CN108154525A - A kind of matched bone fragments joining method of feature based - Google Patents

Abstract

The invention discloses a kind of matched bone fragments joining method of feature based, by the way that being registrated from the three-dimensional bone fragments model of medicine image reconstruction, preoperative guidance is provided for fracture of mandible reparation.This method comprises the following steps：1）Import three-dimensional bone fragments model, the plane of disruption of manual extraction fragment model；2）Based on 3D SIFT algorithms, the key point on the bone fragments model plane of disruption is extracted；3）Point of use feature histogram FPFH algorithms are key point construction feature description of extraction；4）Key point and corresponding description based on extraction establish two bone fragments models（It is referred to as model subject to registration, object module）Initial correspondence；5）Optimize initial correspondence using improved ICP algorithm, obtain final transformation matrix；6）Transformation matrix obtained in the previous step is acted in model subject to registration, obtains the splicing result of two bone fragments models.Method flow is as shown in Figure 1.

Description

A Skeleton Fragment Assembling Method Based on Feature Matching

technical field

The invention relates to the technical field of computer graphics and computer-aided surgical digital design, and mainly relates to a bone fragment splicing method based on feature matching

Background technique

In the early 1990s, domestic and foreign scholars began to apply computer technology such as computer graphics and imaging and virtual reality technology to craniofacial surgery, making traditional craniofacial surgery develop in the direction of digitalization and minimal invasiveness, and become An important direction of development in the medical field. The craniofacial region contains many important blood vessels and nerve tissues, and it also takes into account the beauty of the face. These reasons put forward higher requirements for the accuracy of craniofacial surgery. The traditional surgical method mainly relies on the doctor's personal surgical experience, which results in a longer operation time and greater trauma to the patient. Therefore, computer-aided surgical planning for craniomaxillofacial parts has always been a hot research field in digital surgery.

Aiming at the reduction of craniofacial fracture trauma, the patent of this invention designs a bone fragment repair technology based on three-dimensional model feature extraction and three-dimensional registration technology, which is used to assist doctors in preoperative surgical planning and formulate reasonable surgical plans, thereby reducing The secondary injury to the patient during the operation improves the success rate of the operation.

According to relevant statistics at home and abroad, mandibular fractures account for 25%-28% of the total number of craniofacial injuries and 55%-72% of the total number of maxillofacial bone fractures. Therefore, the patent of the present invention has important application value and broad application prospects for mandibular fracture repair in craniomaxillofacial region.

Contents of the invention

The purpose of the present invention is to provide a preoperative planning and design method for the craniomaxillofacial region, which is most vulnerable to injury, by using relevant computer technology, and assist doctors in formulating a preoperative operation plan.

The technical scheme that the present invention adopts is as follows:

A bone fragment splicing method based on feature matching, the technical scheme is as follows:

1) From the reconstructed three-dimensional bone fragment model, manually extract the fracture surface area; the three-dimensional model used in the present invention is reconstructed from the patient's CT image data.

2) For the extracted fracture surface area, based on the 3D-SIFT feature extraction algorithm, extract the feature points of the fracture area;

In the traditional 3D-SIFT feature extraction algorithm, the principle of selecting feature points is to select data points with larger curvatures in the model. The present invention further improves the traditional algorithm by setting the minimum curvature threshold to remove candidates with smaller curvatures in the model. Key points to obtain feature points of key parts and improve registration accuracy and efficiency; mainly includes three steps:

2.1) Establish scale space and Gaussian pyramid for 3D model;

2.2) Find candidate feature points and remove feature points with low contrast in the scale space;

2.3) By setting the minimum curvature threshold, the candidate key points with smaller curvature in the model are removed, and the remaining candidate feature points are the obtained key feature points.

3) To extract each feature point, calculate and establish a fast point feature histogram FPFH descriptor, the establishment process is as follows:

3.1) For each feature point p, use the formula (1) to calculate the triplet (f1, f2, f3) between it and all its neighbors, and then construct a histogram statistically for all triplets, That is, the simplified feature histogram SPFH of the calculation point p;

(1)

3.2) Query its neighboring points for each point p _i in the neighborhood of p, and calculate the SPFH value of each p _i point, use the SPFH value of the neighboring point p _i to calculate the FPFH value of the query point p, the specific calculation method See formula (2)

(2)

4) Based on the extracted feature points and the FPFH feature descriptors corresponding to the feature points, establish the corresponding point pairs between the two bone fragment models (referred to as the source model and the target model respectively);

The present invention adopts two ways to establish feature correspondences, one is direct query, and the other is circular iterative search for optimality. The direct query algorithm has high accuracy; but the time complexity is also high, resulting in low efficiency. The main idea of the cyclic iterative optimal method is to find the optimal correspondence by continuously reducing the distance difference between the model point sets, and it is an algorithm oriented to the final registration effect. The present invention combines the advantages of these two algorithms to establish the preliminary corresponding relationship before the two models. Through direct query, some feature points with weaker features are excluded, and then the corresponding relationship is established for the remaining feature points based on the optimal method of circular iterative search to ensure accuracy and improve efficiency. The algorithm steps are as follows:

4.1) Randomly select s feature points among the feature points of the model to be registered, and the distance between these s feature points must be greater than a set threshold;

4.2) For each point S _i among the selected s feature points, in the feature description subspace of the target model feature point, select k neighboring points, k is 10, and randomly select one of the 10 points as The relationship points corresponding to point S _i , thus forming s corresponding relationship point pairs;

4.3) According to the obtained s corresponding relationship point pairs, calculate the transformation matrix T by the method of SVD, and use the distance error function value between the two models to evaluate this transformation, if it is better than the last cycle, replace Transformation matrix, otherwise, keep the transformation matrix T in the last loop;

4.4) Repeat the above three steps until the maximum number of cycles is reached; in each cycle, only the optimal value of the transformation matrix obtained in this cycle and the transformation matrix stored in the previous cycle is stored, that is, the difference between the two models The transformation matrix T with smaller distance error function value.

5) Use the improved ICP algorithm to optimize the initial correspondence to obtain the final transformation matrix;

In order to further optimize the registration result, the present invention improves the iterative closest point method (ICP) to further optimize the registration result to obtain a better transformation matrix; the steps are as follows:

5.1) Construct point pairs according to the nearest neighbor selection algorithm: for each point in the model _Mi with registration, search for the nearest point in the target input model T to form a corresponding point pair, and finally find two All corresponding point pairs in the point set. In order to improve the anti-noise capability of the ICP algorithm, the present invention only constitutes a corresponding point pair when the distance between the closest point and the query point is less than the set threshold;

5.2) Calculate the rotation matrix R and translation vector t of the rigid transformation between the two models according to the set of corresponding point pairs calculated in the previous step;

5.3) According to the obtained rotation matrix and translation vector R, t, transform _Mi to obtain a new model Mi ₊₁ ; after that, the traditional method calculates the sum of the squares of the distance between Mi ₊₁ and _Mi to continuously The absolute value of the difference between the sum of squares of the two distances is used as the basis for convergence to determine whether to stop the iteration. The present invention sets the termination condition as the absolute value Epsilon of the error rate of change between two consecutive models, wherein Epsilon=|( _Ec - _Ec-1 )/ _Ec-1 |, _Ec is the difference between the two models The sum of the corresponding point pair errors between them is used to improve the efficiency of the algorithm;

5.4) Repeat the above steps until convergence or reach the set maximum number of iterations, save the transformation matrix.

6) Apply the transformation matrix to the model to be registered to obtain the result of splicing the two models;

By repeating the above steps, the splicing results of multiple fragment models can be obtained, and the correct reset of the fragment models caused by fracture trauma can be realized.

In summary, owing to adopting above-mentioned technical scheme, the beneficial effect of the present invention is:

1. Through 3D feature extraction technology and 3D registration technology, the splicing and repairing of multiple bone fragment models is realized, so that doctors can have a preliminary plan for the restoration position of bone fragments before operation, and provide preoperative guidance for doctors, so as to shorten the It has a very broad application prospect and important application value in the repair of mandibular fractures by reducing the operation time and reducing the secondary injury to the patient during the operation.

Description of drawings

The invention will be illustrated by way of example with reference to the accompanying drawings, in which:

Fig.1 Skeleton fragment model and extracted fracture surface

Figure 2 SPFH feature histogram of feature points

Figure 3 A bone fragment model from different perspectives

Figure 4 The splicing model of the fragment model in Figure 3 from different perspectives

Fig.5 Partial mandibular bone fragment model from different perspectives of the first case

Figure 6 The splicing repair model of the fragment model in Figure 5 from different perspectives

Figure 7 Part of the mandibular bone fragment model from different perspectives in the second case

Figure 8 The splicing repair model of the fragment model in Figure 7 from different perspectives

Fig.9 The complete mandibular fracture model of the third case from different perspectives

Figure 10 The splicing repair model of the fragment model in Figure 9 from different perspectives

Detailed ways

All features disclosed in this specification, or steps in all methods or processes disclosed, may be combined in any manner, except for mutually exclusive features and/or steps. Below in conjunction with Fig. 1-Fig. 10 the present invention is described:

A method for splicing bone fragments based on feature matching, Figure 1 is the overall flow chart of the technology; the detailed steps are as follows:

1) Manually extract the fracture surface area from the reconstructed 3D bone fragment model. The three-dimensional model used in the present invention is reconstructed from CT image data of the patient. As shown in Figure 2, the two imported mandibular fragment models are on the left, and the fracture surface regions extracted from the two models are on the right;

2) To extract the fracture surface area, based on the 3D-SIFT feature extraction algorithm, extract the feature points of the fracture area, as shown in Figure 3, the left side is a fragment model imported, and the right side is the display of the feature points extracted in the fragment model;

3) To extract each feature point, calculate and establish a fast point feature histogram FPFH descriptor; Figure 4 shows the SPFH feature histogram of a feature point;

4) Based on the extracted feature points and the FPFH feature descriptors corresponding to the feature points, establish the corresponding point pairs between the two bone fragment models (referred to as the source model and the target model respectively);

5) In order to further optimize the registration result, the present invention improves the well-known iterative closest point method (ICP), which is used to further optimize the registration result to obtain a better transformation matrix;

6) Apply the transformation matrix to the model to be registered to obtain the result of splicing the two models;

Below is to show the result of the present invention by three groups of data:

Figure 5, Figure 7, and Figure 9 are three sets of experimental data and three sets of bone fragment models

Figure 6, Figure 8, and Figure 10 are the experimental results of three sets of data obtained based on the above steps. Three groups of experiments show that splicing repair methods based on fractured regions can successfully repair and reconstruct common mandibular fractures.

Claims (5)

1. This invention discloses a bone fragment splicing method based on feature matching. By registering between every two fragment models before operation, the splicing of all fracture fragment models is finally realized, and a preoperative fracture repair plan is generated, shortening Operation time, reduce the secondary injury that operation time is too long to patient; It is characterized in that, described method comprises the following steps: 1) Import the 3D bone fragment model, and manually extract the fracture surface of the fragment model; 2) Based on the 3D-SIFT algorithm, extract the key points on the fracture surface of the bone fragment model; 3) Use the point feature histogram FPFH algorithm to construct feature descriptors for the extracted key points; 4) Based on the extracted key points and the corresponding descriptors, establish the initial corresponding relationship between the two bone fragment models (referred to as the registration model and the target model); 5) Use the improved ICP algorithm to optimize the initial correspondence to obtain the final transformation matrix; 6) Apply the transformation matrix obtained in the previous step to the model to be registered to obtain the splicing result of the two bone fragment models. 2. A method for splicing bone fragments based on feature matching according to claim 1, characterized in that said step 2) includes the following steps: 2.1) Establish scale space and Gaussian pyramid for 3D model; 2.2) Calculate candidate feature points and remove feature points with low contrast in the scale space; 2.3) By setting the minimum curvature threshold, the candidate key points with smaller curvature in the model are removed, and the remaining candidate feature points are the obtained key feature points. 3. A method for splicing bone fragments based on feature matching according to claim 1, wherein said step 3) includes the following steps: 3.1) For each feature point p extracted, calculate its simplified feature histogram SPFH; 3.2) Query its neighboring points for each point p _i in the neighborhood of p, and calculate the SPFH value of each point p _i , and use the SPFH value of the neighboring point p _i to calculate the FPFH value of the query point p. 4. A method for splicing bone fragments based on feature matching according to claim 1, characterized in that said step 4) includes the following steps: 4.1) Randomly select s feature points among the feature points of the model to be registered, and the distance between these s feature points must be greater than a set threshold; 4.2) For each point S _i among the selected s feature points, in the feature description subspace of the target model feature point, select k neighboring points, where the value of k is 10, and randomly select one of the 10 points As the relationship point corresponding to the point S _i , thus forming s corresponding relationship point pairs; 4.3) According to the obtained s corresponding relationship point pairs, calculate the transformation matrix T by the method of SVD, and use the distance error function value between the two models to evaluate this transformation, if it is better than the last cycle, replace Transformation matrix, otherwise, keep the transformation matrix T in the last loop; 4.4) Repeat the above three steps until the maximum number of cycles is reached; in each cycle, only the optimal value of the transformation matrix obtained in this cycle and the transformation matrix stored in the previous cycle is stored, that is, the difference between the two models The transformation matrix T with smaller distance error function value. 5. A method for splicing bone fragments based on feature matching according to claim 1, characterized in that said step 5) includes the following steps: 5.1) Construct point pairs according to the nearest neighbor selection algorithm: for each point in the model _Mi to be registered, search for the nearest point in the target input model T to form a corresponding point pair, and finally find two All corresponding point pairs in the point set; in order to improve the anti-noise ability of the ICP algorithm, the present invention forms a corresponding point pair when the distance between the nearest point and the query point is less than the set threshold; 5.2) Calculate the rotation matrix R and translation vector t of the rigid transformation between the two models according to the set of corresponding point pairs calculated in the previous step; 5.3) According to the obtained rotation matrix and translation vector R, t, transform _Mi to obtain a new model Mi ₊₁ ; calculate the absolute value Epsilon of the error change rate between two consecutive models, where Epsilon=|(E _c -E _c-1 )/E _c-1 |, E _c is the sum of the corresponding point pair errors between the two models, and use this as the basis for convergence to determine whether to stop iterations; 5.4) Repeat the above steps until it converges or reaches the maximum number of iterations we set, and save the transformation matrix.