets based on SCI according to claim 3 is characterized in that: After obtaining the final enhanced image, it also includes: Boundary processing is performed; the boundary processing includes edge-preserving filtering.

5. The real-time detection method for minimally invasive spine surgery targets based on SCI according to claim 1, characterized in that: The SE attention mechanism is used for feature enhancement, as follows: First, the feature map of each channel is compressed into a scalar through global average pooling to obtain a global feature description. The specific calculation formula is as follows: ； in, Represents the global description scalar of the cth channel; H represents the height of the feature map, and W represents the width of the feature map; Indicates the input feature map in the cth channel pixel value; Secondly, after obtaining the global description scalar of each channel, the global description is processed through two fully connected layers and activation functions to generate the weight of each channel. The specific calculation formula is as follows: ； in, represents the recalibrated channel weight vector; and Represent the weight matrices of the fully connected layers respectively; Represents the Sigmoid activation function; represents the SiLU activation function; Finally, the obtained channel weights are remapped back to the feature map of each channel, and the features of each channel are reweighted by element-by-element multiplication. The specific calculation formula is as follows: ； in, represents the feature map of the cth channel after recalibration by the SE module; Represents the feature map of the cth channel of the input feature map.

6. The real-time detection method for minimally invasive spine surgery targets based on SCI according to claim 1, characterized in that: There are three alternating modules, which are an upper alternating module, a middle alternating module and a lower alternating module in the order in which the data sets flow in the model.

7. The real-time detection method for minimally invasive spine surgery targets based on SCI according to claim 6, characterized in that: The feature representation obtained from the alternating module is constructed from the distillation loss as follows: First, get the output of the upper alternating module , and the adaptive convolution adjuster is used for adjustment; the calculation formula is as follows: ； in, represents the convolution regulator applied to the upper alternating module, represents the convolution parameters; Represents the output of the upper alternating module after convolution adjustment; Secondly, get the output of the middle-level alternating module , using adaptive convolution adjuster to adjust; the calculation formula is as follows: ； in, represents the convolution regulator applied to the middle layer alternating module, represents the convolution parameters; Represents the output of the middle-layer alternating module after convolution adjustment; Finally, the output of the lower alternating module is obtained ,use Construct the self-distillation loss; the total self-distillation loss is calculated as: ； in, represents the loss from distillation; express Norm.

8. The real-time detection method for minimally invasive spine surgery targets based on SCI according to claim 7, characterized in that: The total loss function of the target detection model is expressed as: ； ； in, represents the total loss function; represents the confidence loss; represents GIoU loss; represents the classification loss; , , , They represent weights respectively; A represents the predicted box; B represents the real box; C represents the minimum closed area containing A and B; Indicates area; Represents the overlapping area of ​​two boxes; Represents the union area of ​​two boxes.

9. The real-time detection method for minimally invasive spine surgery targets based on SCI according to claim 1, characterized in that: The Head part includes several SE attention modules; each SE attention module performs feature enhancement on a dimensional feature map.

10. A real-time detection device for spinal minimally invasive surgery targets based on SCI, comprising: Image acquisition and preprocessing module, which acquires real-time grayscale images of the spinal minimally invasive surgery target, preprocesses the grayscale images and annotates them to produce a data set; Model building module, used to build the target detection model including Backbone part, self-distillation part, Neck part and Head part; A target real-time detection image acquisition module is used to process the data set using a target detection model to obtain images of real-time detection of targets in minimally invasive spine surgery enhanced by self-calibration illumination; in, The Backbone part includes an SCI layer, an SE layer and several alternating modules; the Backbone part obtains a data set, and after the data set is processed by low-level features, it is processed by the SCI layer and the SE layer in sequence, and then the features are extracted by each alternating module in sequence; the feature representation output by each alternating module is output to the self-distillation part and the Neck part respectively; the SCI layer performs self-calibration illumination enhancement; the SE layer uses the SE attention mechanism for feature enhancement; The Neck part obtains feature representation and outputs feature maps of several dimensions; The Head part obtains a dimensional feature map and outputs an image of real-time detection of spinal minimally invasive surgery targets enhanced by self-calibration illumination; The self-distillation part obtains the feature representation from the alternating module to construct the self-distillation loss; the total loss function of the target detection model is based on the GIoU loss and the self-distillation loss.

Images