Puncture guide plate automatic generation method based on deep learning
The automated generation of puncture guides through deep learning technology solves the complex problems of traditional methods, improves the accuracy and efficiency of puncture, and ensures the safety and effectiveness of minimally invasive surgery.
Patent Information
- Application Number
- CN202511090857.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-05
- Publication Date
- 2025-10-21
AI Technical Summary
The traditional method of generating puncture guides is complicated and lacks automation and intelligence, which affects the puncture accuracy and efficiency.
Deep learning technology is used to perform automated multi-target segmentation of cerebral hemorrhage image data and generate a puncture guide, including segmenting cerebral hemorrhage, ears, and noses, automatically calculating the puncture path and positioning features, generating a skin-friendly bridging module, and automatically adjusting the puncture channel posture.
It greatly reduces the manual operation steps, improves the puncture accuracy and efficiency, and ensures the accuracy and safety of the puncture.
Smart Images
Figure CN120823481A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of medical devices and relates to a method for automatically generating a puncture guide based on deep learning. Background Art
[0002] Cerebral hemorrhage is a common disease in neurosurgery, and there are many treatment methods, such as conservative treatment, craniotomy and hematoma removal surgery, hematoma puncture and drainage surgery, and neuroendoscopic hematoma removal surgery.
[0003] Traditional craniotomy for hematoma removal requires extensive incisions and is highly invasive. However, for patients with advanced age, moderate hematoma volume, and multiple underlying medical conditions, minimally invasive surgery offers a safe and reliable treatment option. Accurate and safe puncture is crucial for ensuring efficacy and preventing complications. With the advancement of precision medicine, minimally invasive hematoma puncture and drainage is increasingly being used due to its simplicity, safety, speed, and minimal invasiveness.
[0004] 3D printing guide-assisted puncture and drainage is a precise minimally invasive surgery. It uses CT simulation to locate the puncture with the help of 3D printing guide, which can more accurately place the drainage tube in the center of the hematoma, greatly improving the accuracy and safety of the puncture. The schematic diagram of the puncture guide is shown below. Figure 1 shown.
[0005] Traditional guide plate production methods first load CT images into open-source software such as 3D Slicer, manually paint and circle the cerebral hemorrhage lesion, and identify the target. A skin mask is then generated using threshold segmentation methods. Most of the skin area is manually trimmed and removed, retaining only key features such as the nose and ears. The puncture point is then marked, and the puncture channel position is manually adjusted according to the direction from the target to the puncture point. A bridge module connecting the key features to the puncture channel is then manually painted. The key feature module, bridge module, and puncture channel module are then combined into a guide plate model, exported as an STL file, and loaded into a 3D printer for printing. This manual process is extremely complex, and reducing the steps involved in guide plate model generation and improving automation and intelligence is a pressing issue. Summary of the Invention
[0006] In order to solve the above-mentioned problems in puncture guide generation, the present invention provides a method for automatically generating puncture guides based on deep learning. The specific solution is as follows:
[0007] A method for automatically generating a puncture guide based on deep learning, comprising the following steps:
[0008] Step 1. Deep learning to segment brain hemorrhage, nose, and ears. Specifically, it can be divided into:
[0009] Step 1.1 Isolate the ICH lesion.
[0010] Step 1.2 Separate the ear.
[0011] Step 1.3 Separate the nose part.
[0012] ICH is used to calculate the target point TargetPoint and the puncture point PuncPoint; Nose and Ear are used to generate the positioning feature LocFeature, which can limit the free movement of the guide plate in anatomical directions such as SI and LR, thereby improving the puncture accuracy.
[0013] Step 2. Segment the skin mask. Use threshold segmentation and morphology to obtain the skin mask.
[0014] Step 3. Generate the basic puncture channel BasePuncTube. BasePuncTube can set the channel outer diameter, inner diameter, and length.
[0015] Step 4. Generate the LocFeature. Find the intersection of the Ear output from Step 1.2 or the Nose output from Step 1.3 with the Skin Mask output from Step 2, set the thickness, and generate the LocFeature.
[0016] Step 5. Puncture path recommendation. Calculate the target point and puncture point (PuncPoint) by combining the ICH and Mask. Specifically, calculate the center point of the ICH as the target point, use the principal axis of the ICH's equivalent ellipsoid to determine the puncture direction, and intersect with the skin Mask to obtain candidate puncture points. Among the candidate points, select the point with the shortest Euclidean distance to the target point as the puncture point (PuncPoint).
[0017] Step 6. Calculate the closest point, ClosestPoint, between the skin puncture point and the positioning feature. Calculate the closest point, ClosestPoint, to the LocFeature, from the PuncPoint.
[0018] Step 7. Generate the Bridge. On the Mask, extract the shortest curve from PuncPoint to ClosestPoint to generate the Bridge. Specifically, determine plane α using PuncPoint, ClosestPoint, and TargetPoint. Intersect plane α with the skin Mask surface to obtain the shortest curve from PuncPoint to ClosestPoint. Generate the pipeline model with this curve as the center, and then subtract it from the skin Mask to obtain the Bridge.
[0019] Step 8. Generate the puncture channel PuncTube. Calculate the pose based on the puncture target point TargetPoint and the puncture point PuncPoint, perform pose transformation on the basic puncture channel BasePuncTube, and straighten the puncture channel to obtain PuncTube. Specifically: PuncTube uses PuncPoint as the target reference point and the vector from TargetPoint to PuncPoint as the target vector to establish the target pose; BasePuncTube uses the origin as the reference point and the positive direction of the Z axis as the reference vector to obtain the original pose; the translation vector from the original reference point to the target reference point is obtained as PuncPoint, and then the angle between the reference vector and the target vector is calculated according to the vector angle formula, which is the rotation angle; the translation vector and the rotation angle constitute the pose transformation matrix, and the pose transformation matrix can be used to position PuncTube.
[0020] Step 9. Manually confirm whether to adjust the default parameters of BasePuncTube in Step 3. If the default parameters are adjusted, repeat Step 3 to Step 9.
[0021] Step 10. Merge modules to generate the PunctureGuide. Merge the LocFeature, PuncTube, and Bridge modules to generate the final PunctureGuide and save the guide model.
[0022] Furthermore, Step 1 is specifically as follows: by constructing and training a deep learning convolutional neural network model, the imaging data containing cerebral hemorrhage lesions can be automatically segmented into multiple targets.
[0023] Furthermore, Step 1.1 is specifically as follows: separating the cerebral hemorrhage lesion ICH from the multi-target segmentation results for subsequent calculation of the target point TargetPoint and the puncture point PuncturePoint.
[0024] Furthermore, Step 1.2 is specifically as follows: separating the ear part Ear from the multi-target segmentation result for subsequent selection as the positioning feature LocFeature.
[0025] Furthermore, Step 1.3 specifically separates the nose from the multi-target segmentation results for subsequent selection as a positioning feature. Steps 1.2 and 1.3 limit the guide's deviation from anatomical directions such as SI and LR during use, ensuring its close fit and improving puncture accuracy.
[0026] Furthermore, Step 2 is specifically as follows: skin segmentation, using threshold segmentation and morphology to obtain the skin mask.
[0027] Furthermore, Step 3 is specifically as follows: generating a basic puncture channel BasePuncTube, and setting default parameters of the channel outer diameter, inner diameter, and length.
[0028] Furthermore, Step 4 is specifically as follows: finding the intersection of the Ear output from Step 1.2 or the Nose output from Step 1.3 with the skin Mask output from Step 2, setting the thickness, and generating the positioning feature LocFeature.
[0029] Furthermore, Step 5 is specifically as follows: puncture path recommendation, combining the ICH output from Step 1.1 and the skin mask output from Step 2 to calculate the puncture target point TargetPoint and PuncPoint.
[0030] Furthermore, Step 6 is specifically as follows: calculating the closest point ClosestPoint to PuncPoint on LocFeature.
[0031] Furthermore, Step 7 is specifically as follows: intercepting the shortest curve from PuncPoint to ClosestPoint on the skin Mask to generate a bridge module Bridge.
[0032] Furthermore, Step 8 specifically includes: calculating the posture according to the puncture target and the puncture point, performing posture transformation on the basic puncture channel BasePuncTube, and aligning the puncture channel PuncTube.
[0033] Furthermore, Step 9 is specifically as follows: manual confirmation, merging the LocFeature, PuncTube, and Bridge modules, generating the final PunctureGuide, and saving the guide model.
[0034] The beneficial effects of the present invention are:
[0035] This invention uses deep learning to automatically extract cerebral hemorrhage lesions and localizing features such as the ears and nose; automatically segment the skin using a threshold; automatically calculate a recommended puncture path; automatically calculate the closest point between the localizing features and the puncture point and generate a skin-friendly bridge module; and automatically align the puncture position and generate a puncture channel. This entire guide production process significantly reduces manual application and trimming, while ensuring high puncture accuracy and possessing high practical value. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 Schematic diagram of the puncture guide.
[0037] Figure 2 Flowchart of puncture guide generation.
[0038] Figure 3 UNet++ network structure.
[0039] Figure 4 The automatically generated guide plate. The left picture is the rendering of each module of the guide plate, and the right picture is the model of the guide plate after merging the modules. DETAILED DESCRIPTION
[0040] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.
[0041] A flow chart of a method for automatic generation of puncture guides based on deep learning is shown in the figure. Figure 2 As shown, the following steps are included:
[0042] Step 1. Deep learning to segment brain hemorrhage, nose, and ears. Specifically, it can be divided into:
[0043] Step 1.1 Isolate the ICH lesion.
[0044] Step 1.2 Separate the ear.
[0045] Step 1.3 Separate the nose part.
[0046] ICH is used to calculate the target point TargetPoint and the puncture point PuncPoint; Nose and Ear are used to generate the positioning feature LocFeature, which can limit the free movement of the guide plate in anatomical directions such as SI and LR, thereby ensuring the puncture accuracy.
[0047] Step 2. Segment the skin mask. Use threshold segmentation and morphology to obtain the skin mask.
[0048] Step 3. Generate the basic puncture channel BasePuncTube. BasePuncTube can set the channel outer diameter, inner diameter, and length.
[0049] Step 4. Generate a LocFeature. Find the intersection of the Ear or Nose and the Skin Mask, set the thickness, and obtain the LocFeature.
[0050] Step 5. Recommend puncture path. Calculate the target point and puncture point based on ICH and Mask.
[0051] Step 6. Calculate the closest point, ClosestPoint, between the skin puncture point and the positioning feature. Calculate the closest point, ClosestPoint, to the LocFeature, from the PuncPoint.
[0052] Step 7. Generate the Bridge module. Intercept the closest curve from PuncPoint to ClosestPoint on the Mask to generate the Bridge module.
[0053] Step 8. Generate the puncture channel PuncTube. Calculate the pose based on the puncture target point TargetPoint and the puncture point PuncPoint, transform the pose of the base puncture channel BasePuncTube, and straighten the puncture channel to obtain PuncTube.
[0054] Step 9. Manually confirm and adjust the default parameters.
[0055] Step 10. Merge modules to generate the PunctureGuide. Merge the LocFeature, PuncTube, and Bridge modules to generate the final PunctureGuide and save the guide model.
[0056] In this embodiment, Step 1, deep learning, involves constructing and training a deep learning convolutional neural network model to automatically segment image data containing cerebral hemorrhage lesions. Examples of deep learning models include VoxHRNet, the UNet series, and the DeepLab series.
[0057] The UNet++ network structure is as follows Figure 3 As shown in the figure, it redesigns skip paths (shown in green) based on the UNet network structure; adds more connection and fusion operations between different layers (shown in blue), fully leveraging high-level semantic information and low-level spatial resolution to improve feature richness; adds deep supervision (shown in red), and adjusts model complexity through pruning to achieve a balance between speed and performance. Optionally, by constructing the UNet++ network infrastructure and upgrading 2D convolution to 3D convolution as the network model, it can be used to segment intracerebral hemorrhage lesion images.
[0058] In this embodiment, Step 2 segments the skin mask and obtains the skin mask through a threshold segmentation algorithm and morphological operations. The basic principle of the 3D threshold segmentation method is shown in the formula:
[0059]
[0060] Where: (x, y, z) is the voxel coordinate, g is the output image, f is the input original image, and T is the grayscale threshold.
[0061] Some adaptive threshold segmentation algorithms are based on this basic principle and adaptively calculate the threshold parameter T. Examples include Otsu's method, dynamic thresholding, and thresholding methods based on local window statistics. Alternatively, using a thresholding method based on local window statistics to adaptively determine the threshold parameter T can robustly generate an initial skin mask.
[0062] Furthermore, the initial Mask is subjected to morphological processing. The most basic operations of morphological processing are dilation and erosion operations. By combining dilation and erosion operations to derive opening and closing operations, it is possible to remove burrs and noise and smooth the area.
[0063] The formula for the expansion operation is:
[0064]
[0065] The formula for the corrosion operation is:
[0066]
[0067] Where (x, y, z) is the voxel coordinate of the original image, (x ′ ,y ′ ,z′) is the voxel coordinate of the structural element, element is the input structural element image, f is the input original image, and g is the output result image.
[0068] The opening operation is to perform an erosion operation first and then a dilation operation; the closing operation is to perform a dilation operation first and then an erosion operation. Optionally, the opening operation is used to remove small burrs and isolated voxels, and the closing operation is used to close some small gaps.
[0069] In this embodiment, Step 3 generates a basic puncture channel BasePuncTube. Optionally, the default parameters are set as follows: channel outer diameter 2.5 mm, inner diameter 1.5 mm, and length 5 mm.
[0070] In this embodiment, Step 4 generates a locating feature, LocFeature. The Ear (output from Step 1.2) or the Nose (output from Step 1.3) is intersected with the Skin Mask (output from Step 2), and a thickness is set to generate the locating feature, LocFeature. Optionally, the Nose is used as the default locating feature, with a default thickness of 2 mm.
[0071] In this embodiment, Step 5 recommends the puncture path. Combined with the ICH output from Step 1.1 and the skin Mask output from Step 2, the puncture target points TargetPoint and PuncPoint are calculated. Optionally, by calculating the center point of the ICH as the target point TargetPoint, the puncture direction is determined by the principal axis of the equivalent ellipsoid of the ICH, and the candidate puncture points are obtained by intersecting with the skin Mask. Among the candidate points, the point with the shortest Euclidean distance to the target point is taken as the puncture point PuncPoint. Assuming that the coordinates of point P1 are (x1, y1, z1) and the coordinates of point P2 are (x2, y2, z2), the Euclidean distance d can be calculated by the formula:
[0072]
[0073] In this example, Step 6 calculates the closest point, ClosestPoint, between the skin puncture point and the positioning feature. The ClosestPoint closest to PuncPoint is calculated on the LocFeature. Alternatively, the Euclidean distance between the puncture point and the LocFeature is calculated, and the point with the smallest Euclidean distance is selected as the closest point, ClosestPoint.
[0074] In this example, Step 7 generates the Bridge module. The Bridge module is generated by intercepting the shortest curve from PuncPoint to ClosestPoint on the skin mask. Optionally, the PuncPoint, ClosestPoint, and TargetPoint define a plane α. This plane intersects the skin mask surface to obtain the shortest curve from PuncPoint to ClosestPoint. This curve is then used as the center to generate the pipeline model, and the subtraction is then performed with the skin mask to obtain the Bridge module.
[0075] In this embodiment, Step 8 generates the puncture channel PuncTube. According to the puncture target point TargetPoint and the puncture point PuncPoint, the posture is calculated, the posture of the basic puncture channel BasePuncTube is transformed, and the puncture channel is straightened to obtain PuncTube. Optionally, PuncTube uses PuncPoint as the target reference point and the vector from TargetPoint to PuncPoint as the target vector to establish the target posture; BasePuncTube uses the origin as the reference point and the positive direction of the Z axis as the reference vector to obtain the original posture. The translation vector from the original reference point to the target reference point can be obtained as PuncPoint, and then the angle between the reference vector and the target vector can be calculated according to the vector angle formula, which is the rotation angle. The vector angle formula is as follows:
[0076]
[0077] in and are two vectors, and θ is the angle between them. The translation vector and the rotation angle can form a pose transformation matrix, which can be used to position the PuncTube.
[0078] In this embodiment, Step 9 is manually confirmed to confirm whether to adjust the default parameters of Step 3. If the default parameters are adjusted, Steps 3 to 9 are repeated.
[0079] In this embodiment, Step 10 merges the modules to generate a PunctureGuide for the guide plate. Merge the LocFeature, PuncTube, and Bridge modules to generate the final PunctureGuide and save the guide plate model. Optionally, the guide plate model is saved as an STL file. The guide plate that is automatically generated is as follows: Figure 4 shown.
[0080] The technical means disclosed in the solutions of the present invention are not limited to those disclosed in the above-mentioned embodiments, but also include technical solutions composed of any combination of the above-mentioned technical features. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. A method for automatically generating a puncture guide based on deep learning, characterized in that: The following steps are involved: Step 1. Deep learning to segment brain hemorrhage, nose, and ears; specifically: Step 1.1 Isolate the ICH lesion; Step 1.2 Separate the ear part; Step 1.3 Separate the nose part; Step 2. Segment the skin mask and obtain the skin mask using threshold segmentation and morphology; Step 3. Generate the basic puncture channel BasePuncTube; Step 4. Generate positioning feature LocFeature; Step 5. Recommend puncture path; calculate target point and puncture point based on ICH and Mask; Step 6. Calculate the closest point from the skin puncture point to the positioning feature; Step 7. Generate the bridge module Bridge; Step 8. Generate the puncture channel PuncTube; Step 9. Manually confirm whether to adjust the default parameters of BasePuncTube in Step 3. If so, repeat Step 3 to Step 9. Step 10. Merge modules, generate PunctureGuide, and save the guide model.
2. The method for automatically generating a puncture guide based on deep learning according to claim 1, characterized in that: In Step 1, a deep learning convolutional neural network model is constructed and trained to automatically perform multi-target segmentation on image data containing cerebral hemorrhage lesions. Step 1.1 specifically involves separating the cerebral hemorrhage lesion ICH from the multi-target segmentation results for subsequent calculation of the target point TargetPoint and puncture point PuncturePoint. Step 1.2 specifically involves separating the ear part Ear from the multi-target segmentation results for subsequent selection as the positioning feature LocFeature. Step 1.3 specifically involves separating the nose part Nose from the multi-target segmentation results for subsequent selection as the positioning feature. By constructing the UNet++ network infrastructure, the 2D convolution is upgraded to 3D convolution as the network model for segmenting cerebral hemorrhage lesion images.
3. The method for automatically generating a puncture guide based on deep learning according to claim 2, characterized in that: In Step 2, the formula for the 3D threshold segmentation method is: Where (x, y, z) is the voxel coordinate of the original image, g is the output image, f is the input original image, and T is the grayscale threshold. The threshold parameter T is adaptively determined using a threshold method based on local window statistical features. Morphological processing includes dilation and erosion operations, and the opening and closing operations are derived by combining dilation and erosion operations. The formula for the dilation operation is: The formula for the corrosion operation is: Where (x, y, z) is the voxel coordinate of the original image, (x ′ ,y ′ ,z′) is the voxel coordinate of the structural element, element is the input structural element image, f is the input original image, and g is the output result image; The opening operation is to perform an erosion operation first and then a dilation operation; the closing operation is to perform a dilation operation first and then an erosion operation; specifically, the opening operation is used to remove small burrs and island voxels, and then the closing operation is used to close some small gaps.
4. The method for automatically generating a puncture guide based on deep learning according to claim 1, characterized in that: In Step 3, the puncture channel BasePuncTube can set the default parameters of the channel outer diameter, inner diameter, and length.
5. The method for automatically generating a puncture guide based on deep learning according to claim 1, characterized in that: Step 4 specifically involves finding the intersection of the Ear output from Step 1.2 or the Nose output from Step 1.3 with the skin Mask output from Step 2, setting the thickness, and generating the positioning feature LocFeature.
6. The method for automatically generating a puncture guide based on deep learning according to claim 1, characterized in that: In Step 5, the center point of the ICH is calculated as the target point TargetPoint, the puncture direction is determined by the principal axis of the equivalent ellipsoid of the ICH, and the candidate puncture points are obtained by intersecting with the skin Mask. Among the candidate points, the point with the shortest Euclidean distance to the target point is taken as the puncture point PuncPoint.
7. The method for automatically generating a puncture guide based on deep learning according to claim 1, characterized in that: Step 6 is as follows: Calculate the closest point to PuncPoint on LocFeature.
8. The method for automatically generating a puncture guide based on deep learning according to claim 1, characterized in that: In Step 7, the shortest curve from PuncPoint to ClosestPoint is intercepted on the skin mask to generate the bridge module Bridge. Specifically, the plane α is determined by the three points PuncPoint, ClosestPoint and TargetPoint. The plane α intersects with the surface of the skin mask to obtain the shortest curve from PuncPoint to ClosestPoint. Then, with this curve as the center, the pipeline model is generated and the difference between it and the skin mask is calculated to obtain the bridge module Bridge.
9. The method for automatically generating a puncture guide based on deep learning according to claim 1, characterized in that: In Step 8, the pose is calculated based on the puncture target and puncture point, the pose of the basic puncture channel BasePuncTube is transformed, and the puncture channel PuncTube is aligned. Specifically, PuncTube uses PuncPoint as the target reference point and the vector from TargetPoint to PuncPoint as the target vector to establish the target pose. BasePuncTube uses the origin as the reference point and the positive direction of the Z axis as the reference vector to obtain the original pose. The translation vector from the original reference point to the target reference point is PuncPoint, and then the angle between the reference vector and the target vector is calculated according to the vector angle formula, which is the rotation angle. The translation vector and the rotation angle constitute the pose transformation matrix, and the pose transformation matrix can be used to position PuncTube.
10. The method for automatically generating a puncture guide based on deep learning according to claim 1, characterized in that: Step 10 is as follows: merge the LocFeature, PuncTube, and Bridge modules to generate the final PunctureGuide and save the guide model.