CN105354835A - Method for evaluating medical image quality in combination with phase consistency, gradient magnitude and structural prominence - Google Patents

Abstract

The invention discloses a method for evaluating medical image quality in combination with phase consistency, gradient magnitude and structural prominence. The method comprises the following steps: 1) calculating a phase consistency map of a reference image and a distorted image by using a log-Gabor wavelet; 2) calculating a gradient magnitude map of the reference image and the distorted image by using a Scharr operator; 3) obtaining a local image quality map in combination with the phase consistency map obtained in the step 1) and the gradient magnitude map obtained in the step 2); 4) calculating a structural prominence map, based on a human eye visual feature, of the reference image; 5) performing weighting summation on the image quality map in the step 3) by using the structural prominence map as an aggregation policy to obtain an image quality score; and 6) when the image quality score is higher than a preset score threshold, judging that the medical image quality is relatively high. The method is relatively high in reliability and high in practicality.

Description

In conjunction with the medical image quality evaluation method of phase equalization, gradient magnitude and structure conspicuousness

Technical field

The invention belongs to medical image and evaluate field, especially relate to a kind of medical image quality evaluation method.

Background technology

The medical image of high-quality often contributes to improving doctor diagnosed efficiency, and reduces the generation of malpractice, but medical image quality evaluation study far will lag behind natural image quality assessment research at present.This is mainly limited to the constraint of hardware device, the limited amount in public data source, is engaged in the restriction of medical image staff specialty background.Going back the method that neither one reliability is higher at present, is therefore a necessary job to the research of medical image quality evaluation.

At present, the image quality evaluating method of Corpus--based Method data and the image quality evaluating method (SSIM) of structure based feature are still and use maximum medical image quality evaluation indexes, wherein the image quality evaluating method of Corpus--based Method data mainly contains square error (Mean-squareError, and Y-PSNR (PeakSignal-noiseRatio, RSNR) MSE).These evaluation indexes for measuring the efficiency of the compression algorithm of various image, also for measuring the impact that other various Medical Image Processings produce image.In fact, the consistance of the subjective evaluation result of these evaluation indexes and people is not high, when the quality of diagnosis of evaluation map picture and impracticable yet.

Image, mainly from the angle of mathematical statistics, is regarded as isolated point set, and is ignored the correlativity between local pixel by the evaluation method of Corpus--based Method data.This just determines the method can not provide quality evaluation result accurately when processing structural distortion.Signal errors is equal to the visually-perceptible of quality by it simultaneously, and this does not meet HVS visual characteristic yet.PSNR can make for noise distortion and judging more accurately, but for other types distortion also and inapplicable.The characteristic of SSIM and the structural information mainly extracted in scene based on human visual system's height adaptive of improving one's methods thereof carries out image quality measure.SSIM method carrys out the similarity between reference metric image and distortion from image local brightness, contrast and structure three aspect, adopts the method for averaging to obtain an overall massfraction after obtaining picture quality collection of illustrative plates.Such method performance is better than earlier processes, but does not still reach the reliability desired by practical application.

Summary of the invention

In order to the consistance overcoming the subjective evaluation result of traditional medicine image quality evaluation index and people is not too high, the deficiency that practicality is poor when evaluating quality of diagnosis, the invention provides the medical image quality evaluation method in conjunction with phase equalization, gradient magnitude and structure conspicuousness that a kind of Reliability comparotive is high, practicality is good.

The technical solution adopted for the present invention to solve the technical problems is:

In conjunction with a medical image quality evaluation method for phase equalization, gradient magnitude and structure conspicuousness, comprise the steps:

1) the phase equalization collection of illustrative plates of log-Gabor small echo computing reference image and distorted image is used;

2) the gradient magnitude collection of illustrative plates of Scharr operator computing reference image and distorted image is used;

3) in conjunction with 1) the phase equalization collection of illustrative plates and 2 that obtains) the gradient magnitude collection of illustrative plates that obtains obtains image local quality collection of illustrative plates;

4) the remarkable collection of illustrative plates of the structure based on visual characteristics of human eyes of computing reference image;

5) with described structure conspicuousness collection of illustrative plates as aggregation strategy to 3) in picture quality collection of illustrative plates be weighted summation, structure conspicuousness model SSM is adopted to embody the susceptibility of human eye to image structure information, represent this weight map with SSM (x), image quality score IQS (x) can be expressed as:

$IQS\left(x\right)=\frac{{\mathrm{\Σ}}_{\mathrm{\Ω}}Q\left(x\right)\·SSM\left(x\right)}{{\mathrm{\Σ}}_{\mathrm{\Ω}}SSM\left(x\right)}---\left(9\right)$

Wherein, x represents each pixel of image on the Ω of spatial domain;

6) when image quality score is higher than the score threshold preset, judge that medical image quality is higher.

Further, described step 1) in, in 2D image, the phase equalization of each pixel is defined as:

$PC(x,y)=\frac{\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{n}{\Σ}}{W}_i(x,y)f_p\{A_{ji}(x,y)\[{\mathrm{\Δ\φ}}_{ji}(x,y)\]-T_i\}}{\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{n}{\Σ}}{A}_{ji}(x,y)}---\left(1\right)$

Δφ(x,y)＝cos[φ _n(x,y)-φ(x,y)]

(2)

-|sin[φ _n(x,y)-φ(x,y)]|

Wherein, m and n represents direction yardstick number respectively, Aji (x) and Φ ji (x) are illustrated respectively in amplitude in i-th direction jth logarithmic scale and locally phase deviation, Ti is the estimating noise on i-th direction, Wi is the weighting function on i-th direction, Φ (x, y) is the local phase weighted mean in (x, y) position.

Further again, described step 2) in, image gradient amplitude uses convolution kernel to represent, the gradient operator of use is Scharr operator, and gradient magnitude GM is defined as:

$GM=\sqrt{G_x^2+G_y^2}---\left(3\right)$

Wherein, Gx and Gy represents image f (x) partial derivative in the horizontal and vertical directions respectively.

Further, described step 3) in, suppose that Ref represents reference picture, Dst represents distorted image, PC ₁and PC ₂be respectively the phase equalization collection of illustrative plates of reference image R ef and distorted image Dst, G ₁and G ₂for the gradient amplitude collection of illustrative plates of reference image R ef and distorted image Dst, PC _sbe the similarity collection of illustrative plates of two width phase equalization images, GM _sbe the similarity collection of illustrative plates of two width gradient amplitude figure, Q is the local quality figure of distorted image; Represent the similarity of phase equalization figure with PCs (x), represent the similarity of gradient amplitude figure with GMs (x), PCs is defined as follows:

${\mathrm{PC}}_S\left(x\right)=\frac{2{\mathrm{PC}}_1\left(x\right)\·{\mathrm{PC}}_2\left(x\right)+c_1}{{\mathrm{PC}}_1^2\left(x\right)+{\mathrm{PC}}_2^2\left(x\right)+c_1}---\left(4\right)$

Wherein, c ₁it is a little positive number constant;

GMs is defined as follows:

${\mathrm{GM}}_S\left(x\right)=\frac{2G_1\left(x\right)\·G_2\left(x\right)+c_2}{G_1^2\left(x\right)+G_2^2\left(x\right)+c_2}---\left(5\right)$

Wherein, c ₂it is a little positive number constant;

Local quality collection of illustrative plates Q (x) is expressed as:

Q(x)＝PC _S(x)·GM _S(x)(6)。

Described step 4) in, represent that θ direction, (x, y) position becomes the probability at edge with Pb (x, y, θ), the calculating of Pb based on brightness, color, the Gradient direction information of texture three passages calculates;

For luminance channel, using and be divided into the circle of two halves as window by a diameter, by adjusting the direction of diameter, calculating histogram g and h corresponding to two semicircles, calculate the histogrammic distance in two, (x, y) place with g and h:

${\mathrm{\χ}}^2(g,h)=\frac{1}{2}\underset{i}{\Σ}\frac{{\left(g\right(i)-h(i\left)\right)}^2}{g\left(i\right)+h\left(i\right)}---\left(7\right)$

Texture and color channel also process in the same way;

Then introduce three yardsticks and detect fine structure and coarse structure, finally the local luminance on multiple dimensioned, color and texture information combine, and overall marginal probability is defined as:

$gPb(x,y,\mathrm{\θ})=\underset{s}{\Σ}\underset{i}{\Σ}{\mathrm{\β}}_{i,s}{G}_{i,\mathrm{\σ}(i,s)}(x,y,\mathrm{\θ})+\mathrm{\γ}\·sPb(x,y,\mathrm{\θ})---\left(8\right)$

Wherein, s is the index of yardstick, and i is the index (brightness of feature passage, color, texture), Gi, σ (i, s) in (x, y, θ) Measurement channel i, radius is σ (i, s), the center of circle is after the circle of (x, y) is divided into two semicircles by the diameter that an angle is θ, the difference of the histogram of two semicircles.Pb (x, y, θ) measures the probability becoming edge in θ direction, position, (x, y) place.The value of β i, s and γ obtains by using the gradient of the F-measure of BSDS training image to rise.Structure conspicuousness model SSM is realized by the method for above-mentioned overall marginal probability.

Beneficial effect of the present invention is mainly manifested in: obtain the result consistent with professional assessment height, and improve the reliability of Objective image quality evaluation method, practicality is good.

Accompanying drawing explanation

Fig. 1 is the algorithm flow chart of the medical image quality evaluation method in conjunction with phase equalization, gradient magnitude and structure conspicuousness.

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described.

With reference to Fig. 1, a kind of medical image quality evaluation method in conjunction with phase equalization, gradient magnitude and structure conspicuousness, comprises the steps:

1) the phase equalization collection of illustrative plates of computer vision and the application of perception field log-Gabor small echo computing reference image and distorted image very is widely used in;

In 2D image, the phase equalization of each pixel can be defined as:

$PC(x,y)=\frac{\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{n}{\Σ}}{W}_i(x,y)f_p\{A_{ji}(x,y)\[{\mathrm{\Δ\φ}}_{ji}(x,y)\]-T_i\}}{\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{n}{\Σ}}{A}_{ji}(x,y)}---\left(1\right)$

Δφ(x,y)＝cos[φ _n(x,y)-φ(x,y)]

(2)

-|sin[φ _n(x,y)-φ(x,y)]|

Wherein m and n represents direction yardstick number respectively, and Aji (x) and Φ ji (x) are illustrated respectively in amplitude in i-th direction jth logarithmic scale and locally phase deviation.Ti is the estimating noise on i-th direction, and Wi is the weighting function on i-th direction.Φ (x, y) is the local phase weighted mean in (x, y) position.

2) the gradient magnitude collection of illustrative plates of Scharr operator computing reference image and distorted image is used;

Image gradient amplitude uses convolution kernel to represent, the gradient operator of use is Scharr operator, and gradient magnitude GM is defined as:

$GM=\sqrt{G_x^2+G_y^2}---\left(3\right)$

Wherein Gx and Gy represents image f (x) partial derivative in the horizontal and vertical directions respectively;

3) in conjunction with 1) the phase equalization collection of illustrative plates and 2 that obtains) the gradient magnitude collection of illustrative plates that obtains obtains image local quality collection of illustrative plates;

With reference to Fig. 1, wherein Ref represents reference picture, and Dst represents distorted image, PC ₁and PC ₂be respectively the phase equalization figure calculated by corresponding former figure, G ₁and G ₂for the gradient amplitude figure of corresponding former figure, PC _sbe the similarity collection of illustrative plates of two width phase equalization images, GM _sbe the similarity collection of illustrative plates of two width gradient amplitude figure, Q is the local quality figure of distorted image, represents the similarity of phase equalization figure here with PCs (x), and represent the similarity of gradient amplitude figure with GMs (x), PCs is defined as follows:

${\mathrm{PC}}_S\left(x\right)=\frac{2{\mathrm{PC}}_1\left(x\right)\·{\mathrm{PC}}_2\left(x\right)+c_1}{{\mathrm{PC}}_1^2\left(x\right)+{\mathrm{PC}}_2^2\left(x\right)+c_1}---\left(4\right)$

Wherein, c ₁a little positive number constant with avoid denominator be 0 situation, PC ₁and PC ₂be respectively the phase equalization figure of reference diagram and distortion map, similarly, GMs can be defined as follows:

${\mathrm{GM}}_S\left(x\right)=\frac{2G_1\left(x\right)\·G_2\left(x\right)+c_2}{G_1^2\left(x\right)+G_2^2\left(x\right)+c_2}---\left(5\right)$

Wherein, G ₁and G ₂the gradient amplitude figure of reference diagram and distortion map, c respectively ₂a little positive number constant with avoid denominator be 0 situation, local quality collection of illustrative plates Q (x) can be expressed as:

Q(x)＝PC _S(x)·GM _S(x)(6)

4) the remarkable collection of illustrative plates of the structure based on visual characteristics of human eyes of computing reference image;

Structure conspicuousness model based on visual characteristics of human eyes is realized by the method for overall marginal probability, with Pb (x, y, θ) represent (x, y) θ direction in position becomes the probability at edge, the calculating of Pb is mainly based on brightness, color, the Gradient direction information of texture three passages calculates, and for luminance channel, uses and is divided into the circle of two halves as window by a diameter, by adjusting the direction of diameter, calculate histogram g and h corresponding to two semicircles, calculate the histogrammic distance in two, (x, y) place with g and h:

${\mathrm{\χ}}^2(g,h)=\frac{1}{2}\underset{i}{\Σ}\frac{{\left(g\right(i)-h(i\left)\right)}^2}{g\left(i\right)+h\left(i\right)}---\left(7\right)$

Texture and color channel also process in the same way.Then introduce three yardsticks and detect fine structure and coarse structure, finally the local luminance on multiple dimensioned, color and texture information combine, and final overall marginal probability is defined as:

$gPb(x,y,\mathrm{\θ})=\underset{s}{\Σ}\underset{i}{\Σ}{\mathrm{\β}}_{i,s}{G}_{i,\mathrm{\σ}(i,s)}(x,y,\mathrm{\θ})+\mathrm{\γ}\·sPb(x,y,\mathrm{\θ})---\left(8\right)$

5) with described structure conspicuousness collection of illustrative plates as aggregation strategy to 3) in picture quality collection of illustrative plates be weighted summation, obtain effective image quality score;

6) when image quality score is higher than the score threshold preset, judge that medical image quality is higher.

With reference to the structure conspicuousness collection of illustrative plates that Fig. 1, SSM are reference picture, with it, weighting polymerization is done, to obtain final image massfraction to Quality Map spectrum Q.

When obtaining final mark by quality collection of illustrative plates, we adopt structure conspicuousness model SSM to embody the susceptibility of human eye to image structure information.Represent this weight map with SSM (x), image quality score IQS (x) is expressed as:

$IQS\left(x\right)=\frac{{\mathrm{\Σ}}_{\mathrm{\Ω}}Q\left(x\right)\·SSM\left(x\right)}{{\mathrm{\Σ}}_{\mathrm{\Ω}}SSM\left(x\right)}---\left(9\right)$

Wherein, x represents each pixel of image on the Ω of spatial domain.

Claims (5)

1. in conjunction with a medical image quality evaluation method for phase equalization, gradient magnitude and structure conspicuousness, it is characterized in that: described evaluation method comprises the steps:

1) the phase equalization collection of illustrative plates of log-Gabor small echo computing reference image and distorted image is used;

2) the gradient magnitude collection of illustrative plates of Scharr operator computing reference image and distorted image is used;

3) in conjunction with 1) the phase equalization collection of illustrative plates and 2 that obtains) the gradient magnitude collection of illustrative plates that obtains obtains image local quality collection of illustrative plates;

4) the remarkable collection of illustrative plates of the structure based on visual characteristics of human eyes of computing reference image;

5) with described structure conspicuousness collection of illustrative plates as aggregation strategy to 3) in picture quality collection of illustrative plates be weighted summation, structure conspicuousness model SSM is adopted to embody the susceptibility of human eye to image structure information, represent this weight map with SSM (x), image quality score IQS (x) can be expressed as:

Wherein, x represents each pixel of image on the Ω of spatial domain;

6) when image quality score is higher than the score threshold preset, judge that medical image quality is higher.

2., as claimed in claim 1 in conjunction with the medical image quality evaluation method of phase equalization, gradient magnitude and structure conspicuousness, it is characterized in that: described step 1) in, in 2D image, the phase equalization of each pixel is defined as:

Δφ(x,y)＝cos[φ _n(x,y)-φ(x,y)](2)

-|sin[φ _n(x,y)-φ(x,y)]|

Wherein, m and n represents direction yardstick number respectively, Aji (x) and Φ ji (x) are illustrated respectively in amplitude in i-th direction jth logarithmic scale and locally phase deviation, Ti is the estimating noise on i-th direction, Wi is the weighting function on i-th direction, Φ (x, y) is the local phase weighted mean in (x, y) position.

3. as claimed in claim 1 or 2 in conjunction with the medical image quality evaluation method of phase equalization, gradient magnitude and structure conspicuousness, it is characterized in that: described step 2) in, image gradient amplitude uses convolution kernel to represent, the gradient operator used is Scharr operator, and gradient magnitude GM is defined as:

Wherein, Gx and Gy represents image f (x) partial derivative in the horizontal and vertical directions respectively.

4., as claimed in claim 1 or 2 in conjunction with the medical image quality evaluation method of phase equalization, gradient magnitude and structure conspicuousness, it is characterized in that: described step 3) in, suppose that Ref represents reference picture, Dst represents distorted image, PC ₁and PC ₂be respectively the phase equalization collection of illustrative plates of reference image R ef and distorted image Dst, G ₁and G ₂for the gradient amplitude collection of illustrative plates of reference image R ef and distorted image Dst, PC _sbe the similarity collection of illustrative plates of two width phase equalization images, GM _sbe the similarity collection of illustrative plates of two width gradient amplitude figure, Q is the local quality figure of distorted image; Represent the similarity of phase equalization figure with PCs (x), represent the similarity of gradient amplitude figure with GMs (x), PCs is defined as follows:

Wherein, c ₁it is a little positive number constant;

GMs is defined as follows:

Wherein, c ₂it is a little positive number constant;

Local quality collection of illustrative plates Q (x) is expressed as:

Q(x)＝PC _S(x)·GM _S(x)(6)。

5. as claimed in claim 4 in conjunction with the medical image quality evaluation method of phase equalization, gradient magnitude and structure conspicuousness, it is characterized in that: described step 4) in, with Pb (x, y, θ) represent that θ direction, (x, y) position becomes the probability at edge, the calculating of Pb is based on brightness, color, the Gradient direction information of texture three passages calculates;

For luminance channel, using and be divided into the circle of two halves as window by a diameter, by adjusting the direction of diameter, calculating histogram g and h corresponding to two semicircles, calculate the histogrammic distance in two, (x, y) place with g and h:

Texture and color channel also process in the same way;

Then introduce three yardsticks and detect fine structure and coarse structure, finally the local luminance on multiple dimensioned, color and texture information combine, and overall marginal probability is defined as:

Wherein, s is the index of yardstick, and i is the index (brightness, color, texture) of feature passage, G _{i, σ (i, s)}in (x, y, θ) Measurement channel i, radius is σ (i, s), and the center of circle is after the circle of (x, y) is divided into two semicircles by the diameter that an angle is θ, the difference of the histogram of two semicircles.Pb (x, y, θ) measures the probability becoming edge in θ direction, position, (x, y) place, β _i,sobtain by using the gradient of the F-measure of BSDS training image to rise with the value of γ.