Prostate cancer operation navigation method based on artificial intelligence and fluorescence-like real-time description

Abstract

The invention discloses a prostatic cancer surgical navigation method based on artificial intelligence and fluorescence-like real-time description, which comprises the following steps: S1, acquiring medical image data of a preoperative prostatic cancer patient, and preprocessing the medical image data; s2, generating dynamic fluorescence image data; s3, constructing a first layer of generative adversarial network; s4, a second generator receives the initial image output by the first generator and generates a refined dynamic tumor boundary image; s5, inputting dynamic fluorescence image data captured in real time into the trained double-layer generative adversarial network in the operation process; and S6, according to the generated refined dynamic tumor boundary image and patient data, dynamically adjusting an operation path according to the real-time change of the tumor morphology. According to the invention, the accuracy and safety of the prostatic cancer operation are obviously improved, and the dynamic tracking of tumors and the real-time adjustment of the operation path in the operation process are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of prostate cancer, and in particular to a prostate cancer surgery navigation method based on artificial intelligence and quasi-fluorescence real-time depiction. Background Art

[0002] In the existing technology, radical prostatectomy usually relies on preoperative imaging data, such as magnetic resonance imaging and computer tomography. Preoperative imaging data can provide preliminary morphological and location information of the tumor. However, the identification of the prostate and the protection of surrounding tissues during surgery are particularly important for postoperative recovery. The static preoperative imaging data cannot reflect this dynamic change in real time, resulting in the doctor's inability to obtain accurate navigation information during the operation. Therefore, in actual operations, doctors usually rely on visual observation and touch to judge the boundaries of the prostate, which cannot provide sufficient accuracy and increases the risk of damage to surrounding tissues.

[0003] Existing surgical navigation systems have limitations when dealing with dynamically changing fields of view during surgery. The system lacks the ability to update data in real time, making it difficult to continuously provide accurate prostate boundary information during the operation. This is especially true in prostate cancer surgery, as the prostate is often surrounded by dense nerves and blood vessels. Any accidental injury to these critical structures may lead to serious postoperative complications, such as sexual dysfunction and urinary incontinence. Therefore, how to monitor tumor boundaries in real time during surgery and combine the patient's personalized anatomical data to dynamically adjust the surgical path has become a difficult problem with existing technologies.

[0004] In recent years, fluorescence imaging technology has been used in surgery. The injection of fluorescent markers such as indocyanine green can provide doctors with real-time position and morphological information of target organs. However, fluorescent markers also have limitations. Fluorescence imaging can only provide the phenomenon of rich vascular tissue. It has limited ability to accurately distinguish poorly vascularized tissue such as nerves or capsules, subtle boundary changes, and surrounding healthy tissue. Especially when the boundaries are blurred or mixed with surrounding tissues, fluorescence technology cannot provide sufficient resolution, resulting in doctors still having to rely on experience to make judgments during the operation, and the accuracy of the operation is insufficient.

[0005] In summary, the existing technology has the following main shortcomings: first, surgical navigation based on preoperative images lacks real-time performance and cannot reflect the dynamic changes of the surgical area; second, the existing fluorescence imaging technology has limited accuracy in identifying the prostate and it is difficult to effectively avoid damage to healthy tissue; finally, the existing navigation system cannot dynamically adjust the surgical path according to real-time imaging data during the operation, making it difficult to ensure the safety and accuracy of the operation. Summary of the invention

[0006] One purpose of the present invention is to propose a prostate cancer surgery navigation method based on artificial intelligence and fluorescence-like real-time depiction. The present invention significantly improves the accuracy and safety of prostate cancer surgery, and ensures dynamic tracking of tumors and real-time adjustment of surgical paths during surgery.

[0007] A prostate cancer surgery navigation method based on artificial intelligence and fluorescence-like real-time depiction according to an embodiment of the present invention comprises the following steps:

[0008] S1. Obtaining medical imaging data of preoperative prostate cancer patients and preprocessing the medical imaging data;

[0009] S2, injecting a quasi-fluorescent marker into the patient's prostate region, obtaining real-time tumor location and morphology information through quasi-fluorescent real-time imaging technology, and generating dynamic fluorescent image data;

[0010] S3. Construct the first-layer generative adversarial network. The first generator receives real-time fluorescence image data and preoperative medical imaging data as input, and generates a preliminary dynamic tumor boundary image to simulate the tumor morphological changes under different resection strategies. The first discriminator receives the real fluorescence imaging data and the preliminary dynamic tumor boundary image output by the first generator, evaluates the similarity between the preliminary image and the real image, and feeds back to the first generator for optimization.

[0011] S4, the second generator receives the preliminary image output by the first generator, generates a refined dynamic tumor boundary image, the second discriminator receives the refined dynamic tumor boundary image output by the second generator, evaluates the similarity between the refined dynamic tumor boundary image output by the second generator and the real image, and feeds back to the second generator for optimization;

[0012] S5. During the operation, the dynamic fluorescence image data captured in real time is input into the trained two-layer generative adversarial network. The generator updates the tumor boundary image in real time and provides it to the surgical navigation system for the doctor's reference.

[0013] S6. Based on the generated refined dynamic tumor boundary images and patient data, the artificial intelligence system generates the optimal surgical path recommendation, avoiding important nerve and vascular tissues around the tumor in real time, and dynamically adjusting the surgical path according to real-time changes in tumor morphology.

[0014] Optionally, the S1 step includes the following sub-steps:

[0015] S11. Obtaining three-dimensional magnetic resonance imaging data D of patients with preoperative prostate cancer MRI ;

[0016] S12, using a convolutional neural network based denoising algorithm to 3D magnetic resonance imaging data D MRIPerform denoising to construct denoised three-dimensional magnetic resonance imaging data D' MRI :

[0017]

[0018] Among them, W k (x, y, z) is a weight function used to adjust the degree of denoising at different scales; G k (x, y, z) is the Gaussian kernel function, λ is the smoothing parameter, ∈ is the adjustment term, is the Laplace operator;

[0019] S13, using a deep learning-based segmentation algorithm to denoise the 3D MRI data D' MRI Perform segmentation of the prostate region to generate a preliminary segmentation result S of the prostate tumor tumor :

[0020]

[0021] Among them, σ is the activation function, α n and β are weight coefficients respectively, φ n (D' MRI (x,y,z)) represents the prostate tumor boundary features extracted at different scales; is a gradient operator used to detect edges in an image; γ is a parameter that adjusts the smoothness of the segmentation. MRI ,x,y,z) is the region growing term, which is used to incorporate the morphological information of the prostate tumor region.

[0022] Optionally, the S2 step includes the following sub-steps:

[0023] S21, injecting the fluorescent marker into the prostate area according to a predetermined dose, enhancing the fluorescence signal intensity of the target area through the specific binding of the fluorescent marker with the prostate cancer tissue, and setting the diffusion parameter μ of the fluorescent marker fluor , the diffusion process follows the diffusion equation:

[0024]

[0025] Where C(x,y,z,t) is the concentration distribution of the fluorescent marker at the spatiotemporal position (x,y,z,t), μ fluor is the diffusion coefficient of the fluorescent marker, k is the attenuation coefficient of the fluorescent marker;

[0026] S22, using a fluorescence imaging device to capture the prostate area after the injection of the fluorescent marker in real time, and obtain the fluorescence signal intensity image data I of the tumor fluor(x,y,z,t), the signal intensity of the fluorescence signal intensity image data is proportional to the marker concentration:

[0027] I fluor (x, y, z, t) = η·C(x, y, z, t);

[0028] Wherein, η is the calibration coefficient, which is used to convert the fluorescence signal intensity to the marker concentration;

[0029] S23, fluorescence signal intensity image data I fluor (x, y, z, t) is processed by spatiotemporal filtering to remove noise interference, and the spatial Gaussian filter function G is used. filter (x,y,z) and the time filter function H filter (t) Smoothing the fluorescence image data to obtain filtered dynamic fluorescence image data I' fluor (x,y,z,t):

[0030] I' fluor (x,y,z,t)=G filter (x,y,z)*H filter (t)*I fluor (x,y,z,t);

[0031] Among them, * represents the convolution operation, G filter (x, y, z) is the Gaussian kernel function in three-dimensional space, H filter (t) is the time filter kernel function.

[0032] Optionally, the S3 step includes the following sub-steps:

[0033] S31, constructing a first-layer generative adversarial network, including a first generator G1 and a first discriminator D1; the first generator G1 receives the initial segmentation result S of the prostate tumor tumor (x, y, z) and real-time fluorescence image data I' fluor (x, y, z, t) as input and generate a preliminary dynamic tumor boundary image G1(S tumor ,I' fluor )Simulate the changes in tumor morphology under different resection strategies:

[0034]

[0035] Among them, σ is the activation function, α i is the weight coefficient, φ i (S tumor ,I' fluor ) represents the feature extraction function at multiple scales, extracting tumor boundary features. is the gradient operator, which indicates the directional change of the tumor boundary, σi is the standard deviation of the Gaussian kernel function, is a Gaussian kernel-based similarity measure that measures the difference between the initial tumor segmentation results and the fluorescence image at different locations. j is the weight coefficient, is the Laplace operator, Ω is the integration range of the tumor boundary area;

[0036] S32, the first discriminator D1 receives the preliminary dynamic tumor boundary image G1 (S tumor ,I' fluor ) and real fluorescence image data I' fluor (x,y,z,t), evaluate the similarity between the two:

[0037]

[0038] Where P is the number of sampling points, x k ,y k ,z k ,t k is the sampling point in space and time, λ k is the weight coefficient, is the gradient of the fluorescence image data, |G1(S tumor ,I' fluor )(x k ,y k ,z k ,t k )-I' fluor (x k ,y k ,z k ,t k )| 2 To measure the difference between the generated image and the real image in space and time;

[0039] S33. The first discriminator D1 feeds back the similarity evaluation result to the generator G1. The generator G1 optimizes its generation strategy according to the feedback, iteratively updates the generated preliminary dynamic tumor boundary image, and completes the preliminary dynamic tumor boundary delineation.

[0040] Optionally, the S4 step includes the following sub-steps:

[0041] S41, constructing a second-layer generative adversarial network, including a second generator G2 and a second discriminator D2, wherein the second generator G2 receives the preliminary dynamic tumor boundary image G1 (S tumor ,I' fluor ), generate a refined dynamic tumor boundary image G2(G1(S tumor ,I' fluor)) to further optimize the resolution and accuracy of tumor boundaries:

[0042]

[0043] Among them, σ is the activation function, δ m is the weight coefficient, φ m (G1(S tumor ,I' fluor )) represents the feature extraction function under multi-scale, is the Laplacian operator, which is used to enhance the edge details in the image. m To adjust the weight coefficient of the refined boundary;

[0044] S43, the second discriminator D2 receives the refined dynamic tumor boundary image G2(G1(S tumor ,I' fluor )) and real fluorescence image data I' fluor (x, y, z, t), the similarity between the generated image and the real image is evaluated by the discriminant function:

[0045]

[0046] Where Q is the number of sampling points, x p ,y p ,z p ,t p is the sampling point in space and time, ρ p is the weight coefficient, is the gradient of the fluorescence image data, |G2(G1(S tumor ,I' fluor ))(x p ,y p ,z p ,t p )-I' fluor (x p ,y p ,z p ,t p )| 2 Represents the difference measure between the generated image and the real image;

[0047] S44. The second discriminator D2 feeds back the similarity evaluation result to the second generator G2. The generator G2 optimizes its generation strategy according to the feedback result, iteratively updates the generated refined dynamic tumor boundary image, and finally generates a dynamic tumor boundary image.

[0048] Optionally, the step S6 includes the following sub-steps:

[0049] S61, based on the generated refined dynamic tumor boundary image G2(G1(Stumor ,I' fluor )) and patient-specific data P data , including the patient's anatomical structure, tumor location, and the distribution of peripheral nerves and blood vessels to build a surgical path optimization model. The surgical path optimization objective function is:

[0050]

[0051] Among them, λ1 and λ2 are weight coefficients, is the gradient operator, which indicates the calculation of the boundary change rate of the tumor boundary and important organs in the image space, Ω safe It is a safe area for surgery;

[0052] S62. Optimizing the objective function L based on the surgical path path Generate the optimal surgical path P optimal (t), the optimal surgical path avoids the nerves and blood vessels around the tumor in real time during surgery:

[0053]

[0054] Where P(t) represents the surgical path at time t;

[0055] S63. The artificial intelligence system monitors the changes in tumor boundaries and morphology in real time. When the tumor morphology changes, the system automatically updates and refines the dynamic tumor boundary image G2(G1(S tumor ,I' fluor ,t)), and recalculate the surgical path. The updated path optimization objective function is:

[0056]

[0057] Among them, λ'1 and λ'2 are the updated weight coefficients, Ω' safe To update the safe surgical area, the system calculates a new surgical path based on the new path optimization objective function;

[0058] S64, the artificial intelligence system generates a new optimal surgical path P' based on the updated optimization model optimal (t), and dynamically adjust surgical recommendations.

[0059] Optionally, the step S62 includes the following sub-steps:

[0060] S621. Optimizing the objective function L based on the surgical path path Select the initial surgical path P initial (t), the initial path is based on the patient's personalized data P data The safe path generated by the distribution of nerve and vascular tissues and the tumor boundary position is expressed as:

[0061]

[0062] Among them, λ1 and λ2 are the weight coefficients of path optimization, is a gradient operator, which indicates the calculation of the boundary change rate of the tumor boundary and the neurovascular distribution in the personalized data;

[0063] S622, based on the initial path P initial (t), the artificial intelligence system performs real-time path optimization to generate the current optimal surgical path P optimal (t):

[0064]

[0065] Among them, λ3 and λ4 are weight coefficients adjusted during the optimization process. is the gradient change of the initial surgical path, is the spatial change rate of the refined dynamic tumor boundary;

[0066] S623. During the operation, the artificial intelligence system updates the surgical path according to the real-time changes of the tumor boundary. When the tumor boundary changes significantly, the path optimization objective function L is adjusted based on the latest refined dynamic tumor boundary image and personalized data. path The weight coefficients λ1,λ2 in the equation are used to recalculate the optimal surgical path P optimal (t).

[0067] The beneficial effects of the present invention are:

[0068] (1) The present invention adopts a two-layer generative adversarial network architecture. Through the collaborative work of the first generator and the second generator, it can dynamically identify and optimize the boundary information of the tumor during the operation. The first generator receives the preoperative image and the real-time fluorescence image data during the operation to generate a preliminary tumor boundary image, while the second generator further refines the boundary to make the generated tumor contour more accurate. The traditional surgical navigation system relies on the preoperative static image data and cannot respond to the changes of the tumor boundary in real time, resulting in large errors in tumor identification during the operation. The present invention dynamically updates the generative adversarial network to continuously adjust the generated tumor boundary according to the real-time data during the operation, which greatly improves the accuracy and precision of boundary identification. Experimental data show that the Dice coefficient of the present invention is improved by more than 10% when processing complex tumor boundaries, which significantly improves the accuracy of tumor resection.

[0069] (2) The present invention combines the patient's personalized anatomical data, including tumor location, nerve and blood vessel distribution, and dynamically adjusts the surgical path through an adaptive path planning algorithm to fuse the real-time updated tumor boundary with the personalized anatomical data to generate the optimal surgical path, thereby ensuring that important nerves and blood vessels are effectively avoided during the operation. Unlike traditional fixed surgical paths, the path planning algorithm of the present invention can monitor changes in tumor location in real time and automatically adjust the path to avoid damage to surrounding healthy tissues. The nerve injury rate during surgery is reduced by 90%, and the number of surgical path adjustments is reduced by 75%, greatly improving the safety and efficiency of the operation.

[0070] (3) The present invention introduces fluorescence imaging technology combined with a deep learning model to capture the dynamic morphological changes of the tumor in real time. Fluorescence imaging can mark the real-time position of the tumor and combine these data with preoperative images through a deep learning algorithm to generate an accurate tumor boundary image, which solves the static limitations of traditional surgery that relies on preoperative imaging data and provides a real-time tumor feedback mechanism. By using this feedback mechanism, the present invention can update the surgical path and adjust the resection strategy in real time to ensure that the surgery can still accurately locate and resect the tumor when it changes. Experimental data show that the boundary recognition response time of the present invention during surgery is shortened to 2 seconds, which greatly improves the surgical efficiency compared to the 10 seconds of the traditional method. At the same time, the accuracy of tumor boundary recognition is improved by more than 15%. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0072] Figure 1 A flowchart of a prostate cancer surgical navigation method based on artificial intelligence and fluorescence-like real-time depiction proposed by the present invention;

[0073] Figure 2 This is a response schematic diagram of dynamically adjusting the surgical path in a prostate cancer surgical navigation method based on artificial intelligence and quasi-fluorescence real-time depiction proposed by the present invention. DETAILED DESCRIPTION

[0074] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0075] refer to Figure 1-2 , a prostate cancer surgical navigation method based on artificial intelligence and real-time fluorescence depiction, comprising the following steps:

[0076] S1. Obtaining medical imaging data of preoperative prostate cancer patients and preprocessing the medical imaging data;

[0077] S2, injecting a quasi-fluorescent marker into the patient's prostate region, obtaining real-time tumor location and morphology information through quasi-fluorescent real-time imaging technology, and generating dynamic fluorescent image data;

[0078] S3. Construct the first-layer generative adversarial network. The first generator receives real-time fluorescence image data and preoperative medical imaging data as input, and generates a preliminary dynamic tumor boundary image to simulate the tumor morphological changes under different resection strategies. The first discriminator receives the real fluorescence imaging data and the preliminary dynamic tumor boundary image output by the first generator, evaluates the similarity between the preliminary image and the real image, and feeds back to the first generator for optimization.

[0079] S4, the second generator receives the preliminary image output by the first generator, generates a refined dynamic tumor boundary image, the second discriminator receives the refined dynamic tumor boundary image output by the second generator, evaluates the similarity between the refined dynamic tumor boundary image output by the second generator and the real image, and feeds back to the second generator for optimization;

[0080] S5. During the operation, the dynamic fluorescence image data captured in real time is input into the trained two-layer generative adversarial network. The generator updates the tumor boundary image in real time and provides it to the surgical navigation system for the doctor's reference.

[0081] S6. Based on the generated refined dynamic tumor boundary images and patient data, the artificial intelligence system generates the optimal surgical path recommendation, avoiding important nerve and vascular tissues around the tumor in real time, and dynamically adjusting the surgical path according to real-time changes in tumor morphology.

[0082] In this implementation, step S1 includes the following sub-steps:

[0083] S11. Obtaining three-dimensional magnetic resonance imaging data D of patients with preoperative prostate cancer MRI ;

[0084] S12, using a convolutional neural network based denoising algorithm to 3D magnetic resonance imaging data D MRI Perform denoising to construct denoised three-dimensional magnetic resonance imaging data D' MRI :

[0085]

[0086] Among them, W k (x, y, z) is a weight function used to adjust the degree of denoising at different scales; G k (x, y, z) is the Gaussian kernel function, λ is the smoothing parameter, ∈ is the adjustment term, is the Laplace operator;

[0087] S13, using a deep learning-based segmentation algorithm to denoise the 3D MRI data D' MRI Perform segmentation of the prostate region to generate a preliminary segmentation result S of the prostate tumor tumor :

[0088]

[0089] Among them, σ is the activation function, α n and β are weight coefficients respectively, φ n (D' MRI (x,y,z)) represents the prostate tumor boundary features extracted at different scales; is a gradient operator used to detect edges in an image; γ is a parameter that adjusts the smoothness of the segmentation. MRI ,x,y,z) is the region growing term, which is used to incorporate the morphological information of the prostate tumor region.

[0090] In this implementation, step S2 includes the following sub-steps:

[0091] S21, injecting the fluorescent marker into the prostate area according to a predetermined dose, enhancing the fluorescence signal intensity of the target area through the specific binding of the fluorescent marker with the prostate cancer tissue, and setting the diffusion parameter μ of the fluorescent marker fluor , the diffusion process follows the diffusion equation:

[0092]

[0093] Where C(x,y,z,t) is the concentration distribution of the fluorescent marker at the spatiotemporal position (x,y,z,t), μ fluor is the diffusion coefficient of the fluorescent marker, k is the attenuation coefficient of the fluorescent marker;

[0094] S22, using a fluorescence imaging device to capture the prostate area after the injection of the fluorescent marker in real time, and obtain the fluorescence signal intensity image data I of the tumor fluor (x,y,z,t), the signal intensity of the fluorescence signal intensity image data is proportional to the marker concentration:

[0095] I fluor (x, y, z, t) = η·C(x, y, z, t);

[0096] Wherein, η is the calibration coefficient, which is used to convert the fluorescence signal intensity to the marker concentration;

[0097] S23, fluorescence signal intensity image data I fluor(x, y, z, t) is processed by spatiotemporal filtering to remove noise interference, and the spatial Gaussian filter function G is used. filter (x,y,z) and the time filter function H filter (t) Smoothing the fluorescence image data to obtain filtered dynamic fluorescence image data I' fluor (x,y,z,t):

[0098] I' fluor (x,y,z,t)=G filter (x,y,z)*H filter (t)*I fluor (x,y,z,t);

[0099] Among them, * represents the convolution operation, G filter (x, y, z) is the Gaussian kernel function in three-dimensional space, H filter (t) is the time filter kernel function.

[0100] In this implementation, step S3 includes the following sub-steps:

[0101] S31, constructing a first-layer generative adversarial network, including a first generator G1 and a first discriminator D1; the first generator G1 receives the initial segmentation result S of the prostate tumor tumor (x, y, z) and real-time fluorescence image data I' fluor (x, y, z, t) as input and generate a preliminary dynamic tumor boundary image G1(S tumor ,I' fluor )Simulate the changes in tumor morphology under different resection strategies:

[0102]

[0103] Among them, σ is the activation function, α i is the weight coefficient, φ i (S tumor ,I' fluor ) represents the feature extraction function at multiple scales, extracting tumor boundary features. is the gradient operator, which indicates the directional change of the tumor boundary, σ i is the standard deviation of the Gaussian kernel function, is a Gaussian kernel-based similarity measure that measures the difference between the initial tumor segmentation results and the fluorescence image at different locations. j is the weight coefficient, is the Laplace operator, Ω is the integration range of the tumor boundary area;

[0104] S32, the first discriminator D1 receives the preliminary dynamic tumor boundary image G1 (Stumor ,I' fluor ) and real fluorescence image data I' fluor (x,y,z,t), evaluate the similarity between the two:

[0105]

[0106] Where P is the number of sampling points, x k ,y k ,z k ,t k is the sampling point in space and time, λ k is the weight coefficient, is the gradient of the fluorescence image data, |G1(S tumor ,I' fluor )(x k ,y k ,z k ,t k )-I' fluor (x k ,y k ,z k ,t k )| 2 To measure the difference between the generated image and the real image in space and time;

[0107] S33. The first discriminator D1 feeds back the similarity evaluation result to the generator G1. The generator G1 optimizes its generation strategy according to the feedback, iteratively updates the generated preliminary dynamic tumor boundary image, and completes the preliminary dynamic tumor boundary delineation.

[0108] In this implementation, step S4 includes the following sub-steps:

[0109] S41, constructing a second-layer generative adversarial network, including a second generator G2 and a second discriminator D2, wherein the second generator G2 receives the preliminary dynamic tumor boundary image G1 (S tumor ,I' fluor ), generate a refined dynamic tumor boundary image G2(G1(S tumor ,I' fluor )) to further optimize the resolution and accuracy of tumor boundaries:

[0110]

[0111] Among them, σ is the activation function, δ m is the weight coefficient, φ m (G1(S tumor ,I' fluor )) represents the feature extraction function under multi-scale, is the Laplacian operator, which is used to enhance the edge details in the image. m To adjust the weight coefficient of the refined boundary;

[0112] S43, the second discriminator D2 receives the refined dynamic tumor boundary image G2(G1(S tumor ,I' fluor )) and real fluorescence image data I' fluor (x, y, z, t), the similarity between the generated image and the real image is evaluated by the discriminant function:

[0113]

[0114] Where Q is the number of sampling points, x p ,y p ,z p ,t p is the sampling point in space and time, ρ p is the weight coefficient, is the gradient of the fluorescence image data, |G2(G1(S tumor ,I' fluor ))(x p ,y p ,z p ,t p )-I' fluor (x p ,y p ,z p ,t p )| 2 Represents the difference measure between the generated image and the real image;

[0115] S44. The second discriminator D2 feeds back the similarity evaluation result to the second generator G2. The generator G2 optimizes its generation strategy according to the feedback result, iteratively updates the generated refined dynamic tumor boundary image, and finally generates a dynamic tumor boundary image.

[0116] In this implementation, step S6 includes the following sub-steps:

[0117] S61, based on the generated refined dynamic tumor boundary image G2(G1(S tumor ,I' fluor )) and patient-specific data P data , including the patient's anatomical structure, tumor location, and the distribution of peripheral nerves and blood vessels to build a surgical path optimization model. The surgical path optimization objective function is:

[0118]

[0119] Among them, λ1 and λ2 are weight coefficients, is the gradient operator, which indicates the calculation of the boundary change rate of the tumor boundary and important organs in the image space, Ω safe It is a safe area for surgery;

[0120] S62. Optimizing the objective function L based on the surgical path path Generate the optimal surgical path P optimal (t), the optimal surgical path avoids the nerves and blood vessels around the tumor in real time during surgery:

[0121]

[0122] Where P(t) represents the surgical path at time t;

[0123] S63. The artificial intelligence system monitors the changes in tumor boundaries and morphology in real time. When the tumor morphology changes, the system automatically updates and refines the dynamic tumor boundary image G2(G1(S tumor ,I' fluor ,t)), and recalculate the surgical path. The updated path optimization objective function is:

[0124]

[0125] Among them, λ'1 and λ'2 are the updated weight coefficients, Ω' safe To update the safe surgical area, the system calculates a new surgical path based on the new path optimization objective function;

[0126] S64, the artificial intelligence system generates a new optimal surgical path P' based on the updated optimization model optimal (t), and dynamically adjust surgical recommendations.

[0127] In this implementation, step S62 includes the following sub-steps:

[0128] S621. Optimizing the objective function L based on the surgical path path Select the initial surgical path P initial (t), the initial path is based on the patient's personalized data P data The safe path generated by the distribution of nerve and vascular tissues and the tumor boundary position is expressed as:

[0129]

[0130] Among them, λ1 and λ2 are the weight coefficients of path optimization, is a gradient operator, which indicates the calculation of the boundary change rate of the tumor boundary and the neurovascular distribution in the personalized data;

[0131] S622, based on the initial path P initial(t), the artificial intelligence system performs real-time path optimization to generate the current optimal surgical path P optimal (t):

[0132]

[0133] Among them, λ3 and λ4 are weight coefficients adjusted during the optimization process. is the gradient change of the initial surgical path, is the spatial change rate of the refined dynamic tumor boundary;

[0134] S623. During the operation, the artificial intelligence system updates the surgical path according to the real-time changes of the tumor boundary. When the tumor boundary changes significantly, the path optimization objective function L is adjusted based on the latest refined dynamic tumor boundary image and personalized data. path The weight coefficients λ1,λ2 in the equation are used to recalculate the optimal surgical path P optimal (t).

[0135] Embodiment 1:

[0136] In order to verify the feasibility of the present invention in prostate cancer surgical navigation, the experimenters conducted a large number of simulation experiments and comparative analyses, using the public medical imaging dataset ProstateX and a prostate cancer intraoperative quasi-fluorescence imaging dataset for experimental verification. The ProstateX dataset includes magnetic resonance imaging data and detailed pathological information of patients, and the quasi-fluorescence imaging dataset covers fluorescently labeled prostate tumor data from real surgical procedures. The experimenters divided the two datasets into training set and test set in a ratio of 7:3.

[0137] ProstateX dataset: Contains MRI image data of 500 prostate cancer patients. The image resolution of each patient is 256x256, and the voxel size of the 3D data is 1mm. 3 , in addition, it also provides detailed pathological diagnosis and annotated information on the tumor location.

[0138] Fluorescence imaging dataset: contains fluorescence image data collected during 100 real prostate cancer surgeries. The image resolution is 1024x1024, covering the entire process of dynamic changes of tumors. The image sequence of each surgery contains about 500 frames, covering the process from tumor marking, resection to completion of the surgery.

[0139] The researchers used MRI data to perform preliminary segmentation of the tumor and aligned it with the intraoperative fluorescence image. In the data fusion step, a generative adversarial network-based method was used to match the MRI data with the fluorescence image to generate preliminary dynamic tumor boundaries. The researchers used the generator and discriminator models to train and optimize the images.

[0140] In the experiment, the experimenters used the ResNet50 model to extract features from MRI images and combined it with the BERT model to extract dynamic information from real-time fluorescence images during surgery. In the fusion model, the embedding dimension of the MRI image is 1024, and the embedding dimension of the fluorescence image is 512. The experimenters processed the two parts of data uniformly and trained them using a deep learning model based on the Adam optimizer. The training lasted for 500 epochs, with a learning rate of 0.0001 and a batch size of 64.

[0141] During the surgical navigation process, the experimenters dynamically generated the tumor boundary based on the real-time fluorescent image data and updated the surgical path in real time based on the trained model. Specifically, when the system recognizes that the tumor boundary has changed, the path optimization model will immediately recalculate the resection path to ensure that the surgical instruments avoid important nerves and blood vessels.

[0142] In order to verify the effectiveness of the system, the experimenters used a variety of evaluation indicators to measure the performance of surgical navigation:

[0143] Accuracy: used to evaluate the overall correctness of the system in identifying tumor boundaries.

[0144] Boundary overlap coefficient: evaluates the overlap between the generated tumor boundary and the true boundary.

[0145] Path planning accuracy: measures whether the generated surgical path avoids critical tissue.

[0146] Real-time response time: How quickly the system responds to changes in the tumor.

[0147] The experimenters conducted a comparative experiment on the surgical navigation system of the present invention and the traditional navigation system based on preoperative MRI. The specific results are shown in Tables 1 and 2.

[0148] Table 1 Comparison of the proposed method and traditional methods on the ProstateX dataset

[0149] index Traditional navigation system Method of the present invention Accuracy 85％ 97％ Boundary overlap coefficient (Dice) 0.78 0.92 Path planning accuracy 83％ 96％ Real-time response time (seconds) 10 2

[0150] Table 2 Comparison of the proposed method and traditional methods on a fluorescence imaging dataset

[0151]

[0152]

[0153] In order to more intuitively demonstrate the training process of the surgical navigation system, the experimenters selected a typical prostate cancer case in the ProstateX dataset for simulation training. In the case, the patient's tumor was close to the nerve bundle and the tumor boundary was mixed and blurred with the surrounding healthy tissue. During the training process, the system first processed the MRI data of the case, generated the preliminary boundary of the tumor through the generative adversarial network, and fused it with the fluorescent image data to refine the tumor boundary. In the first 100 epochs of training, the system's boundary recognition accuracy increased by about 15%, reaching a Dice coefficient of 0.85. In the next 400 epochs, the system further optimized the path planning model to ensure real-time dynamic update of the surgical path.

[0154] In the simulated surgical scenario, the system updates the surgical path in real time through fluorescence-like imaging technology based on the patient's personalized anatomical data and the dynamic boundary of the tumor. During the operation, when the system detects changes in the tumor boundary, it immediately adjusts the surgical path and generates a new resection strategy. The system's response time in surgical navigation is 2 seconds, significantly better than the 10 seconds of traditional systems.

[0155] Through this simulated training and surgical scenario, the feasibility of the application of the present invention in complex prostate cancer cases was further verified. The surgical navigation system of the present invention can not only accurately identify tumor boundaries in real time, but also dynamically adjust the surgical path according to real-time data to ensure the safety and accuracy of the operation.

[0156] The experimental results show that the method of the present invention is significantly superior to traditional surgical navigation methods in multiple indicators. In actual surgical scenarios, the system can obtain tumor boundary information in real time through fluorescence-like imaging and generate the optimal surgical path based on the artificial intelligence system, which significantly improves the accuracy and safety of the surgery. At the same time, the real-time response capability of the system greatly reduces the risk of damage to healthy tissues during surgery, further verifying the practical application value of the present invention in prostate cancer surgery.

[0157] The present invention adopts a two-layer generative adversarial network architecture. Through the collaborative work of the first generator and the second generator, it can dynamically identify and optimize the boundary information of the tumor during the operation. The first generator receives preoperative images and real-time fluorescence image data during the operation to generate a preliminary tumor boundary image, while the second generator further refines the boundary to make the generated tumor contour more accurate. Traditional surgical navigation systems rely on preoperative static image data and cannot respond to changes in tumor boundaries in real time, resulting in large errors in tumor identification during surgery. The present invention dynamically updates the generative adversarial network to continuously adjust the generated tumor boundaries according to real-time data during the operation, greatly improving the accuracy and precision of boundary identification. Experimental data show that the Dice coefficient of the present invention is improved by more than 10% when processing complex tumor boundaries, significantly improving the accuracy of tumor resection.

[0158] The present invention combines the patient's personalized anatomical data, including tumor location, nerve and blood vessel distribution, and dynamically adjusts the surgical path through an adaptive path planning algorithm to fuse the real-time updated tumor boundary with the personalized anatomical data to generate the optimal surgical path, thereby ensuring that important nerve and vascular tissues are effectively avoided during the operation. Unlike traditional fixed surgical paths, the path planning algorithm of the present invention can monitor changes in tumor location in real time and automatically adjust the path to avoid damage to surrounding healthy tissues. The nerve injury rate during surgery is reduced by 90%, and the number of surgical path adjustments is reduced by 75%, greatly improving the safety and efficiency of the operation.

[0159] The present invention introduces fluorescence imaging technology combined with a deep learning model to capture the dynamic morphological changes of the tumor in real time. Fluorescence imaging can mark the real-time position of the tumor, and combine these data with preoperative images through a deep learning algorithm to generate an accurate tumor boundary image, which solves the static limitations of traditional surgery that relies on preoperative imaging data and provides a real-time tumor feedback mechanism. By using this feedback mechanism, the present invention can update the surgical path and adjust the resection strategy in real time to ensure that the surgery can still accurately locate and resect when the tumor changes. Experimental data show that the boundary recognition response time during the operation of the present invention is shortened to 2 seconds, which greatly improves the surgical efficiency compared to the 10 seconds of the traditional method. At the same time, the accuracy of tumor boundary recognition is improved by more than 15%.

[0160] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A prostate cancer surgical navigation method based on artificial intelligence and real-time fluorescence-like depiction, characterized in that: The steps include: S1. Obtain medical imaging data of preoperative prostate cancer patients and pre-process the medical imaging data; S2, injecting a quasi-fluorescent marker into the patient's prostate region, obtaining real-time tumor location and morphology information through quasi-fluorescent real-time imaging technology, and generating dynamic fluorescent image data; S3. Construct the first-layer generative adversarial network. The first generator receives real-time fluorescence image data and preoperative medical imaging data as input, and generates a preliminary dynamic tumor boundary image to simulate the tumor morphological changes under different resection strategies. The first discriminator receives the real fluorescence imaging data and the preliminary dynamic tumor boundary image output by the first generator, evaluates the similarity between the preliminary image and the real image, and feeds back to the first generator for optimization. S4, the second generator receives the preliminary image output by the first generator, generates a refined dynamic tumor boundary image, the second discriminator receives the refined dynamic tumor boundary image output by the second generator, evaluates the similarity between the refined dynamic tumor boundary image output by the second generator and the real image, and feeds back to the second generator for optimization; S5. During the operation, the dynamic fluorescence image data captured in real time is input into the trained two-layer generative adversarial network. The generator updates the tumor boundary image in real time and provides it to the surgical navigation system for the doctor's reference. S6. Based on the generated refined dynamic tumor boundary images and patient data, the artificial intelligence system generates the optimal surgical path recommendation, avoiding important nerve and vascular tissues around the tumor in real time, and dynamically adjusting the surgical path according to real-time changes in tumor morphology.

2. The prostate cancer surgical navigation method based on artificial intelligence and real-time fluorescence depiction according to claim 1, characterized in that: The S1 step includes the following sub-steps: S11. Obtaining three-dimensional magnetic resonance imaging data D of patients with preoperative prostate cancer MRI ; S12, using a convolutional neural network based denoising algorithm to 3D magnetic resonance imaging data D MRI Perform denoising to construct denoised three-dimensional magnetic resonance imaging data D' MRI : Among them, W k (x, y, z) is a weight function used to adjust the degree of denoising at different scales; G k (x, y, z) is the Gaussian kernel function, λ is the smoothing parameter, ∈ is the adjustment term, is the Laplace operator; S13, using a deep learning-based segmentation algorithm to denoise the 3D MRI data D' MRI Perform segmentation of the prostate region to generate a preliminary segmentation result S of the prostate tumor tumor : Among them, σ is the activation function, α n and β are weight coefficients respectively, φ n (D' MRI (x,y,z)) represents the prostate tumor boundary features extracted at different scales; is a gradient operator used to detect edges in an image; γ is a parameter that adjusts the smoothness of the segmentation. MRI ,x,y,z) is the region growing term, which is used to incorporate the morphological information of the prostate tumor region.

3. The prostate cancer surgical navigation method based on artificial intelligence and real-time fluorescence depiction according to claim 1, characterized in that: The S2 step includes the following sub-steps: S21, injecting the fluorescent marker into the prostate area according to a predetermined dose, enhancing the fluorescence signal intensity of the target area through the specific binding of the fluorescent marker with the prostate cancer tissue, and setting the diffusion parameter μ of the fluorescent marker fluor , the diffusion process follows the diffusion equation: Where C(x,y,z,t) is the concentration distribution of the fluorescent marker at the spatiotemporal position (x,y,z,t), μ fluor is the diffusion coefficient of the fluorescent marker, k is the attenuation coefficient of the fluorescent marker; S22, using a fluorescence imaging device to capture the prostate area after the injection of the fluorescent marker in real time, and obtain the fluorescence signal intensity image data I of the tumor fluor (x,y,z,t), the signal intensity of the fluorescence signal intensity image data is proportional to the marker concentration: I fluor (x,y,z,t)=η·C(x,y,z,t); Wherein, η is the calibration coefficient, which is used to convert the fluorescence signal intensity to the marker concentration; S23, fluorescence signal intensity image data I fluor (x, y, z, t) is processed by spatiotemporal filtering to remove noise interference, and the spatial Gaussian filter function G is used. filter (x,y,z) and the time filter function H filter (t) Smoothing the fluorescence image data to obtain filtered dynamic fluorescence image data I' fluor (x,y,z,t): I′ fluor (x,y,z,t)=G filter (x,y,z)*H filter (t)*I fluor (x,y,z,t); Among them, * represents the convolution operation, G filter (x, y, z) is the Gaussian kernel function in three-dimensional space, H filter (t) is the time filter kernel function.

4. The prostate cancer surgical navigation method based on artificial intelligence and real-time fluorescence depiction according to claim 1, characterized in that: The S3 step includes the following sub-steps: S31, constructing a first-layer generative adversarial network, including a first generator G1 and a first discriminator D1; the first generator G1 receives the initial segmentation result S of the prostate tumor tumor (x, y, z) and real-time fluorescence image data I' fluor (x, y, z, t) as input and generate a preliminary dynamic tumor boundary image G1(S tumor ,I' fluor )Simulate the changes in tumor morphology under different resection strategies: Among them, σ is the activation function, α i is the weight coefficient, φ i (S tumor ,I' fluor ) represents the feature extraction function at multiple scales, extracting tumor boundary features. is the gradient operator, which indicates the directional change of the tumor boundary, σ i is the standard deviation of the Gaussian kernel function, is a Gaussian kernel-based similarity measure that measures the difference between the initial tumor segmentation results and the fluorescence image at different locations. j is the weight coefficient, is the Laplace operator, Ω is the integration range of the tumor boundary area; S32, the first discriminator D1 receives the preliminary dynamic tumor boundary image G1 (S tumor ,I' fluor ) and real fluorescence image data I' fluor (x,y,z,t), evaluate the similarity between the two: Where P is the number of sampling points, x k ,y k ,z k ,t k is the sampling point in space and time, λ k is the weight coefficient, is the gradient of the fluorescence image data, |G1(S tumor ,I' fluor )(x k ,y k ,z k ,t k )-I' fluor (x k ,y k ,z k ,t k )| 2 To measure the difference between the generated image and the real image in space and time; S33. The first discriminator D1 feeds back the similarity evaluation result to the generator G1. The generator G1 optimizes its generation strategy according to the feedback, iteratively updates the generated preliminary dynamic tumor boundary image, and completes the preliminary dynamic tumor boundary delineation.

5. The prostate cancer surgical navigation method based on artificial intelligence and real-time fluorescence depiction according to claim 1, characterized in that: The S4 step includes the following sub-steps: S41, constructing a second-layer generative adversarial network, including a second generator G2 and a second discriminator D2, wherein the second generator G2 receives the preliminary dynamic tumor boundary image G1 (S tumor ,I' fluor ), generate a refined dynamic tumor boundary image G2(G1(S tumor ,I' fluor )) to further optimize the resolution and accuracy of tumor boundaries: Among them, σ is the activation function, δ m is the weight coefficient, φ m (G1(S tumor ,I' fluor )) represents the feature extraction function under multi-scale, is the Laplacian operator, which is used to enhance the edge details in the image. m To adjust the weight coefficient of the refined boundary; S43, the second discriminator D2 receives the refined dynamic tumor boundary image G2(G1(S tumor ,I' fluor )) and real fluorescence image data I' fluor (x, y, z, t), the similarity between the generated image and the real image is evaluated by the discriminant function: Where Q is the number of sampling points, x p ,y p ,z p ,t p is the sampling point in space and time, ρ p is the weight coefficient, is the gradient of the fluorescence image data, |G2(G1(S tumor ,I' fluor ))(x p ,y p ,z p ,t p )-I' fluor (x p ,y p ,z p ,t p )| 2 Represents the difference measure between the generated image and the real image; S44. The second discriminator D2 feeds back the similarity evaluation result to the second generator G2. The generator G2 optimizes its generation strategy according to the feedback result, iteratively updates the generated refined dynamic tumor boundary image, and finally generates a dynamic tumor boundary image.

6. The prostate cancer surgical navigation method based on artificial intelligence and real-time fluorescence depiction according to claim 1, characterized in that: The S6 step includes the following sub-steps: S61, based on the generated refined dynamic tumor boundary image G2(G1(S tumor ,I' fluor )) and patient-specific data P data , including the patient's anatomical structure, tumor location, and the distribution of peripheral nerves and blood vessels to build a surgical path optimization model. The surgical path optimization objective function is: Among them, λ1 and λ2 are weight coefficients, is the gradient operator, which indicates the calculation of the boundary change rate of the tumor boundary and important organs in the image space, Ω safe It is a safe area for surgery; S62. Optimizing the objective function L based on the surgical path path Generate the optimal surgical path P optimal (t), the optimal surgical path avoids the nerves and blood vessels around the tumor in real time during surgery: Where P(t) represents the surgical path at time t; S63. The artificial intelligence system monitors the changes in tumor boundaries and morphology in real time. When the tumor morphology changes, the system automatically updates and refines the dynamic tumor boundary image G2(G1(S tumor ,I' fluor ,t)), and recalculate the surgical path. The updated path optimization objective function is: Among them, λ'1 and λ'2 are the updated weight coefficients, Ω' safe To update the safe surgical area, the system calculates a new surgical path based on the new path optimization objective function; S64, the artificial intelligence system generates a new optimal surgical path P' based on the updated optimization model optimal (t), and dynamically adjust surgical recommendations.

7. The prostate cancer surgical navigation method based on artificial intelligence and real-time fluorescence depiction according to claim 1, characterized in that: The step S62 includes the following sub-steps: S621. Optimizing the objective function L based on the surgical path path Select the initial surgical path P initial (t), the initial path is based on the patient's personalized data P data The safe path generated by the distribution of nerve and vascular tissues and the tumor boundary position is expressed as: Among them, λ1 and λ2 are the weight coefficients of path optimization, is a gradient operator, which indicates the calculation of the boundary change rate of the tumor boundary and the neurovascular distribution in the personalized data; S622, based on the initial path P initial (t), the artificial intelligence system performs real-time path optimization to generate the current optimal surgical path P optimal (t): Among them, λ3 and λ4 are weight coefficients adjusted during the optimization process. is the gradient change of the initial surgical path, is the spatial change rate of the refined dynamic tumor boundary; S623. During the operation, the artificial intelligence system updates the surgical path according to the real-time changes of the tumor boundary. When the tumor boundary changes significantly, the path optimization objective function L is adjusted based on the latest refined dynamic tumor boundary image and personalized data. path The weight coefficients λ1,λ2 in the equation are used to recalculate the optimal surgical path P optimal (t).

Images