A kind of method obtaining three-dimensional tracheae tree from lung CT image
Technical field
The present invention relates to computer image processing technology field, particularly a kind of method obtaining three-dimensional tracheae tree from lung CT image.
Background technology
Obtain the basis that accurate intratracheal tree is the automatic diagnosis of lung qi pipe relevant diseases parameter, the accurate extraction of lung qi Guan Shu is significant for the computer-aided diagnosis system of pulmonary disease.Utilize preoperative lung three-dimensional CT image, integrative medicine image processing techniques, computer graphics techniques and modern electronic technology reconstruct three-dimensional tracheae tree, realize guiding in real time in art, the probability met accident when greatly can reduce patient's row bronchoscopy.
In order to rebuild more multistage bronchus, many researchers propose and utilize the Anatomical Structure information of bronchial tree and the method for relevant information.These class methods can be roughly divided into five classes: the 1) method of knowledge based or rule; 2) template matching method; 3) morphological method; 4) shape analysis method; 5) mixed method.The method of knowledge based or rule attempts to introduce when bronchial tree is rebuild the knowledge such as the anatomy relationship priori of lung's bronchus and blood vessel, tracheae geometry, local image characteristics, fuzzy logic and connectivity.The principle of template matching method is according to bronchus anatomical structure priori predefine one group of mask or template varied in size, the extraction of aided two-dimensional or three-dimensional image space mesobronchus architectural feature.Such as, Kaftan then attempts the stay in place form of tree path for bronchial reconstruction.Morphological method is often used to the three-dimensional bronchial tree of refinement original reconstruction (such as by 3D region growth algorithm), and its design philosophy attempts utilizing various morphological operator, connects or the two-dimensional/three-dimensional bronchi image region of mixed fracture.Such as, Aykac etc. propose the same area utilizing expansion form operator to be connected to adjacent image layers, to improve the bronchial reconstruction accuracy rate of single layer image.In the literature, Fetita etc. propose the Mathematical Morphology Method based on connecting cost function, detect bronchiolar region.Local image characteristics method is the feature being usually expressed as pipeline configuration according to bronchus, utilizes the method solving the eigenwert of Hessian matrix, is strengthened set with reconstruction of three-dimensional tubular trachea by the second derivative analyzing tracheae border.
Those skilled in the art are clear, and revealing and blocking is the two large main bugbears that current bronchial tree is rebuild.Causing the main cause revealed with blocking to be that CT image exists partial volume effect, causing the lumen contrast of air in the bronchial tube wall of lung and air flue to reduce.Reveal and the tracheae tree causing rebuilding is merged with its periphery lung tissue (such as, pulmonary parenchyma); Block the tracheae tree then causing rebuilding rupture and rebuild the discontinuous of tracheae.In addition, picture noise, image artifacts and imaging time the respiratory movement motion artifacts or image blurring that causes, bring huge challenge all can to the reconstruction of bronchial tree.Such as, and in time running into some pulmonary disease, chronic obstructive pulmonary disease or interstitial lung disease, the difficulty of reconstruction can be more obvious.And leakage and obstruction are the conflicts that three-dimensional tracheae tree is extracted, this contradiction impossible complete under single algorithm frame.
Therefore, a kind of method obtaining three-dimensional tracheae tree from lung CT image of the impact that can reduce to reveal and block is needed badly.
Summary of the invention
The object of the present invention is to provide a kind of a kind of method obtaining three-dimensional tracheae tree from lung CT image that can reduce the impact of revealing and blocking in tracheal reconstruction process.The method comprises the following steps: adopt adaptive three-dimensional spatial area growth method to extract main bronchus; The method of optimized image feature extraction is adopted to extract other segmenta bronchopulmonalias except main bronchus; Adopt fuzzy connectedness algorithm main bronchus and described segmenta bronchopulmonalia to be carried out " stitching ", obtain three-dimensional tracheae tree.
As a kind of preferred version, described extract the step of main bronchus before, first to the smoothing process of CT image, by the second order arrangement architecture of the CT value of certain tissue points and the local brightness variation around it in binding analysis CT image, and analyze described tissue points and whether belong to tubular structure, filter out and belong to bronchial tissue points, belong to bronchial tissue points described in collecting all and obtain one-level tracheae pretreatment image.
As a kind of preferred version, described extract the step of main bronchus before, described one-level tracheae pretreatment image is carried out closed operation, first the structural element of described one-level tracheae pretreatment image is expanded, and then the image structural element after expanding is corroded, for filling duck eye, making the edge smoothing of object, obtaining secondary tracheae pretreatment image.
As a kind of preferred version, in the step of described extraction main bronchus, specifically comprise the following steps: choose initial Seed Points; The criterion selecting adaptive local adjacent thresholds method to increase as region obtains threshold value, and the tissue points being more than or equal to described threshold value is integrated into seed region as Seed Points; After all Seed Points of periphery are all integrated into seed region, obtain main bronchus image.
As a kind of preferred version, extracting in the step of other segmenta bronchopulmonalias except main bronchus described, specifically comprising the following steps: extract some characteristics of image for building the energy term in cost function; Adopt the method for Multiple Kernel Learning, use the compound kernel function of described feature to be embedded in three-dimensional bronchial Seed Points extraction algorithm, obtain the Seed Points forming described segmenta bronchopulmonalia; Each independently segmenta bronchopulmonalia is obtained continuously according to the Seed Points of continuity to the described segmenta bronchopulmonalia of composition.
As a kind of preferred version, described some characteristics of image comprise 3D pipeline feature coefficient, local phase, SIFT feature, low-definition version Haar-like feature based on multiple dimensioned Hessian matrix.
As a kind of preferred version, the Seed Points of described for described composition segmenta bronchopulmonalia is being carried out, in continuous print step, have employed the method based on Snake Spline Model.
As a kind of preferred version, described, main bronchus and described segmenta bronchopulmonalia are carried out in the step of " stitching ", first by described in main bronchus and adjacent one end independently segmenta bronchopulmonalia sew up, then carry out iterative computation until whole segmenta bronchopulmonalias " stitching " to together.
As a kind of preferred version, described, main bronchus and described segmenta bronchopulmonalia are carried out in the step of " stitching ", utilize central line pick-up algorithm, extract the center line of described main bronchus, wherein one or more degree of sentencing in general space distance, local image characteristics, bronchus anatomical structure, structure Fuzzy connected degree function, from the end of described main bronchus, utilize three-dimensional fuzzy connection algorithm, main bronchus and described segmenta bronchopulmonalia are coupled together.
As a kind of preferred version, described, main bronchus and described segmenta bronchopulmonalia are carried out in the step of " stitching ", utilize central line pick-up algorithm, extract the center line of each segmenta bronchopulmonalia, wherein one or more degree of sentencing in general space distance, local image characteristics, bronchus anatomical structure, structure Fuzzy connected degree function, utilizes three-dimensional fuzzy connection algorithm, iteration connects each described segmenta bronchopulmonalia, until reconstruct whole bronchial tree.
Implement the present invention, a kind of method obtaining three-dimensional tracheae tree from lung CT image can be obtained, reduce in tracheae tree process of reconstruction owing to revealing and blocking the impact caused.
Accompanying drawing explanation
Fig. 1 is a kind of Technology Roadmap obtaining the method for three-dimensional tracheae tree from lung CT image provided by the invention;
Fig. 2 is the person's windpipe tree construction simulation drawing adopting a kind of method obtaining three-dimensional tracheae tree from lung CT image provided by the invention to obtain;
Fig. 3 is that the main bronchus adopting a kind of method obtaining three-dimensional tracheae tree from lung CT image provided by the invention to obtain extracts result simulation drawing;
Fig. 4 is the segmenta bronchopulmonalia leaching process simulation drawing adopting a kind of method obtaining three-dimensional tracheae tree from lung CT image provided by the invention to obtain;
Fig. 5 is the extraction result simulation drawing of the segmenta bronchopulmonalia adopting a kind of method obtaining three-dimensional tracheae tree from lung CT image provided by the invention to obtain;
Fig. 6 adopts a kind of method obtaining three-dimensional tracheae tree from lung CT image provided by the invention to carry out the result schematic diagram of " stitching ".
Embodiment
With reference to figure 1, Fig. 2, the invention provides a kind of method obtaining three-dimensional tracheae tree from lung CT image, mainly comprise the following steps: as shown in arrow 1, adopt adaptive three-dimensional spatial area growth method to extract main bronchus, for ease of describing, by this step referred to as step S101; As shown in arrow 2 and arrow 3, the method for optimized image feature extraction is adopted to extract other segmenta bronchopulmonalias except main bronchus, by this step referred to as step S103; As shown in arrow 4, adopt fuzzy connectedness algorithm main bronchus and segmenta bronchopulmonalia to be carried out " stitching ", obtain the tracheae tree of three-dimensional as shown in Figure 2, by this step referred to as step S105.
When carrying out step S101, preferably first to the smoothing process of CT image of three-dimensional, by the second order arrangement architecture of the CT value of certain tissue points in binding analysis CT image and the local brightness variation around it, and analyze this tissue points and whether belong to tubular structure, filter out and belong to bronchial tissue points, collect whole bronchial tissue points that belongs to and obtain one-level tracheae pretreatment image.Concrete grammar is by analyzing Hessian matrix stress release treatment to the impact of extracting tracheae.
Under table 1 three-dimensional situation, various possibility structure is with the relation table of Hessian proper value of matrix
Upper table be under three-dimensional situation various may structure with the relation of Hessian proper value of matrix, λ k represents the eigenwert that a kth amplitude is minimum, 3 eigenvalue λ 1 of Hessian matrix, λ 2, λ 3 (| λ 1|≤| λ 2|≤| λ 3|) in, amplitude maximum eigenwert characteristic of correspondence vector represents the maximum direction of certain tissue points curvature, and amplitude minimum eigenwert characteristic of correspondence vector represent the minimum direction of certain tissue points curvature.In CT image, tracheae is always dark, so should be that λ 1 is less in the eigenwert of the said three-dimensional body vegetarian refreshments at lung CT image tracheae place, be almost 0, λ 2 and λ 3 is positive number.CT image is calculated to the Hessian matrix of each tissue points, and calculate its eigenwert, determine whether tracheae voxel.
Implement after above-mentioned steps, the air pipe structure after enhancing has more clearly showed out.But because the gray scale that may be subject to the pipe that phlegm or other human body fluids cause is inconsistent, and the tracheae that tracheae branch causes is undetected, or the reason such as other local noises, make the main bronchus extracted still may there is discontinuous phenomenon.So carry out closed operation to this one-level tracheae pretreatment image, namely first the structural element of one-level tracheae pretreatment image is expanded, all background dots with object contact are merged in this object, border is externally expanded, and final result is that integral image expands a circle, and then corrodes the image structural element after expanding, eliminate frontier point, border is internally shunk, and the process of whole closed operation makes the tracheae edge smoothing of CT image, obtains secondary tracheae pretreatment image.
Step S101 is carried out on the basis obtaining secondary tracheae pretreatment image, specifically comprises the following steps:
Choose initial Seed Points, method for optimizing is manually, can more accurately find more quickly accurately main bronchus 100 voxel as initial Seed Points;
The criterion selecting adaptive local adjacent thresholds method to increase as region obtains threshold value, and the tissue points being more than or equal to this threshold value is integrated into seed region as Seed Points;
Until after all Seed Points of periphery are all integrated into seed region, obtain main bronchus 100 image as shown in Figure 3.
Specifically comprise the following steps in step S103:
Extract some characteristics of image as 3D pipeline feature coefficient, local phase, SIFT feature, the low-definition version Haar-like feature based on multiple dimensioned Hessian matrix, when cost function energy term form is determined, adopt the method for Multiple Kernel Learning, the compound kernel function of described feature is used to be embedded in three-dimensional bronchial Seed Points extraction algorithm, obtain the weight that different characteristic is applicable to, the energy term built in cost function is out of shape to drive shape;
Obtain the Seed Points forming described segmenta bronchopulmonalia 200;
Snake Spline Model method is adopted to obtain each independently segmenta bronchopulmonalia 200 continuously according to the Seed Points of continuity to the described segmenta bronchopulmonalia 200 of composition, as shown in Figure 4, Figure 5.
Image characteristics extraction algorithm is adopted both to avoid bronchus, in growth structure process, " fracture " occur---growth stops, and turn avoid generation " leakage "---to grow into outside lung qi pipe and organize, the regions such as such as alveolar.According to clinical experience, trickle segmenta bronchopulmonalia 200, partial structurtes is have significant multi-scale image feature, but can not ensures to communicate with each other, therefore need implementation step S105 to carry out " stitching ".
Specifically comprise the following steps in step S105:
First carried out " stitchings " by the main bronchus 100 independently segmenta bronchopulmonalia 200 with adjacent one end, concrete grammar carries out " fuzzy " connection based on the centerline direction of local main bronchus 100 to adjacent segmenta bronchopulmonalia 200.After main bronchus 100 and the independently segmenta bronchopulmonalia 200 of adjacent one end are sewed up, the end deblurring recycling the segmenta bronchopulmonalia 200 that this is sewed up connects adjacent segmenta bronchopulmonalia 200, then iterative computation is adopted until whole segmenta bronchopulmonalias 200 " stitching " to together, obtain measurements of the chest, waist and hips tracheae tree as shown in Figure 6.
Central line pick-up algorithm is have employed in above-mentioned stitching step, extract the center line of main bronchus 100, extract the center line of each segmenta bronchopulmonalia 200 adopting to use the same method, the degree of sentencing such as general space distance, local image characteristics, bronchus anatomical structure, structure Fuzzy connected degree function, from the end of described main bronchus 100, utilizes three-dimensional fuzzy connection algorithm, iteration connects each described segmenta bronchopulmonalia 200, until reconstruct whole bronchial tree.
The largest benefit of " stitching " is to solve bronchus tube wall is thinning, liquid fills, bronchus subsides and the clinical pulmonary lesion such as pathological change causes local bronchiole characteristics of image and disappears and the problem that cannot be communicated with.Although final local 3D tracheae result has certain geometric unsharpness, in clinical tolerance interval, do not affect the whole structure of bronchoscopic Surgical Evaluation and surgical simulation training.
In processing scheme of the present invention, first extract main bronchus and segmenta bronchopulmonalia 200 respectively, efficiently solve the problem that main bronchus in conventional art easily blocks, segmenta bronchopulmonalia 200 is easily revealed.Subsequently, the main bronchus extracted, segmenta bronchopulmonalia 200 are combined into one, obtain complete three-dimensional tracheae tree.
For the person of ordinary skill of the art, without departing from the inventive concept of the premise, can also make some distortion and improvement, these all belong to protection scope of the present invention.Therefore, the protection domain of patent of the present invention should be as the criterion with claims.