Three-dimensional imaging reconstruction system in minimally invasive interventional operation based on adaptive neural network
Through adaptive neural network technology, combined with CNN and GAN, the radiation risks, limitations of image processing and insufficient real-time performance of three-dimensional imaging reconstruction systems in minimally invasive intervention are solved, and high-quality three-dimensional image reconstruction and surgical instrument position are achieved Tracking improves the safety and accuracy of the surgery.
Patent Information
- Application Number
- CN202411053263.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-02
- Publication Date
- 2024-11-26
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing three-dimensional imaging reconstruction system in minimally invasive interventional surgery has problems such as radiation risks, limitations of image processing algorithms, insufficient real-time performance and limited adaptability to dynamic changes.
Adopting an adaptive neural network based system, combining convolutional neural network CNN and generative adversarial network GAN, the training data set is constructed through the pairing and data augmentation technology of preoperative imaging data and three-dimensional models, process two-dimensional slice data in real time and generate high-quality The three-dimensional image uses a variety of sensors and computer vision technologies to track the location of surgical instruments and adjust the image transparency and color according to surgical needs.
It improves the detailed performance and accuracy of the image, enhances the real-time and safety of the surgery, reduces the risk of radiation exposure, and ensures accurate tracking of the position of the surgical instrument and adaptability to tissue deformation.
Smart Images

Figure CN119027585A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a three-dimensional imaging reconstruction system during minimally invasive interventional surgery, and in particular to a three-dimensional imaging reconstruction system during minimally invasive interventional surgery based on an adaptive neural network. Background Art
[0002] The main challenges and shortcomings of the current three-dimensional imaging reconstruction system for minimally invasive interventional surgery can be elaborated in detail by comparing and analyzing the method proposed in the Chinese invention patent CN117726672A. The patent proposes a method for real-time three-dimensional imaging of interventional equipment, including imaging the region of interest at different angles and times by a radiographic imaging device, and determining the real-time three-dimensional coordinate information of the interventional equipment based on multi-angle images and epipolar geometry constraint algorithms. Although this method has certain advantages, there are still some significant shortcomings and disadvantages in practical applications, as follows: First, the use of radiographic imaging equipment has radiation risks. During minimally invasive surgery, the increased exposure time of patients and medical teams to radiation poses health risks. This risk is particularly significant in long or multiple surgeries. Although current imaging technologies attempt to reduce risks by reducing radiation intensity or using intermittent imaging, they still cannot completely eliminate health problems caused by radiation exposure. In contrast, the use of non-invasive imaging technologies, including ultrasound imaging and magnetic resonance imaging MRI, can reduce this risk, but such technologies are not as good as radiographic imaging in terms of real-time performance and image resolution.
[0003] Secondly, image processing that relies on epipolar geometry constraint algorithms has limitations. Epipolar geometry algorithms rely on high-quality, clear multi-angle images, but in actual operations, images are blurred or distorted due to various factors including equipment movement, patient movement, and instrument occlusion. The decline in image quality directly affects the accuracy and real-time performance of three-dimensional coordinates. This dependence leads to the limitation of the stability and reliability of the algorithm in some complex surgical environments, resulting in erroneous position information, which affects the safety and success rate of surgical operations. In addition, radiographic imaging equipment has deficiencies in real-time and continuity. Although this method calculates through images at different time points, there is a time interval between the capture and processing of these images, which will have a lag effect on the rapidly changing surgical environment. The lack of real-time performance causes doctors to be unable to obtain the latest position information in time during the operation, affecting the smoothness and accuracy of the operation. The lack of continuity will also result in the inability to provide continuous visual feedback at critical moments of the operation, increasing the difficulty and risk of the operation. Finally, this method has limited ability to handle real-time dynamic changes and adapt to tissue deformation. During the operation, the tissue will deform due to reasons such as operation or patient breathing, and this deformation requires real-time adjustment and registration in the calculation of three-dimensional coordinates. Epipolar geometry algorithms and traditional imaging methods are slow to respond to these dynamic changes and cannot provide real-time adjustments and feedback, affecting the accuracy of surgery. Summary of the invention
[0004] The purpose of the present invention is to provide a three-dimensional imaging reconstruction system for minimally invasive interventional surgery based on an adaptive neural network, thereby solving some of the drawbacks and shortcomings pointed out in the background technology.
[0005] The present invention solves the above-mentioned technical problems by adopting the following technical solutions:
[0006] S1. Design an adaptive neural network structure that combines convolutional neural network (CNN) and generative adversarial network (GAN), with the network input as a two-dimensional image slice and the output as a reconstructed three-dimensional image;
[0007] S1.1. Use preoperative imaging data to pair with known 3D models to construct a training data set, and apply data enhancement technology to increase data diversity; convert 2D slice data layer by layer into 3D images through an adaptive neural network, and use the network to identify the structural relationship between layers and perform 3D reconstruction;
[0008] S1.1.1. The training data set construction process: annotate the paired two-dimensional images and three-dimensional models, record the corresponding relationship, use a combination of intelligent annotation tools and manual verification, and convert the paired data into the format required for neural network training, and control the network to learn the mapping relationship from two-dimensional images to three-dimensional models;
[0009] S1.1.2. Apply image transformation technology to enhance two-dimensional image data, including rotation, scaling, translation, mirror flipping and gamma correction, to increase data diversity and prevent neural network overfitting;
[0010] S1.1.3. Generate new 2D image data using generative adversarial networks (GANs) and combine them with existing data to expand the size and diversity of datasets.
[0011] S2. During the operation, the 3D preview image of the contour is quickly generated through downsampling technology, and the high-resolution image display is restored after the interactive operation is stopped; the position and movement trajectory of the surgical instrument are tracked in real time using a variety of sensors and computer vision technology, and the 3D point cloud data of the surgical instrument is rigidly and non-rigidly registered with the 3D model reconstructed before the operation;
[0012] S3. According to the surgical requirements, adjust the transparency and color of different tissues in the reconstructed image; use the Dijkstra algorithm to calculate the shortest path between two points in the surgical area to mark the surgical area, and apply the region growing algorithm and threshold segmentation technology to control the segmentation and presentation of the surgical area.
[0013] Furthermore, the adaptive neural network construction includes: a convolutional neural network CNN is used to extract features from two-dimensional slices, and a generative adversarial network GAN is used to generate three-dimensional images, by defining a loss function:
[0014]
[0015] R represents reconstruction loss, D represents adversarial loss, f and g are mapping functions of CNN and GAN respectively, α and β are coefficients to adjust the influence of the two losses, and x i is a 2D slice of the input, y i and z j is the desired output three-dimensional image;
[0016] The slice sequence is processed by recurrent neural network RNN or long short-term memory network LSTM, and the sequence processing formula is defined:
[0017]
[0018] where s i represents the i-th slice, h is the recursive function of RNN or LSTM, γ i is the weighting coefficient, reflecting the importance of each slice, and λ is the regularization parameter;
[0019] The processed features are mapped to three-dimensional space, and high-quality three-dimensional images are generated layer by layer. The feature comparison mechanism inside the network is used to identify the structural relationship between layers, and the structural relationship identification formula is defined:
[0020]
[0021] Where k is the kernel function, describing the similarity between two consecutive layers, x is the feature of the layer, represents the feature gradient, dx is the inter-layer difference, a and b are the upper and lower limits of the integral, representing the beginning and end of the entire sequence.
[0022] Furthermore, the training data set construction process is as follows:
[0023] S1. Use image registration technology to pair the 2D image with the corresponding 3D model, and perform geometric adjustment through the spatial transformation network STN to define the image registration function:
[0024] P(x,y)=∫φ(x,y,z)dz
[0025] Where φ is the mapping function from two-dimensional to three-dimensional space, x, y are the two-dimensional plane coordinates, and z is the depth;
[0026] S2. Standardize and normalize the paired images, apply advanced filtering and enhancement techniques to optimize image quality, and define data preprocessing functions:
[0027]
[0028] Where k is the slope of enhancement and μ is the average brightness;
[0029] S3. Develop intelligent annotation tools to identify and annotate key structures and feature points in 2D images, and enable medical experts to perform manual verification and correction through the interface, and define annotation verification functions:
[0030]
[0031] Where L and M represent the annotated and manually corrected label sets, respectively, and λ n is the weight of different labels, and p is the regularization exponent;
[0032] S4. Convert the labeled and verified data into the format required for neural network training, and design a loss function to ensure that the neural network learns the mapping relationship from two-dimensional images to three-dimensional models. The loss function is defined as:
[0033]
[0034] f and g are the forward and reverse neural network functions for reconstruction, respectively. i ,y i are the training input and output, z j ,w j is the reverse verification input and output, ω i and σ j is the weight coefficient, α and β are the coefficients that adjust the contribution of the two losses.
[0035] Furthermore, the image transformation technology includes the following steps: firstly, the image registration technology is used, and the spatial transformation network STN is combined to rotate, scale and translate according to the content and geometric properties of the two-dimensional image, and the image registration formula is defined:
[0036]
[0037] Where I(u,v) is the original two-dimensional image, θ is the rotation angle, and δ is the Diracdelta function;
[0038] Gamma correction is then performed to adapt to lighting changes during surgery, using the adaptive gamma correction formula:
[0039] G(I;γ)=I γ
[0040] Where I is the input image intensity, and γ is the gamma value adjusted according to the image brightness;
[0041] Diversity enhancement is performed on the data after image transformation and gamma correction, and the diversity enhancement formula is applied:
[0042] S(x;α,β)=αx+βx 2
[0043] Where x is the transformed image data, α and β are dynamically adjusted parameters;
[0044] Finally, the enhanced dataset is used to train the adaptive neural network to learn the mapping relationship from two-dimensional images to three-dimensional models, and the loss function is defined as:
[0045]
[0046] Where f represents the mapping function of the neural network, x i and i are the training input and expected output respectively, w1 and w2 are the weight coefficients.
[0047] Furthermore, the generative adversarial network process is as follows:
[0048] A custom generative adversarial network (GAN) is used to design and define adversarial loss functions to optimize the generator and discriminator:
[0049]
[0050] The generator G(z;θ G ) uses a random noise vector z to generate a visually realistic two-dimensional image, and the discriminator D(x; θ D ) Determine the authenticity of the image;
[0051] The generator runs in real time during the surgery, dynamically generates or adjusts the image data according to the real-time situation of the surgery, and applies a composite transformation function:
[0052] T(x;α,β,γ)=αx+βx 2 +γcos(x)
[0053] Where x is the original image data, α, β, and γ are parameters adjusted according to real-time surgical needs to enhance the contrast and detail of the image;
[0054] Merge the generated new image data with the existing real image data and pass the quality assessment function:
[0055]
[0056] Where f(x) represents the intensity function of the image, is the Laplacian operator, which evaluates the consistency of the quality of the generated image with the real image;
[0057] Finally, the enhanced dataset is used to train an adaptive neural network to learn the mapping relationship from two-dimensional images to three-dimensional models, and the loss function is defined as:
[0058]
[0059] Where f represents the mapping function of the neural network, x i and i are the training input and expected output respectively, w1 and w2 are weight coefficients, and a and b define the range of integration.
[0060] Furthermore, the downsampling technology is implemented as follows:
[0061] S1, using dynamic resolution adjustment technology, during the operation, the spatial resolution of the image is reduced to generate a 3D preview image according to the current operation stage and the doctor's needs;
[0062] S2. Integrate multiple sensors to capture the position and motion trajectory of surgical instruments in real time. Use computer vision and machine learning technology to perform rigid and non-rigid registration of the three-dimensional point cloud data of surgical instruments captured in real time with the three-dimensional model reconstructed before surgery to adapt to the movement of instruments and deformation of tissues.
[0063] Furthermore, the method for implementing the dynamic resolution adjustment technology includes: firstly, using an intelligent resolution adjustment algorithm to adjust the image resolution by real-time analysis of the surgical stage and the doctor's needs, so as to achieve dynamic conversion between the three-dimensional preview and the high-resolution image, and using the algorithm:
[0064]
[0065] x is the real-time surgical data obtained from the sensor, α and β are parameters adjusted according to the surgical stage to control the downsampling and restoration of the image;
[0066] Integrate GPU accelerated image rendering technology, optimize image rendering process, and apply rendering functions:
[0067]
[0068] Where x is the image data to be processed, γ and δ are optimization parameters that adjust the speed and quality of image rendering;
[0069] The generated new image data is merged with the existing real image data, and the image quality is controlled by the quality assessment function:
[0070]
[0071] where f is the synthetic image function, λ is the evaluation parameter, and a and b define the range of the integration.
[0072] Furthermore, the computer vision and machine learning technology construction method is as follows:
[0073] S1. Integrate multiple sensors, including optical, magnetic and inertial sensors, to capture the position and motion trajectory of surgical instruments in real time, and use data fusion algorithms to:
[0074] F(p i ,p i-1 ; α, β) = αp i +(1-α)βp i-1
[0075] where p i is the currently captured position data, p i-1 is the previous position data, α and β are coefficients for adjusting the real-time response and the smoothness of historical data;
[0076] S2. Apply the deep learning model to process the captured 3D point cloud data of surgical instruments in real time, and perform rigid and non-rigid registration with the preoperatively reconstructed 3D model, using the registration function:
[0077]
[0078] Where x, y are points in the point cloud data, and θ, φ are transformation parameters.
[0079] Furthermore, the method for realizing transparency and color of different tissues in the reconstructed image includes:
[0080] Develop a dynamic visual enhancement module, use an adaptive neural network to adjust the transparency and color of different tissues in the 3D reconstructed image according to the surgical stage and specific needs, and apply the transparency adjustment function:
[0081]
[0082] c represents the current color intensity of the tissue, ν represents the color target value adjusted according to surgical requirements, and k is the adjustment coefficient;
[0083] The Dijkstra algorithm, region growing algorithm and threshold segmentation technology are combined to perform accurate segmentation and path planning of the surgical area, where the shortest path calculation uses the path planning function:
[0084]
[0085] v i and v j is the image pixel represented by the vertex in the graph, E is the edge set between the vertices, and w ijis the weight of the edge, indicating the distance between pixels;
[0086] At the same time, the control of region growth is performed through the partition function:
[0087] S(x;θ)=1 / (1+e -θ(x-μ) )
[0088] x is the grayscale value of the image, and θ and μ are parameters for adjusting the segmentation sensitivity.
[0089] Beneficial effects of the present invention:
[0090] First, through the adaptive neural network, the system processes the two-dimensional slice data in real time and reconstructs high-quality three-dimensional images. The adaptive learning ability of the neural network enables it to accurately identify and reconstruct complex anatomical structures, improve the detail and accuracy of the image, and help doctors locate and operate more accurately during surgery, reducing errors.
[0091] The system then uses image processing algorithms, including convolutional neural networks (CNNs) and generative adversarial networks (GANs), to improve the resolution and contrast of the image. By dynamically adjusting the transparency and color of the image, the system can optimize the visual effect in real time according to the surgical stage and specific needs, allowing doctors to observe different tissues in the surgical area more clearly.
[0092] In addition, the system has powerful real-time processing capabilities, which can capture and process the 3D point cloud data of surgical instruments in real time during the operation, and perform rigid and non-rigid registration with the preoperative reconstructed 3D model. This real-time performance ensures accurate tracking of the position and motion trajectory of surgical instruments, and adapts to the deformation of tissues and movement of instruments during the operation.
[0093] Finally, compared with traditional three-dimensional imaging methods based on radiography, this system can integrate non-invasive imaging technologies, including ultrasound imaging and magnetic resonance imaging (MRI), to reduce the risk of radiation exposure to patients and medical teams during surgery and improve safety. BRIEF DESCRIPTION OF THE DRAWINGS
[0094] Figure 1 This is a flow chart of the three-dimensional imaging reconstruction system during minimally invasive intervention based on an adaptive neural network of the present invention.
[0095] Figure 2 This is a diagram of the adaptive neural network building module of the present invention.
[0096] Figure 3 A module diagram is constructed for the training data set of the present invention.
[0097] Figure 4 This is a flow chart for implementing the downsampling technology of the present invention. DETAILED DESCRIPTION
[0098] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.
[0099] View attached Figure 1 The adaptive neural network-based intraoperative 3D imaging reconstruction system for minimally invasive interventional surgery adopts a structure that integrates convolutional neural networks (CNNs) and generative adversarial networks (GANs), with the goal of generating 3D images from 2D image slices in real time. First, preoperative 2D images are paired with known 3D models to establish a training dataset. The dataset increases the diversity of the data by applying data enhancement techniques such as rotation, scaling, translation, mirror flipping, and gamma correction. In addition, a generative adversarial network (GAN) is used to generate new 2D image data, which is combined with existing data to further expand the scale and diversity of the dataset.
[0100] The construction of the training data set involves annotating paired 2D images and 3D models and recording the correspondence between them. Intelligent annotation tools are used with manual verification to ensure data accuracy and convert the data into a format suitable for neural network training. In actual surgery, the system can quickly generate a 3D preview image through downsampling technology. When the surgical operation stops interactive operations, the system will restore the image to a high-resolution display for detailed observation by the surgical team. At the same time, the system integrates a variety of sensors and computer vision technologies to track the position and movement trajectory of surgical instruments in real time, and rigidly and non-rigidly align the 3D point cloud data of surgical instruments with the preoperatively reconstructed 3D model to adapt to the movement of instruments and deformation of tissues.
[0101] In addition, the system can adjust the transparency and color of different tissues in the reconstructed image according to surgical needs, use the Dijkstra algorithm to calculate the shortest path between two points in the surgical area, and accurately control the segmentation and presentation of the surgical area through the region growth algorithm and threshold segmentation technology, thereby improving the accuracy and efficiency of the surgery. The integration of these technologies not only improves the accuracy of surgical operations, but also enhances the dynamic adaptability and real-time response capabilities of the system, making the surgical process safer and more effective.
[0102] Embodiment 1:
[0103] View attached Figure 2 , Embodiment: To perform spinal surgery, it is necessary to understand the three-dimensional structure of the spine to avoid damaging key nerves. Before the operation, the patient underwent a multi-angle CT scan to generate multiple two-dimensional image slices. The image slices were input into the adaptive neural network system to generate a three-dimensional spine model.
[0104] Convolutional neural network (CNN) is used to extract features from the input two-dimensional slices, including but not limited to edges, textures, and other anatomical markers; Generative adversarial network (GAN) is used to generate three-dimensional images from the features extracted by CNN. The generator creates three-dimensional images, and the discriminator distinguishes the generated images from the real preoperative three-dimensional model. The loss function is defined as:
[0105]
[0106] in:
[0107] R represents the reconstruction loss, which quantifies the difference between the generated 3D image and the true model;
[0108] D represents adversarial loss, which evaluates the discriminator’s ability to distinguish generated images;
[0109] f and g are the mapping functions of CNN and GAN respectively;
[0110] α and β are coefficients for adjusting the effects of the two losses, with values ranging from α = 0.5 to 1.0 and β = 0.1 to 0.5;
[0111] x i is a 2D slice of the input, y i and z j is the desired output 3D image.
[0112] Assuming that there are n = 100 2D slices, the average reconstruction loss R is 0.02, the average adversarial loss D is 0.05, α = 0.8 and β = 0.3, the loss function can be calculated as:
[0113]
[0114] L=-312.96+0.3·0.5·10
[0115] L=-312.96+1.5
[0116] L=-311.46
[0117] The computational results show that the neural network is reducing the reconstruction and adversarial losses during the training process, which is achieved by quantitatively adjusting the α and β coefficients to optimize the loss function.
[0118] The embodiment processes a sequence of continuous two-dimensional slices through a recurrent neural network RNN, and each two-dimensional slice is regarded as an element s in the sequence i , where i represents the position in the sequence. These slices are processed in turn by the recursive function h:
[0119]
[0120] s irepresents the i-th slice;
[0121] h is the RNN recursive function, which is responsible for processing the slices and outputting the relevant three-dimensional data;
[0122] γ i is the weighting factor, which can be adjusted according to the clinical importance of each slice and ranges from 0.1 to 1.0;
[0123] λ is a regularization parameter used to adjust the model's sensitivity to changes in the slice sequence, and its value ranges from 0.01 to 0.1.
[0124] Assume that there are n=50 slices, each slice is processed by RNN, and γ i =0.5 (same for all slices), λ=0.05.
[0125] Set the recursive function h to be i ,s i-1 )’s output difference is 2, that is, h(s i ,s i-1 )=2.
[0126] Settings i+1 -s i =1 for all i.
[0127] Calculate the value of S:
[0128]
[0129] set up is approximately 1 (i.e., sensitive to small changes), then:
[0130] S=50+0.05·49·1
[0131] S = 50 + 2.45 = 52.45
[0132] Then, the embodiment maps the processed two-dimensional slice features to three-dimensional space to generate high-quality three-dimensional images layer by layer, and uses the internal network mechanism to identify the structural relationship between different layers. The identification formula is as follows:
[0133]
[0134] Parameter settings:
[0135] k is the kernel function, which evaluates the similarity between two consecutive layers;
[0136] x represents the features of the image layer, the high-dimensional features extracted by CNN;
[0137] Represents the gradient of the feature, which is used to capture the spatial changes of the feature;
[0138] dx is the difference between adjacent layers, which appears as a small change between layers;
[0139] a and b define the range of the integration, representing the beginning and end of the entire image sequence.
[0140] By processing a series of spine CT slices from bottom to top, each slice is mapped into specific three-dimensional data. In this case, the feature x includes edges, textures and other morphological features, and the gradient of these features The inter-layer difference dx reflects the slight morphological changes between adjacent vertebrae.
[0141] Assume that there are 50 slices and evaluate the structural relationship between the 25th and 26th layers. Set the upper and lower limits of the integral a=25, b=26, and set The specific expression is Such a kernel function expresses the exponential decay of the similarity between two layers as the features change.
[0142] set up (indicates a slight change), then the kernel function The structural relationship identification formula becomes:
[0143]
[0144] The results show that there is a high degree of similarity between layers 25 and 26, indicating that the reconstruction system is effectively identifying and maintaining the continuity of the spine.
[0145] Embodiment 2:
[0146] View attached Figure 3 In the embodiment, a training data set is implemented in a 3D imaging reconstruction system for minimally invasive intervention based on an adaptive neural network. In the preparation stage of surgery, the patient's 2D image (CT scan) is paired with the corresponding 3D model (from a previous MRI scan or a standard model generated by specialized software). The spatial transformation network STN is used to perform geometric adjustment of the image, and the registration function is performed through the function:
[0147] P(x,y)=∫φ(x,y,z)dz
[0148] Where φ represents the mapping function from two-dimensional to three-dimensional space, x, y are the two-dimensional plane coordinates, and z is the depth.
[0149] When processing a set of two-dimensional images, STN determines the rotation angle applied to be 15° and the translation 10mm upward, standardizes and normalizes the paired images, and uses advanced filtering and enhancement techniques to optimize image quality. Through the data preprocessing function:
[0150]
[0151] where k is the slope of the enhancement and μ is the average brightness.
[0152] Set the average brightness μ of an image to 100, enhance the contrast of the image, and select slope k = 0.1. For a pixel in the image, its original grayscale value is 120, and its new grayscale value is calculated using the preprocessing function:
[0153]
[0154] This indicates that the grayscale value of the pixel after processing is improved, thereby enhancing the contrast of the image and making the details clearer.
[0155] The embodiment further develops an intelligent annotation tool, uses machine learning technology to identify key structures and feature points in two-dimensional images, including identifying different vertebrae and key nerve pathways in spinal surgery, and then provides them to medical experts through a user interface for manual verification and necessary corrections to ensure the accuracy and completeness of the annotations. The annotation verification function:
[0156]
[0157] L and M represent the label sets after automatic annotation and manual correction, respectively;
[0158] λ n is the weight of different labels, ranging from 0.1 to 1.0, and is adjusted according to the importance of the label in the surgery;
[0159] p is the regularization exponent, with a common value of 2, which is used to amplify the impact of labeling errors.
[0160] Assume that there are N = 10 key structures to be annotated, and the weight of each structure is λ n =0.5, the difference between automatic annotation and manual correction is |L n -M n |=0.05:
[0161]
[0162] This value indicates that the annotation error is very small, which means that the annotation tool performs well and the corrections made by the medical experts are effective.
[0163] Then, data conversion and neural network training are performed to convert the labeled and verified data into a format suitable for neural network training, including feature extraction and label encoding.
[0164] Neural network loss function:
[0165]
[0166] α and β adjust the coefficients of the two loss contributions, with values of α=1 and β=0.5.
[0167] ω i and σ j is the weight coefficient, which takes a value of 1, indicating that each data point has equal importance in the loss function.
[0168] Assuming n = 100 and m = 50, the average error of the training data |f(x i )-y i |=0.03 and |g(z j )-w j |=0.02:
[0169]
[0170] L=1×100×0.0009+0.5×50×0.0004
[0171] L = 0.09 + 0.01 = 0.1
[0172] This low loss value indicates that the neural network effectively learns the mapping relationship from 2D images to 3D models, demonstrating the feasibility and effectiveness of the training scheme.
[0173] Embodiment 3:
[0174] Embodiment Spinal surgery requires three-dimensional images to navigate surgical instruments. The patient's preoperative CT scan provides a series of two-dimensional slices that need to be processed by image transformation technology to generate a three-dimensional model. Before three-dimensional reconstruction, all two-dimensional slices need to be registered to ensure their consistency in space. Image registration includes rotation, scaling, and translation to correct image deviations caused by different shooting angles and positions. Use the spatial transformation network STN to adjust according to the content and geometric properties of the two-dimensional image, and apply appropriate transformations by learning the features of the image. The specific formula is as follows:
[0175]
[0176] Where I(u,v) is the original two-dimensional image, θ is the rotation angle, and δ is the Diracdelta function.
[0177] Set the pixel value of the two-dimensional image I(u,v) to 150 at u=10,v=20. If the rotation angle θ=30°, the position of the point in the new image calculated by STN is:
[0178] x=10cos 30°-20sin 30°≈10×0.866-20×0.5=8.66-10=-1.34
[0179] y=10sin 30°+20cos 30°≈10×0.5+20×0.866=5+17.32=22.32
[0180] According to the new coordinates, the pixel value of the point is mapped to the new position (-1.34, 22.32), realizing the rotation transformation of the image.
[0181] In order to adapt to the lighting changes during surgery, it is necessary to adjust the brightness and contrast of the image, use gamma correction, and enhance the visual effect of the image through nonlinear transformation. The formula is as follows:
[0182] G(I;γ)=I γ
[0183] I is the input image intensity and γ is the gamma value.
[0184] Set the input intensity of the image to I = 120, select the gamma value γ = 0.8 according to the current lighting conditions, and calculate the intensity after gamma correction adjustment as:
[0185] G(120;0.8)=120 0.8 ≈79.43
[0186] This adjustment makes the image brightness more in line with the actual needs of the surgical environment, ensuring that image details are clearly visible during surgery.
[0187] The embodiment further performs diversity enhancement on the data after image transformation and gamma correction to improve the generalization ability and accuracy of the neural network, and uses a diversity enhancement formula to dynamically adjust parameters α and β to maintain the overall characteristics of the image while enhancing image details.
[0188] S(x;α,β)=αx+βx 2
[0189] x is the transformed image data;
[0190] α and β are dynamically adjusted parameters, usually in the range of α∈[0.5,1.5] and β∈[0.01,0.1].
[0191] Set the image data point x = 100, select parameters α = 1.2 and β = 0.05 for diversity enhancement:
[0192] S(100;1.2,0.05)=1.2×100+0.05×100 2
[0193] S(100;1.2,0.05)=120+0.05×10000
[0194] S(100;1.2,0.05)=120+500=620
[0195] After diversity enhancement, the value of this data point becomes 620, which significantly enhances the diversity and richness of the data.
[0196] Then define the loss function of the neural network and improve the prediction accuracy of the model by optimizing the loss function.
[0197]
[0198] f represents the mapping function of the neural network;
[0199] x i and i are the training input and expected output respectively;
[0200] w1 and w2 are weight coefficients, usually in the range of w1∈[0.5,1.0] and w2∈[0.01,0.1].
[0201] Assume that there are n=100 data points in the training set and the training input x i =620, expected output y i =600, when calculating the loss at this point, we take w1=0.8 and w2=0.05, and set
[0202]
[0203]
[0204] L=1610
[0205] The calculation results show how to optimize the model performance by adjusting the weights and loss functions during the training process, so that the neural network can learn the characteristics of the three-dimensional model from the enhanced two-dimensional image data.
[0206] Embodiment 4:
[0207] In the three-dimensional imaging reconstruction system for minimally invasive intervention based on adaptive neural network, a generative adversarial network GAN is used to generate and adjust two-dimensional image data, including a generator G and a discriminator D. Generator G(z; θ G ) uses a random noise vector z to generate a visually realistic two-dimensional image, while the discriminator D(x; θ D ) to determine the authenticity of the image.
[0208] Define the adversarial loss function used to optimize the generator and discriminator:
[0209]
[0210] Generator G(z;θ G ) Generate an image deception discriminator;
[0211] Discriminator D(x;θ D ) to distinguish between real images and generated images.
[0212] The probability of the discriminator D outputting a real image in the training iteration is set to 0.9, and the probability of the generated image is set to 0.1. The input range of the random noise z is between -1 and 1, and the image generated by the generator is 128x128 pixels.
[0213]
[0214] The generator runs in real time during the operation, dynamically generating or adjusting image data according to the real-time surgical situation.
[0215] A composite transformation function is then applied to enhance the contrast and detail of the image:
[0216] T(x;α,β,γ)=αx+βx 2 +γcos(x)
[0217] x is the original image data;
[0218] α, β, and γ are parameters adjusted according to real-time surgical requirements, and their value ranges are α∈[0.8,1.2], β∈[0.01,0.1], and γ∈[0.5,1.5].
[0219] Set the original image data x = 120, and select parameters α = 1.1, β = 0.05, γ = 1.0 according to real-time surgical requirements:
[0220] T(120;1.1,0.05,1.0)=1.1×120+0.05×120 2 +1.0cos(120)
[0221] T(120;1.1,0.05,1.0)=132+0.05×14400+1.0cos(120°)
[0222] T(120;1.1,0.05,1.0)=132+720+1.0×(-0.5)
[0223] T(120;1.1,0.05,1.0)=132+720-0.5
[0224] T(120;1.1,0.05,1.0)=851.5
[0225] The next step of the embodiment is to evaluate the consistency between the quality of the generated image and the real image through the quality evaluation function:
[0226]
[0227] f(x) represents the intensity function of the image; is the Laplacian operator, used to evaluate the edges and details of an image; is a Gaussian weight that ensures smoothing of edges and details.
[0228] Assume that the intensity function f(x) of an image data is approximately x 2 , through the Laplace operator Calculate the quality assessment function:
[0229]
[0230] because is a Gaussian function, and the integral result is:
[0231]
[0232] This shows that the quality of the image data is high and has good consistency with the real data.
[0233] Then, we optimize the neural network by defining the loss function to achieve the mapping relationship from two-dimensional images to three-dimensional models:
[0234]
[0235] f represents the mapping function of the neural network; x i and i are the training input and expected output respectively; w1 and w2 are weight coefficients, and their value ranges are w1∈[0.5,1.0] and w2∈[0.01,0.1]; a and b define the range of integration.
[0236] Assume that there are n=100 data points, and the input x of a data point is i =620, expected output y i =600, w1=0.8 and w2=0.05 are taken when calculating the loss, and In the interval 0,1:
[0237]
[0238]
[0239] L=32020
[0240] This loss value shows how to optimize the model performance by adjusting the weights and loss functions during the training process, so that the neural network can learn the characteristics of the three-dimensional model from the enhanced two-dimensional image data.
[0241] Embodiment 5:
[0242] View attached Figure 4 In the spinal surgery, it is necessary to view the three-dimensional image of the spine in real time. Through the intelligent resolution adjustment algorithm, a three-dimensional preview image is generated during the surgery to restore the high-resolution image. The implementation of the dynamic resolution adjustment technology relies on the following algorithms:
[0243]
[0244] x is the real-time surgical data obtained from the sensor;
[0245] α and β are parameters adjusted according to the surgical stage, and their value ranges are α∈[0.5,1.5] and β∈[0.01,0.1].
[0246] Assume that during the operation, the real-time surgical data x obtained by the sensor is 10. According to the current surgical stage, select parameters α = 1.2 and β = 0.05 for calculation:
[0247]
[0248] R(10;1.2,0.05)=1.2e -5
[0249] R(10;1.2,0.05)≈1.2×0.0067
[0250] R(10;1.2,0.05)≈0.00804
[0251] The calculation results show that in the current stage of surgery, by adjusting the resolution, the image data is significantly downsampled, thereby speeding up the image processing speed and providing a real-time three-dimensional preview.
[0252] The embodiment further applies the rendering function to optimize the image rendering process, and controls the rendering speed and quality by adjusting the parameters γ and δ:
[0253]
[0254] x is the image data to be processed;
[0255] γ and δ are optimization parameters, and their value ranges are γ∈0.5,2.0 and δ∈[0.01,0.1].
[0256] Set the image data to be processed x=400, and select parameters γ=1.5 and δ=0.05 for rendering calculation according to the current surgical requirements:
[0257]
[0258] G(400;1.5,0.05)=1.5×4.472
[0259] G(400;1.5,0.05)≈6.708
[0260] The last embodiment combines the generated new image data with the existing real image data, and performs image quality control through a quality assessment function, using the quality assessment function:
[0261]
[0262] f is the synthetic image function;
[0263] λ is the evaluation parameter, and its value range is λ∈[0.1,1.0];
[0264] a and b define the range of the integration, which are the start and end points of the image.
[0265] Set the synthetic image function f(x) = x 2 , the evaluation parameter λ = 0.5, the integration range is a = 0 and b = 10, and the quality evaluation function is calculated:
[0266]
[0267] because
[0268]
[0269] Q(f;λ)=2×(10-0)
[0270] Q(f;λ)=20
[0271] The calculation results show that the image quality parameter after evaluation by the quality assessment function is 20, indicating that the quality of the synthetic image is high and has good consistency with the real data.
[0272] Embodiment 6:
[0273] In the embodiment, multiple sensors are integrated in spinal surgery to capture the position and motion trajectory of surgical instruments in real time, and these data are registered with the preoperative reconstructed 3D model through a data fusion algorithm:
[0274] F(p i ,p i-1 ; α, β) = αp i +(1-α)βp i-1
[0275] p i is the currently captured position data;
[0276] p i-1is the previous location data;
[0277] α and β are coefficients for adjusting the real-time response and the smoothness of historical data, and their value ranges are α∈[0.5,1.0] and β∈[0.1,0.5].
[0278] Set the current captured position data p i =100, previous position data p i-1 =90, select parameters α=0.8 and β=0.3 for calculation:
[0279] F(100,90;0.8,0.3)=0.8×100+(1-0.8)×0.3×90
[0280] F(100,90;0.8,0.3)=80+0.2×27
[0281] F(100,90;0.8,0.3)=80+5.4
[0282] F(100,90;0.8,0.3)=85.4
[0283] Through the data fusion algorithm, the current location information is smoothly combined with the historical data, which improves the stability and accuracy of the location data.
[0284] The embodiment further applies a deep learning model to process the captured three-dimensional point cloud data of surgical instruments in real time, and performs rigid and non-rigid registration with the preoperatively reconstructed three-dimensional model. The registration function is defined as follows:
[0285]
[0286] x, y are points in the point cloud data; θ, φ are transformation parameters, usually in the range of θ, φ∈[0,10].
[0287] Assume that the 3D point cloud data captured by the system contains N = 5 points, and the point coordinates are (x1, y1) = (2, 3), (x2, y2) = (3, 4), (x3, y3) = (5, 6), (x4, y4) = (7, 8), (x5, y5) = (9, 10). Select the transformation parameters θ = 4 and φ = 5 for the registration calculation:
[0288]
[0289]
[0290] R(x,y;4,5)=e -4-4 +e -1-1 +e -1-1 +e -9-9 +e -25-25
[0291] R(x,y;4,5)=e -8 +e -2 +e -2 +e -18 +e -50
[0292] R(x,y;4,5)≈3.354626×10 -4 +0.135335+0.135335+1.522997×10 -8
[0293] +1.928749×10 -22
[0294] R(x,y;4,5)≈0.2707037
[0295] The calculation results show that after the registration function calculation, the point cloud data is accurately registered with the preoperative three-dimensional model, and the obtained registration value R provides a measure of the consistency between the point cloud data and the model.
[0296] Embodiment 7:
[0297] In the three-dimensional imaging reconstruction system for minimally invasive intervention based on adaptive neural network, the transparency and color of different tissues in the reconstructed image are adjusted. By developing a dynamic visual enhancement module, the transparency and color of different tissues in the three-dimensional reconstructed image are adjusted according to the surgical stage and specific needs to ensure that the doctor obtains the best visual information support. The transparency adjustment function is used to adjust the transparency and color of the tissue:
[0298]
[0299] c represents the current color intensity of the tissue;
[0300] ν represents the color target value adjusted according to surgical needs;
[0301] k is the adjustment coefficient, and its value range is k∈[0.1,2.0].
[0302] Assume that during the operation, the current color intensity of the tissue is c=120, the color target value ν=100 adjusted according to the operation requirements, and the adjustment coefficient k=1.5 is selected for calculation:
[0303]
[0304] Due to e -30 Very close to 0,
[0305]
[0306] The calculation results show that the transparency and color adjustment of the current tissue are close to their target values, ensuring the clear display of the tissue in the image.
[0307] The embodiment then uses the Dijkstra algorithm to perform the shortest path calculation to determine the optimal path for the surgical instrument from the entrance to the target area. The path planning function is defined as follows:
[0308]
[0309] v i and v j It is the vertex in the graph, representing the image pixel;
[0310] E is the set of edges between vertices;
[0311] w ij is the weight of the edge, indicating the distance between pixels, and its value range is w ij ∈[1,10].
[0312] Combined with the threshold segmentation technology, starting from the starting point, the region is gradually increased according to the similarity of pixel values and adjacent pixels to segment the surgical area. Suppose there are several pixels in the image, representing different parts of the spine, and the shortest path from the entry point v1 to the target point v4 needs to be calculated. The edge weight between vertices: w 12 =2; w 13 =6; w 24 =3; w 34 =1.
[0313] Construct the graph and apply Dijkstra's algorithm to calculate the shortest path:
[0314]
[0315] P(v1,v4)=min(2+3,6+1)
[0316] P(v1,v4)=min(5,7)
[0317] The shortest path is v1→v2→v4, and the total distance is 5.
[0318] From the starting point v2 obtained by path planning, the region growing algorithm is applied to segment the surgical area.
[0319] Set the pixel value threshold T = 50, and perform segmentation based on the similarity of adjacent pixels. Set the starting point pixel value to 100, and the adjacent pixel values to 95 and 45 respectively, and add the pixels that meet the conditions to the segmentation area. Then continue to check the adjacent pixels and repeat the above steps until the surgical area segmentation is completed.
[0320] The final example uses a partitioning function to control the region growing process, defined as follows:
[0321]
[0322] x is the grayscale value of the image;
[0323] θ and μ are parameters for adjusting the segmentation sensitivity, and their value ranges are θ∈[0.1,2.0] and μ∈[0,255].
[0324] Assume that during the operation, the grayscale value of a certain image pixel is x=150, the target segmentation grayscale value μ=100, and the adjustment coefficient θ=1.0 is selected for calculation:
[0325]
[0326] Due to e -50 Very close to 0,
[0327]
[0328] The calculation results show that the pixels with a grayscale value of 150 are effectively segmented, ensuring the accurate segmentation of the surgical area.
[0329] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.
Claims
1. A three-dimensional imaging reconstruction system for minimally invasive interventional surgery based on an adaptive neural network, characterized by The following steps are involved: S1. Design an adaptive neural network structure that combines convolutional neural network (CNN) and generative adversarial network (GAN), with the network input as a two-dimensional image slice and the output as a reconstructed three-dimensional image; S1.
1. Use preoperative imaging data to pair with known 3D models to construct a training data set, and apply data enhancement technology to increase data diversity; convert 2D slice data layer by layer into 3D images through an adaptive neural network, and use the network to identify the structural relationship between layers and perform 3D reconstruction; S1.1.
1. The training data set construction process: annotate the paired two-dimensional images and three-dimensional models, record the corresponding relationship, use a combination of intelligent annotation tools and manual verification, and convert the paired data into the format required for neural network training, and control the network to learn the mapping relationship from two-dimensional images to three-dimensional models; S1.1.
2. Apply image transformation technology to enhance two-dimensional image data, including rotation, scaling, translation, mirror flipping and gamma correction, to increase data diversity and prevent neural network overfitting; S1.1.
3. Generate new 2D image data using generative adversarial networks (GANs) and combine them with existing data to expand the size and diversity of datasets. S2. During the operation, the 3D preview image of the contour is quickly generated through downsampling technology, and the high-resolution image display is restored after the interactive operation is stopped; the position and movement trajectory of the surgical instrument are tracked in real time using a variety of sensors and computer vision technology, and the 3D point cloud data of the surgical instrument is rigidly and non-rigidly registered with the 3D model reconstructed before the operation; S3. According to the surgical requirements, adjust the transparency and color of different tissues in the reconstructed image; use the Dijkstra algorithm to calculate the shortest path between two points in the surgical area to mark the surgical area, and apply the region growing algorithm and threshold segmentation technology to control the segmentation and presentation of the surgical area.
2. The three-dimensional imaging reconstruction system based on adaptive neural network for minimally invasive interventional surgery according to claim 1, characterized in that The adaptive neural network construction includes: a convolutional neural network (CNN) for extracting features from two-dimensional slices, and a generative adversarial network (GAN) for generating three-dimensional images, by defining a loss function: R represents reconstruction loss, D represents adversarial loss, f and g are mapping functions of CNN and GAN respectively, α and β are coefficients to adjust the influence of the two losses, and x i is a 2D slice of the input, y i and z j is the desired output three-dimensional image; The slice sequence is processed by recurrent neural network RNN or long short-term memory network LSTM, and the sequence processing formula is defined: where s i represents the i-th slice, h is the recursive function of RNN or LSTM, γ i is the weighting coefficient, reflecting the importance of each slice, and λ is the regularization parameter; The processed features are mapped to three-dimensional space, and high-quality three-dimensional images are generated layer by layer. The feature comparison mechanism inside the network is used to identify the structural relationship between layers, and the structural relationship identification formula is defined: Where k is the kernel function, describing the similarity between two consecutive layers, x is the feature of the layer, represents the feature gradient, dx is the inter-layer difference, a and b are the upper and lower limits of the integral, representing the beginning and end of the entire sequence.
3. The three-dimensional imaging reconstruction system based on adaptive neural network for minimally invasive interventional surgery according to claim 1, characterized in that The training data set construction process is as follows: S1. Use image registration technology to pair the 2D image with the corresponding 3D model, and perform geometric adjustment through the spatial transformation network STN to define the image registration function: Where φ is the mapping function from two-dimensional to three-dimensional space, x, y are the two-dimensional plane coordinates, and z is the depth; S2. Standardize and normalize the paired images, apply advanced filtering and enhancement techniques to optimize image quality, and define data preprocessing functions: Where k is the slope of enhancement and μ is the average brightness; S3. Develop intelligent annotation tools to identify and annotate key structures and feature points in 2D images, and enable medical experts to perform manual verification and correction through the interface, and define annotation verification functions: Where L and M represent the annotated and manually corrected label sets, respectively, and λ n is the weight of different labels, and p is the regularization exponent; S4. Convert the labeled and verified data into the format required for neural network training, and design a loss function to ensure that the neural network learns the mapping relationship from two-dimensional images to three-dimensional models. The loss function is defined as: f and q are the forward and reverse neural network functions for reconstruction, respectively, and x i ,y i are the training input and output, z j ,w j is the reverse verification input and output, ω i and σ j is the weight coefficient, α and β are the coefficients that adjust the contribution of the two losses.
4. The three-dimensional imaging reconstruction system based on adaptive neural network for minimally invasive interventional surgery according to claim 1, characterized in that The image transformation technology includes the following steps: firstly, the image registration technology is used, and the spatial transformation network STN is combined to rotate, scale and translate according to the content and geometric properties of the two-dimensional image, and the image registration formula is defined: Where I(u,v) is the original two-dimensional image, θ is the rotation angle, and δ is the Diracdelta function; Gamma correction is then performed to adapt to lighting changes during surgery, using the adaptive gamma correction formula: G(I;γ)=I γ Where I is the input image intensity, and γ is the gamma value adjusted according to the image brightness; Diversity enhancement is performed on the data after image transformation and gamma correction, and the diversity enhancement formula is applied: S(x;α,β)=αx+βx 2 Where x is the transformed image data, α and β are dynamically adjusted parameters; Finally, the enhanced dataset is used to train the adaptive neural network to learn the mapping relationship from two-dimensional images to three-dimensional models, and the loss function is defined as: Where f represents the mapping function of the neural network, x i and i are the training input and expected output respectively, w1 and w2 are the weight coefficients.
5. The three-dimensional imaging reconstruction system based on adaptive neural network for minimally invasive interventional surgery according to claim 1, characterized in that The generative adversarial network process is as follows: A custom generative adversarial network (GAN) is used to design and define adversarial loss functions to optimize the generator and discriminator: The generator G(z;θ G ) uses a random noise vector z to generate a visually realistic two-dimensional image, and the discriminator D(x; θ D ) Determine the authenticity of the image; The generator runs in real time during the surgery, dynamically generates or adjusts the image data according to the real-time situation of the surgery, and applies a composite transformation function: T(x;α,β,γ)=αx+βx 2 +γcos(x) Where x is the original image data, α, β, and γ are parameters adjusted according to real-time surgical needs to enhance the contrast and detail of the image; Merge the generated new image data with the existing real image data and pass the quality assessment function: Where f(x) represents the intensity function of the image, is the Laplacian operator, which evaluates the consistency of the quality of the generated image with the real image; Finally, the enhanced dataset is used to train an adaptive neural network to learn the mapping relationship from two-dimensional images to three-dimensional models, and the loss function is defined as: Where f represents the mapping function of the neural network, x i and i are the training input and expected output respectively, w1 and w2 are weight coefficients, and a and b define the range of integration.
6. The three-dimensional imaging reconstruction system based on adaptive neural network for minimally invasive interventional surgery according to claim 1, characterized in that The downsampling technology is implemented as follows: S1, using dynamic resolution adjustment technology, during the operation, the spatial resolution of the image is reduced to generate a 3D preview image according to the current operation stage and the doctor's needs; S2. Integrate multiple sensors to capture the position and motion trajectory of surgical instruments in real time. Use computer vision and machine learning technology to perform rigid and non-rigid registration of the three-dimensional point cloud data of surgical instruments captured in real time with the three-dimensional model reconstructed before surgery to adapt to the movement of instruments and deformation of tissues.
7. The adaptive neural network-based 3D imaging reconstruction system for minimally invasive interventional surgery according to claim 6, characterized in that The method for implementing the dynamic resolution adjustment technology includes: firstly, using an intelligent resolution adjustment algorithm to adjust the image resolution by real-time analysis of the surgical stage and the doctor's needs, and realizing dynamic conversion between three-dimensional preview and high-resolution image, using the algorithm: x is the real-time surgical data obtained from the sensor, α and β are parameters adjusted according to the surgical stage to control the downsampling and restoration of the image; Integrate GPU accelerated image rendering technology, optimize image rendering process, and apply rendering functions: Where x is the image data to be processed, γ and δ are optimization parameters that adjust the speed and quality of image rendering; The generated new image data is merged with the existing real image data, and the image quality is controlled by the quality assessment function: where f is the synthetic image function, λ is the evaluation parameter, and a and b define the range of the integration.
8. The three-dimensional imaging reconstruction system based on adaptive neural network for minimally invasive interventional surgery according to claim 6, characterized in that The computer vision and machine learning technology construction method is as follows: S1. Integrate multiple sensors, including optical, magnetic and inertial sensors, to capture the position and motion trajectory of surgical instruments in real time, and use data fusion algorithms to: F(p i ,p i-1 ;a,b)=αp i +(1-α)βp i-1 where p i is the currently captured position data, p i-1 is the previous position data, α and β are coefficients for adjusting the real-time response and the smoothness of historical data; S2. Apply the deep learning model to process the captured 3D point cloud data of surgical instruments in real time, and perform rigid and non-rigid registration with the preoperatively reconstructed 3D model, using the registration function: Where x, y are points in the point cloud data, and θ, φ are transformation parameters.
9. The three-dimensional imaging reconstruction system based on adaptive neural network for minimally invasive interventional surgery according to claim 1, characterized in that The method for realizing transparency and color of different tissues in the reconstructed image comprises: Develop a dynamic visual enhancement module, use an adaptive neural network to adjust the transparency and color of different tissues in the 3D reconstructed image according to the surgical stage and specific needs, and apply the transparency adjustment function: c represents the current color intensity of the tissue, ν represents the color target value adjusted according to surgical requirements, and k is the adjustment coefficient; The Dijkstra algorithm, region growing algorithm and threshold segmentation technology are combined to perform accurate segmentation and path planning of the surgical area, where the shortest path calculation uses the path planning function: v i and v j is the image pixel represented by the vertex in the graph, E is the edge set between the vertices, and w ij is the weight of the edge, indicating the distance between pixels; At the same time, the control of region growth is performed through the partition function: S(x;θ)=1 / (1+e -θ(x-μ) ) x is the grayscale value of the image, and θ and μ are parameters for adjusting the segmentation sensitivity.
Citation Information
Cited By
Surgical robot resection navigation method based on multi-modal image registration
CN120070423A
