胸部X线影像增强研究多尺度Retinex技术

Chest Radiographic image Enhancement based on Multi-scale Retinex TechniqueChen Shuyue, Zou LingSchool of Information Science and Engineering,Jiangsu Polytechnic UniversityChangzhou, Chinacsyue2000@Abstrac t-Retinex technique has been widely employed in medical and other fields. Chest radiographic image is often suffered from its poor dynamic rang due to x-ray radiography of different chest tissues densities. To compress the dynamic range and enhance contrast for improving the visibility of the dark regions on chest radiograph, an enhancement method based on multi-scale Retinex was presented. Two-scale Retinex with different weighted factors was utilized and compared its result with several methodologies including three-scale Retinex, contrast-limited adaptive histogram equalization, Frankle-McCann and McCann Retinex. It is shown that the given method is more effective for chest radiographic image enhancement.Keywords-image enhancement; chest radiograph; Retinex; medical image processingI.I NTRODUCTIONThe main goal of Retinex theory is to compensate for local brightness in images. It tries to flatten the gap between the local bright and dark region, and provides with a better perception effect, especially on the fine features of ROI. Retinex model for the computation of illumination was introduced by Land and McCann[1]. Since that time Land and his colleagues have described several variants on the original method[2-5]. The output of multi-scale Retinex(MSR) is the weighted summation of each single-scale Retinex(SSR), which is capable of enhancing some particular characteristic of the input image, such as narrow surrounds highlighting the fine features and wide surrounds holding all the tonal information. The MSR methods are investigated in [6-8]. Recently, Bo Sun et al. chosen an adaptive Gaussian function whose shape follows edges of image to reduce the artifacts along high contrast edges[9]. Jinhua Wang et al. used three Gaussian constants for a single filter instead of a traditional fixed Gaussian constant to get timesaving[10]. Meylan and Susstrunk applied an adaptive filter for reducing halo artifacts[11].Chest radiographic images are often suffered from their poor dynamic rang due to x-ray radiography of different chest tissues densities. The dark zones contain amount of disease information of lungs, but an uneasy set-up condition of machine or films digitizing may cause the image to be difficult to distinguish. Nagesha and Kuma presented a novel (recursive) algorithm that tailors the required amount of contrast enhancement based on a combination of the optimal phase representation and the theory of projection onto a convex set[12]. Jyh-Shyan et al. used wavelet processing technique to segment a chest radiograph into different gray level regions, then applied different degrees of contrast enhancement to process them[13]. We normalized the chest radiographic image with histogram equalization then applied an established piecewise linear transformation model to obtain an automatic procedure in our previous work[14].In the paper, we employed two-scale Retinex method associated with Gaussian surround functions. Two different weighted coefficients were used, where the larger one is engaged in the whole bright framework of chest radiographic image, the other enhances the dark region details. Some methods, including three-scale Retinex, contrast-limited adaptive histogram equalization(CLAHE), Frankle-McCann and McCann Retinex methodologies were compared through experiments.II.R ETINEX THEORYA.Single Scale RetinexLand proposed Retinex theory of the brightness and chromaticness models for human perception, which reveals the principle how to sense a scene with eyes vision system and receive a consistency sensation confirming to distinguish different objects under different illumination. Land introduced an image center/surround space[4] based on human vision. Jobson and his team defined a single scale Retinex algorithm drawing from Land’s research. SSR can be described as)],(),(log[),(log),(yxIyxFyxIyxR∗−= (1) Where R(x,y) is image output function, I(x,y) is input image distribution function. ‘*’ and log imply convolution operator and nature logarithm respectively. Surround function F(x,y) includes some forms. Hulbert et al. provided with a Gaussian function:))(exp(),(222σyxAyxF+−⋅= (2) Where σ is standard deviation in Gaussian function. Choosing A should satisfy with below condition:∫∫=1),(dxdyyxF(3)The image enhancement effect is determined by the standard deviation directly, that controls how many fine details are left. We should balance the final effects between the dynamic range compression and consistency sensation.This work is sponsored by Qing Lan Project of Jiangsu ProvinceB. Multi-scale RetinexThe basic form of the MSR is given by [6,7]Ni y x I y x F y x I Wy x R i Kk k i k i ,,1)]),(*),(log[),((log ),(1"=−=∑= (4)where the sub-index i represents the i -th spectral band, N isthe number of spectral bands, N = 1 for grayscale images andN = 3 for typical color images. I is the input image, R is the MSR output, F k represents the k -th surround function, W k isthe weight associated with F k , K is the number of surroundfunctions, or scales, and ‘*’ represents the convolution operator. The surround functions F k are given as:)2)22(exp(),(k y x A y x k F σ+−⋅= (5)where σk are the scales that control the extent of thesurround, smaller values of σk lead to narrower surrounds.III. C HEST RADIOGRAPH ENHANCEMENT BY MULTI -SCALER ETINEXWe utilize the following steps to process chest radiographic images based on MSR theory. (1) Determine standard deviations of Gaussian surroundfunction:According to our tests for chest radiographs, we select twostandard deviations: σ1 =10, σ2 =270. （2）Convolution operation under two scales in (4).（3）Calculate image output by means of (4).Two weighted factors are selected : W 1=0.8, W 2=0.2, correspond to two scales under σ1 and σ2 respectively.（4）Stretch gray-level rang bymin max minf offset ffoffset inI outI −+−+= (6) Where I out is the final output of the image, offset is theaverage of input image I in . f max , f min can be obtained bydiv avg f ⋅+=αmax (7)div avg f ⋅−=αmin (8)Where avg is equal to offset , div is the standard deviation of image I in , α is a coefficient ranging from 1.5 to 3.IV. R ESULTS AND DISCUSSIONThe processed results of original image(Fig.1) are shown in Fig.2-3 by use of above MSR algorithm steps with three-scale and two-scale Retinex. To compare with the enhancement effects, we processed the same chest radiographic image with Frankle-McCann and McCann Retinex, contrast-limited adaptive histogram equalization (Matlab Toolbox) as shown in Fig.4-6.All of the processed images contrast and brightness are adjusted depending on the rule of lung’s regions to be clear at post-processing stage. McCann and Frankle-McCann Retinex loose some textures information, especially within the darker areas(Fig.4-5).Figure 1. Original imageFigure 2. Three-scale MSR (σ1 =10, σ2 =150, σ3 =270, w 1=w 2=w 3=1/3)Figure 3. Two-scale MSR (σ1 =10, σ2 =270,w 1=0.8, w 2=0.2)CLAHE can give all tissue features including the spine, but it fades the bright nodule(marked in circle in Fig.6) referring to Fig.1. Two-scale MSR is better than three-scale MSR as shown in Fig.3 and Fig.2 because amount of low frequency components participate in the synthesis in three-scale MSR. The presented technique based on MSR is of a better ability to deal with chest radiograph and less time consuming.V.C ONCLUSIONIn the paper, an image enhancement method for chest radiographs was proposed on the basis of MSR. In order to highlight the lung’s features, two-scale Retinex was applied, in which the chosen weighted factors and the standard deviations of Gaussian surround function plays an important role. It is shown that the given method can improve local and whole contrast and be more suitable for physician to diagnose through comparing with some methodologies.R EFERENCES[1] Land, Edwin and McCann, John, “Lightness and Retinex Theory”,Journal of the Optical Society of America, 61(1), January 1971.[2] McCann, John, McKee, Suzanne and Taylor, Thomas, "QuantitativeStudies in Retinex Theory, A Comparison between Theoretical Predictions and Observer Responses to the Color Mondrian Experiments", Vision Research, 16, pp.445-458, 1976.[3] Land, Edwin, "The Retinex Theory of Color Vision", ScientificAmerican, 237, pp.108 - 128, 1977.[4] Land, Edwin, “An alternative technique for the computation of thedesignator in the retinex theory of color vision”, Proc. of the National Academy of Science USA, 83, pp.2078-3080, May 1986.[5] McCann, John, “Lessons Learned from Mondrians Applied to RealImages and Color Gamuts”, Proc. IS&T/SID Seventh Color Imaging Conference, pp. 1-8, 1999.[6] D. J. Jobson, Z. Rahman, and G. A. Woodell, “A multi-scale Retinex forbridging the gap between colorimages and the human observation of scenes,” IEEE Transactions on Image Processing: Special Issue on Color Processing 6, pp. 965–976, July 1997.[7] Z. Rahman, D. J. Jobson, and G. A. Woodell, “Multiscale retinex forcolor rendition and dynamic range compression,” Applications of Digital Image Processing XIX, A. G. Tescher, ed., Proc. SPIE 2847, 1996.[8] Z. Rahman, D. J. Jobson, and G. A. Woodell, “Retinex processing forautomatic image enhancement,” Journal of Electronic Imaging 13(1), pp. 100–110, 2004.[9] Bo Sun, Weifang Chen, Hongyu Li ,Wenjing Tao , Jiang Li, “ModifiedLuminance Based Adaptive MSR”, Proceedings of the Fourth International Conference on Image and Graphics, pp.116-120, 2007. [10] Jinhua Wang, De Xu, Bing Li, “A novel tone mapping method based onRetinex theory”, Proceedings of the 7th WSEAS International Conference on Signal, Speech and Image Processing, Beijing, China, pp.146-149, 2007.[11] L. Meylan and S. Süsstrunk, “High Dynamic Range Image RenderingUsing a Retinex-Based Adaptive Filter”, IEEE Transactions on Image Processing, 15(9), pp. 2820-2830, 2006.[12] Nagesha, G. Hemantha Kuma, “A level crossing enhancement schemefor chest radiograph images”, Computers in Biology and Medicine,37(10), pp.1455-1460, 2007.[13] Jyh-Shyan Lin, Huai Li, M. T. Freedman, “Region-based enhancementof digital chest radiographs”, Proceedings of the Acoustics, Speech, and Signal Processing. on Conference Proceedings, 1996 IEEE International Conference - Volume 04, pp.2211-2214, 1996. [14] Chen Shuyue, Hou Honghua, Zeng Yanjun,et al. “Study of AutomaticEnhancement for Chest Radiograph”, Journal of Digital Imaging, 19(4), pp. 371-375 ,December 2006.Figure 4. McCann Retinex resultFigure 5. Frankle-McCann Retinex resultFigure 6. CLAHE result。