CN103034989B - A kind of low dosage CBCT image de-noising method based on high-quality prior image - Google Patents

Abstract

The present invention relates to a kind of low dosage CBCT image de-noising method based on high-quality prior image, need the high-quality high s/n ratio three-dimensional data first obtaining a width same scan position, using this qualitative picture as the prior imformation of denoising.After each CBCT scanning, in the image rebuild, voxel processes one by one.Treating denoising voxel for some, select the neighboring area centered by it and form image block, by calculating SSIM value, in high-quality prior image, finding the image block similar to it.All similar image block center voxels are weighted on average, thus obtain the result after this voxel denoising.Carry out this one by one to all voxels to calculate, then finally obtain the CBCT 3-D view after denoising.The present invention is owing to employing high s/n ratio qualitative picture as priori, denoising effect can be made to be improved significantly, and do not need in denoising process to carry out registration, and each scanning all can use same width prior image to process, and enhances the feasibility of method.

Description

A kind of low dosage CBCT image de-noising method based on high-quality prior image

Technical field

The present invention relates to radiant image and technical field of image processing, relate to CBCT image denoising disposal route, particularly a kind of low dosage CBCT image de-noising method based on high-quality prior image.

Background technology

Noise affects CBCT(Cone-beamCT, conical beam CT) key factor of picture quality, especially at low dosage, (low dosage is herein without official definition clear and definite especially, it is generally acknowledged, during the normal dose used lower than clinical diagnosis CT, namely can be called low dosage) when scanning, noise becomes more serious, now it may bury in oblivion focus, has influence on diagnosis, therefore studies suitable Image Denoising Technology very important.At present, the denoising thinking for low-dose CT mainly comprises projection domain denoising, choose reasonable filter function or iterative constrained item in process of reconstruction, rebuilds image area denoising.Specific to denoising method, the denoising method comprising wavelet multiresolution comparatively popular at present, the minimized denoising method of TV, denoising method etc. based on partial differential equation.These methods can reduce the noise level of image effectively, improve the SNR value of image, but also inevitably bring image blurring and loss that is details simultaneously, as shown in Figure 1.Therefore, study new image processing algorithm keeps detailed information very important for lifting CBCT picture quality simultaneously.

In recent years, along with the development of mathematical theory and image analysis technology, the denoising method based on redundant data and image similarity obtains rapidly the concern of people.Most representative be exactly BuadesA propose non-local mean method (see A.Buades, B.Coll, J.M.Morel.Areviewofimagedenoisingalgorithms, withanewone.MultiscaleModel.Simul, 2005,4 (2): 490-530.), it have employed image natural redundancies, utilize the similar area in every width image to carry out denoising, good reservation can be obtained to details and marginal information.The method may be summarized as follows: for given Noise image v={v (i) | and i ∈ I}, I represents image coordinate system.So non-local mean algorithm is used to estimate that the weighted mean of an all pixel of full figure is as the estimated value after this denoising for pixel i:

$\mathrm{NL}\left[v\right]\left(i\right)=\frac{1}{C_i}\underset{j\∈I}{\mathrm{\Σ}}w(i,j)v\left(j\right)$

Wherein, weighted value w (i, j) depends on the similarity between pixel i and pixel j.C _iit is normalization coefficient.Similarity how between decision block and block choose reasonable weight are the keys of this algorithm, and general similarity can be determined by the Gauss's weighted euclidean distance between neighborhood gray value vectors: if distance more closely shows between two pixels more similar, then weighting weight is larger.Can arrange weighting function is: w (i, j)=exp (d (i, j)/h ²).Here h is filtering parameter, the speed of control characteristic function decay.This algorithm can reduce the noise level in image effectively, and keeps detailed information and marginal information.But for the data that noise amplitude is strong, its denoising effect is still limited.

The introducing of prior imformation has very important meaning for image denoising.If the similar image of a width high-quality can be obtained, just effectively can recover Noise information.Someone proposes a kind of ODCT(Oracle-basedDCTfilter) hard-threshold filtering noise reduction algorithm, one-level dct transform threshold denoising is first used to obtain clean image, carry out a secondary filter again, thus obtain clean image (see OnurG.Guleryuz, WeightedAveragingforDenoisingWithOvercompleteDictionarie s. [J] .IEEETrans.ImageProcess., 2007.16 (12): 3020-3034).In addition, also have a kind of method based on dictionary learning, first train dictionary (a kind of sparse expression of image) by clean image, then denoising is carried out to noisy image.But in Medical Image Processing, high-quality image is often difficult to obtain.In addition, even if obtain high-quality image by again scanning, also need to carry out the process such as registration, ensure that structure and edge mate completely and just can be used for denoising operation, difficulty is very large, lacks practicality.

At present, the patented claim of the image de-noising method of non-local mean is utilized to have a lot, as " the non-local mean image de-noising method of integrated structure information " (application number: 201110091450.2), " non-local mean de-noising method for natural image " (application number: 201010271499.1), " the non-local mean denoising method based on associating similarity " (application number: 201110282126.9) etc.But these methods all carry out denoising based on the redundancy of image self, do not introduce suitable prior image.When prior image is introduced, the effect of denoising effectively can be promoted.For in the denoising method of low-dose CT, invention " data for projection recovers the non local average low-dose CT method for reconstructing of guiding " (application number: 201010517537.7) propose a kind of disposal route, first pass through BM3D denoising by low-dose CT projection sinogram and rebuild the reconstruction image of acquisition one width low noise, with this width image, average filter denoising is weighted for the low dosage strong noise image directly rebuild, thus realizes the effect of low-dose CT reconstruction.Invention " the low-dose CT image processing method based on the filtering of wavelet space directivity " (application number: 201010595896.4) propose a kind of image de-noising method based on wavelet space carried out in projection domain.With the present invention similar comprise " the low-dose CT image rebuilding method based on standard-dose image redundant information " (application number: 201010555893.8).The method provides the image of a secondary normal dose by previous scan-data, then carries out a low-dose CT scanning.The image of normal dose and low-dose CT are carried out registration, and then carries out denoising by the similarity between image block.The method requires that prior image and low dosage image are twice different scanning of same object, and higher for the requirement of registration.And application claims prior image position similar to low dosage image scanning, from different patients, and can not need accurate registration process.

Summary of the invention

The technology of the present invention is dealt with problems: overcome the deficiencies in the prior art, a kind of low dosage CBCT image de-noising method based on high-quality prior image is provided, have employed high s/n ratio qualitative picture as priori, utilize in priori data with treat the block similarity matching of denoising low dosage CBCT, denoising effect is improved significantly, solves the problem that low dosage CBCT picture noise is excessive, improve picture quality, improve the signal to noise ratio (S/N ratio) of image, do not lose detailed information and marginal texture information simultaneously as far as possible.

The technology of the present invention solution: a kind of principle of the low dosage CBCT image de-noising method based on high-quality prior image is, but in the CBCT image scanned for Different Individual same area, because the difference, scanning position, scanning mode, scanning angle etc. of human body are all distinct, although two width images are similar positions, its one-piece construction difference is still very large.But, when being divided into little image block to two width images and processing and mate, wherein there is the part that many height are similar.The present invention then make use of the similarity of these image blocks, for treating that denoising low dosage CBCT image carries out similar block search by tissue points in prior image, and is weighted average treatment to the similar image block searched, thus reaches the object of denoising.Because prior image has good picture quality, therefore the introducing of its information effectively can improve picture quality.Meanwhile, present invention, avoiding the step of image registration, do not need the exact matching carrying out prior image He treat denoising image, reduce the requirement for prior image, avoid the error that registration out of true is brought.

The inventive method specific implementation step is as follows:

The first step: the acquisition of high-quality prior image

Use the prior imformation that the image of high s/n ratio calculates as denoising.The following aspects is required to include to this image: (1) to treat that denoising image has similar structural information and similar intensity profile.Therefore, require similar to treating denoising image scanning position, thus ensure that CBCT rebuilds the volume data that obtains and treats that denoising image has higher similarity.In general, for the CBCT equipment of same use, as oral cavity CBCT, all gather the oromaxillo-facial region data of zone similarity.(2) this image needs good picture quality, comprises high s/n ratio, more rich detailed structure information.

For these two needs, prior image can be got by following several mode and as denoising database.

(1) higher X-ray tube electric current is selected to scan;

(2) each projection angle Multiple-Scan superposition;

(3) slow down rotational speed, in certain Rotary District, gather the more projected angle number of degrees;

(4) other can improve the means of CBCT picture quality.

Second step: image denoising

(1) for prior image P and treat that denoising data carry out pre-service, simple location matches is carried out.This matching process can manually complete, and also can use the automatic calculating method of image centroid and main shaft coupling, make organ site that two width images are corresponding and angle close.

(2) treating, in denoising CBCT data reconstruction, to choose and treat denoising tissue points x.Choose the image block v(x that the size centered by x is n × n).

(3) search in prior image.Block-by-block calculate in prior image centered by tissue points y image block v(y) and structural similarity index SSIM (x, y) v(x).

$\mathrm{SSIM}(x,y)=\frac{\left({2\mathrm{\μ}}_x{\mathrm{\μ}}_y\right)({2\mathrm{\σ}}_{\mathrm{xy}}+c_2)}{({\mathrm{\μ}}_x^2+{\mathrm{\μ}}_y^2+c_1)({\mathrm{\σ}}_x^2+{\mathrm{\σ}}_y^2+c_2)}$

Wherein μ _x, μ _ybe respectively the voxel mean value of two image blocks, σ _x, σ _ybe respectively the standard deviation of two image block voxel values, σ _xybe the covariance of two image block voxel values, c ₁=(k ₁l) ², c ₂=(k ₂l) ², L is the dynamic range of two image block gray-scale values to be calculated;

(4) standard that threshold epsilon judges as similarity is set.As SSIM (x, y) > ε, then think that two image block centrosome vegetarian refreshments are similar, then this tissue points can be used for treating handling body vegetarian refreshments x and be weighted average denoising.Otherwise, and think do not use two tissue points dissmilarities in weighted mean procedure.

(5) choosing all similitude data, for treating that denoising tissue points x is weighted on average, obtaining the data after denoising:

$v^{\′}\left(x\right)=\frac{1}{C\left(x\right)}\underset{y\∈P}{\mathrm{\Σ}}w(x,y)v\left(y\right)$

Y is the tissue points in all similar areas searched in noisy image, and w (x, y) is weighting function, and C (x) is weight normalized factor, and its computing method are:

$C\left(x\right)=\underset{y\∈P}{\mathrm{\Σ}}w(x,y).$

The computing method of weighting function w (x, y) are also according to SSIM(x, y) carry out conversion and obtain.Due to SSIM(x, y) value be [-1,1], and when its value absolute value close to 1 time, two image blocks are more similar.And traditional non-local mean adds temporary, when distance is 0, two image blocks are more similar, and weighted value is maximum.Therefore, Similarity Structure parameter is defined: S (x, y)=1-|SSIM (x, y) |.Weighting function can be written as:

w(i,j)=exp(S(x,y)/h ²)

Wherein h is filtering parameter, generally becomes certain linear relationship with prior image noise criteria difference σ.

The method directly can carry out denoising to three-dimensional data.In addition, judge in second step that the similarity of image block also can utilize other similar approach, be not confined to computation structure index of similarity SSIM, the present invention is just as example.

In addition, use the method, a kind of CBCT image de-noising method based on database can be obtained.Namely by clinical scan data, set up a series of patients database, the image in database all uses normal dose to obtain, and has good signal to noise ratio (S/N ratio) and picture quality.After this Database, when carrying out low dosage CBCT image denoising, can carry out in a database choosing image the most similar, flow process processes according to the method described above.This process can be completed online by network.

The present invention's advantage is compared with prior art:

(1) present invention employs high s/n ratio qualitative picture as priori, combining image denoising process simultaneously, denoising effect is improved significantly, solve the problem that CBCT picture noise is excessive, improve picture quality, improve the signal to noise ratio (S/N ratio) of image, do not lose detailed information and marginal texture information simultaneously as far as possible.

(2) the present invention requires lower for prior image, not Structure of need information accuracy registration, only need scanning position close, do not need in denoising process to carry out registration, and each scanning all can use same width prior image to process, enhance the feasibility of method, reduce the difficulty of enforcement.In actual applications, the prior data bank of a set of high-quality can first be set up.In clinical practice, the scanning for different patient can share a set of prior image information and carry out denoising.

Accompanying drawing explanation

Fig. 1 is the confirmatory experiment carried out the inventive method, and upper figure is used high-quality prior image, and figure below is the low-dose CT image treating denoising;

Fig. 2 is the Comparative result figure that use the inventive method carries out before and after denoising;

Fig. 3 is the partial enlargement comparison diagram that use the inventive method carries out before and after denoising;

Fig. 4 is realization flow figure of the present invention;

Fig. 5 is the confirmatory experiment using oral cavity CBCT clinical data to carry out, and left figure is used high-quality prior image, and right figure is the low dosage CBCT image treating denoising;

Fig. 6 uses the inventive method to carry out the Comparative result figure before and after denoising to oral cavity CBCT low dose imaging data;

Fig. 7 uses the inventive method to carry out the partial enlargement comparison diagram before and after denoising to oral cavity CBCT low dose imaging data.

Embodiment

As shown in Figure 3, the inventive method is divided into two steps, specific as follows:

The first step is the acquisition of high-quality prior image, and second step is the image denoising carried out based on this prior image.CBCT for different patient scans, if scanning area and organ-tissue similar, same width prior image generally can be adopted to process.Due to CBCT scanning obtain for three-dimensional data, therefore for be that three-dimensional data processes.But clear and easy in order to what describe, use two-dimentional tomography to be described the method below.The method directly can extend to three-dimensional situation.

Be specially:

The first step: the acquisition of high-quality prior image

This step needs the prior imformation using the image of high s/n ratio to calculate as denoising.The following aspects is required to include to this image: (1) to treat that denoising image has similar structural information and similar intensity profile.Therefore, require similar to treating denoising image scanning position, thus ensure that CBCT rebuilds the volume data that obtains and treats that denoising image has higher similarity.In general, scan same position and obtain prior image, such as oral cavity CBCT, two width images all gather oromaxillo-facial region data.(2) this image needs good picture quality, comprises high s/n ratio, more rich detailed structure information.

For these two needs, prior image can be got by following several mode and as denoising database.

(1) higher X-ray tube electric current is selected to scan;

(2) each projection angle Multiple-Scan superposition;

(3) slow down rotational speed, in certain Rotary District, gather the more projected angle number of degrees;

(4) use multi-detector diagnosis CT to generate three-dimensional data, and it is corrected, CT value and space voxel size are corresponded in corresponding CBCT system, as prior image.

Second step: image denoising

(1) for prior image P and treat that denoising data carry out pre-service, carry out simple position and slightly mate.This matching process can manually complete, and also can use the automatic calculating method of image reform and main shaft coupling, make organ site that two width images are corresponding and angle close, thus the size of region of search can be reduced.

(2) treating, in denoising CBCT data reconstruction, to choose and treat denoising tissue points x.Choose the image block v(x that the size centered by x is n × n).

(3) search in prior image.Block-by-block calculate in prior image centered by voxel y image block v(y) and structural similarity index SSIM (x, y) v(x).

$\mathrm{SSIM}(x,y)=\frac{\left({2\mathrm{\μ}}_x{\mathrm{\μ}}_y\right)({2\mathrm{\σ}}_{\mathrm{xy}}+c_2)}{({\mathrm{\μ}}_x^2+{\mathrm{\μ}}_y^2+c_1)({\mathrm{\σ}}_x^2+{\mathrm{\σ}}_y^2+c_2)}$

Wherein μ _x, μ _ybe respectively the voxel mean value of two image blocks, σ _x, σ _ybe respectively the standard deviation of two image block voxel values, σ _xybe the covariance of two image block voxel values, c ₁=(k ₁l) ², c ₂=(k ₂l) ², L is the dynamic range of two image block gray-scale values to be calculated;

(4) standard that threshold epsilon judges as similarity is set.As SSIM (x, y) > ε, then think that two image blocks are set of metadata of similar data, the voxel block in prior image can be used for carrying out denoising to tissue points x.

(5) choosing all similitude data, for treating that denoising tissue points x is weighted on average, obtaining the data after denoising:

$v^{\′}\left(x\right)=\frac{1}{C\left(x\right)}\underset{y\∈P}{\mathrm{\Σ}}w(x,y)v\left(y\right)$

Y is the tissue points in all similar areas searched in noisy image, and w (x, y) is weighting function, and C (x) is weight normalized factor.Wherein, the computing method of weighting function are also according to SSIM(x, y) carry out conversion and obtain.Due to SSIM(x, y) value be [-1,1], and when its value absolute value close to 1 time, two image blocks are more similar.And traditional non-local mean adds temporary, when distance is 0, two image blocks are more similar, and weighted value is maximum.Therefore, Similarity Structure parameter is defined: S (x, y)=1-|SSIM (x, y) |.Weighting function can be written as:

w(i,j)=exp(S(x,y)/h ²)

Wherein h is filtering parameter, generally becomes certain linear relationship with the noise criteria difference σ in prior image.

Judge in second step that the similarity of image block also can utilize other similar approach, be not confined to computation structure index of similarity SSIM, as used the modes such as the Euclidean distance of Gauss's weighting.The present invention is just as example.

Applicant uses CT database to carry out confirmatory experiment.Fig. 1 is the prior image used and the low-dose CT experiment treating denoising, and upper figure is high-quality prior image, and figure below is the low-dose CT image treating denoising.What it scanned is similar position, but comes from different patients.Fig. 2 is the contrast of image after denoising and low dosage original image, and left figure is original image, and right figure is the image after denoising.Fig. 3 is the contrast of partial enlargement image, and left figure is original low dosage image, and right figure is the image after denoising.Can be found out by Detail contrast, use the inventive method, can restraint speckle effectively, improve picture quality.

In addition, applicant also makes the oral cavity CBCT of LifeView instrument development carry out the denoising experiment of clinical data.With one group of X-ray tube voltage 100Kev, tube current be the oromaxillo-facial region scan rebuilding data of 4mA as prior image, meanwhile, this image uses traditional NonLocalMeans method to carry out denoising, ensure that signal to noise ratio (S/N ratio) and the quality of prior image.Treat denoising image be adopt 100Kev tube voltage, 2mA tube current carries out the low dosage CBCT image arrived that scans.In Fig. 5, left figure is the prior image used, and right figure is the low dosage CBCT image treating denoising.Fig. 6 is the contrast of image and the former low dosage CBCT image using this kind of method denoising to obtain, and can find to use this kind of method to ensure that detailed information while denoising.Fig. 7 is topography's contrast of amplifying before and after denoising.Can find, use the present invention effectively can improve picture quality under CBCT low-dose scanning and signal to noise ratio (S/N ratio), more traditional method can not be brought image blurring, saves image detail and edge.

The present invention is relative to former various denoising methods, and owing to introducing the prior image of high s/n ratio, therefore its noise removal capability is stronger, and make use of the similarity of image-region, and the preservation for detailed information is better.Prior image of the present invention is selected more flexible, and each scanning for different patient all can use identical prior image, and practicality is stronger.

Non-elaborated part of the present invention belongs to techniques well known.

The above; be only part embodiment of the present invention, but protection scope of the present invention is not limited thereto, any those skilled in the art are in the technical scope that the present invention discloses; the change that can expect easily or replacement, all should be encompassed within protection scope of the present invention.

Claims (4)

1., based on a low dosage CBCT image de-noising method for high-quality prior image, it is characterized in that performing step is as follows:

The first step: the acquisition of the prior image P of high-quality

The image of high s/n ratio is adopted to calculate prior imformation as denoising; The image request of described high s/n ratio comprises following: (1) to treat that denoising image has similar structural information and similar intensity profile, similar to treating denoising image scanning position, thus ensure that CBCT rebuilds the volume data that obtains and treats that denoising image has high similarity; (2) this image needs high s/n ratio and abundant detailed structure information; Described high s/n ratio is through experiment, should than the low dosage CBCT figure image height more than 50% treating denoising;

Use spiral CT carry out scanning and obtain the prior image of high s/n ratio, and obtained CT value and spatial discrimination are corrected, correspond to and treat in the CBCT system of denoising;

Second step: image denoising

(21) for prior image P and treat that denoising data carry out pre-service, carry out position slightly to mate, this matching process can manually complete, and also can use the automatic calculating method of image centroid and main shaft coupling, make organ site that two width images are corresponding and angle close;

(22) treating, in denoising CBCT data reconstruction, to choose and treat denoising voxel x, then choosing image block v (x) that the size centered by x is n × n × n;

(23) search in prior image P, image block v (y) in block-by-block calculating prior image P centered by voxel y and structural similarity index SSIM (x, y) of v (x):

$SSIM(x,y)=\frac{\left(2{\mathrm{\μ}}_x{\mathrm{\μ}}_y\right)(2{\mathrm{\σ}}_{xy}+c_2)}{({\mathrm{\μ}}_x^2+{\mathrm{\μ}}_y^2+c_1)({\mathrm{\σ}}_x^2+{\mathrm{\σ}}_y^2+c_2)}$

Wherein μ _x, μ _ybe respectively the voxel mean value of two image blocks, σ _x, σ _ybe respectively the standard deviation of two image block voxel values, σ _xybe the covariance of two image block voxel values, c ₁=(k ₁l) ², c ₂=(k ₂l) ², L is the dynamic range of two image block gray-scale values to be calculated,

K ₁, k ₂constant, general k ₁=0.01, k ₂=0.03;

(24) set the standard that threshold epsilon judges as similarity, as SSIM (x, y) > ε, then think that two image blocks are similar, its center voxel can carry out denoising by weighted mean to tissue points x; Otherwise, then think and two image block dissmilarities do not adopt when weighted average calculation;

(25) use the center voxel of all similar image blocks searched in previous step for treating that denoising tissue points x is weighted on average, thus obtain the voxel value after denoising, the formula of weighted average calculation is as follows:

$v^{\′}\left(x\right)=\frac{1}{C\left(x\right)}\underset{y\∈P}{\Σ}w(x,y)v\left(y\right)$

Wherein, w (x, y) is weighting function, the computing method of weighting function are carried out conversion according to SSIM (x, y) and are obtained, due to SSIM (x, y) value is [-1,1], and when its value absolute value close to 1 time, two image blocks are more similar, definition Similarity Structure parameter: S (x, y)=1-|SSIM (x, y) |, weighting function is written as:

w(i,j)＝exp(-S(x,y)/h ²)

Wherein h is filtering parameter, and linear with noise criteria difference σ, C (x) is weight normalized factor, and its computing method are: $C\left(x\right)=\underset{y\∈P}{\Σ}w(x,y).$

2. a kind of low dosage CBCT image de-noising method based on high-quality prior image according to claim 1, is characterized in that: described method directly can carry out denoising to the volume data of three-dimensional.

3. a kind of low dosage CBCT image de-noising method based on high-quality prior image according to claim 1, is characterized in that: judge in described second step that the similarity of image block also can utilize Euclidean distance computing method.

4. a kind of low dosage CBCT image de-noising method based on high-quality prior image according to claim 1, is characterized in that: can also get prior image by one of following several mode in the described first step:

(1) X-ray tube electric current is selected to scan;

(2) each projection angle Multiple-Scan superposition;

(3) slow down rotational speed, in certain Rotary District, gather the more projected angle number of degrees.