Prostate cancer focus accurate positioning method based on fusion navigation technology

Abstract

The invention provides a prostate cancer focus accurate positioning method based on a fusion navigation technology, and relates to the technical field of data processing, and the method comprises the steps: collecting a plurality of prostate images; selecting a control point from the prostate image in combination with a sparse regularization algorithm; determining an optimal deformation field according to the control points, and performing elastic registration on the prostate image based on the optimal deformation field; differential multi-resolution fusion is carried out on the registered prostate images, and a prostate three-dimensional model is reconstructed; constructing a lesion dynamic extraction model comprising a three-dimensional convolutional network, a pooling layer, an interested layer and a bounding box regression layer which are connected in sequence; a loss function of the focus dynamic extraction model is established, and the loss function comprises an intersection-to-union ratio loss item and a distributed focus loss item; under the supervision of the loss function, training the focus dynamic extraction model; and positioning a lesion point in the prostate three-dimensional model through the trained lesion dynamic extraction model. And the focus point positioning speed and accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to a precise positioning method for prostate cancer lesions based on fusion navigation technology. Background Art

[0002] Prostate cancer lesions refer to cancerous tumor areas that develop in prostate tissue. These areas typically exhibit abnormal cell proliferation and appear as regions with different densities or signals compared to surrounding tissues in imaging examinations. Prostate cancer is one of the common types of cancer in men. Early identification of lesions can help formulate treatment plans and improve the pertinence and effectiveness of treatment. Precise localization of lesions can assist doctors in determining the exact location, size, and relationship with key anatomical structures of the tumor, which is crucial for selecting the most suitable treatment methods (such as surgery, radiotherapy, or drug therapy). In addition, accurate lesion localization can maximize the protection of surrounding healthy tissues, reduce complications during treatment, and thus improve the quality of life of patients.

[0003] Existing prostate cancer lesion localization methods often roughly select possible lesion points at the two-dimensional level based on image differences in various regions of prostate images. They do not have depth features, cannot accurately locate lesion points, are inefficient and inaccurate, and affect the accuracy of patient treatment. Summary of the Invention

[0004] To solve the technical problems in the existing technology that prostate cancer lesion localization methods often roughly select possible lesion points at the two-dimensional level based on image differences in various regions of prostate images, do not have depth features, cannot accurately locate lesion points, are inefficient and inaccurate, and affect the accuracy of patient treatment, the present invention provides a precise positioning method for prostate cancer lesions based on fusion navigation technology.

[0005] The technical solutions provided by the embodiments of the present invention are as follows: A precise positioning method for prostate cancer lesions based on fusion navigation technology provided by an embodiment of the present invention includes: S1: Collect multiple prostate images.

[0006] S2: Select control points from the prostate images in combination with the sparse regularization algorithm.

[0007] S3: Determine the optimal deformation field based on the control points, and perform elastic registration on the prostate images based on the optimal deformation field.

[0008] S4: Perform differential multi-resolution fusion on the registered prostate images to reconstruct a three-dimensional prostate model.

[0009] S5: Construct a lesion dynamic extraction model including a three-dimensional convolutional network, a pooling layer, a region of interest layer, and a bounding box regression layer connected in sequence.

[0010] S6: Establish a loss function for the lesion dynamic extraction model, where the loss function includes an intersection over union loss term and a distributed focal loss term.

[0011] S7: Train the lesion dynamic extraction model under the supervision of the loss function.

[0012] S8: Locate the lesion points in the prostate three-dimensional model through the trained lesion dynamic extraction model.

[0013] The beneficial effects brought by the technical solutions provided in the embodiments of the present invention at least include: In the present invention, the sparse regularization algorithm is combined to selectively determine the control points with greater influence, and the optimal deformation field is determined based on these control points, thereby completing the elastic registration of each prostate image. This process reduces the complexity of the registration process while retaining important control points, excluding the registration interference of abnormal control points, improving the registration speed and accuracy. Then, differential multi-resolution fusion is performed on the registered images, which retains the detailed content of the images while reducing the complexity of detail extraction. The prostate three-dimensional model reconstructed based on this data has a fast reconstruction speed, high clarity, and can accurately reflect the location of the lesion points. Moreover, it also makes up for the shortcoming of consuming a large amount of computing resources in the traditional three-dimensional model reconstruction process, effectively reducing the resource consumption during the reconstruction process, enabling the reconstruction process of the prostate three-dimensional model to be applied to preoperative reference and intraoperative navigation, improving the application scenarios and practicality. Based on the improvement of the construction speed of this three-dimensional model, the lesion dynamic extraction model can dynamically lock the lesion points through the region of interest layer and the bounding box regression layer, and automatically and accurately output the location of the lesion points with depth features in the three-dimensional structure. The introduction of the loss function with the intersection over union loss term and the distributed focal loss term conforms to the characteristic that the true distribution of the lesion points usually is not too far from the labeled position, and the model should focus on the values near the labeled position during training, enabling the lesion dynamic extraction model to accurately learn the lesion features, further improving the positioning accuracy of the lesion point positions, and being able to efficiently assist the patient treatment process. Description of the Drawings

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0015] Figure 1A flow chart showing a method for accurately positioning prostate cancer lesions based on a fusion navigation technology provided by an embodiment of the present invention; Figure 2 A structural diagram showing a dynamic lesion extraction model provided by an embodiment of the present invention. Detailed implementation manners

[0016] The following describes the technical solutions in the present invention with reference to the accompanying drawings.

[0017] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as an "example" in the present invention should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Exactly speaking, the use of the word "example" is intended to present concepts in a specific manner. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one of the two.

[0018] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when their differences are not emphasized, their intended meanings are the same. "(of)", "corresponding", and "corresponding" can sometimes be used interchangeably. It should be noted that when their differences are not emphasized, their intended meanings are the same.

[0019] In the embodiments of the present invention, sometimes subscripts such as W1 may be miswritten as non-subscript forms such as W1. When their differences are not emphasized, their intended meanings are the same.

[0020] To make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments.

[0021] Refer to the accompanying drawings of the specification Figure 1 , which shows a flow chart of a method for accurately positioning prostate cancer lesions based on a fusion navigation technology provided by an embodiment of the present invention.

[0022] The embodiments of the present invention provide a method for accurately positioning prostate cancer lesions based on a fusion navigation technology. This method can be implemented by a device for accurately positioning prostate cancer lesions based on a fusion navigation technology. The device for accurately positioning prostate cancer lesions based on a fusion navigation technology can be a terminal or a server. The processing flow of the method for accurately positioning prostate cancer lesions based on a fusion navigation technology can include the following steps: S1: Collect multiple prostate images.

[0023] In a possible implementation, the prostate image is an MRI image, a CT image, an ultrasound image, a PET image, or a prostate image collected in real time by a probe.

[0024] In a possible implementation, after S1, it further includes: Perform preprocessing on the prostate image, including contrast enhancement, image denoising, and image intensity normalization.

[0025] It should be noted that contrast enhancement can improve the visibility of the region of interest in the image, especially the contrast between the prostate and cancerous tissues. Image denoising can reduce image noise and improve image quality. Due to differences in image intensity that may be caused by different devices or different scanning parameters, image intensity normalization can help unify the intensity ranges of different images, making model training and testing more consistent.

[0026] S2: Select control points from the prostate image in combination with a sparse regularization algorithm.

[0027] Among them, the sparse regularization algorithm is a mathematical method for image registration. Its purpose is to optimize the selection and use of control points in the image registration process. Sparse regularization encourages the algorithm to reduce the number of non-zero parameters by adding a regularization term, which means selecting fewer but key control points in practical applications. The result of this is to reduce the computational burden and accelerate the entire image registration process while maintaining a high-precision registration result. Control points in image registration refer to the key points used to guide image deformation. In the scenario of prostate image registration, control points are usually selected from the significant feature positions of the image, such as organ boundaries or specific anatomical structures. These points serve as the basis for the deformation field, guiding the entire image to deform according to the reference or fixed image to achieve precise overlap and alignment. Through the sparse regularization algorithm, the system only selects the most influential control points, thus avoiding overly complex calculations and improving efficiency, which is particularly important for real-time or resource-constrained medical environments.

[0028] It should be noted that by introducing the sparse regularization technique in the optimization process, the number of control points that actually need to be optimized can be effectively reduced, thereby reducing the computational complexity of subsequent registration. By constructing an optimization objective function including mutual information and a sparse regularization term, high-precision registration and fusion of the images are achieved. This method not only improves the similarity between images but also reduces the number of control points by introducing sparse regularization, reducing the computational complexity, thus maximizing the alignment accuracy of the images while maintaining the rationality of deformation and greatly improving the accuracy of lesion localization.

[0029] In a possible implementation, S2 specifically includes: S201: Establish an objective function including the mutual information value between prostate images and a sparse regularization term: ; wherein, represents the objective function value under the deformation field guided by the control points , represents the moving prostate image, represents the moving prostate image after being deformed by the deformation field, represents the reference prostate image and the fidelity between the deformed moving prostate image, R represents the sparse regularization term including the L1 regularization term and the L2 regularization term, represents the regularization parameter for controlling the sparsity intensity, wherein the data fidelity is the mutual information value.

[0030] wherein, the deformation field is used to describe how to deform the moving prostate image for alignment with the reference image.

[0031] When attempting to "minimize" the entire objective function in the optimization algorithm, in fact, an attempt is being made to find a balance point that can both maximize the similarity between the two images, i.e., maximize the mutual information value, and maintain the rationality of the transformation. This balance is achieved by adjusting the value of, ensuring that the transformation is neither overly simplified nor overly adapted to noise or irrelevant details.

[0032] It should be noted that minimizing the difference between images does not simply mean making the two images look visually similar, but rather adjusting these images so that their corresponding points in the same coordinate system match as closely as possible. Minimizing the difference between images actually ensures that in the common coordinate system, the corresponding parts of all images are superimposed as precisely as possible. To avoid information inconsistencies caused by the time or space acquisition process.

[0033] S202: Introduce auxiliary variables, solve the objective function with the goal of minimizing the objective function value, and select the non-zero value auxiliary variables as the selected control points: ; wherein, ρ represents the penalty coefficient, z represents the auxiliary variable, k represents the number of iteration steps, represents the Lagrange multiplier, represents after k +1 number of iteration steps to obtain the auxiliary variable, represents after k +1 number of iteration steps to obtain the Lagrange multiplier, represents the L1 norm, represents the square of the L2 norm, represents the soft threshold operator, Denotes the sign function, Denotes taking the variable The absolute value of minus And the larger value compared with 0.

[0034] It should be noted that the process of updating the auxiliary variable updates the auxiliary variable by minimizing the regularized L1 norm and the difference between the adjusted auxiliary quantity and the new deformation field to make it sparser. The process of updating the Lagrange multiplier ensures better alignment of the deformation field and the auxiliary variable in the next iteration. The introduction of the soft threshold operation function compresses or sets it to zero while maintaining x The original symbol. The main purpose of the soft threshold operation is to achieve and enhance sparsity, that is, to make many components of the variable become zero. In image registration, especially when using a deformation model, not all control points are necessary to achieve high-quality and low-complexity registration. The soft threshold can effectively reduce the number of parameters that need to be adjusted in the model, thereby simplifying the model and reducing the risk of overfitting, and increasing the positioning speed and accuracy.

[0035] S3: Determine the optimal deformation field according to the control points, and perform elastic registration on the prostate image based on the optimal deformation field.

[0036] Among them, the optimal deformation field is a mathematical model used to describe how to non-rigidly deform an image (moving image) through a spatial transformation so as to achieve the best alignment with another image (fixed image, also known as reference image). This deformation field indicates how each pixel or voxel position in the moving image should move to match the corresponding position of the fixed image. The optimal deformation field is obtained by minimizing an objective function, which usually includes a data fidelity term (such as the similarity between images) and a regularization term (such as smoothness or sparsity). The goal is to find a deformation field that can both maintain the authenticity of the image content and ensure the smoothness and physical feasibility of the deformation. Elastic registration is an image registration technique used to handle non-rigid (i.e., non-linear) deformations between images, which may be caused by the natural movement of the target object (such as breathing or heartbeat) or mechanical operations (such as movement during a medical operation). Elastic registration not only adjusts the position and rotation of the image, but also adjusts the local scaling and distortion of the image, so that the corresponding structures in the image are as accurately matched as possible. It can help doctors more accurately compare and analyze the anatomical structures that change over time, so as to perform more accurate diagnosis and treatment planning. For example, in the treatment of prostate cancer, elastic registration can help accurately locate the tumor position and provide important spatial positioning information for surgery or radiotherapy.

[0037] In a possible implementation manner, S3 specifically includes: S301: Update the deformation field with the goal of minimizing the weighted sum of the fidelity and the sparse regularization term to obtain the optimal deformation field: ; Wherein, represents the deformation field obtained after k +1 iteration steps, represents the L2 norm, represents the convergence threshold.

[0038] S302: Determine the reference prostate image from the prostate images, and the remaining ones are the moving prostate images.

[0039] S303: Map each point in each moving prostate image to the corresponding point in the reference prostate image according to the optimal deformation field, so as to reposition each pixel point in the moving prostate image and complete the elastic registration of each prostate image.

[0040] Specifically, the result obtained through the deformation field mapping is a deformed image, which is a version of the original moving image adjusted by the deformation field mapping. This deformed image should be closer or aligned with the fixed image in terms of spatial position and structure. This alignment makes the corresponding features or structures of the two images in the same or similar positions, thus achieving the purpose of registration. The main feature of the deformed image is that its spatial structure is adjusted to match the fixed image. In medical imaging, this means that the scan results taken at different time points of the patient are adjusted so that the corresponding anatomical structures (such as organs, bones, etc.) are aligned. The deformed image obtained through the deformation field mapping can provide visually more accurate spatially aligned information.

[0041] It should be noted that the process of updating the deformation field is to update the deformation field by minimizing the data fidelity term and the difference between the adjusted deformation field and the auxiliary variable. Minimizing the data fidelity term actually means "minimizing the dissimilarity between the two images" or "maximizing the similarity between the two images". Update the deformation field by minimizing the weighted sum of the data fidelity term and the sparse regularization term to ensure that the deformation field can accurately map the moving image to the fixed image, while maintaining the smoothness and simplicity of the deformation field, optimizing the deformation field, and ensuring that the distribution of the control points can cover the important features in the image without being overly dense resulting in waste of computing resources.

[0042] It should be noted that those skilled in the art can set the size of the convergence threshold according to actual needs, and the present invention does not limit it here.

[0043] S4: Perform differential multi-resolution fusion on the registered prostate images to reconstruct the three-dimensional prostate model.

[0044] Among them, the advantage of differential multi-resolution fusion is that it can effectively combine information from multiple sources at different resolution levels, retain key details such as edges and textures, while reducing noise and unnecessary details.

[0045] It should be noted that the reconstruction of the three-dimensional model is very important. The three-dimensional model can provide detailed information about the size, shape and tumor location of the prostate, assisting doctors in making more accurate diagnoses. Before performing prostate surgery or other treatments, a detailed three-dimensional model can help doctors plan the surgical path and reduce risks. The three-dimensional models used for pre- and post-contrast can evaluate the treatment effects, such as the reduced tumor volume and changed tissue structure. Using this differential multi-resolution fusion, while maintaining the key information of the image, the computational amount is effectively reduced. By decomposing the image layer by layer and performing fusion at different resolution levels, it can effectively retain the detailed information from different modalities, improve the fusion accuracy, and further improve the positioning accuracy.

[0046] In a possible implementation manner, S4 specifically includes: S401: Reconstruct the Laplacian pyramid in combination with an approximation algorithm.

[0047] ; Among them, I represents the input image, that is, the registered prostate image, represents the l -th layer of the image after Gaussian filtering, and respectively represent the standard deviations of the adaptive Gaussian kernels of the l -th layer and the l +1-th layer, and respectively represent Gaussian kernels with standard deviations of and , represents the prostate image of the l -th layer obtained after Gaussian filtering, represents the upsampling of the prostate image of the l +1-th layer to the size of the l -th layer, represents the convolution operation, represents the l -th layer of the Laplacian component obtained by performing differential calculation on each layer of the prostate image. Among them, the Laplacian components of each layer form the Laplacian pyramid.

[0048] Among them, the Laplacian component captures image details by recording the difference information between resolution levels. During downsampling and Gaussian blurring, high-frequency details of the image (such as edges, textures, etc.) may be lost, and each layer of the Laplacian pyramid stores these detail differences. Using this approximation method, while maintaining the key information of the image, the number of layers and computational complexity of constructing the Laplacian pyramid are effectively reduced.

[0049] S402: Perform multi-resolution fusion on the registered prostate images using the Laplacian pyramid obtained by reconstruction to obtain a prostate fusion image.

[0050] Specifically, first, construct a Laplacian pyramid for each image to be fused. In this process, first create a Gaussian pyramid (by repeated Gaussian blurring and downsampling), and then calculate the difference between adjacent Gaussian layers to form Laplacian layers. Using this approximation method, while maintaining the key information of the image, the number of layers and computational complexity of constructing the Laplacian pyramid are effectively reduced.

[0051] S403: Dynamically voxelize the prostate fusion image based on image information entropy.

[0052] ; Among them, represents the voxel size of the prostate fusion image at the coordinate ( x , y ), β represents the adjustment coefficient, θ represents the information entropy threshold, exp represents the natural exponential function, represents the image information entropy value at the coordinate ( x , y ).

[0053] Among them, β is a positive coefficient that controls the steepness of the curve. In the formula, it controls the sensitivity and speed of the voxel size change when the information entropy deviates from its threshold. A larger value will make the curve steeper, that is, the voxel size is more sensitive to small changes in the information entropy.

[0054] It should be noted that the resolution setting in the voxelization process has a great impact on the details of the final model. If the voxels are too large, it may lead to loss of details; if they are too small, it may result in excessive noise and increase the processing difficulty. Dynamically adjust the voxel size according to the importance and degree of change of the image content. Use smaller voxels in areas with rich details or key regions, and larger voxels in relatively uniform or less important regions. By intelligently adjusting the voxel size, it is possible to effectively balance the retention of image details and the use of computing resources, thereby improving the overall efficiency of image processing and analysis. Specifically, adjust the voxel size according to the information entropy. The information entropy of the lesion area is significantly different from that of the normal area, so that the voxel size is smaller in information-dense regions (such as the structures around the prostate or the lesion area) to capture more details. While in information-sparse regions (such as the background or uniform regions), the voxel size is larger to reduce the computational complexity. This method is applicable to multi-scale voxelization in image processing, especially when performing 3D reconstruction or advanced image analysis, which can effectively balance the retention of details and processing efficiency. The accurate division of voxel points can fully extract image information and improve the accuracy of subsequent 3D model reconstruction.

[0055] It should be noted that those skilled in the art can set the size of the information entropy threshold according to actual needs, and the present invention does not make any limitations in this regard.

[0056] S404: Reconstruct the 3D model of the prostate through the marching cubes algorithm based on the voxelized prostate fusion image.

[0057] Among them, the marching cubes algorithm is a commonly used 3D surface reconstruction algorithm that constructs a 3D surface from voxel data. The algorithm traverses all voxels and determines the path of the surface passing according to the threshold on the boundary of the voxels.

[0058] Specifically, first, by using differential multi-resolution fusion and reconstructed Laplacian pyramids, the process effectively captures the details lost at each resolution level, retains the key image information, and at the same time significantly reduces the computational requirements and processing time. Further dynamic voxelization is based on image information entropy, allowing the system to dynamically adjust the voxel size according to the local content density of the image. In this way, more information can be retained in areas with rich details, while in information-sparse areas, the storage and processing speed are optimized. Finally, 3D reconstruction is performed through the marching cubes algorithm, which can accurately reconstruct the complex 3D shape of the prostate from the voxel data, providing accurate diagnostic information and surgical planning support for doctors. The whole process not only improves the accuracy of diagnosis and treatment, but also significantly enhances the efficiency of the workflow through intelligent data processing. This method is particularly suitable for situations that require processing a large amount of or complex medical image data, optimizing the entire chain from image acquisition to final diagnosis.

[0059] Refer to the attached Figure 2, showing a schematic structural diagram of a lesion dynamic extraction model provided by an embodiment of the present invention.

[0060] S5: Construct a lesion dynamic extraction model including a three-dimensional convolutional network, a pooling layer, a region of interest layer, and a bounding box regression layer connected in sequence.

[0061] Among them, the three-dimensional convolutional network can capture information in the depth direction of the image and is used to extract spatial features from volume data. The pooling layer is mainly used to reduce the spatial dimension of the output of the convolutional layer, thereby reducing the computational amount and increasing the abstraction level of the features. In the three-dimensional convolutional network, the pooling layer is also executed in a three-dimensional form to maintain the three-dimensional structure of the data. The region of interest layer is used to accurately extract predetermined regions of interest from complex backgrounds, and these regions may contain lesions or other important medical features. The task of the bounding box regression layer is to accurately locate the position and size of the object in the image. In medical image processing, this means being able to accurately mark the specific position and boundary of the lesion. By extracting key features from the original image, reducing the dimension, and locating the lesion, valuable diagnostic information is provided for clinical use. Through such a model, doctors can obtain a detailed view of the location and shape of the lesion, thereby making more accurate judgments in surgical planning and treatment decisions.

[0062] For example, the lesion dynamic extraction model can first process the input image through a three-dimensional convolutional layer with 32 output channels and a 3x3x3 convolutional kernel, with a stride of 1 and padding of 1 to keep the spatial dimensions unchanged. Then, a 2x2x2 max pooling layer is used for downsampling with a stride of 2 and no padding. After that, the image passes through a second three-dimensional convolutional layer with 64 output channels and the same-sized convolutional kernel, and then passes through a pooling layer with the same configuration again. Finally, an ROI alignment layer is used to accurately extract and align the regions of interest, providing a high-quality feature representation for subsequent lesion localization and analysis. This model combines the powerful capabilities of deep learning and the requirements of three-dimensional image analysis, ensuring high precision and efficiency. Finally, a bounding box regression layer is introduced as the terminal layer of the model, specifically for accurately locating the spatial coordinates of the lesion. This layer receives the features from the region of interest layer and outputs the specific position and size of the lesion, including the three-dimensional coordinates (x, y, z) and the corresponding dimensions.

[0063] In a possible implementation manner, the lesion dynamic extraction model further includes a region proposal network layer, and the region proposal network is respectively connected to the pooling layer and the region of interest layer.

[0064] It should be noted that integrating the Region Proposal Network (RPN) into the lesion dynamic extraction model and connecting it with the pooling layer and the Region of Interest (ROI) layer greatly improves the object detection efficiency and accuracy of the model. By automatically generating high-quality region proposals, RPN directly affects the feature extraction of the ROI layer, thus effectively focusing on the most relevant parts of the image, reducing background noise interference, optimizing the use of computing resources, and enhancing the overall model performance. This design not only accelerates the processing process but also enhances the application adaptability and scalability of the model in complex medical image tasks through precise region localization.

[0065] S6: Establish the loss function of the lesion dynamic extraction model.

[0066] Among them, the loss function includes the Intersection over Union (IoU) loss term and the Distributional Focal Loss term.

[0067] Among them, the IoU loss term is a commonly used performance metric for evaluating the overlap between the predicted bounding box and the ground truth bounding box. The IoU loss term value is calculated as the ratio of the intersection of the two bounding boxes to their union. The IoU loss term value is 1 minus this ratio, aiming to maximize the overlapping area between the predicted box and the ground truth box. In medical image analysis, this loss term helps the model to be more accurate when generating the lesion localization bounding box, reducing errors and improving the accuracy of localization. The Distributional Focal Loss term is an advanced loss function that focuses on optimizing the probability distribution of the model prediction. In implementation, it is usually used to optimize the distribution difference between the prediction result and the ground truth value. Especially when the gap between the predicted value and the actual value is large, the Distributional Focal Loss term increases the model's attention to these regions, thereby improving the model's learning efficiency and the accuracy of the results, helping the model to focus on the precise location and size of the lesion. Especially in the detection of lesions with unclear boundaries or irregular shapes, it can better adapt to the distribution characteristics of the target. The combined use of the two loss functions can significantly improve the performance of the lesion detection model in medical images, especially in terms of accuracy and robustness, enabling the model to not only accurately locate the lesion but also effectively handle various variations and abnormal cases.

[0068] In a possible implementation, the loss function L is specifically: ;

[0069] Among them, represents the IoU loss term, represents the Distributional Focal Loss term, A and B represent the predicted box and the ground truth box respectively, represent the center points of the predicted box and the ground truth box respectively, represent the widths of the predicted box and the ground truth box respectively, respectively represent the height of the predicted bounding box and the height of the ground truth bounding box, represents the normalized distance between the center point of the predicted bounding box and the center point of the ground truth bounding box, v represents the penalty term for shape inconsistency, represents the overlap loss weight factor, c represents the diagonal length of the minimum closed region between the predicted bounding box and the ground truth bounding box, also known as the normalization factor, and arctan represents the arctangent function, π represents pi, y represents the ground truth label, y i and y i+1 respectively represent the i th predicted label and the i +1 th predicted label that are closest to the ground truth label, represents the relative approximation of the ground truth label to the i +1 th predicted label, represents the relative approximation of the ground truth label to the i th predicted label.

[0070] S7: Under the supervision of the loss function, train the lesion dynamic extraction model.

[0071] In a possible implementation manner, S7 specifically includes: S701: Obtain a prostate image sample data set with lesion point labels.

[0072] Among them, each image in the prostate image sample set is also a three-dimensional image, such as MRI image, CT image, ultrasound image, PET image, etc. However, this kind of image is a three-dimensional reconstruction through this solution, and the lesion point features are not very accurate. The specific positions and sizes of the lesion points in these images are manually labeled, and the annotation information can be directly selecting the lesion points by a bounding box.

[0073] S702: Input the sample data set as the training set into the lesion dynamic extraction model, and train the lesion dynamic extraction model until the loss function value is less than the preset loss function value.

[0074] It should be noted that during the training process, under the supervision of the loss function, the features learned by the model can fully learn the features of the lesion points in the training set according to the requirements of the loss function. After training, the model can accurately identify the lesion point features and can quickly and automatically select and mark the lesion points on a more accurate prostate three-dimensional model, thereby completing the accurate and automatic localization of the lesion points.

[0075] It should be noted that those skilled in the art can set the size of the preset loss function value according to actual needs, and the present invention does not limit this here.

[0076] S8: Locate the lesion points in the prostate three-dimensional model through the trained lesion dynamic extraction model.

[0077] It should be noted that this process simplifies the processes with complex data processing and high computational resource consumption in medical images, reduces resource consumption, improves the localization speed and accuracy of prostate cancer lesion localization, enables the obtained lesion points to be applied to real-time intraoperative navigation, and expands the application scope. The three-dimensional convolutional network in the prostate three-dimensional model can directly extract spatial features from the three-dimensional model, and then convert the three-dimensional data into a two-dimensional data plane to adapt to the features learned from two-dimensional images during the training process, and accurately locate the lesion points in the prostate three-dimensional model while retaining the depth information.

[0078] In a possible implementation manner, after S8, it further includes: Collect the accuracy rate of the located lesion points. When the accuracy rate of the lesion points is less than the preset accuracy rate of the lesion points, return to step S7.

[0079] It can be understood that the process of collecting the accuracy rate of lesion point localization and returning to step S7 when the accuracy rate does not meet the expectation is a quality control and iterative optimization mechanism. It can ensure that the model meets or exceeds the expected performance standards in actual applications, and avoid diagnostic errors or treatment planning mistakes caused by insufficient model performance. This not only enhances the practicability of the model, but also improves the reliability and trust of the entire system.

[0080] It should be noted that those skilled in the art can set the size of the preset accuracy rate of the lesion points according to actual needs, and the present invention does not limit this here.

[0081] The beneficial effects brought by the technical solution provided by the embodiments of the present invention at least include: In the present invention, the sparse regularization algorithm is combined to selectively determine the control points with greater influence, and based on these control points, the optimal deformation field is determined, thereby completing the elastic registration of each prostate image. This process reduces the complexity of the registration process while retaining the important control points, excluding the registration interference of abnormal control points, and improving the registration speed and accuracy. After that, differential multi-resolution fusion is performed on the registered images, which retains the detailed content of the images while reducing the complexity of detail extraction. The prostate three-dimensional model reconstructed based on this data has a fast reconstruction speed, high clarity, and can accurately reflect the location of the lesion points. Moreover, it also makes up for the shortcoming of the traditional three-dimensional model reconstruction process consuming a large amount of computing resources, effectively reducing the resource consumption during the reconstruction process, enabling the reconstruction process of the prostate three-dimensional model to be applied to preoperative reference and intraoperative navigation, improving the application scenario and practicality. Based on the improvement of the construction speed of this three-dimensional model, the lesion dynamic extraction model can dynamically lock the lesion points through the region of interest layer and the bounding box regression layer, and automatically and accurately output the location of the lesion points with depth features under the three-dimensional structure. The introduction of the loss function with the intersection over union loss term and the focal loss term conforms to the characteristic that the true distribution of the lesion points usually is not too far from the labeled position, and the model should focus on the values near the labeled position during training, enabling the lesion dynamic extraction model to accurately learn the lesion features, further improving the positioning accuracy of the lesion points, and being able to efficiently assist the patient treatment process.

[0082] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

[0083] The following points need to be explained: (1) The attached drawings of the embodiments of the present invention only relate to the structures involved in the embodiments of the present invention, and other structures can refer to the general design.

[0084] (2) For clarity, in the attached drawings used to describe the embodiments of the present invention, the thickness of the layers or regions is enlarged or reduced, that is, these drawings are not drawn according to the actual ratio. It can be understood that when an element such as a layer, film, region, or substrate is referred to as being "on" or "under" another element, the element can be "directly" on or under the other element or there can be an intermediate element.

[0085] (3) Without conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other to obtain new embodiments.

[0086] The above are only the specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. The protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for accurately locating prostate cancer lesions based on fusion navigation technology, characterized in that: Methods include: S1: Acquire multiple prostate images; S2: selecting control points from the prostate image in combination with a sparse regularization algorithm; S3: determining an optimal deformation field according to the control points, and elastically registering the prostate image based on the optimal deformation field; S4: Perform differential multi-resolution fusion on the registered prostate images to reconstruct the prostate 3D model; S5: construct a lesion dynamic extraction model including a three-dimensional convolutional network, a pooling layer, an interest layer and a bounding box regression layer connected in sequence; S6: establishing a loss function of the dynamic lesion extraction model, wherein the loss function includes an intersection-over-union loss term and a distributed focus loss term; S7: Under the supervision of the loss function, training the lesion dynamic extraction model; S8: Locating the lesion point in the prostate three-dimensional model through the trained lesion dynamic extraction model.

2. The method for accurately locating prostate cancer lesions based on fusion navigation technology according to claim 1, characterized in that: The prostate image is an MRI image, a CT image, an ultrasound image, a PET image, or a prostate image acquired in real time by a probe.

3. The method for accurately locating prostate cancer lesions based on fusion navigation technology according to claim 1, characterized in that: After S1, the method further includes: The prostate image is preprocessed including contrast enhancement, image denoising and image intensity normalization.

4. The method for accurately locating prostate cancer lesions based on fusion navigation technology according to claim 1, characterized in that: The S2 specifically includes: S201: establishing an objective function including a mutual information value between prostate images and a sparse regularization term; S202: Introduce auxiliary variables to solve the objective function with the goal of minimizing the objective function value, and select auxiliary variables with non-zero values ​​as selected control points.

5. The method for accurately locating prostate cancer lesions based on fusion navigation technology according to claim 4, characterized in that: The S3 specifically includes: S301: updating the deformation field with the goal of minimizing the weighted sum of fidelity and sparse regularization terms to obtain the optimal deformation field; S302: determining a reference prostate image from the prostate images, and the remaining ones are moving prostate images; S303: Mapping each point in each moving prostate image to a corresponding point in the reference prostate image according to the optimal deformation field, so as to reposition each pixel point in the moving prostate image and complete elastic registration of each prostate image.

6. The method for accurately locating prostate cancer lesions based on fusion navigation technology according to claim 1, characterized in that: The S4 specifically includes: S401: Reconstruct the Laplacian pyramid by combining the approximation algorithm; S402: performing multi-resolution fusion on the registered prostate image through the reconstructed Laplacian pyramid to obtain a prostate fusion image; S403: Dynamically voxelizing the prostate fusion image based on image information entropy; S404: Reconstructing the prostate three-dimensional model by using a marching cubes algorithm according to the voxelized prostate fusion image.

7. The method for accurately locating prostate cancer lesions based on fusion navigation technology according to claim 1, characterized in that: The lesion dynamic extraction model also includes a region proposal network layer, and the region proposal network is connected to the pooling layer and the layer of interest respectively.

8. The method for accurately locating prostate cancer lesions based on fusion navigation technology according to claim 1, characterized in that: The S7 specifically includes: S701: Acquire a prostate image sample dataset with lesion point labels; S702: Input the sample data set as a training set into the dynamic lesion extraction model, and train the dynamic lesion extraction model until the loss function value is less than a preset loss function value.

9. The method for accurately locating prostate cancer lesions based on fusion navigation technology according to claim 1, characterized in that: After S8, the method further includes: The accuracy of the lesion point located is collected. When the accuracy of the lesion point is less than the preset accuracy of the lesion point, the process returns to step S7.

Images