CN103473745A - Low-dosage CT image processing method based on distinctive dictionaries - Google Patents

Abstract

The invention discloses a low-dosage CT image processing method based on distinctive dictionaries. According to the method, first, single-layer decomposition is carried out on a low-dosage CT image to be processed through stationary wavelet transform, then, corresponding distinctive dictionary expression is carried out on high-frequency detail images in the horizontal direction, the vertical direction and the diagonal direction respectively, image artifacts are removed, and therefore information strength of the star-and-strip-shaped image artifacts in the high-frequency detail images in the different directions can be restrained; second, reverse stationary wavelet transform is carried out according to processed high-frequency detail images in the horizontal direction, the vertical direction and the diagonal direction and original low-frequency images to obtain a CT image with the image artifacts restrained through reconstruction; third, the CT image is further processed with an existing and traditional sparse and redundant representation method over learned dictionaries so as to remove residual image artifacts and noise. The method can effectively restrain the star-and-strip-shaped image artifacts and the noise in the low-dosage CT image, improve quality of the low-dosage CT image, and make the low-dosage CT image meet quality requirements in clinical diagnosis.

Description

A kind of low dosage CT image processing method based on the distinctiveness dictionary

Technical field

The present invention relates to a kind of low dosage CT image processing method, relate in particular to a kind of low dosage CT image processing method based on the distinctiveness dictionary, belong to the computerized tomograph technical field.

Background technology

As the effective clinical diagnose instrument of current a kind of routine, X ray computer fault imaging (X-ray Computerized Tomography, CT) can on mm-scale, obtain clearly the dampening information of human body different tissues for X ray, thereby provide abundant 3 D human body organ-tissue information for clinician's diagnosis and prevention.Today, as a kind of maturation and universally recognized inspection method clinically, CT has become one of indispensable main tool in the radiodiagnosis field.Yet, popularizing along with the CT tomoscan in clinical diagnosis especially routine inspection, radiation dose problem in CT scan has caused that people more and more pay close attention to, and a large amount of clinical researches shows that paranormal CT radiation dose easily brings out the diseases such as human body pathobolism and even cancer.Yet, in current CT equipment, often pass through adjustable pipe electric current and voltage to reduce CT scan dosage, will increase like this block distortion of reconstructed image and the star strip artifact with directivity, reduce the quality that CT rebuilds, thereby affect the diagnosis rate of clinician to abnormal structure.

The method of current raising low dosage CT picture quality mainly is divided into based on the projector space data and two large classes based on the image space data, method based on the projector space data mainly provides noise data for projection still less by the recovery denoising of CT data for projection for reconstruction, to improve the image reconstruction quality under the low dosage condition, current (the Filtered Back projection of filtered back projection generally used in Clinical CT is rebuild for example, FBP) algorithm, process to suppress artifact and noise by built-in frequency domain filtering exactly, also have some scholars to propose set up data model and set up recovery algorithms based on this and suppress the noise in the CT data for projection under the low dosage condition at projector space.The research of these class methods, due to the restriction that is subject to commercial CT data for projection and can't obtains, is difficult in reality be used widely.Another kind of method is directly by the image space treatment technology, to improve the low dosage picture quality of having rebuild, have advantages of do not rely on the raw projections data and processing speed fast, usually the denoising of using nonlinear disposal route to be kept image edge information, method as anisotropic filter (nonlinear or anisotropic filter) or small echo (wavelet), yet this quasi-nonlinear side is the local message based on image mainly, be difficult to obtain good treatment effect, for example, can't effectively suppress block distortion and star strip artifact in low dosage CT image, also can in processing, produce new artifact.

The rarefaction representation based on dictionary learning (the Sparse and Redundant Representations over Learnd Dictionaries) Image denoising algorithm proposed recently belongs to the Equations of The Second Kind method.This method first splits into target image a lot of little segments, then each segment is encoded, and makes it be meaned by linear combination with seldom several in a complete dictionary of mistake.In this course, by controlling parameter, can make normal configuration be meaned, and the artifact noise can not get meaning preferably, thereby reach the denoising purpose.Finally, by these segments superposed average again, further increase denoising effect.At first this method is found the best overall situation and is crossed complete dictionary, and each former segment is expressed as to the linear combination of vector (atom) in dictionary.Coefficient in linear combination can calculate by the sparse coding process.Rarefaction representation method target based on dictionary learning is for addressing the problem:

$\underset{x,D,\mathrm{\α}}{\min }{\left|\right|x-y\left|\right|}_2^2+\mathrm{\μ}\underset{\mathrm{ij}}{\mathrm{\Σ}}{\left|\right|R_{\mathrm{ij}}x-{\mathrm{D\α}}_{\mathrm{ij}}\left|\right|}_2^{2}s.t.{\left|\right|{\mathrm{\α}}_{\mathrm{ij}}\left|\right|}_0\≤T\∀i,j---\left(1\right)$

Wherein x and y mean respectively pending image and the original low dosage CT image of m pixel; Subscript ij has indicated the pixel index (i, j) in the image; R _ijit is the segment x of n * n (center is at (i, j)) that size is extracted in expression from image x _ijoperator; Dictionary D is the matrix of a n * K, K n-dimensional vector atom (column vector), consists of.Corresponding n * n the segment of each n dimensional vector; α means the coefficient sets { α of the rarefaction representation of all _ij} _ij, each segment x _ijcan be by linear combination D α _ijcarry out approximate representation; || α _ij|| ₀mean l ⁰norm, be used for compute vector α _ijin the nonzero term number; T is the sparse extent index of presetting, and is used for limiting α _ijthe number of middle nonzero term.The problem solved in (1) comprises following (2) and (3) two subproblems:

$\underset{D,\mathrm{\α}}{\min }\mathrm{\μ}\underset{\mathrm{ij}}{\mathrm{\Σ}}{\left|\right|R_{\mathrm{ij}}x-{\mathrm{D\α}}_{\mathrm{ij}}\left|\right|}_2^{2}s.t.{\left|\right|{\mathrm{\α}}_{\mathrm{ij}}\left|\right|}_0\≤T\∀i,j---\left(2\right)$

$\underset{x}{\min }{\left|\right|x-y\left|\right|}_2^2+\mathrm{\μ}\underset{\mathrm{ij}}{\mathrm{\Σ}}{\left|\right|R_{\mathrm{ij}}x-{\mathrm{D\α}}_{\mathrm{ij}}\left|\right|}_2^2---\left(3\right)$

Wherein, the purpose of (2) is to train sparse factor alpha and dictionary D from a series of segments, and this problem can be replaced x with known image y, utilize K average svd (K-SVD) to solve.Originate in an initial dictionary (for example DCT dictionary), the K-SVD method is estimated α and dictionary D by two steps that hocket: the sparse coding step based on quadrature matched transform algorithm (OMP) and the dictionary updating step of decomposing based on SVD.By the dictionary D and the factor alpha that obtain, by the first order derivative of asking (3), obtain output image x afterwards:

$x={(I+\mathrm{\μ}\underset{\mathrm{ij}}{\mathrm{\Σ}}{R}_{\mathrm{ij}}^T{R}_{\mathrm{ij}})}^{-1}(y+\mathrm{\μ}\underset{\mathrm{ij}}{\mathrm{\Σ}}{R}_{\mathrm{ij}}^T{\mathrm{D\α}}_{\mathrm{ij}})---\left(4\right)$

Rarefaction representation method based on dictionary learning has been proved to be its treatment effect in the low dosage abdominal CT images, be reduced to abdominal CT images under original 1/5th condition at tube current and still can obtain image recovery effects preferably in processing, consider the linear relationship of dosage and tube current, the rarefaction representation method based on dictionary learning can make patient, in abdomen scanning, suffered dosage is reduced to original 1/5th.Yet this kind of method has certain limitation, easily assign the star strip artifact in CT image under the low dosage condition as the information in image, thereby can't effectively suppress the star strip artifact be prone in the CT image under the low-dose scanning condition, the generation of these star strip artifacts is because tissue has larger decay to the X ray projection of some angle, generally appear in the CT scan image at the more position of high density (as bone), as positions such as shoulder, chest and vertebras.Under the low dosage condition of tube current or lower voltage, due to the penetrability of X ray and the decline of energy, the star strip artifact will be more serious.

Summary of the invention

Goal of the invention: technical matters to be solved by this invention is to overcome the problem that can not effectively suppress the star strip artifact in the CT image under the low dosage condition that existing low dosage CT image processing method exists, a kind of low dosage CT image processing method that can effectively suppress the star strip artifact is provided, be called artifact and suppress dictionary learning method (Artifact Suppressed Dictionary Learning, ASDL).The present invention improves on the rarefaction representation method based on dictionary learning, has proposed the concept of distinctiveness dictionary.Comprise artifact atom and characteristic atomic in the distinctiveness dictionary, while with the distinctiveness dictionary, the low dosage image being expressed, artifact composition major part in image is expressed by the artifact atom, and the normal structure constituent in image is most of expressed by characteristic atomic.Afterwards the part of artifact expressed atom in the image after expressing is artificially removed, reached the purpose of removal of images artifact.

Technical scheme: a kind of low dosage CT image processing method based on the distinctiveness dictionary comprises the following steps:

Step 1, use static wavelet transformation to some low dosages and normal dose CT image f ^tdo the decomposition of individual layer, the level after being decomposed respectively, vertical and to the high frequency detail pictures of angular direction

and low-frequency image

afterwards respectively to the high frequency detail pictures of low dosage and normal dose carry out dictionary learning, obtain three distinctiveness dictionaries

(this step only need carry out once getting final product, and the dictionary trained can remove Reusability in artifact afterwards);

If step 2, use static wavelet transformation pending low dosage CT image f to be done to the decomposition of individual layer, the level after being decomposed respectively, vertical and to the high frequency detail pictures f of angular direction _chd, f _cvd, f _cddand low-frequency image f _ca, to level, vertical and to the high frequency detail pictures f of angular direction _chd, f _cvd, f _cdduse respectively corresponding distinctiveness dictionary

with

meaned, afterwards the part of artifact expressed atom in corresponding distinctiveness dictionary is removed, the level after being processed, vertical and to the high frequency detail pictures of angular direction

(this step can be carried out twice to strengthen going the artifact effect);

Step 3, to the level after processing, vertical and to the high frequency detail pictures of angular direction

and original low-frequency image f _cacarry out contrary static reconstruction with wavelet, obtain the CT image that artifact suppresses

Further, technique scheme can be combined with the sparse expression method of existing tradition based on dictionary learning, thus the treatment effect of raising low dosage CT image, and concrete is exactly to increase step 4, the image that step 3 is obtained in technique scheme after step 3

utilize the sparse expression method of existing tradition based on dictionary learning to be further processed image, to remove residual artifact and noise.

Beneficial effect: compared with prior art, at first the inventive method is used static wavelet transformation to carry out the individual layer decomposition to pending low dosage CT image, then to level, vertical and the high frequency detail pictures of angular direction is meaned with corresponding distinctiveness dictionary respectively, and the artifact atom is meaned to part eliminates to suppress the information strength of star strip artifact in the high frequency detail pictures of different directions, then according to the level after processing, vertical and the high frequency detail pictures of angular direction and original low-frequency image are carried out to contrary static wavelet transformation and rebuild and obtain the CT image that artifact is inhibited, then utilize the existing rarefaction representation method based on dictionary learning to be further processed image.The inventive method can effectively suppress star strip artifact and the noise in low dosage CT image, improves low dosage CT picture quality, makes it meet the quality requirements of clinical diagnosis.

The accompanying drawing explanation

Fig. 1 is low dosage CT image in the embodiment of the present invention;

Fig. 2 is normal dose CT image in the embodiment of the present invention;

The low dosage CT image of Fig. 3 for using traditional dictionary method to process;

Fig. 4 is the low dosage CT image that the inventive method ASDL processes.

Be specially embodiment

Below in conjunction with specific embodiment, further illustrate the present invention, should understand these embodiment only is not used in and limits the scope of the invention for the present invention is described, after having read the present invention, those skilled in the art all fall within the application's claims limited range to the modification of the various equivalent form of values of the present invention.

ASDL low dosage CT image processing method comprises the following steps:

Step 1, use static wavelet transformation to some low dosages and normal dose CT image f ^tdo the decomposition of individual layer, the level after being decomposed respectively, vertical and to the high frequency detail pictures of angular direction

and low-frequency image

afterwards respectively to the high frequency detail pictures of low dosage and normal dose

carry out dictionary learning, obtain three distinctiveness dictionaries

with

(this step only need carry out once getting final product, and the dictionary trained can remove Reusability in artifact afterwards);

Concrete, the high and low frequency wave filter of use Haar small echo, along low dosage CT image f ^tLthe horizontal and vertical direction carry out respectively high and low frequency filtering, the image obtained is got to the zone line of original image size, obtain the high frequency detail pictures of horizontal direction vertical and horizontal direction along image f is done respectively high and low frequency filtering, the image obtained is got to the zone line of original image size, obtains the high frequency detail pictures of horizontal direction

along image f ^tLvertical and horizontal direction do respectively High frequency filter, the image obtained is got to the zone line of original image size, obtain the high frequency detail pictures to angular direction the artificial extraction

the artifact segment of middle flat site, and utilize the K-SVD algorithm to carry out the dictionary training, obtain three artifact dictionaries corresponding to level, vertical and diagonal detail

with

to high-dose images f ^tHcarry out equally wavelet decomposition, obtain level, vertical and diagonal detail image

directly it is carried out to the training of K-SVD dictionary, obtain three characteristics dictionaries corresponding to level, vertical and diagonal detail with

two kinds of dictionaries are merged, $D_{\mathrm{chd}}^d=\left[D_{\mathrm{chd}}^a\right|D_{\mathrm{chd}}^f],D_{\mathrm{chd}}^d=[D_{\mathrm{chd}}^a\left|D_{\mathrm{chd}}^f\right],D_{\mathrm{cdd}}^d=\left[D_{\mathrm{cdd}}^a\right|D_{\mathrm{cdd}}^f],$ Obtain three distinctiveness dictionaries of level, vertical and diagonal detail

with .

If step 2, use static wavelet transformation pending low dosage CT image f to be done to the decomposition of individual layer, the level after being decomposed respectively, vertical and to the high frequency detail pictures f of angular direction _chd, f _cvd, f _cddand low-frequency image f _ca, to level, vertical and to the high frequency detail pictures f of angular direction _chd, f _cvd, f _cdduse respectively corresponding distinctiveness dictionary

with meaned, afterwards the part of artifact expressed atom in corresponding distinctiveness dictionary is removed, the level after being processed, vertical and to the high frequency detail pictures of angular direction

(this step can be carried out twice to strengthen going the artifact effect);

Concrete, for level detail image f _chd, the f of (1) formula, make dictionary be below substitution

$\underset{\mathrm{\α}}{\min }\underset{\mathrm{ij}}{\mathrm{\Σ}}{\left|\right|{\mathrm{\α}}_{\mathrm{ij}}\left|\right|}_{0}s.t.{\left|\right|R_{\mathrm{ij}}f-{\mathrm{D\α}}_{\mathrm{ij}}\left|\right|}_2^2\≤\mathrm{\ϵ}\∀i,j---\left(1\right)$

utilize the OMP algorithm to try to achieve sparse factor alpha _h, afterwards the coefficient of the artifact atom of M row before corresponding distinctiveness dictionary in sparse coefficient is set to zero, be about to α _hin before M is capable is set to zero, the coefficient after being processed

again will

with f _chdenter the level detail image that (2) formula obtains pseudo-movie queen below generation

$x={(I+\mathrm{\μ}\underset{\mathrm{ij}}{\mathrm{\Σ}}{R}_{\mathrm{ij}}^T{R}_{\mathrm{ij}})}^{-1}(y+\mathrm{\μ}\underset{\mathrm{ij}}{\mathrm{\Σ}}{R}_{\mathrm{ij}}^T{\mathrm{D\α}}_{\mathrm{ij}})---\left(2\right)$

In like manner, to f _cvdand f _cddutilize

with

take identical step, obtain pseudo-movie queen's vertical detail image

with the diagonal detail image

Step 3, to the level after processing, vertical and to the high frequency detail pictures of angular direction

and original low-frequency image f _cacarry out contrary static reconstruction with wavelet, obtain the CT image that artifact suppresses

For contrary static wavelet reconstruction, concrete is calculated as: the high and low frequency wave filter of same use front Haar small echo, and at first to top untreated wavelet space low-frequency image component f _cacarrying out respectively low frequency and low frequency filtering along the horizontal and vertical direction obtains

wavelet space horizontal direction high frequency detail pictures component after step 2 is processed carrying out respectively high and low frequency filtering along the horizontal and vertical direction obtains

wavelet space vertical direction high frequency detail pictures component after step 2 is processed

carrying out respectively high and low frequency filtering along the horizontal and vertical direction obtains

wavelet space after step 2 is processed is to angular direction high frequency detail pictures component

same carry out respectively high and low frequency filtering along the horizontal and vertical direction and obtain

finally will obtain

with

addition, get the zone line of original image size to the image obtained, just can obtain contrary static wavelet reconstruction, the image that artifact is inhibited

Step 4, the image that step 3 is obtained

carry out the sparse expression method of tradition based on dictionary learning and suppress noise;

Concrete, definition with

mean respectively the low dosage CT figure of the artifact inhibition that preceding step three obtains and the final image after the inventive method processing, computing formula is as follows:

$\underset{D,\mathrm{\α}}{\min }\mathrm{\μ}\underset{\mathrm{ij}}{\mathrm{\Σ}}{\left|\right|R_{\mathrm{ij}}\tilde{f}-{\mathrm{D\α}}_{\mathrm{ij}}\left|\right|}_2^{2}s.t.{\left|\right|{\mathrm{\α}}_{\mathrm{ij}}\left|\right|}_0\≤T\∀i,j---\left(3\right)$

$\hat{f}={(I+\mathrm{\μ}\underset{\mathrm{ij}}{\mathrm{\Σ}}{R}_{\mathrm{ij}}^T{R}_{\mathrm{ij}})}^{-1}(\tilde{f}+\mathrm{\μ}\underset{\mathrm{ij}}{\mathrm{\Σ}}{R}_{\mathrm{ij}}^T{\mathrm{D\α}}_{\mathrm{ij}})---\left(4\right)$

Wherein (3) formula can solve and obtain dictionary D and sparse factor alpha by the K-SVD algorithm, and substitution can be tried to achieve the net result image in (4)

5. recruitment evaluation criterion

At first obtain the low dosage CT image (Fig. 1) at same position, normal dose CT image (Fig. 2), use the low dosage CT image (Fig. 3) of traditional dictionary method (Dictionary Learning, DL) processing and the low dosage CT image (Fig. 4) of processing with the inventive method ASDL.Using CT equipment in experiment is one 16 row CT (Somatom Sensation16), the condition of scanning is 120kVp and 5mm bed thickness, adopt the FBP method to rebuild, other parameters adopt machine default values, and the CT image of normal dose and the CT image of low dosage are respectively by being made as 270mA to the tube current parameter and 70mA obtains.

5.1 visual assessment

By the low dosage of observation Fig. 1 to Fig. 4 and the CT image of normal dose, and the low dosage CT image of DL method and the inventive method processing, can see that the DL method can't effectively suppress the star strip artifact when suppressing noise, and using the picture quality after the inventive method is processed obviously to improve, star strip artifact and noise have all obtained effective inhibition.

5.2 quantitative evaluation

For the validity of checking the inventive method of quantizing, we have calculated the standard deviation in the background area (zone shown in red frame in Fig. 1 to Fig. 4) of certain selected intensity profile homogeneous, and standard deviation is defined as here:

$\mathrm{STD}=\sqrt{\frac{1}{\mathrm{np}-1}{\mathrm{\Σ}}_{j=1}^{\mathrm{np}}{(x_j-\stackrel{\‾}{x})}^2}---\left(10\right)$

Wherein np represents the number of pixels in selection area, x _jwith

represent respectively single pixel CT value and mean CT-number (Housfield units in this selection area, HU), can see from following table 1 standard deviation that disposal route can the original low dosage CT of decrease image, obtain and approach normal dosage CT image in the standard deviation of selecting the homogeneous zone.

Table 1

Can see from above-mentioned experiment, adopt the inventive method can effectively suppress star strip artifact and the noise in low dosage CT image, obtain the CT image that approaches the normal dose level under the low dosage condition, meet the quality requirements of clinical diagnosis.And the inventive method restriction of not can't obtain by commercial CT data for projection, there is the larger scope of application.

Claims (4)

1. the low dosage CT image processing method based on the distinctiveness dictionary, is characterized in that, comprises the following steps:

Step 1, use static wavelet transformation pending low dosage CT image f to be done to the decomposition of individual layer, the level after being decomposed respectively, vertical and to the high frequency detail pictures f of angular direction _chd, f _cvd, f _cddand low-frequency image f _ca;

Step 2, to level, vertical and to the high frequency detail pictures f of angular direction _chd, f _cvd, f _cddmeaned with corresponding distinctiveness dictionary respectively, afterwards the part of artifact expressed atom in corresponding distinctiveness dictionary is removed, the level after being processed, vertical and to the high frequency detail pictures of angular direction

Step 3, to the level after processing, vertical and to the high frequency detail pictures of angular direction

and original low-frequency image f _cacarry out contrary static reconstruction with wavelet, obtain the CT image that artifact suppresses

2. remove as claimed in claim 1 the low dosage CT image processing method of artifact based on the wavelet space directivity, it is characterized in that, also comprise after step 3:

Step 4, the image that step 3 is obtained

the sparse expression method of utilization based on dictionary learning is further processed image, to remove residual artifact and noise.

3. remove as claimed in claim 1 the low dosage CT image processing method of artifact based on the wavelet space directivity, it is characterized in that, in step 2, distinctiveness dictionary used is to be merged and form with " characteristics dictionary " by " artifact dictionary ", front M atom pair in each distinctiveness dictionary answered the atom in the artifact dictionary, and atom afterwards is the atom in characteristics dictionary; Wherein the artifact dictionary is to obtain from the artifact sample training of artificial extraction, and characteristics dictionary is from the training of high dose CT image pattern and obtain.

4. remove as claimed in claim 1 the low dosage CT image processing method of artifact based on the wavelet space directivity, it is characterized in that, distinctiveness dictionary one in step 2 has three, respectively from level, vertical and to training the high frequency detail pictures of angular direction, and be applied to the high frequency detail pictures of correspondence direction in going artifact.

Images