A kind of low-dose CT image processing method based on distinctiveness dictionary
Technical field
The present invention relates to a kind of low-dose CT image processing method, particularly relate to a kind of based on distinctiveness dictionary
Low-dose CT image processing method, belongs to computerized tomograph technical field.
Background technology
As a kind of conventional effective clinical diagnose instrument, X ray computer fault imaging (X-ray
Computerized Tomography, CT) can obtain clearly on mm-scale human body different tissues for
The dampening information of X-ray, thus the diagnosis and prevention for clinician provides the 3 D human body organ group enriched
Knit information.Today, as a kind of maturation and the most universally recognized inspection method, CT has become as
One of indispensable main tool in radiodiagnosis field.But, along with CT tomoscan is in clinical diagnosis
Especially popularizing in routine examination, the radiation dose problem in CT scan has caused people and has got more and more
Concern, substantial amounts of clinical research shows that the CT radiation dose exceeding normal range easily induces human metabolism
The disease such as exception or even cancer.But, in current CT equipment, often through regulation tube current and voltage
To reduce CT scan dosage, so will increase the block distortion rebuilding image and have directive star strip puppet
Shadow, reduces the quality that CT rebuilds, thus affects clinician's diagnosis rate to abnormal structure.
The current method improving low-dose CT picture quality is broadly divided into based on projection space data and based on figure
The big class of the two of image space data, method based on projection space data is mainly gone by the recovery of CT data for projection
Making an uproar provides noise less data for projection for rebuilding, to improve the image reconstruction quality under the conditions of low dosage, example
Such as the filtered back projection (Filtered Back projection, FBP) currently commonly used in Clinical CT is rebuild
Algorithm, it is simply that processed by built-in frequency domain filtering and suppress artifact and noise, also has some scholars to propose throwing
Shadow space is set up data model and sets up recovery algorithms based on this and suppress under the conditions of low dosage in CT data for projection
Noise.The research of this type of method is limited owing to cannot be obtained by commercial CT data for projection, difficult in reality
To be used widely.Another kind of method is directly to be improved low dose rebuild by image space treatment technology
Amount picture quality, has and is independent of Raw projection data and the fast advantage of processing speed, generally use nonlinear
Processing method carries out keeping the denoising of image edge information, such as anisotropic filter (nonlinear or
Anisotropic filter) or the method for small echo (wavelet), but this quasi-nonlinear side is based primarily upon image
Local message, it is difficult to obtain good treatment effect, such as, it is impossible to effectively in suppression low-dose CT image
Block distortion and star strip artifact, also can produce new artifact in processes.
Rarefaction representation based on dictionary learning (the Sparse and Redundant Representations being recently proposed
Over Learnd Dictionaries) Image denoising algorithm belongs to Equations of The Second Kind method.This method is first by target image
Split into the least segment, then each segment encoded, make its with in a complete dictionary of mistake very
Few several pieces are indicated by linear combination.In this course, by controlling parameter, can make normal
Structure is represented, and artifact noise can not get preferably representing, thus reaches denoising purpose.Finally by these
Segment superposed average again, increases denoising effect further.It is the most complete that this method first looks for the optimal overall situation
Dictionary, and each artwork block is expressed as in dictionary the linear combination of vector (atom).Linear combination is
Number can be calculated by sparse coding process.Sparse representation method target based on dictionary learning is as follows for solving
Problem:
Wherein x and y represents the pending image of m pixel and original low-dose CT image respectively;Subscript ij indicates
Pixel index in image (i, j);RijRepresent that extracting size from image x is that (center is at the segment of (i, j)) for n × n
xijOperator;Dictionary D is the matrix of a n × K, is made up of K n-dimensional vector atom (column vector).
Corresponding n × n the segment of each n dimensional vector;α represents the coefficient sets { α of the rarefaction representation of all piecesij}ij,
Each segment xijCan be by linear combination D αijCarry out approximate representation;||αij||0Represent l0Norm, is used for calculating
Vector αijIn nonzero term number;T is default sparse extent index, is used for limiting αijThe number of middle nonzero term.
The problem in (1) that solves comprises following (2) and (3) two subproblems:
Wherein, the purpose of (2) is to train sparse coefficient α and dictionary D from a series of segments, and this problem can be by
X known image y is replaced, and utilizes K average singular value decomposition (K-SVD) to solve.Originate at the beginning of one
Beginning dictionary (such as DCT dictionary), K-SVD method estimates α and dictionary D by alternately two steps:
Sparse coding step based on orthogonal matched transform algorithm (OMP) and the dictionary updating step decomposed based on SVD.
Afterwards by obtain dictionary D and factor alpha, by the first derivative asking (3) obtain export image x:
Wherein I is unit matrix, and μ is weight coefficient.
Sparse representation method based on dictionary learning has been proved to its process in low dosage abdominal CT images
Effect, the abdominal CT images under conditions of tube current is reduced to original 1/5th still can obtain in processing
Obtain preferable image recovery effects, it is contemplated that dosage and the linear relationship of tube current, based on dictionary learning sparse
Method for expressing can make patient suffered decrease in dose to original 1/5th in abdomen scanning.But this kind
Method has certain limitation, easily the star strip artifact in CT image under the conditions of low dosage as image
In information, thus cannot effectively suppress in CT image the star strip easily occurred under the conditions of low-dose scanning
Artifact, the generation of these star strip artifacts is owing to the x-ray projection of some angle is had bigger by tissue
Decay, be generally present in the CT scan image at the more position of high density (such as skeleton), as shoulder,
The position such as chest and vertebra.Under the conditions of the low dosage that tube current or voltage reduce, due to wearing of X-ray
Property and the decline of energy thoroughly, star strip artifact will be more serious.
Summary of the invention
Goal of the invention: the technical problem to be solved is to overcome existing low-dose CT image processing method
The problem of the star strip artifact that can not effectively suppress under the conditions of low dosage in CT image existed, it is provided that Yi Zhongneng
Enough low-dose CT image processing methods effectively suppressing star strip artifact, referred to as artifact suppression dictionary learning method
(Artifact Suppressed Dictionary Learning,ASDL).The present invention is based on dictionary learning sparse
Improved on method for expressing, it is proposed that the concept of distinctiveness dictionary.Distinctiveness dictionary comprises artifact atom
With characteristic atomic, artifact composition major part when low dosage image being expressed with distinctiveness dictionary, in image
Expressed by artifact atom, the then major part of the normal organization composition in image is by expressed by characteristic atomic.
In image after expressing afterwards, the part of artifact expressed atom is artificially removed, and reaches to eliminate image artifacts
Purpose.
Technical scheme: a kind of low-dose CT image processing method based on distinctiveness dictionary, comprises the following steps:
Step 1, use stationary wavelet transform are to some low dosages and high dose CT image fTDo the decomposition of monolayer,
Respectively obtain the level after decomposition, vertical and diagonally opposed high frequency detail imageAnd it is low
Frequently imageHigh frequency detail image to the low dosage in training set image library and high dose the most respectively Carry out dictionary learning, obtain three distinctiveness dictionariesWith(this step only need to be carried out
Once, the dictionary trained can remove Reusability in artifact afterwards);
If pending low-dose CT image f is done the decomposition of monolayer by step 2, use stationary wavelet transform,
Respectively obtain the level after decomposition, vertical and diagonally opposed high frequency detail image fchd、fcvd、fcddAnd it is low
Frequently image fca, to level, vertical and diagonally opposed high frequency detail image fchd、fcvd、fcddRespectively with right
The distinctiveness dictionary answeredWithIt is indicated, afterwards by artifact atom in correspondence distinctiveness dictionary
The part expressed is removed, the level after being processed, vertical and diagonally opposed high frequency detail image (this step can carry out twice to strengthen going artifact effect);
Step 3, the level to after processing, vertical and diagonally opposed high frequency detail imageWith
And original low-frequency image fcaCarry out inverse stationary wavelet transform to rebuild, obtain the CT image of artifact suppression
Further, technique scheme can be tied with existing tradition sparse expression method based on dictionary learning
Close, thus improve the treatment effect of low-dose CT image, particularly as be in technique scheme step 3 it
Rear increase step 4, the image that step 3 is obtainedUtilize existing tradition sparse expression based on dictionary learning
Image is further processed by method, to remove artifact and the noise of residual.
Beneficial effect: compared with prior art, the inventive method first by stationary wavelet transform to pending
Low-dose CT image carries out monolayer decomposition, then divides level, vertical and diagonally opposed high frequency detail image
It is not indicated with corresponding distinctiveness dictionary, and represents to partially remove by artifact atom and suppress star strip artifact
Information strength in the high frequency detail image of different directions, then according to the level after processing, vertical and diagonal angle
The high frequency detail image in direction and original low-frequency image carry out rebuilding against stationary wavelet transform obtaining artifact
The CT image being inhibited, then utilizes existing sparse representation method based on dictionary learning to carry out image
Process further.The inventive method can effectively suppress the star strip artifact in low-dose CT image and noise,
Improve low-dose CT picture quality so that it is meet the prescription of clinical diagnosis.
Accompanying drawing explanation
Fig. 1 is low-dose CT image in the embodiment of the present invention;
Fig. 2 is high dose CT image in the embodiment of the present invention;
Fig. 3 is the low-dose CT image using tradition dictionary method to process;
Fig. 4 is the low-dose CT image that the inventive method ASDL processes.
It is specially embodiment
Below in conjunction with specific embodiment, it is further elucidated with the present invention, it should be understood that these embodiments are merely to illustrate this
Invention rather than restriction the scope of the present invention, after having read the present invention, those skilled in the art are to this
The amendment of the bright various equivalent form of values all falls within the application claims limited range.
ASDL low-dose CT image processing method, comprises the following steps:
Step 1, use stationary wavelet transform are to some low dosages and high dose CT image fTDo the decomposition of monolayer,
Respectively obtain the level after decomposition, vertical and diagonally opposed high frequency detail imageAnd it is low
Frequently imageHigh frequency detail image to low dosage and high dose the most respectivelyCarry out dictionary
Study, obtains three distinctiveness dictionariesWith(this step only need to be carried out once, trains
Dictionary can remove Reusability in artifact afterwards);
Concrete, use the high and low frequency wave filter of Haar small echo, along low-dose CT image fTLLevel
Carry out high and low frequency filtering with vertical direction respectively, the image obtained taken the zone line of original image size,
Obtain the high frequency detail image of horizontal directionAlong image fTLDo the most respectively high frequency and
Low frequency filtering, takes the zone line of original image size to the image obtained, and obtains the high frequency detail figure of vertical direction
PictureAlong image fTLDo High frequency filter the most respectively, the image obtained is taken original image
The zone line of size, obtains diagonally opposed high frequency detail imageArtificial extractionIn
The artifact segment of flat site, and utilize K-SVD algorithm to carry out dictionary training, obtain corresponding to level, hang down
Straight and three artifact dictionaries of diagonal detailWithTo high-dose images fTHCarry out small echo equally
Decompose, obtain level, vertical and diagonal detail imageDirectly it is carried out K-SVD
Dictionary training, obtains corresponding to level, vertical and three characteristics dictionaries of diagonal detailWith
Two kinds of dictionaries are merged,
Obtain level, vertical and three distinctiveness dictionaries of diagonal detailWith
If pending low-dose CT image f is done the decomposition of monolayer by step 2, use stationary wavelet transform,
Respectively obtain the level after decomposition, vertical and diagonally opposed high frequency detail image fchd、fcvd、fcddAnd it is low
Frequently image fca, to level, vertical and diagonally opposed high frequency detail image fchd、fcvd、fcddRespectively with right
The distinctiveness dictionary answeredWithIt is indicated, afterwards by artifact atom table in correspondence distinctiveness dictionary
The part reached is removed, the level after being processed, vertical and diagonally opposed high frequency detail image (this step can carry out twice to strengthen going artifact effect);
Concrete, for level detail image fchd, substitute into the f of following (1) formula,
By dictionaryBring the D in (1) formula into, utilize OMP algorithm to ask (1) formula to obtain sparse coefficient α, afterwards
The coefficient of the artifact atom of M row before distinctiveness dictionary corresponding in sparse coefficient is set to zero, will in α before M
Row is set to zero, the factor alpha after being processedp, then by αpWith fchdFor following enter α with f in (2) formula obtain
To the level detail image going pseudo-movie queen
In like manner, to fcvdAnd fcddUtilizeWithTake identical step, obtain the vertical detail figure of pseudo-movie queen
PictureWith diagonal detail image
Step 3, the level to after processing, vertical and diagonally opposed high frequency detail image
And original low-frequency image fcaCarry out inverse stationary wavelet transform to rebuild, obtain the CT image of artifact suppression
Rebuilding for inverse stationary wavelet, concrete is calculated as: the high frequency of same use above Haar small echo and
Low-frequency filter, first to untreated wavelet space low-frequency image component f abovecaAlong horizontal and vertical side
To carrying out low frequency respectively and low frequency filtering obtainsWavelet space horizontal direction high frequency after step 2 is processed
Detail pictures componentObtain along carrying out high and low frequency filtering the most respectivelyTo step 2
Wavelet space vertical direction high frequency detail picture content after processAlong carrying out the most respectively
High and low frequency filtering obtainsWavelet space diagonally opposed high frequency detail image after processing step 2 divides
AmountSame edge carries out high and low frequency filtering the most respectively and obtainsFinally will obtain
'sWithIt is added, the image obtained is taken the zone line of original image size, it is possible to
Obtain inverse stationary wavelet to rebuild, the image that i.e. artifact is inhibited
Step 4, the image that step 3 is obtainedCarry out tradition sparse expression method based on dictionary learning to press down
Noise processed;
Concrete, definitionWithRespectively represent preceding step three obtain artifact suppression low-dose CT figure and
Final image after the inventive method process, computing formula is as follows:
Wherein (3) formula can be solved by K-SVD algorithm and obtain dictionary D and sparse coefficient α, substitutes in (4)
Final result image can be tried to achieve
5. recruitment evaluation criterion
First obtain the low-dose CT image (Fig. 1) at same position, high dose CT image (Fig. 2), use
Tradition dictionary method (Dictionary Learning, DL) the low-dose CT image (Fig. 3) that processes and with
The low-dose CT image (Fig. 4) that inventive method ASDL processes.Used in experiment, CT equipment is one
16 row CT (Somatom Sensation 16), the condition of scanning is 120kVp and 5mm thickness, uses FBP
Method is rebuild, and other parameters use machine default value, and the CT image of high dose and the CT image of low dosage divide
Do not obtain by tube current parameter is set to 270mA and 70mA.
5.1 visual assessment
By observing the low dosage of Fig. 1 to Fig. 4 and the CT image of high dose, and DL method and the present invention
The low-dose CT image that method processes, it can be seen that DL method cannot effectively press down while suppression noise
Star strip artifact processed, and the picture quality after using the inventive method to process significantly improves, star strip artifact and making an uproar
Sound has all obtained effective suppression.
5.2 quantitative evaluation
The effectiveness of the checking the inventive method in order to quantify, the intensity profile that we calculate certain selected is homogeneous
Standard deviation in background area (region shown in red frame in Fig. 1 to Fig. 4), the definition of standard deviation here is:
Number of pixels in wherein np represents selection area, xjWithRepresent the single pixel in this selection area respectively
Point CT value and mean CT-number (Housfield units, HU), from table 1 below it can be seen that processing method can be big
The standard deviation of amplitude reduction original low-dose CT image, it is thus achieved that close to normal dosage CT image in selected homogeneous district
The standard deviation in territory.
Table 1
From above-mentioned experiment it will be seen that use the inventive method can effectively suppress the star low-dose CT image
Strip artifact and noise, obtain the CT image close to high dose level under the conditions of low dosage, meet clinic
The prescription of diagnosis.And the inventive method cannot not obtained by commercial CT data for projection and be limited, and has
The bigger scope of application.