KIND OF X-RAY CHEST IMAGE RIB SUPPRESSION METHOD BASED ON POISSON MODEL

Abstract

A X-ray chest image rib suppression method based on Poisson model. It conducts contourlet transformation on the image and utilizes transformation coefficient correlation between different scales to conduct texture enhancement on the image; it designs strip-type detection filter in accordance with the Hessian matrix eigenvalue to the image and detects the area where the rib locates in; it combines enhanced texture and rib area information, establishes and solves rib suppression Poisson model, realizing the rib suppression in the image. Anisortropy and contourlet transformation multi-direction feature is utilized, scale and coefficients direction information are combined and distinction degree between texture and noise improves, enhancing texture while restraining noise; it realizes ribs suppression through solving the Poisson model, which does not need to conduct accurate segmentation on the rib, prevents unnatural transition problem of edges resulted from explicit ribs suppression and effectively suppress the ribs, improving observation effect of X-ray chest image.

Claims

1. A X-ray chest image rib suppression method based on Poisson model, comprises the steps as follows: Step 1: Read the original Chest Image I and conduct contourlet transformation on the Chest Image I; Step 2: Stretch the contrast of the low frequency coefficient after contourlet transformation; Step 3: Conduct gain transformation on the high frequency coefficient after contourlet transformation and reconstitute the Chest Image E based on the transformed coefficients for texture enhancement; Step 4: Conduct Gaussian filtering on the original Chest Image I at multiple scales, calculate the Hessian matrix of the filtered image and its eigenvalue, and establish the strip-type detection filter; the detection filter detects and locates the rib to obtain the binary mask image; Step 5: In accordance with the Chest Image E gotten from Step 3 and the binary mask image of the rib area gotten from Step 4, obtain the output Image I*, which means to conduct convolution operation on the Check Image E by adopting laplacian template, gaining the divergence value $\mathrm{DIV}\left(E\right)=\frac{\partial E_x}{\partial x}+\frac{\partial E_y}{\partial y}$ of the gradient field for that image and establishing the Poisson equation ΔG=DIV(E) s.t. G|.sub.∂Ω=I|.sub.∂Ω; wherein, G is the gray value of pixel within the rib area, I is the original chest image and ∂Ω is the rib area boundary in the binary mask image of Step 4.

2. The X-ray chest image rib suppression method based on Poisson model in claim 1, wherein the mentioned Step 1, certain amount of translation in the directions of row and column to the original Chest Image I shall be conducted respectively to overcome the Pseudo-Gibbs problem before conducting contourlet transformation.

3. The X-ray chest image rib suppression method based on Poisson model in claim 1, wherein the mentioned contrast stretching formula is Ĉ.sub.0(m,n)=(h1−h2)(k(C.sub.0(m,n)−M)+M); wherein, Ĉ.sub.0(m, n) is the coefficient after adjustment, h1, h2 are the high frequency gain and low frequency gain respectively, M is the mean of the low frequency coefficients, k∈[0,1] is the contrast adjustment factor.

4. The X-ray chest image rib suppression method based on Poisson model in claim 3, wherein the mentioned gain transformation formula is ${\hat{C}}_{j,k}\left(m,n\right)=\{\begin{array}{cc}{\left(W_k\left(m,n\right)/T_g\right)}^p.Math.C_{j,k}\left(m,n\right)& .Math.C_{j,k}\left(m,n\right).Math.\ge T_g\\ \left(1-{\left(\left(1-W_k\left(m,n\right)\right)/\left(1-T_g\right)\right)}^p\right).Math.C_{j,k}\left(m,n\right)& .Math.C_{j,k}\left(m,n\right).Math.<T_g\end{array}$ wherein, Ĉ.sub.j,k(m,n) is the transform domain coefficient after enhancement, p∈(1, ∞) is the gain factor, T.sub.g is the gain threshold and W.sub.k is the gain coefficient; the mentioned gain coefficient W.sub.k is expressed as $W_k\left(m,n\right)=\underset{j=1}{\overset{j}{.Math.}}.Math..Math.C_{j,k}^{\prime }\left(m,n\right).Math./\max .Math.\left\{\underset{j=1}{\overset{j}{.Math.}}.Math..Math.C_{j,k}^{\prime }\left(m,n\right).Math.\right\}$ in detail; wherein, C′.sub.j,k is the matrix gotten through interpolation from the transform domain coefficient matrix C.sub.j,k in scale J and direction k; M, n are the rows/columns index and J is the maximum scale level, for which 0<j≦J.

5. The X-ray chest image rib suppression method based on Poisson model in claim 1 wherein the number of the mentioned scales in Step 4 is 3.

6. The X-ray chest image rib suppression method based on Poisson model in claim 1, wherein the mentioned strip-type detection filter is F=e.sup.(λ.sup.2.sup./λ.sup.1.sup.).sup.2(1−e.sup.(λ.sup.1.sup.+λ.sup.2.sup.).sup.2); wherein the output result of Filter F, adopt Otsu algorithm to conduct self-adaptive threshold segmentation to gain the binary mask image of the rib area.

7. The X-ray chest image rib suppression method based on Poisson model in claim 2, wherein the mentioned contrast stretching formula is Ĉ.sub.0(m,n)=(h1−h2)(k(C.sub.0(m,n)−M)+M), Therein, Ĉ.sub.0(m,n) is the coefficient after adjustment, h1, h2 are the high frequency gain and low frequency gain respectly, M is the mean of the low frequency coefficients, k∈[0,1] is the contrast adjustment factor.

8. The X-ray chest image rib suppression method based on Poisson model in claim 5, wherein the mentioned strip-type detection filter is F=e.sup.(λ.sup.2.sup./λ.sup.1.sup.).sup.2(1−e.sup.(λ.sup.1.sup.+λ.sup.2.sup.).sup.2); wherein the output result of Filter F, adopt Otsu algorithm to conduct self-adaptive threshold segmentation to gain the binary mask image of the rib area.

9. The X-ray chest image rib suppression method based on Poisson model in claim 7, wherein the mentioned gain transformation formula is ${\hat{C}}_{j,k}\left(m,n\right)=\{\begin{array}{cc}{\left(W_k\left(m,n\right)/T_g\right)}^p.Math.C_{j,k}\left(m,n\right)& .Math.C_{j,k}\left(m,n\right).Math.\ge T_g\\ \left(1-{\left(\left(1-W_k\left(m,n\right)\right)/\left(1-T_g\right)\right)}^p\right).Math.C_{j,k}\left(m,n\right)& .Math.C_{j,k}\left(m,n\right).Math.<T_g\end{array};$ wherein, Ĉ.sub.j,k (m,n) is the transform domain coefficient after enhancement, p∈(1, ∞) is the gain factor, T.sub.g is the gain threshold and W.sub.k is the gain coefficient; the mentioned gain coefficient W.sub.k is expressed as $W_k\left(m,n\right)=\underset{j=1}{\overset{j}{.Math.}}.Math..Math.C_{j,k}^{\prime }\left(m,n\right).Math./\max .Math.\left\{\underset{j=1}{\overset{j}{.Math.}}.Math..Math.C_{j,k}^{\prime }\left(m,n\right).Math.\right\}$ in detail; wherein, C′.sub.j,k is the matrix gotten through interpolation from the transform domain coefficient matrix C.sub.j,k in scale J and direction k; M, n are the rows/columns index and J is the maximum scale level, for which 0<j≦J.

Description

SPECIFICATION OF THE ATTACHED FIGURES

[0019] FIG. 1 shows the realization process schematic diagram for the method of this invention.

[0020] FIG. 2 shows the comparison diagram for the original image and the texture enhanced image of this invention.

[0021] FIG. 3 shows the detection effect diagram of the rib area in this invention.

[0022] FIG. 4 shows the comparison diagram for the original image and the rib suppression figure in this invention.

SPECIFIC IMPLEMENTATION METHOD

[0023] Further specific specification is given as follows by combining the attached figures and the implementation case in detail. As shown in FIG. 1, a kind of X-ray chest image rib suppression method based on Poisson model includes the steps as follows:

[0024] Step 1: Read the original Chest Image I and conduct Contourlet transformation on the Chest Image I. In the Contourlet transformation, the signal will be processed by up-sampling and down-sampling, which will make the coefficient change obviously. Such situation is called as the Pseudo-Gibbs phenomena. That means the Contourlet transformation does not have translation invariance and larger amplitude will occur near the edge points of the reconstructed image after enhancement. To overcome the Pseudo-Gibbs phenomena, conduct certain amount of translation in the directions of row and column to the image before Contourlet transformation.

[0025] Step 2: Stretch the contrast of the low frequency coefficient in accordance with Formula (1) after Contourlet transformation:

Ĉ.sub.0(m,n)=(h1−h2)(k(C.sub.0(m,n)−M)+M)   (1)

[0026] Therein, Ĉ.sub.0(m,n) is the coefficient after adjustment, h1, h2 are the high-frequency gain and low-frequency gain respectively, M refers to the mean of low-frequency gain and k∈[0,1] is the contrast adjustment factor. Enhance the partial contrast on the basis of keeping the original appearance of the image.

[0027] Step 3: Conduct gain transformation on the high frequency coefficient in accordance with Formula (2) after Contourlet transformation and reconstitute the Chest Image E for texture enhancement;

[00004]$\begin{array}{cc}{\hat{C}}_{j,k}\left(m,n\right)=\{\begin{array}{cc}{\left(W_k\left(m,n\right)/T_g\right)}^p.Math.C_{j,k}\left(m,n\right)& .Math.C_{j,k}\left(m,n\right).Math.\ge T_g\\ \left(1-{\left(\left(1-W_k\left(m,n\right)\right)/\left(1-T_g\right)\right)}^p\right).Math.C_{j,k}\left(m,n\right)& .Math.C_{j,k}\left(m,n\right).Math.<T_g\end{array}& \left(2\right)\end{array}$

[0028] Therein, Ĉ.sub.j,k(m,n) is the transform domain coefficient after enhancement, p∈(1, ∞) is the gain factor, T.sub.g is the gain threshold and W.sub.k is the gain coefficient. Comprehensively consider about the correlation between the signal and the noise within the scale to calculate the gain coefficient. Transform the coefficient module value on every decomposition direction to the finest scale through quadratic interpolation and then conduct the summation and normalization.

[00005]$\begin{array}{cc}{W}_k\left(m,n\right)=\underset{j=1}{\overset{j}{.Math.}}.Math..Math.C_{j,k}^{\prime }\left(m,n\right).Math./\max .Math.\left\{\underset{j=1}{\overset{j}{.Math.}}.Math..Math.C_{j,k}^{\prime }\left(m,n\right).Math.\right\}& \left(3\right)\end{array}$

[0029] Therein, C′.sub.j,k is the matrix gotten through interpolation from the transform domain coefficient matrix C.sub.j,k in Scale J and Direction k. M, n are the rows/columns index and J is the maximum scale level. As only the high frequency coefficient is considered here, thus 0<j≦J. W.sub.k is the coefficient matrix after normalization, whose element value range is [0,1].

[0030] Step 4: Read the original chest image and conduct Gaussian filtering on the image at multiple scales to obtain I.sub.S. The larger the scale number is, the more accurate the algorithm will be. However, the computation will also increase. With overall consideration, normally it is better to select 3 scales. Size of the scale is determined based on the standard deviation of Gaussian kernel, which can be evaluated in accordance with the proportion of the rib in the image. In Image I.sub.s with scale as s, calculate the Hessian matrix H.sub.i of all pixel points. Calculate the eigenvalue λ.sub.1, λ.sub.2 of H.sub.i and construct the strip-type detection filter F:

F=e.sup.(λ.sup.2.sup./λ.sup.1.sup.).sup.2(1−e.sup.(λ.sup.1.sup.+λ.sup.2.sup.).sup.2)   (4)

[0031] Use the filter to strengthen the ribs at multiple scales and take the maximum value at multiple scales as the output result, Use Otsu algorithm to conduct self-adaptive threshold segmentation on the output result, obtaining the binary mask image of the rib area.

[0032] Step 5: in accordance with the gradient information in the Texture Enhancement Image E and the boundary information in the rib area, obtain the finally output Image I* through solving the Poisson's equation. Conduct convolution operation on the Texture Enhancement Image E gotten in Step 3 with laplacian template to obtain the divergence value

[00006]$\mathrm{DIV}\left(E\right)=\frac{\partial E_x}{\partial x}+\frac{\partial E_y}{\partial y}$

of the gradient field for that image and establish the Poisson equation:

ΔG=DIV(E) s.t. G|.sub.∂Ω=I|.sub.∂Ω  (5)

[0033] Therein, G is the gray value of pixel within the rib area of output Image I*, I is the original chest image and ∂Ωis the rib area boundary, i.e. the rib area boundary in the binary mask image of Step 4.

[0034] FIG. 2 shows the comparison diagram between the original image and the texture enhancement image. In the figure, the original image shows a local image of lung and the enhanced image highlights the detail feature of the edge. And the image contrast is also improved. FIG. 3 shows the binary mask image obtained through rib detection, in which the mask includes the approximate area where the rib locates in. FIG. 4 shows the comparison diagram in the lung field between the original image and the rib suppression image, from which the texture of the overlap area is clear and visible after rib suppression.