DBS electrode implantation path planning method based on deep learning

Abstract

The invention provides a DBS electrode implantation path planning method based on deep learning. The DBS electrode implantation path planning method comprises the steps of collecting a to-be-processed image of a to-be-operated person; segmenting a to-be-processed region in the to-be-processed image through the first deep network to obtain a first mask; performing placeholder identification on nerve fibers and blood vessels in the to-be-processed image through a second deep network to obtain a second mask; and pre-training a path planning network, and inputting the first mask and the second mask into the path planning network to obtain an electrode implantation planning path. According to the method, the nucleus subregion structure which is difficult to segment can be accurately adjusted, the segmentation precision is improved, the implantation precision of a DBS operation is improved, an individual STN subregion structure of a patient can be provided, and planning of an operation path can be guided.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and in particular to a DBS electrode implantation path planning method based on deep learning. Background Art

[0002] During deep brain stimulation (DBS) surgery, the target nuclei need to be located based on the patient's preoperative MR and other imaging data. The subthalamic nucleus (STN) is the preferred target for DBS treatment of Parkinson's disease. However, the ambiguity of the STN subdivision anatomical structure, individual disease differences, and uncertainty in efficacy pose new challenges to current DBS treatment technology.

[0003] Accurate implantation of DBS electrodes into sub-regions is crucial to maximizing the therapeutic effect while minimizing side effects. The target confirmation of the imaging data of the current MR+CT positioning system needs to be done manually, but this process depends largely on the image resolution and the doctor's experience. It is often difficult for the doctor to distinguish the sub-regions with the naked eye alone.

[0004] In addition, the process of planning the optimal implantation path in most cases still relies on the clinician's experience to manually avoid important blood vessels, brain grooves and ventricles. This process increases the operation time, and planning deviations may also lead to side effects such as movement disorders and the risk of complications such as bleeding. Summary of the invention

[0005] The present invention aims to solve at least one of the technical problems existing in the related art. To this end, the present invention provides a DBS electrode implantation path planning method based on deep learning.

[0006] The present invention provides a DBS electrode implantation path planning method based on deep learning, comprising: S1: Segmenting the region to be processed in the image to be processed by a first deep network to obtain a first mask; S2: performing space-occupancy identification on nerve fibers and blood vessels in the image to be processed by a second deep network to obtain a second mask; S3: pre-training a path planning network, inputting the first mask and the second mask into the path planning network, and obtaining an electrode implantation planning path.

[0007] According to a deep learning-based DBS electrode implantation path planning method provided by the present invention, the area to be operated on in step S1 is the subthalamic nucleus.

[0008] According to a deep learning-based DBS electrode implantation path planning method provided by the present invention, before step S1, the method further includes: S111: collecting images of the patient to be operated on; S112: marking the two-dimensional region to be processed on the image to be processed.

[0009] According to a DBS electrode implantation path planning method based on deep learning provided by the present invention, the pre-training process of the first deep network in step S1 includes: S121: training a two-dimensional segmentation network using the image to be processed with the two-dimensional region to be processed labeled; S122: Segmenting the image to be processed into a region to be operated on based on the two-dimensional segmentation network to obtain a two-dimensional segmented image; S123: marking sub-regions of the area to be operated on in the preliminary segmented image by using diffusion tensor imaging technology; S124: segmenting the two-dimensional segmentation image by marking sub-regions of the area to be operated on, and obtaining a sub-structure image; S125: reconstructing the continuous substructure images into a three-dimensional image according to the data space information corresponding to the sub-region annotations of the area to be operated on; S126: Training a three-dimensional segmentation network using the three-dimensional image to obtain a pre-trained first deep network.

[0010] According to a DBS electrode implantation path planning method based on deep learning provided by the present invention, the loss function of the first deep network in step S1 is expressed as:

[0011]

[0012]

[0013] in, is the total loss of the first deep network, is the focal loss, For Tversky's loss, is the number of subdivision categories of the area to be operated on, is the category index value of the area to be segmented. is the total number of voxels in the input image, is the voxel index value of the input image, Voxel The true value of The fine segmentation category is Voxels The predicted value of is the balance correction factor, Indicates that the fine segmentation category is correctly identified as , Indicates that the fine segmentation category is The judgment of was missed. Indicates that the fine segmentation category is misclassified as , The fine segmentation category is The penalty coefficient for discriminating missed reports, The fine segmentation category is misclassified as The penalty coefficient.

[0014] According to a DBS electrode implantation path planning method based on deep learning provided by the present invention, step S2 specifically includes: S21: obtaining a nerve fiber mask and a blood vessel mask from the image to be processed through the second deep network recognition; S22: superimposing the nerve fiber mask and the blood vessel mask to obtain a placeholder mask, wherein 1 indicates the presence of a placeholder voxel, and 0 indicates the absence of a placeholder voxel; S23: Reconstruct the placeholder mask through the second deep network to obtain a second mask.

[0015] According to a DBS electrode implantation path planning method based on deep learning provided by the present invention, the second deep network in step S2 is a CNN network established based on an occupancy network architecture, and in step S23, the placeholder mask is reconstructed through the occupancy network.

[0016] According to a DBS electrode implantation path planning method based on deep learning provided by the present invention, step S23 further includes: S231: performing boundary point inference on the placeholder mask using any voxel point in the placeholder mask by using a marching cubes algorithm to obtain a placeholder boundary point; S232: After enlarging the resolution of the placeholder boundary points, perform boundary point reasoning until the resolution is enlarged to a preset multiple, thereby obtaining a second mask.

[0017] According to a DBS electrode implantation path planning method based on deep learning provided by the present invention, the expression of the loss function for boundary point reasoning in step S231 is:

[0018] in, is the loss of boundary point reasoning, is the training batch size, is the network for boundary point reasoning, are the model parameters of the network used for boundary point inference, is the output dimension, is the starting pixel for boundary point reasoning, is the observation point in the cube centered at the starting pixel. for occupancy status.

[0019] According to a DBS electrode implantation path planning method based on deep learning provided by the present invention, the pre-training process of the path planning network in step S3 specifically includes: S31: Obtaining a planning entry point and a planning target point by annotating the image to be processed; S32: Migrating weight parameters of the first deep network to the path planning network; S33: Performing migration training on the path planning network based on the weight parameters of the first deep network through the three-dimensional coordinates of the planning entry point and the three-dimensional coordinates of the planning target point.

[0020] The present invention provides a DBS electrode implantation path planning method based on deep learning, which is a deep learning-based system for subthalamus nucleus (STN) subdivision and surgical path planning during deep brain stimulation (DBS) electrode implantation. The method can provide a patient-specific STN subdivision structure and guide the planning of the surgical path. The present invention is based on deep learning, multimodal imaging technology and individualized fiber bundle nucleus reconstruction technology, and uses an optimized U-Net model to segment key brain structures, thereby achieving accurate segmentation of each subregion of the STN nucleus, and by combining a feature extraction method for a region of interest, accurately adjusts the structure of the nucleus subregion that is difficult to segment, thereby improving the segmentation accuracy and the implantation accuracy of the DBS surgery; secondly, an intelligent navigation planning architecture is also provided, which can accurately avoid blood vessels, cerebral sulci and ventricles on the path, and achieve intelligent STN nucleus subregion accurate implantation, providing key technical support for DBS electrode implantation planning, and reducing DBS implantation side effects and complications caused by experience factors.

[0021] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0023] Figure 1It is a flowchart of a DBS electrode implantation path planning method based on deep learning provided by an embodiment of the present invention; Figure 2 It is a schematic flow chart of a pre-training method of a first deep network provided in an embodiment of the present invention; Figure 3 is a schematic flow chart of a method for obtaining a second mask provided by an embodiment of the present invention; Figure 4 It is a schematic diagram of a process flow of a pre-training method for a path planning network provided by an embodiment of the present invention; Figure 5 It is a schematic diagram of the principle of a DBS electrode implantation path planning method based on deep learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0025] Combine the following Figures 1 to 5 Embodiments of the present invention are described.

[0026] like Figure 1 As shown, the present invention provides a DBS electrode implantation path planning method based on deep learning, comprising: S1: Segmenting the region to be processed in the image to be processed by a first deep network to obtain a first mask; The purpose of step S1 is to train the encoder and decoder of 3D U-net so that 3D U-net can input the image to be processed and the annotation, and output the STN sub-class segmentation mask - Mask STN, which is the first mask mentioned above. The area to be operated on in step S1 is the subthalamic nucleus.

[0027] Furthermore, the above-mentioned images to be processed include MRI images and CT images inputted together, such as Figure 5 As shown, Figure 5 The red parts in the MRI images and CT images marked in the middle are the marked subthalamic nucleus.

[0028] Furthermore, the purpose of the present invention is to provide a method for surgical path planning during deep brain stimulation electrode implantation. During the electrode implantation process, the subthalamic nucleus (STN) is the preferred target for DBS treatment. Subsequently, the STN subdivision structure is divided and the surgical path planning is guided.

[0029] Wherein, before step S1, the method further includes: S111: collecting images of the patient to be operated on; Furthermore, this stage provides an imaging basis for the subsequent STN subdivision structure segmentation by collecting images to be processed. MRI can provide high-resolution images, especially for soft tissues such as brain and spinal cord structures. In addition, MRI can perform multi-parameter and multi-sequence imaging, which can provide rich diagnostic information.

[0030] S112: marking the two-dimensional region to be processed on the image to be processed.

[0031] In step S112, the area to be operated on in the image, that is, the subthalamic nucleus area, needs to be manually annotated to form two-dimensional annotation data. Generally, a special image annotation tool is used to outline the boundary of the area to be operated on on the slice of the MRI image.

[0032] Furthermore, in steps S111 to S112, the original training tomographic MRI image is first obtained, and then the image is manually calibrated and verified to obtain the annotation of the two-dimensional STN area. The annotated image is used for training the deep network in the subsequent step S121.

[0033] like Figure 2 As shown, the pre-training process of the first deep network in step S1 includes: S121: training a two-dimensional segmentation network using the image to be processed with the two-dimensional region to be processed labeled; Furthermore, after steps S111 to S112, a set of MRI image data containing the target area to be operated on can be collected, and then a two-dimensional convolutional neural network (CNN) architecture suitable for MRI image segmentation is selected, and the prepared MRI images and corresponding annotated data are used as input and labels to train the two-dimensional segmentation network. Through multiple iterative training and adjusting the network parameters, the above-mentioned two-dimensional segmentation network can be obtained.

[0034] S122: Segmenting the image to be processed into a region to be operated on based on the two-dimensional segmentation network to obtain a two-dimensional segmented image; The network obtained in step S121 can learn the mapping relationship from the MRI image to the segmentation result of the area to be operated on, so in this stage, the MRI image to be segmented is input into the trained two-dimensional segmentation network, and the probability or score of each pixel or pixel block of the input MRI image belonging to the area to be operated on is calculated through the forward propagation of the network. According to the probability or score output by the network, the pixels or pixel blocks in the MRI image are divided into the area to be operated on and the non-area to be operated on using threshold processing or other post-processing methods, thereby generating a two-dimensional segmented image.

[0035] S123: marking sub-regions of the area to be operated on in the preliminary segmented image by using diffusion tensor imaging technology; Diffusion tensor imaging technology is a method based on the principle of magnetic resonance imaging (MRI) to describe and analyze the diffusion characteristics of water molecules in biological tissues. Therefore, diffusion tensor imaging technology can be used to study the structure, connection and function of white matter fiber bundles in the brain in neuroscience research. Therefore, in this stage, DTI technology is used to analyze the diffusion tensor of water molecules in the subthalamic nucleus to determine the different subregions in the area to be operated on. These subregions represent different fiber bundles or tissue types, so that the sub-divisions of the area to be operated on can be marked for fine segmentation through diffusion tensor imaging technology.

[0036] S124: segmenting the two-dimensional segmentation image by marking sub-regions of the area to be operated on, and obtaining a sub-structure image; Based on the subregion annotation of the subthalamic nucleus in step S123, the precise boundaries of each subthalamic nucleus subregion are identified to ensure the accuracy of the boundaries. Based on the identified subregion boundaries of the subthalamic nucleus, the initially segmented two-dimensional image is segmented more finely to obtain substructure images representing different subthalamic nucleus subregions. These substructure images can more clearly show the detailed structure within the subthalamic nucleus.

[0037] S125: reconstructing the continuous substructure images into a three-dimensional image according to the data space information corresponding to the sub-region annotations of the area to be operated on; In step S125, the images are first registered and corrected to ensure their accurate positions in the three-dimensional space and to ensure that all substructure images are consistent in spatial information. Then, three-dimensional reconstruction technology is used to stack continuous two-dimensional substructure images to form a three-dimensional image.

[0038] S126: Training a three-dimensional segmentation network using the three-dimensional image to obtain a pre-trained first deep network.

[0039] Furthermore, in steps S121 to S126, through the aforementioned annotation of the STN region, an automatic segmentation network based on 2DU-Net is first trained to enable the 2D U-Net network to achieve high-precision segmentation of the STN region structure of the tomographic MRI image.

[0040] After the STN regional structure segmentation, the STN sub-region segmentation standard is provided by combining clinical experience and anatomical experience. According to this standard, the STN mask generated by 2D U-Net is re-segmented to generate a sub-structure mask. Finally, the continuous tomographic sub-structure mask is reconstructed into a 3D mask based on the data space information of MRI-DTI (diffusion tensor imaging, DTI) to train the 3D U-Net network model.

[0041] The expression of the loss function of the first deep network in step S1 is:

[0042]

[0043]

[0044] in, is the total loss of the first deep network, is the focal loss, For Tversky's loss, is the number of subdivision categories of the area to be operated on, is the category index value of the area to be segmented. is the total number of voxels in the input image, is the voxel index value of the input image, Voxel The true value of The fine segmentation category is Voxels The predicted value of is the balance correction factor, Indicates that the fine segmentation category is correctly identified as , Indicates that the fine segmentation category is The judgment of was missed. Indicates that the fine segmentation category is misclassified as , The fine segmentation category is The penalty coefficient for discriminating missed reports, The fine segmentation category is misclassified as The penalty coefficient.

[0045] In the above formula, the total loss function is the sum of focal loss (Focal LOSS) and Tversky loss (Tversky LOSS). Focal LOSS reduces the learning rate of easy-to-predict voxels in network training, and relatively improves the learning rate of difficult-to-predict voxels, such as the boundaries of fine-category partitions, in training; Tversky LOSS is an extension of DICE LOSS, which solves the problem of voxel category imbalance in segmentation.

[0046] For the first deep network in step S1, it is generally necessary to form an iterative closed loop of data, model and manual intervention. Specifically, after the 3D-Unet is trained, the verification MRI-DTI image is input, and the 3D-Unet generates a detailed three-dimensional structure of the STN. The STN three-dimensional structure generated by the manual automatic segmentation effect evaluation is returned to the 3D-Unet model after intervention for the next round of iteration, thereby realizing the iterative closed loop of data-model-manual intervention of the 3D-Unet model at work. This closed-loop and iterative training method enables the 3D-Unet to fully evolve and iterate based on the feedback of doctors' professional knowledge.

[0047] S2: performing space-occupancy identification on nerve fibers and blood vessels in the image to be processed by a second deep network to obtain a second mask; Among them, the second deep network in step S2 is a CNN network established based on an occupation network architecture, and in step S23, the placeholder mask is reconstructed through the occupation network.

[0048] In step S2, Occupancy networks are used to output the second mask. Occupancy networks are a binary classification network constructed by connecting feature extraction using a fully connected layer. It is a CNN network used to finely render discrete 3D occupancy information. The output of this network has better continuity and integrity than 3D masks in the form of voxels, points, and grids. like Figure 3 As shown, step S2 specifically includes: S21: obtaining a nerve fiber mask and a blood vessel mask from the image to be processed through the second deep network recognition; First, you need to prepare the labeled image data to be processed, including the label information of nerve fibers and blood vessels, as the training set of the deep network. Then the deep network model trains and processes the training set. The model will output two masks: one represents nerve fibers (nerve fiber mask) and the other represents blood vessels (blood vessel mask). The value of each pixel or voxel represents the probability that the location belongs to a nerve fiber or blood vessel.

[0049] S22: superimposing the nerve fiber mask and the blood vessel mask to obtain a placeholder mask, wherein 1 indicates the presence of a placeholder voxel, and 0 indicates the absence of a placeholder voxel; Furthermore, the nerve fiber mask and blood vessel mask obtained in step S22 are superimposed by performing a voxel-by-voxel logical connection operation on the two mask images, that is, if the value of a certain position in the nerve fiber mask or the blood vessel mask is 1 (indicating existence), then the value of the position in the placeholder mask is also 1. In the obtained placeholder mask, a pixel or voxel with a value of 1 indicates that a nerve fiber or blood vessel (i.e., a placeholder voxel) exists at the position, and a pixel or voxel with a value of 0 indicates that a nerve fiber or blood vessel does not exist at the position.

[0050] S23: Reconstruct the placeholder mask through the second deep network to obtain a second mask.

[0051] In step S23, the placeholder mask obtained in steps S21 to S22 is used as input and forward propagated through the trained second deep network. The network will output a reconstructed mask, i.e., a second mask. The second mask should contain more details and more accurate information than the original placeholder mask.

[0052] Furthermore, in step S21 to step S23, the nerve fiber 3D mask MASK A and the blood vessel 3D mask MASK B are superimposed through the CNN network-based placeholder mask generation system, the presence of the mask voxel is represented as 1, and the absence of the voxel is represented as 0, and then the occupancy 3D mask MASK C can be output, and finally MASK C is input into the 3D occupancy networks to reconstruct the rendered fine occupancy 3D mask MASK OCCUPIED.

[0053] Among them, the above-mentioned existence mask voxel indicates that there are space-occupying lesions in the nerve fibers and blood vessels at the voxel point. Space-occupying lesions is a special term in medical imaging diagnosis, which usually appears in the examination results of X-ray, B-ultrasound, CT, etc., and there are redundant substances in the examined part. Specifically, it indicates the overlap of nerve fibers and blood vessels at the voxel point.

[0054] Wherein, step S23 further comprises: S231: performing boundary point inference on the placeholder mask using any voxel point in the placeholder mask by using a marching cubes algorithm to obtain a placeholder boundary point; The loss function for boundary point reasoning in step S231 is expressed as:

[0055] in, is the loss of boundary point reasoning, is the training batch size, is the network for boundary point reasoning, are the model parameters of the network used for boundary point inference, is the output dimension, is the starting pixel for boundary point reasoning, is the observation point in the cube centered at the starting pixel. for occupancy status.

[0056] S232: After enlarging the resolution of the placeholder boundary points, perform boundary point reasoning until the resolution is enlarged to a preset multiple, thereby obtaining a second mask.

[0057] In the training batch, we take a voxel point P and construct a cube with the position of P as the center. Then, we use the voxels in the cube as observation information, calculate the network output predicted occupancy probability, and the true value of the voxel. Then, we use the marching cubes algorithm to use the trained network to Reasoning is performed on the occupancy 3D mask MASK C to find the boundary points of the placeholder; the boundary points are enlarged by 2 times the resolution and the grid is subdivided, and the boundary of the placeholder is continued to be reasoned and refined. The above steps are repeated until the enlargement is 32 times, that is, the refined occupancy 3D mask MASK OCCUPIED is output, which is the second mask mentioned above.

[0058] The marching cubes algorithm mentioned above is a method for representing 3D objects, or a method for volume rendering or surface reconstruction. It is an algorithm that rasterizes a three-dimensional object and then represents it using multiple cubes (voxels).

[0059] S3: pre-training a path planning network, inputting the first mask and the second mask into the path planning network, and obtaining an electrode implantation planning path.

[0060] like Figure 4 As shown, the pre-training process of the path planning network in step S3 specifically includes: S31: Obtaining a planning entry point and a planning target point by annotating the image to be processed; In step S31, the positions of the planning entry point and the planning target are first marked on the MRI image. The planning position is determined based on the doctor's diagnosis, surgical requirements or treatment plan, and then the three-dimensional coordinates of the planning entry point and the planning target are extracted from the marked MRI image.

[0061] S32: Migrating weight parameters of the first deep network to the path planning network; In step S32, the extracted weight parameters are migrated to the corresponding layer of the path planning network, so that the path planning network can use the knowledge and experience of the first deep network for subsequent training and optimization.

[0062] S33: Performing migration training on the path planning network based on the weight parameters of the first deep network through the three-dimensional coordinates of the planning entry point and the three-dimensional coordinates of the planning target point.

[0063] In this stage, the first weight parameter obtained by migration in S32 is used as the initial value to train the path planning network. During the training process, the network will learn how to start from the planning entry point, reach the planning target point, and plan the optimal or feasible path.

[0064] Furthermore, the path planning network in this stage is PLANet. During the knowledge transfer process of the path planning network, PLANnet adopts a one-layer encoder-decoder structure, in which the encoder loads the Encoder A weights of the 3D U-net in the STN fine segmentation system, and trains PLANnet based on the labeled preoperative planning entry points and planning target 3D coordinates.

[0065] After obtaining the path planning network, the first mask and the second mask, namely the mask MASK STN generated by the STN fine segmentation system and the 3D fine occupancy mask MASK OCCUPIED, are combined and input into the feature fusion module to output the alignment features guided by the avoidance area knowledge, namely the fusion of spatial information. Then, the path planning network is used to process the fused alignment features to obtain the optimal electrode implantation path.

[0066] In general, the principle of the DBS electrode implantation path planning method based on deep learning provided by the present invention is as follows: Figure 5 As shown, Figure 5 The colored part in the STN3D mask diagram is the reconstructed part of the subthalamic nucleus. By inputting the mask into the placeholder network, the 3D spatial collision information obtained by the placeholder network processing is fused with the original MRI image and CT image, and then input into the PLANet trained in advance, the planned implantation path can be obtained.

[0067] The above original MRI images and CT images are fused through the cross-attention mechanism. The specific expression is:

[0068] in, is the query vector, which comes from the target sequence and represents the information or features that need to be paid attention to or queried at present. is the key vector, which comes from the source sequence and is used to calculate the attention weight. is a value vector, coming from the source sequence, the actual information to be extracted or integrated, is the key vector dimension, that is, the regularization parameter, is the matrix transpose, Represents the matrix multiplication of the query vector and the key vector.

[0069] When performing specific fusion, there are different fusion methods with different dependency biases for different regional features. If the task is to let the features of MRI images guide the extraction of CT image information, then the features of MRI images may be used as query vectors, and the features of CT images as key vectors and value vectors. Conversely, if the task is to let the features of CT images guide the extraction of MRI image information, then the features of CT images are used as query vectors, and the features of MRI images are used as key vectors and value vectors.

[0070] The present invention provides a DBS electrode implantation path planning method based on deep learning, which is implemented based on a deep learning system for deep brain electrical stimulation electrode implantation planning. The system consists of a STN fine segmentation system based on 3D U-net, a fine occupancy mask generation system based on 3D Occupancy Networks architecture CNN, and an implantation path planning system including Cross-Attention and feature fusion modules.

[0071] Specifically, the STN fine segmentation system based on 3D U-net can generate the STN sub-partition fine segmentation mask MASKSTN, and the fine occupancy mask generation system based on the 3D Occupancy Networks architecture CNN can generate the occupancy mask MASK OCCUPIED of important side-effect related anatomical structures such as blood vessels and nerves. Both are input into the PLANet network to finally output the target and entry point of electrode implantation, so as to reduce the probability of complications and side effects and increase the clinical benefits of patients.

[0072] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0073] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A DBS electrode implantation path planning method based on deep learning, It is characterized in that include: S1: Segmenting the region to be processed in the image to be processed by a first deep network to obtain a first mask; S2: performing space-occupancy identification on nerve fibers and blood vessels in the image to be processed by a second deep network to obtain a second mask; S3: pre-training a path planning network, inputting the first mask and the second mask into the path planning network, and obtaining an electrode implantation planning path.

2. A DBS electrode implantation path planning method based on deep learning according to claim 1, It is characterized in that The area to be operated on in step S1 is the subthalamic nucleus.

3. A DBS electrode implantation path planning method based on deep learning according to claim 1, It is characterized in that Before step S1, the method further includes: S111: collecting images of the patient to be operated on; S112: marking the two-dimensional region to be processed on the image to be processed.

4. A DBS electrode implantation path planning method based on deep learning according to claim 3, It is characterized in that The pre-training process of the first deep network in step S1 includes: S121: training a two-dimensional segmentation network using the image to be processed with the two-dimensional region to be processed labeled; S122: Segmenting the image to be processed into a region to be operated on based on the two-dimensional segmentation network to obtain a two-dimensional segmented image; S123: marking sub-regions of the area to be operated on in the preliminary segmented image by using diffusion tensor imaging technology; S124: segmenting the two-dimensional segmentation image by marking sub-regions of the area to be operated on, and obtaining a sub-structure image; S125: reconstructing the continuous substructure images into a three-dimensional image according to the data space information corresponding to the sub-region annotations of the area to be operated on; S126: Training a three-dimensional segmentation network using the three-dimensional image to obtain a pre-trained first deep network.

5. The DBS electrode implantation path planning method based on deep learning according to claim 1, It is characterized in that The expression of the loss function of the first deep network in step S1 is: in, is the total loss of the first deep network, is the focal loss, For Tversky's loss, is the number of subdivision categories of the area to be operated on, is the category index value of the area to be segmented. is the total number of voxels in the input image, is the voxel index value of the input image, Voxel The true value of The fine segmentation category is Voxels The predicted value of is the balance correction factor, Indicates that the fine segmentation category is correctly identified as , Indicates that the fine segmentation category is The judgment of was missed. Indicates that the fine segmentation category is misclassified as , The fine segmentation category is The penalty coefficient for discriminating missed reports, The fine segmentation category is misclassified as The penalty coefficient.

6. A DBS electrode implantation path planning method based on deep learning according to claim 1, It is characterized in that Step S2 specifically includes: S21: obtaining a nerve fiber mask and a blood vessel mask from the image to be processed through the second deep network recognition; S22: superimposing the nerve fiber mask and the blood vessel mask to obtain a placeholder mask, wherein 1 indicates the presence of a placeholder voxel, and 0 indicates the absence of a placeholder voxel; S23: Reconstruct the placeholder mask through the second deep network to obtain a second mask.

7. A DBS electrode implantation path planning method based on deep learning according to claim 6, It is characterized in that The second deep network in step S2 is a CNN network established based on an occupation network architecture, and in step S23, the placeholder mask is reconstructed by the occupation network.

8. The DBS electrode implantation path planning method based on deep learning according to claim 6, It is characterized in that Step S23 further comprises: S231: performing boundary point inference on the placeholder mask using any voxel point in the placeholder mask by using a marching cubes algorithm to obtain a placeholder boundary point; S232: After enlarging the resolution of the placeholder boundary points, perform boundary point reasoning until the resolution is enlarged to a preset multiple, thereby obtaining a second mask.

9. A DBS electrode implantation path planning method based on deep learning according to claim 8, It is characterized in that The expression of the loss function for boundary point reasoning in step S231 is: in, is the loss of boundary point reasoning, is the training batch size, is the network for boundary point reasoning, are the model parameters of the network used for boundary point inference, is the output dimension, is the starting pixel for boundary point reasoning, is the observation point in the cube centered at the starting pixel. for occupancy status.

10. The DBS electrode implantation path planning method based on deep learning according to claim 1, It is characterized in that The pre-training process of the path planning network in step S3 specifically includes: S31: Obtaining a planning entry point and a planning target point by annotating the image to be processed; S32: Migrating weight parameters of the first deep network to the path planning network; S33: Performing migration training on the path planning network based on the weight parameters of the first deep network through the three-dimensional coordinates of the planning entry point and the three-dimensional coordinates of the planning target point.

Images