A medical image abdominal contour model generation method and system

Abstract

The application discloses a medical image abdominal contour model generation method and system, and belongs to the technical field of medical image processing; the method comprises the following steps: firstly, marking the connected regions in three-dimensional image data, and filtering out the regions smaller than a preset volume threshold; then, performing layer-by-layer slice processing, extracting the outer contour of each layer through Gaussian blurring, Canny edge detection and morphological operation, and generating a three-dimensional contour based on multi-direction contour combination; removing duplicate points through neighbor search using a KD tree, then performing voxelization down-sampling, and extracting the outer contour of an irregular shape; removing noise points and screening out effective clusters; finally, generating a three-dimensional abdominal contour model through triangular meshing processing; through processing of medical image data, the application generates a three-dimensional model, can accurately reflect the contour structure of the abdomen, can provide help for image-guided puncture operation path planning and accurate positioning of a target lesion, and is helpful to improving the success rate and safety of puncture operation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, and more particularly to a medical image abdominal contour model generation method and system. BACKGROUND

[0002] At present, with the continuous development of medical imaging technology, especially the wide application of computed tomography (CT) and magnetic resonance imaging (MRI) technology, the precision and clarity of medical image data have been significantly improved. At the same time, the processing and analysis of medical image data play an increasingly important role in clinical diagnosis and treatment, especially in some complex surgical operations and precision medicine fields, such as abdominal puncture operations.

[0003] However, although the traditional image-guided puncture method can provide certain visual assistance, the operation of the doctor still faces many challenges under complex anatomical structures, for example: the puncture path needs to select a suitable needle entry point on the body surface, and in the process of reaching the lesion, important blood vessels and other organs need to be avoided, increasing the difficulty of operation; the puncture operation requires very high precision, and slight deviation may cause the puncture to not reach the target area.

[0004] Therefore, although the existing medical imaging technology can already provide high-resolution image data, how to accurately extract the three-dimensional contour of the body surface range, the lesion area, the important organs or the target site from these data and convert it into a three-dimensional model that can be directly applied is still a challenging task. The existing traditional medical image processing method mainly focuses on the processing of two-dimensional images, and the processing of three-dimensional data often faces problems such as large data volume, complex processing, and much noise.

[0005] Therefore, how to accurately reconstruct the three-dimensional contour of each part from medical images and effectively remove noise to improve processing accuracy and efficiency is a technical problem that needs to be solved in the field of medical image processing. SUMMARY

[0006] Therefore, the present application provides a medical image abdominal contour model generation method and system, which can solve the problem of limited precision in constructing a three-dimensional abdominal contour model in the prior art, and can accurately reconstruct a three-dimensional model of the abdominal contour structure through processing and analysis of medical image data, thereby providing accurate visual reference for clinical puncture operations. It is compatible with existing medical imaging equipment and has high application value and promotion prospects.

[0007] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0008] In a first aspect, the present application provides a medical image abdominal contour model generation method, comprising the following steps:

[0009] S10, label the connected regions in the three-dimensional image data, generate labeled three-dimensional volume data, and filter out regions smaller than a preset volume threshold;

[0010] S20, perform layer-by-layer slicing processing on the retained three-dimensional image data, extract the outer contour of each layer through Gaussian blur, Canny edge detection and morphological operation, and generate a three-dimensional contour based on multi-directional contour combination;

[0011] S30, use KD tree to perform neighbor search to remove duplicate points of the three-dimensional contour;

[0012] S40, voxelize and downsample the point cloud data of the three-dimensional contour after the deduplication processing to reduce the number of point clouds;

[0013] S50, extract the irregular outer contour of the three-dimensional point cloud data after voxelization and downsampling processing through Delaunay triangulation, boundary screening and DBSCAN clustering algorithm;

[0014] S60, use the DBSCAN algorithm again to remove noise points in the point cloud and screen out effective clusters;

[0015] S70, perform triangular meshing processing on the point cloud data in the effective cluster to generate a final three-dimensional abdominal contour model.

[0016] Further, the step S10 comprises:

[0017] Use the 8-neighborhood algorithm to identify and label the connected regions of the three-dimensional image data;

[0018] Filter out small regions and remove background regions according to the preset volume threshold.

[0019] Further, the step S20 comprises:

[0020] Obtain the size of the three-dimensional image data of the retained connected regions;

[0021] Perform normalization operation on the three-dimensional image data and map to the range of 0-255;

[0022] Iterate through each slice of the three-dimensional image data layer by layer, extract the outer contour of each layer through Gaussian blur, Canny edge detection and morphological operation;

[0023] Draw the detected edges and outer contour information;

[0024] Extract the contour along the axial, coronal and sagittal directions respectively and combine into a complete three-dimensional contour.

[0025] Further, the step S30 comprises:

[0026] The point cloud data extracted from the axial, coronal and sagittal directions is merged into a single data set;

[0027] KD tree is used to find the nearest two neighbors of each point; when the distance between the point and the two neighbors is less than the preset distance threshold, the corresponding neighbor is removed.

[0028] Further, the step S40 comprises:

[0029] The point cloud data of the three-dimensional contour after deduplication is divided into a plurality of cubic grids; each small cube is a voxel;

[0030] For each voxel, only one representative point in the voxel is retained; the representative point is the centroid, randomly selected, or a point with a significant feature.

[0031] Further, the step S50 comprises:

[0032] The three-dimensional point cloud data after voxelization and downsampling is subjected to Delaunay triangulation;

[0033] Each triangle generated by the triangulation is traversed, and the three edges of the triangle are obtained by indexing to form a non-repeating edge set;

[0034] The Euclidean distance between the two endpoints of each edge is calculated, and the edges with a distance less than the parameter alpha are retained to generate an effective triangle edge set;

[0035] Each edge in the effective triangle edge set is traversed, and the initial boundary point set of the point cloud is generated according to the start point and the end point of the edge;

[0036] The initial boundary point set is clustered by DBSCAN clustering to remove noise points and select effective boundary points.

[0037] Further, the step S70 comprises:

[0038] Delaunay triangulation is performed on the point cloud data in the effective cluster, and triangles with an inscribed circle radius less than Alpha are selected;

[0039] The grid surface is smoothed by averaging the coordinates of each vertex and its surrounding adjacent vertices, the normal vector of each vertex of the grid is calculated, and the final three-dimensional abdominal contour model is output.

[0040] In a second aspect, the embodiments of the present application also provide a medical image abdominal contour model generation system, comprising:

[0041] An image preprocessing module: labeling connected regions in three-dimensional image data, generating labeled three-dimensional volume data, and filtering out regions smaller than a preset volume threshold;

[0042] A contour scanning module: performing layer-by-layer slicing processing on the retained three-dimensional image data, extracting the outer contour of each layer through Gaussian blur, Canny edge detection, and morphological operation, and generating a three-dimensional contour based on multi-directional contour combination;

[0043] A removal of overlapping module: using KD tree for neighbor search to remove duplicate points of the three-dimensional contour;

[0044] A grid voxel downsampling module: voxelizing and downsampling the point cloud data of the three-dimensional contour after the removal of duplicates to reduce the number of point clouds;

[0045] An Alpha shape module: extracting the irregular outer contour of the three-dimensional point cloud data after voxelization and downsampling through Delaunay triangulation, boundary screening, and DBSCAN clustering algorithm;

[0046] An intra-abdominal impurity removal module: removing noise points in the point cloud again using the DBSCAN algorithm and screening out effective clusters;

[0047] A model output module: performing triangular meshing processing on the point cloud data in the effective clusters to generate a final three-dimensional abdominal contour model.

[0048] According to the technical solution described above, compared with the prior art, the present application has the following technical advantages:

[0049] By obtaining high-resolution medical image data, the present application generates a three-dimensional model using advanced image processing techniques to accurately reflect the contour structure of the abdomen. For example: through multi-directional contour combination, the missing of local contour information caused by blurred scanning images and other reasons can be effectively recovered; through Alpha shape processing, the local details of the surface contour can be maximally preserved; through voxel downsampling and DBSCAN clustering, the computational load can be greatly reduced, and modeling or real-time model updating can be completed faster, which facilitates doctors to perform puncture operations under dynamic observation, thereby significantly reducing the risk of surgery. The innovative technology proposed in the present application is expected to provide more reliable image support for abdominal puncture, thereby assisting in improving the treatment effect and medical experience of patients. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute a part of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on the provided drawings.

[0051] Figure 1 The flow chart of the medical image abdominal contour model generation method provided by the present application.

[0052] Figure 2 The principle diagram of finding a connected region using an 8-neighbor algorithm provided by the present application.

[0053] Figure 3 The principle diagram of KD tree.

[0054] Figure 4 The principle diagram of grid voxelization provided by the present application.

[0055] Figure 5 The block diagram of the medical image abdominal contour model generation system provided by the present application. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0057] In the related art, although some methods can perform image segmentation and contour extraction of the abdominal region, they mostly rely on traditional two-dimensional image processing methods, lack comprehensive processing ability for three-dimensional data, and mostly have problems such as long processing time, large consumption of computing resources, low precision, etc. In addition, the subsequent optimization (such as removing overlapping points, noise points, generating grid models, etc.) of point cloud data in the prior art is not perfect, which limits the application effect of the finally generated three-dimensional model in visualization, analysis and actual operation.

[0058] Therefore, the embodiments of the present application disclose a medical image abdominal contour model generation method, which is widely used in the field of medical imaging. By comprehensively processing and modeling the image data (such as axial, coronal and sagittal positions), the present application can generate a high-precision three-dimensional medical model, which provides important support for clinical diagnosis, surgical planning and medical teaching. For example Figure 1 as shown, comprising steps S10-S70:

[0059] S10, label the connected regions in the three-dimensional image data, generate the labeled three-dimensional volume data, and filter out regions smaller than a preset volume threshold.

[0060] In this step, the three-dimensional image data is usually generated by scanning the entire body or a specific region by a medical imaging device (such as CT, MRI, etc.), and the abdomen may only be a part of the body. In actual application, the image data is usually obtained from the entire body of the patient (including the head, chest, abdomen, limbs, etc.), so the abdominal region needs to be extracted for processing through image segmentation or region selection.

[0061] Specifically in step S10, it includes:

[0062] S101, identify and label the connected regions in the three-dimensional image data using an 8-neighbor algorithm;

[0063] S102, filter out small regions according to a preset volume threshold, and remove background regions.

[0064] First, label the connected regions in the three-dimensional image data volume, and return a labeled volume label_volume and the number of regions num_labels. Assign a unique label to each connected region. Each position in the labeled volume label_volume stores the label of the connected region it belongs to.

[0065] That is, identify those "connected" regions from the medical image. Simply put, it is to mark the parts in the image that are connected to each other and belong to the same object, such as different tissue structures in the abdomen. This labeling process is equivalent to labeling each region in the image with a label.

[0066] As shown in Figure 2 , find the connected regions using 8-neighbor (1 is the object, 0 is the background); then calculate the volume (number of pixel / voxel points) of each connected region, i.e. the number of times each label appears in label_volume. Flatten the three-dimensional label_volume into a one-dimensional array, then use bincount (a statistical function) to calculate the count of each label, and get the size (i.e. volume) of each connected region.

[0067] That is, once the labeling is complete, the "volume" of each connected region will be calculated. The volume here is not the actual spatial volume, but the number of pixels (or voxels) occupied by these regions in the image. For example, if a region occupies 100 pixel points in the image, its volume is 100.

[0068] The connected regions whose volume is larger than the minimum threshold are marked by comparing with min_size (a pre-set volume threshold). label_sizes is an array of the size of each connected region, which is compared with min_size to generate a Boolean array large_labels. Each value in the array indicates whether the corresponding connected region is larger than min_size.

[0069] That is, for each region marked, its volume size is compared to see if it meets the pre-set volume threshold requirement. If the volume of a region is too small, it may be just a noise point or an irrelevant small block, and it will be filtered out and not retained.

[0070] The common pre-set volume threshold can range from tens to hundreds of voxels, but this threshold usually needs to be adjusted according to the actual image and processing requirements. For example:

[0071] For general medical image data, the minimum volume threshold can be set between 100-500 voxels.

[0072] For the identification of some small structures, a smaller threshold can be set.

[0073] Secondly, large_labels[0] = false (Boolean false value) sets the background region (label 0 region) as not retained. This is because, in the connected region marking, 0 is usually used to represent the background. In order to ensure that the background is not considered as a connected region, large_labels[0] is set to false, indicating that even if the background region is larger than min_size, it will not be retained.

[0074] Finally, a filtered volume data is generated, which only retains the connected regions that meet the conditions. It is ensured that only the abdominal related regions (i.e. relatively large connected regions) are concerned, and not the irrelevant background.

[0075] In this step, the abdominal related regions are found from the image, the size of these regions is calculated, and those too small or irrelevant parts are removed to ensure that only useful regions are retained. In order to improve the calculation efficiency in subsequent processing, and to ensure that the subsequent contour extraction and model generation are more accurate.

[0076] S20, perform layer-by-layer slicing on the retained three-dimensional image data, extract the outer contour of each layer through Gaussian blur, Canny edge detection and morphological operation, and generate a three-dimensional contour based on multi-directional contour combination.

[0077] Specifically, the following steps are included:

[0078] S201, obtain the three-dimensional image data size of the reserved connected region.

[0079] Image data reading and size determination: through the input multi-dimensional image data, obtain its three-dimensional size, i.e. height, width and depth, so as to perform subsequent slice contour detection in different directions.

[0080] S202, perform normalization operation on the three-dimensional image data, and map to the range of 0-255, i.e. single-channel gray image;

[0081] Perform normalization operation on the image data to eliminate brightness and contrast difference, so that the data is in a unified range, facilitating subsequent contour detection. First, obtain the global minimum and maximum values of the input data, and standardize to the range of 0-255 through the following steps, facilitating subsequent contour detection processing:

[0082] 1) Calculate the difference (denom) between the minimum value (slice_min) and the maximum value (slice_max) of each slice. If the difference is zero, generate a full zero matrix directly.

[0083] 2) Otherwise, map the slice data to the range of 0-255 through linear transformation.

[0084] After normalization, the gray value of each layer of image data is between 0 and 255, ensuring the consistency and processability of the data.

[0085] S203, traverse each slice of the three-dimensional image data layer by layer, and extract the outer contour of each layer through Gaussian blur, Canny edge detection and morphological operation.

[0086] Among them, Gaussian blur smoothing processing is used for slice image to reduce noise interference.

[0087] Use Canny algorithm for edge detection to extract strong edge response in the image.

[0088] Then use morphological dilation operation to dilate the edge, and then perform erosion operation to remove internal redundant edges, and only keep the outer contour.

[0089] S204, draw the detected edge and outer contour information;

[0090] Scan the input image pixel by pixel, locate the white pixel point (value 255), then form the contour based on the connected region in the edge map, and finally optimize the point set of each contour to compress the points on the straight line into the start and end points.

[0091] Output slice_contours: detected contour array, each contour is a set of points. slice_contours = {C1, C2, …, Cn} Ci represents the i-th contour, in the format of point set: Ci = {(x1, y1), (x2, y2), …, (xi, yi)} k ,y k )}

[0092] S205, extract the contours along the axial, coronal and sagittal directions respectively and combine them into a complete three-dimensional contour.

[0093] In the application scenario of the present application, the above contour detection function is called along the axial, coronal and sagittal directions respectively (step S203), and the slice contour of each direction is extracted. After obtaining the coronal and sagittal contours, the two images are rotated and aligned to the axial direction to generate the final result of the three-dimensional combined contour.

[0094] In this step, the contours of each layer are accurately extracted by processing the image data layer by layer, combined with Gaussian blur and Canny edge detection algorithm. This step uses a variety of algorithm combinations to ensure efficient extraction of the real abdominal contour in complex image data. In addition, through scanning and contour extraction along different directions (axial, coronal and sagittal), the multi-angle accurate capture of the abdominal contour is ensured, which can effectively construct a three-dimensional model.

[0095] S30, using KD tree to search for neighbors to remove duplicate points of the three-dimensional contour;

[0096] When processing medical image data, we usually extract some points (for example, key points or feature points of abdominal contour) from the image. Sometimes, these points may have some overlap, that is, the same position may have multiple points very close to each other. Such duplicate points will cause the model to be inaccurate, and even the contour to be discontinuous. Step S30 specifically includes:

[0097] S301, merging the point cloud data extracted from the axial, coronal and sagittal directions into a single data set;

[0098] S302, using KD tree to find the nearest two neighbors of each point; when the distance between the point and the two neighbors is less than a preset distance threshold (in this case, 1e-3), the corresponding neighbor is removed.

[0099] The main purpose of this step is to process the duplicate points in the point cloud data to ensure that each point is independent and unique. This process is to ensure that in the subsequent steps, the accuracy and quality of the model will not be affected by duplicate points.

[0100] This step specifically combines the three sets of axis arrays into one set of arrays to remove duplicate points. As shown in Figure 3 KD tree finds the two closest points to each point and stores the distances of each point to its two nearest neighbors in the array distances, and indices is the array of indices of the corresponding neighbor points. A boolean array is used to store the distance of all points to their second nearest neighbor (distances[:, 1]), if the distance is greater than the threshold, it is True, otherwise it is False. Finally, filter out the points that meet the conditions, that is, the distance between the point and its nearest neighbor is greater than 1e-3 (the distance between two points is less than 0.001, then it is an overlapping point), these points can be considered as unique points.

[0101] For example, assume that we are processing the contour extraction data of an image, and the following points are obtained: Figure 3 The distance between point B and point A is 0.0005 (very close).

[0102] The distance between point B and point D is 0.002 (relatively far).

[0103] If the threshold is set to 1e-3 (i.e. 0.001), then point B and point A will be considered as duplicates because their distance is less than the threshold. The distance between point B and point D is greater than the threshold, so they will be considered as two different points.

[0104] The purpose of this step is to filter out the duplicate points that are very close to each other by calculating the distance between each point and other points, and remove them. Finally, only unique points are retained, which ensures that only one representative point is used in subsequent model generation and contour drawing, avoiding the interference of duplicate points, so as to improve the accuracy and quality of the model.

[0105] S40, voxelizing and down-sampling the point cloud data of the three-dimensional contour after the deduplication processing to reduce the number of point clouds;

[0106] As shown in

[0107] , the point cloud is divided into a series of voxel grids, and for each voxel, only one representative point within the voxel is retained. This can reduce the number of point clouds, thereby reducing the computational complexity. The general information of the original shape is still retained, but the density of points is reduced, and the original features are still guaranteed not to be destroyed. Figure 4 The selection of the representative point can be selected as follows:

[0108] 1) Select the "average position" of all points within the voxel as the representative point (i.e. calculate the centroid of these points).

[0109] 2) Select the point with the maximum distance within the voxel as the representative point.

[0110] 3) Select the point with the minimum distance within the voxel as the representative point.2) Randomly select a point in the voxel as a representative point.

[0111] 3) Select a point with a significant feature as a representative point (e.g., a point on the contour edge).

[0112] In this step, the number of point cloud data is reduced by dividing the three-dimensional point cloud data into small grid voxels, and only one representative point is retained in each voxel. This makes the subsequent calculation more efficient and reduces the processing time and computational burden without losing too much important information.

[0113] S50, on the voxelized down-sampling processed three-dimensional point cloud data, the irregular shape of the outer contour is extracted by Delaunay triangulation, boundary screening and DBSCAN clustering algorithm;

[0114] This step extracts complex boundary shapes from point cloud data, which is suitable for generating irregular, tortuous contours or surfaces. Its function is to determine the shape boundary of point cloud data through an Alpha shape algorithm, and represent these point sets as a complete geometric shape (such as polygon or polyhedron). Specifically, it includes:

[0115] S501, on the voxelized down-sampling processed three-dimensional point cloud data, using Delaunay triangulation;

[0116] Delaunay triangulation is an algorithm that divides a set of points into triangles, ensuring that the generated triangles satisfy the condition that any two triangles share a side that forms a circumcircle containing no other triangle vertices. This maximizes the minimum angle and avoids the generation of excessively thin triangles, improving the quality of triangulation.

[0117] Find the convex hull in the input point set to determine the initial triangle. Insert the point set into the current triangle network and adjust the triangle division according to the Delaunay condition to ensure the circumcircle property. Check the boundary conditions of all triangles to ensure that no points are missed and handle possible numerical errors.

[0118] S502, traverse each triangle generated by triangulation to obtain three edges of the triangle to form a non-repeating edge set;

[0119] Traverse each triangle generated by triangulation to obtain three edges of the triangle. Each triangle has three edges, and the indices of the triangle vertices (i, j, k) are combined in the loop to form a non-repeating edge set edges. This step helps to avoid redundancy caused by multiple calculations of edges.

[0120] S503, Calculate the Euclidean distance between the two endpoints of each edge, and keep the edges with a distance less than the parameter alpha to generate a set of valid triangle edges.

[0121] Suppose the two endpoints of an edge are (x1, y1, z1) and (x2, y2, z2), calculate the Euclidean distance of each edge , and keep the edges with a distance less than the parameter alpha as outer_edges (outer edge set). This operation aims to remove edges with excessively large lengths to meet the Alpha Shape boundary condition.

[0122] During the triangulation process, the algorithm generates some circumscribed circles for the triangles (each triangle has a circle around it). If the circumscribed circle radius of a certain triangle is very large, it may no longer be "truly" part of the point cloud's outer contour. At this time, the algorithm controls through a parameter called alpha, and only keeps triangles with smaller circumscribed circle radii, removing those with excessively large radii.

[0123] This alpha parameter controls the smoothness of the "shape", the smaller the alpha, the more refined the boundary retained; the larger the alpha, the smoother the boundary obtained, and the higher the simplification level.

[0124] S504, Traverse each edge in the set of valid triangle edges, and generate an initial set of boundary points of the point cloud according to the start and end points of the edge;

[0125] Traverse each edge in outer_edges, and include the start and end points of the edge in the set outer_points (edge start and end point set) to generate a preliminary set of boundary points of the point cloud. Use set deduplication to ensure that each point appears only once.

[0126] S505, Cluster the initial boundary point set by DBSCAN clustering to remove noise points and select valid boundary points.

[0127] Use the DBSCAN clustering algorithm to cluster outer_points, set the clustering radius eps to alpha, and define the minimum sample size min_samples as 2 to remove outliers and ensure the continuity of the boundary. DBSCAN can effectively identify core boundary points in the point cloud while excluding noise points (points with label -1).

[0128] In the clustering results, filter out noise points and extract different boundary point sets according to the clustering labels. For each cluster, arrange the boundary points in the cluster in coordinate order and return them, ensuring that the returned boundary point set meets a specific arrangement order.

[0129] In this step, by combining Delaunay triangulation, boundary screening and DBSCAN clustering algorithm, the boundary profile of complex point cloud data is efficiently extracted, which is suitable for generating irregularly shaped outer contour.

[0130] S60, using DBSCAN algorithm to remove noise points in the point cloud, and screening out effective clusters;

[0131] The purpose of this step is to remove irrelevant noise points that may occur during processing and retain real points related to the abdominal profile. In other words, it identifies and removes "stray" points unrelated to the target shape, making the final point cloud cleaner and more accurate.

[0132] Specifically, DBSCAN clustering algorithm is used to divide the point cloud into connected regions. The point cloud data is divided into different clusters by eps and min_points parameters. Eps is the neighborhood radius of each point, and min_points is the minimum number of points in each cluster. The number of points in each cluster is counted, and only effective clusters are calculated. When a point has more than min_points neighbor points, it is classified as a core point, otherwise it is considered as a noise point.

[0133] According to the min_size parameter, clusters with a point number greater than min_size are screened out. A Boolean mask is created to mark points belonging to large clusters, and the remaining points are stored, so that all points belonging to large clusters can be obtained. Finally, this point cloud containing only large cluster points is returned.

[0134] Assuming that this step extracts some abdominal point cloud data from medical images, these point clouds contain correct abdominal profile points and some random noise points (such as points generated by scanning errors). Through DBSCAN clustering:

[0135] The real abdominal profile points will be clustered into one or more clusters, with a large number of points, which are valid data;

[0136] Some outliers (such as interference points or noise points during scanning) will be identified as noise, with a small number of points and a long distance from other points, which will be removed.

[0137] In this step, the removal of impurities in the abdomen is achieved by using DBSCAN clustering algorithm to identify and remove noise points that do not belong to the abdominal profile. It retains those dense and relevant points and removes those isolated and irrelevant noise points, ensuring that the final point cloud data is cleaner and more accurate. This process can greatly improve the accuracy and quality of subsequent model generation.

[0138] S70, for the point cloud data in the effective cluster, a three-dimensional abdominal profile model is generated through triangular meshing processing.

[0139] The purpose of this step is to convert the processed point cloud data into a three-dimensional mesh model and further smooth and optimize this model for better display and analysis. This step is the final step of the entire process, which collects scattered points into a complete three-dimensional shape, which can be used for visualization, printing or further analysis.

[0140] The point cloud data generates a triangular mesh, which is converted into a mesh according to the model accuracy requirements. Among them, the construction of Delaunay triangulation, first of all, to find the convex hull, to determine the initial triangle, then insert the point set into the current triangle network, adjust the triangle division according to the Delaunay condition, ensure to meet the circumcircle property, filter out those containing the circumcircle radius greater than the value of the triangle, so as to extract the shape boundary. Then smooth the mesh surface by averaging the coordinates of each vertex and its adjacent vertices, reduce the grid noise and irregularity, so that the model is more smooth, calculate the normal vector of each vertex of the grid and output the model file.

[0141] Specifically includes:

[0142] S701, Delaunay triangulation is adopted for the point cloud data in the effective cluster, and the triangle with a circumcircle radius less than Alpha is selected;

[0143] Point cloud data itself is just a bunch of points distributed in three-dimensional space, representing the shape of the object, but these points are not directly connected. In order to "connect" these points into a complete surface, first of all, Delaunay triangulation is needed. This algorithm will connect these points into triangles according to certain rules. In this way, a triangular mesh is obtained, representing the surface of the object.

[0144] This triangular mesh is the most basic way to represent the shape of a three-dimensional object. Each triangle has three corners, each representing a vertex of the mesh, and the edges of the triangle represent the connection between these vertices.

[0145] Remove unqualified triangles: in the generated mesh, some triangles may not meet the required shape. For example, the circumcircle radius of some triangles is too large, indicating that these triangles may be at the "edge" or "away" from the object, and should be removed.

[0146] Therefore, the algorithm will filter these triangles according to a threshold value (called Alpha). Only those triangles with a circumcircle radius less than Alpha will be retained. In this way, unnecessary and excessively distorted triangles can be removed, and only the part that truly represents the outer contour of the object is retained.

[0147] S702, smoothing the mesh surface by averaging the coordinates of each vertex and its surrounding adjacent vertices, calculating the normal vector of each vertex of the mesh, and outputting the final three-dimensional abdominal contour model.

[0148] Once the mesh is obtained, the next task is to make the mesh smoother and remove sharp corners or irregular surfaces. The smoothing process is usually achieved by calculating the average position of each vertex and its surrounding adjacent vertices. By this method, the coordinates of each vertex will be adjusted to be closer to the position of the neighbor vertices, and finally a smoother mesh surface is formed.

[0149] In order to make the mesh look more realistic, it is usually necessary to calculate the normal vector for each triangular face. The normal vector is a vector perpendicular to the triangular face, indicating the direction of the face. Through the calculation of the normal vector, the lighting effect of the mesh can be obtained, so that the final three-dimensional model looks more natural when displayed.

[0150] Finally, the triangular mesh after smoothing and optimization is saved as a standard three-dimensional model format (such as STL, OBJ, etc.). This file can be used for three-dimensional visualization, 3D printing, or further analysis and processing.

[0151] The present application can accurately and quickly reconstruct the three-dimensional contour of the abdomen from the original medical image data, and generate a high-quality three-dimensional model, which provides accurate visual reference for clinical surgery, puncture operation, etc., and has important practical significance and application prospect.

[0152] Based on the same inventive concept, the embodiments of the present application also provide a medical image abdominal contour model generation system. Since the principles of the problems solved by the system are similar to those of the aforementioned medical image abdominal contour model generation method, the implementation of the device can be referred to the implementation of the aforementioned method, and the repeated parts will not be described here.

[0153] Referring to Figure 5 A medical image abdominal contour model generation system, comprising:

[0154] An image preprocessing module: marking the connected regions in the three-dimensional image data to generate labeled three-dimensional volume data, and filtering out regions smaller than a preset volume threshold;

[0155] A contour scanning module: performing layer-by-layer slice processing on the retained three-dimensional image data, extracting the outer contour of each layer through Gaussian blur, Canny edge detection and morphological operation, and generating a three-dimensional contour based on multi-directional contour combination;

[0156] A removal overlap module: using KD tree to perform neighbor search to remove duplicate points of the three-dimensional contour;

[0157] Grid voxel downsampling module: voxelize and downsample the point cloud data of the three-dimensional profile after deduplication processing to reduce the number of point clouds;

[0158] Alpha shape module: extract the irregular shape of the outer contour of the three-dimensional point cloud data after voxelization and downsampling processing through Delaunay triangulation, boundary screening and DBSCAN clustering algorithm;

[0159] Abdominal impurity removal module: use the DBSCAN algorithm again to remove noise points in the point cloud and screen out effective clusters;

[0160] Model output module: through triangular meshing processing on the point cloud data in the effective cluster, the final three-dimensional abdominal profile model is generated.

[0161] The medical image abdominal profile model generation system provided by the application has the following advantages:

[0162] 1) Optimize precision and efficiency: through a series of modular processing steps, optimize data layer by layer, remove noise, fine extract abdominal profile, and finally generate accurate three-dimensional model.

[0163] 2) Efficient calculation and storage optimization: use downsampling and deduplication technology to reduce redundant data and improve calculation efficiency, suitable for large-scale data processing.

[0164] 3) Strong compatibility and adaptability: the system can be compatible with different medical image devices, and the ability to process irregular shapes makes it widely applicable, with strong popularization prospects.

[0165] The overall processing flow has significant technical advantages in accuracy, efficiency, adaptability and scalability, and can provide strong support for clinical medical image analysis. In addition, the system can update the model in real time, which is convenient for doctors to perform puncture operation under dynamic observation, thereby significantly reducing the risk of complications.

[0166] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same and similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the related parts can be referred to the method part.

[0167] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating a medical image abdominal contour model, characterized in that: The following steps are involved: S10, marking connected areas in the three-dimensional image data to generate marked three-dimensional volume data, and filtering out areas smaller than a preset volume threshold; S20, slicing the retained three-dimensional image data layer by layer, extracting the outer contour of each layer through Gaussian blur, Canny edge detection and morphological operations, and generating a three-dimensional contour based on a combination of multi-directional contours; S30, performing neighbor search using a KD tree to remove duplicate points of the three-dimensional contour; S40, performing voxel downsampling on the point cloud data of the three-dimensional contour after deduplication processing to reduce the number of point clouds; S50, extracting the outer contour of the irregular shape from the three-dimensional point cloud data after voxel downsampling through Delaunay triangulation, boundary screening and DBSCAN clustering algorithm; S60. Use the DBSCAN algorithm again to remove noise points in the point cloud and filter out valid clusters. Use the DBSCAN algorithm to divide the point cloud into connected areas and divide the point cloud into different clusters. According to the min_size parameter, filter out clusters with points larger than min_size. Create a Boolean mask to mark points belonging to large clusters, and return a point cloud containing only points in large clusters. S70, generating a final three-dimensional abdominal contour model by triangulating the point cloud data within the valid cluster; Wherein, the step S20 includes: Acquire the size of the three-dimensional image data retaining the connected area; perform slice contour detection in different directions according to the size of the three-dimensional image data; Performing a normalization operation on the three-dimensional image data to map it to a range of 0 to 255; Traverse each slice of the 3D image data layer by layer, and extract the outer contour of each layer through Gaussian blur, Canny edge detection and morphological operations; Draw the detected edge and outer contour information; Contours are extracted along the axial, coronal and sagittal directions and combined into a complete three-dimensional contour; The step S30 includes: Merge the point cloud data extracted from the axial, coronal, and sagittal directions into a single dataset; Use KD tree to find each point and its two nearest neighbors; when the distance between the point and two neighbors is less than the preset distance threshold, the corresponding neighboring point is removed; The step S50 includes: Delaunay triangulation is used for the 3D point cloud data after voxel downsampling; Traverse each triangle generated by triangulation and obtain the three edges of the triangle by index to form a non-repeated edge set; Calculate the Euclidean distance between the two endpoints of each edge, retain the edges with a distance less than the parameter alpha, and generate a valid triangle edge set; Traversing each edge in the valid triangle edge set, and generating an initial boundary point set of the point cloud according to the starting point and the end point of the edge; The initial boundary point set is clustered by DBSCAN clustering to remove noise points and screen effective boundary points.

2. The method for generating a medical image abdominal contour model according to claim 1, wherein: The step S10 includes: Use the 8-nearest neighbor algorithm to identify 3D image data and mark connected areas; The regions that are too small are filtered out according to the preset volume threshold, and the background regions are removed.

3. The method for generating a medical image abdominal contour model according to claim 1, wherein: The step S40 includes: Dividing the point cloud data of the three-dimensional contour after deduplication processing into a plurality of cubic grids; each small cube is a voxel; For each voxel, only one representative point within the voxel is retained, and the representative point is the centroid, a random selection, or a point with significant features.

4. The method for generating a medical image abdominal contour model according to claim 1, wherein: The step S70 includes: Delaunay triangulation is used for the point cloud data within the valid cluster, and triangles with a circumscribed circle radius less than Alpha are selected; The mesh surface is smoothed by averaging the coordinates of each vertex with its surrounding adjacent vertices, the normal vector of each vertex of the mesh is calculated, and the final three-dimensional abdominal contour model is output.

5. A medical image abdominal contour model generation system, characterized in that: include: Image preprocessing module: Marks the connected areas in the 3D image data, generates marked 3D volume data, and filters out areas smaller than a preset volume threshold; Contour scanning module: Slice the retained 3D image data layer by layer, extract the outer contour of each layer through Gaussian blur, Canny edge detection and morphological operations, and generate a 3D contour based on the combination of multi-directional contours; Overlap removal module: using KD tree to perform neighbor search to remove duplicate points of the three-dimensional contour; Grid voxel downsampling module: performs voxel downsampling on the point cloud data of the three-dimensional contour after deduplication processing to reduce the number of point clouds; Alpha shape module: extracts the outer contours of irregular shapes from voxelized downsampled 3D point cloud data through Delaunay triangulation, boundary screening, and DBSCAN clustering algorithms; Abdominal Decongestion Module: This module uses the DBSCAN algorithm again to remove noise points from the point cloud and select valid clusters. It also uses the DBSCAN algorithm to divide the point cloud into connected regions and separate them into clusters. Clusters with a point count greater than min_size are selected based on the min_size parameter. A Boolean mask is created to mark points belonging to large clusters, and a point cloud containing only points from large clusters is returned. Model output module: generates the final three-dimensional abdominal contour model by triangulating the point cloud data within the valid cluster; The contour scanning module is specifically used to: obtain the size of the three-dimensional image data retaining the connected area; perform slice contour detection in different directions according to the size of the three-dimensional image data; Performing a normalization operation on the three-dimensional image data to map it to a range of 0 to 255; Traverse each slice of the 3D image data layer by layer, and extract the outer contour of each layer through Gaussian blur, Canny edge detection and morphological operations; Draw the detected edge and outer contour information; Contours are extracted along the axial, coronal and sagittal directions and combined into a complete three-dimensional contour; The overlap removal module is specifically used to merge the point cloud data extracted from the axial, coronal and sagittal directions into a single data set; Use KD tree to find each point and its two nearest neighbors; when the distance between the point and two neighbors is less than the preset distance threshold, the corresponding neighboring point is removed; The Alpha shape module is specifically used to: use Delaunay triangulation on the 3D point cloud data after voxel downsampling; Traverse each triangle generated by triangulation and obtain the three edges of the triangle by index to form a non-repeated edge set; Calculate the Euclidean distance between the two endpoints of each edge, retain the edges with a distance less than the parameter alpha, and generate a valid triangle edge set; Traversing each edge in the valid triangle edge set, and generating an initial boundary point set of the point cloud according to the starting point and the end point of the edge; The initial boundary point set is clustered by DBSCAN clustering to remove noise points and screen effective boundary points.

Images