WIRELESS CAPSULE ENDOSCOPY IMAGES ENHANCEMENT BASED ON ADAPTIVE ANISOTROPIC DIFFUSION Lei Li 1, Y. X. ZOU 1* and Yi Li 1 ADSPLAB/ELIP, Shool of ECE, Peking Universit, Shenzhen 518055, China Shenzhen JiFu Tehnolog Ltd. {*zoux@pkusz.edu.n} ABSTRACT Wireless apsule endosop (WCE) is a new innovative solution for gastrointestinal disease detetion. The image qualit of WCE is not satisfator for medial appliations sine some of them are dark and low-ontrast. The WCE image enhanement is a hallenge task, mainl beause the diversit of the WCE images of different people and the need to preserve the loal fine details of WCE images. Hene, it is diffiult to appl the traditional image enhanement methods based on global information to WCE image enhanement. In this paper, motivated b the apabilit of the anisotropi diffusion method in preserving loal fine details, a new adaptive ontrast anisotropi diffusion method has been developed to enhane WCE images, where the Hessian matrix is used to obtain better representation of the ontrast information of WCE images meanwhile the diffusion parameter of the diffusion oeffiient funtion is automatiall determined aording to the loal harateristi. Experimental results have demonstrated the effetiveness of the proposed image enhanement method. Index Terms wireless apsule endosop, anisotropi diffusion, adaptive, image enhanement, loal harateristi 1. INTRODUCTION In the past ear, digestive sstem aner has been the seond aner-related killer in U.S.A. and aused 61,950 death[1]. Earl detetion and treatment of gastrointestinal disease are ver important. Wireless Capsule Endosop (WCE) was invented in ear 000 and put in use in 001 []. When a patient swallows a WCE, it will be propelled b peristalsis though the entire gastrointestinal trat, while it reords and sends the aptured images to a devie attahed to the patient wirelessl. The main advantages of WCE are that patients an avoid ross infetion and suffer no pain. WCE solution brings big benefits to elder and weak patients. The whole examination proess of WCE will last for 8 hours and produe about 5,000-10,000 images per person. However, the images taken b WCE sstem are not as distint as those taken b traditional endosop due to the following reasons. Firstl, due to the volume restrition of the WCE, the batter apait is limited. Hene, some images are not taken under suffiient illumination. Seondl, the amera used in WCE is low-foal-length amera whih ma produe blurring images. Thirdl, ompliated irumstane of gastrointestinal trat and moving imaging method also lead to a poor image qualit. Fourthl, the image-data must be ompressed in high ratio before transmission outside. As a result, the image enhanement is highl demanded in WCE medial appliations. There are man famous image enhanement methods. Histogram equalization and its derived methods [3, 4] are ommon approahes to enhane low-luminosit and low-ontrast images, whih are effetive and fast. Filter based methods [5] and transform domain based methods [6] are also useful in image enhanement. However, researh shows that the WCE images have rather smooth and low-ontrast properties. Experimental results showed that, the methods mentioned above don t have desired effet in preserving loal fine details in the enhanement proess whih are important for medial diagnosis. In 1990, anisotropi diffusion, also alled Perona Malik diffusion, was proposed, whih is a tehnique aiming at reduing image noise without removing signifiant parts of the image ontent, tpiall edges, lines or other details [7]. Anisotropi diffusion has been widel used in medial image enhanement and other image proessing appliations [8, 9]. In this paper, motivated b the favorite properties of the anisotropi diffusion, we put forward a new adaptive ontrast anisotropi diffusion method for WCE image enhanement b exploring the loal harateristis of WCE images. Referring to [10], we took the advantage of the propert of Hessian matrix to obtain better representation of the ontrast information of WCE images. The experimental results validated the effetiveness of the proposed method.. IMAGE ENHANCEMENT USING ANISOTROPIC DIFFUSION The anisotropi diffusion proposed b Perona and Malik is based on the following equation: 978-1-4799-1043-4/13/$31.00 013 IEEE 73 ChinaSIP 013
I ( xt,, )/ t= div[ ( xt,, ) I( xt,, )] (1) I( x,, t = 0 ) = I ( x, ) 0 where x, is the absissa and ordinate of the point in image respetivel, t is the diffusion time. I(x,,t) is the gra level of point (x,) at time t, and I 0 (x,) is the original gra level; div and are the divergene and gradient operators, respetivel; (x,,t) is the diffusion oeffiient funtion ontrolling the diffusion rate and is usuall hosen as a funtion of the image gradient to preserve edges in the image. Anisotropi diffusion is a ontinuous proess in whih the value of eah pixel of the image is diffused to its neighborhood pixels. Eah pixel adds its neighbors diffused value to form a new urrent value iterativel. From (1), we an see that aurate measurement of gradient information of the image is ruial to the edge preserving during the anisotropi diffusion proess. But in WCE images, the edges are rather weak and the gradients are diffiult to estimate auratel. To ahieve better performane, Li and Meng proposed a ontrast desription of the original image using Hessian matrix [10] to stress the gradient information. Suppose I is a grasale image, its Hessian matrix at (x,) under sale σ is defined as: I xx I x H σ = I x I () where I xx, I, I x are the seond-order derivative of the image along diretion x,, x respetivel. The seondorder derivative is more sensitive to edges than the firstorder derivative, so it is more appropriate to detet details in low ontrast image like WCE images. Sine the Hessian matrix is a matrix, it has two eigenvalues λ 1 and λ and the are proportional to the intensit variations of the image. Hene, the flat image region orresponds to small eigenvalue and the image region inluding edges and more details will ield larger eigenvalue. Based on above observations, a ontrast desription of I at sale σ is as I x, = λ1 x, + λ x, σ (3) Following the definition given in (1), the anisotropi diffusion of I an be formulated as the following: I ( x,, t ) / t = div [ ( x,, t ) (,, )] I x t (4) I ( x,, t = 0 ) = I ( x, ) 0 Assume we get a resultant image D after the anisotropi diffusion, then we will normalize it to get the enhaned image using the following equation: D( x, ) min( D( x, )) I ( x, ) = 55 (5) result max ( Dx (, )) min( Dx (, )) Disussions: the image enhanement tehnique based on anisotropi diffusion using (1) or (4) signifiantl depends on the hoie of the diffusion oeffiient funtion (x,,t). In (a) (b) () Fig. 1. Anisotropi diffusion with different parameter k: (a) original image (b) k=10, iteration=30 () k=30, iteration=30 (a) (b) Fig.. (a) WCE image; (b) its gra sale image order to ahieve smoothness in non-edge regions and preserving edges and loal details in inter-edges regions, the funtion (x,,t) should have the following properties: for the-non-edge or flat regions, (x,,t) should be of large value; for the inter-edge and rih detail regions, (x,,t) should be of small value. Two famous funtions for the diffusion oeffiient are shown below 1/ 1 ( x/ k) x = + (6) and ( x) = exp ( x/ k) (7) where k is an edge sensitivit ontrolling parameter and is usuall hosen experimentall or as a funtion of the image loal properties. Therefore, the seletion of k is ver ruial to the qualit of the enhaned image I result. Experimental results illustrated that the larger k leads to the smoother image and the smaller k leads to a sharper image. One example is shown in Fig. 1. Careful evaluation shows that the low-ontrast regions in the WCE image ontain less medial information. Conversel, the high-ontrast regions ontain muh more medial information, whih is important to diagnose and should be preserved. Therefore, it is desired that the large k is seleted for the low-ontrast regions and small one for the high-ontrast regions. 3. SELECTION OF DIFFUSION THRESHOLD PARAMETER BASED ON IMAGE LOCAL DETAILS As we disussed in Setion, the parameter k in diffusion oeffiient funtion plas a ke role. Obviousl, it is optimal to selet k aording to its loal information and it 74
should be big where the loal ontrast is low and be small where the loal ontrast is high. Bearing this idea, let s define a data-related parameter k x, = K / L x, (8) 0 ( dd ) where K 0 is a onstant. L dd is a loal detail desriptor, whih should meet the following requirements: For the flat regions of the image, the L dd should be small. For the high ontrast regions, the L dd should be of large value. With the definition in (8), the formulation of the loal desriptor L dd is a ke. In this paper, we proposed to use the loal omplexit (LC) termed as L and the loal variane (LV) named as L v to jointl form L dd. Essentiall, L and L v measure the loal image entrop and dnami variation, respetivel. Intuitivel, flat regions have poor unertaint and less information, on the ontrar, edges and fine details tend to have massive unertaint and information. Image entrop proposed b Pun in 1980 an appropriatel desribe the unertaint of an image [11]. Considering an image region entered at (x, ) with a fixed m n window, the image loal entrop an be defined as L x L 1, = P log P (9) e j= 0 j j where P n / ( m n) j = (10) j where L is the gra levels, n j is the number of pixels of the level j. From (9), it is lear to see that the loal entrop is a good loal unertaint desriptor, but it suffers from high omputation omplexit. Aiming at reduing the omplexit, an alternative approah is adopted [1]: L 1 k = 0 L x, = sgn( k) (11) where k is the gra levels of the image between 0 and L-1. Simulation studies showed that either L e (x,) given in (9) or L (x,) given in (11) is not good enough to represent the loal desriptor required in (8). This is beause L e (x,) and L (x,) are onl able to measure the overall unertaint of the region, but annot measure the dnami hange in the region. It is well-known that the data variane is an appropriate measure of the image gra level spread range. Thus, the gra level loal variane of the image region entered at (x, ) with a fixed m n window is defined as: m n 1 L ( x, ) = ( f ( x, v ) m f ) (1) m n i= 1 j= 1 where m f is the mean gra level of the pixels in the m n neighborhood window. Taking both loal omplexit and loal variane into onsideration, the proposed loal detail desriptor L dd an be formulated as follows: (, ) (, ) log ( (, )) L x = L x L x (13) dd v The reasons for formulating L dd in (13) are as follows: For low-ontrast regions, sine L (x,) and L v (x,) are small, then L dd (x,) is small as a onsequene. In high-ontrast regions, sine L (x,) and L v (x,) are large and L dd (x,) is large as well. It is noted that L v (x,) in (1) is sensitive to noise and an impulsive noise ma lead to a large L v (x,). But L (x,) in (11) is muh robust to impulsive-like noise. Therefore, for noise regions, L (x,) is small whih makes L dd (x,) smaller than high-ontrast regions. It is noted that the log operator in (13) is able to eliminate the effet of impulsive-like noise in some measure. Generall, the variation of L (x,) is muh smaller than that of L v (x,). For example, for a WCE image shown in Fig., the L (x,) ranges from to 84 but L v (x,) ranges from 1.13 to 3150, meanwhile the (log(l v (x,))) is in the range of [0.15, 64.89] whih nearl mathes the variation range of L (x,). In onlusion, L dd proposed in (13) is able to distinguish highontrast regions from low-ontrast regions and alleviate the adverse impat of the impulsive-like noise. Extensive experiments have been arried out with real WCE images to validate the formulation of the proposed loal detail desriptor L dd in (13). It is onfident to onlude that L dd given in (13) is able to appropriatel desript the loal details of WCE images and meanwhile is robust to noise. 4. EXPERIMENTAL RESULTS In order to evaluate the performane of the proposed image enhanement algorithm, we onduted several experiments. The WCE images used are from Shenzhen JiFu Tehnolog Ltd., whih were aptured data from trial patients. The size of the WCE images is 480 480. Sine the aptured WCE images are RGB olor images, therefore, we diretl applied the proposed image enhanement method on R, G, B images respetivel then merged them to the enhaned olor WCE images. We ompared the results of the proposed image enhanement algorithm with the anisotropi diffusion method (ADM) [7] and the Dnami Histogram Equalization (DHE) image enhanement algorithm [3]. For ADM method and our proposed method, (6) is taken as the diffusion oeffiient funtion onerning its low omputing omplexit. For ADM method, parameter k in (6) is set to 10. For our proposed method, the parameter K 0 in (8) is also set to be 10. Loal window size used in (11) is set to 11 11. The results are obtained after 30 iterations of diffusion. Iteration time is also a ke fator to the result. It is noted that with the inrease of the iterations, the enhaned images tend to onverge to the best result and then start to degenerate. Hene, 30 iteration is an empirial hoie after the omparison with 10, 0 40, and 50 iterations. The optimal hoie of the iteration is also our future researh ontent. It is noted that it is diffiult to evaluate the performane of a olor image enhanement algorithm objetivel. Aording to our knowledge, there is 75
no authoritative quantitative measurement. So we will estimate the proposed approah in a subjetive wa. Due to the paper length limitation, we onl present three sets of the experimental results here. Fig. 3 demonstrated a set of experimental results. Fig. 3 (a) is the original WCE image of some vessels. Fig. 3 (b) to Fig. 3 (d) shows the enhaned WCE image b DHE algorithm, the anisotropi diffusion method (ADM), and our proposed method, respetivel. As we an see, the original image is blurred slightl. Some of the edges of this WCE image are too weak to detet. The ADM algorithm failed to enhane the image and produed an even more blurred image. The DHE algorithm produed aeptable result but magnified some noise. In the image enhaned b our proposed method, we an learl see the vessels without noise amplifiation. Fig. 4 gives another set of experimental results, where the original WCE image in Fig. 4 (a) is muh dark ompared with the one in Fig. 3 (a). Our proposed algorithm produed a lear result even in the dark regions refleting image details learl, whih is shown in Fig. 4 (d). Result produed b the ADM algorithm shown in Fig. 4 () is also inferior to the original image. Result from DHE algorithm is also dark, and aused hrominane hange as we an see in Fig. 4 (b). In Fig. 5 (a), the original image is suspet of jejunum erosion disease. The result from DHE algorithm produed some hrominane hange in the abnormal region as shown in Fig. 5 (b). Our proposed approah enhaned the image properl, as shown in Fig. 5 (d), espeiall at the abnormal region where muh more image details an be viewed for more aurate judgment. From the experimental results illustrated, we an onlude that our proposed method has the abilit to effetivel enhane the olor WCE images. 5. CONCLUSION This paper works on the WCE images enhanement tehnique. With the analsis of the loal harateristi of WCE images, we proposed an effiient image enhanement method based on the anisotropi diffusion tehnique. A new approah of automatiall seleting the parameter of the diffusion oeffiient based on the loal information of the WCE image has been developed. Extensive simulations with the real WCE images show that the proposed image enhanement method is able to effetivel enhane various WCE images with the good preservation of the loal details. Future stud will fous on reduing the omputational omplexit of the proposed approah and automati nidus detetion based on the enhaned images. ACKNOWLEDGMENT This work is partiall supported b the funded projet of the Shenzhen Government and the ollaboration researh projet funded b Shenzhen JiFu Tehnolog Ltd. and the () Anisotropi diffusion (d) Proposed approah Fig. 3. Enhanement result of different approahes () Anisotropi diffusion (d) Proposed approah Fig. 4. Enhanement result of different approahes () Anisotropi diffusion (d) Proposed approah Fig. 5. Enhanement result of different approahes authors would thank Shenzhen JiFu Tehnolog Ltd. for providing us the WCE testing images. 76
6. REFERENCES [1] Siegel, R., D. Naishadham, and A. Jemal, Caner statistis, 01. CA: a aner journal for liniians, 01. [] Iddan, G., et al., Wireless apsule endosop. Nature. 405: p. 417, 000. [3] Abdullah-Al-Wadud, M., et al., A Dnami Histogram Equalization for Image Contrast Enhanement., IEEE Transations on Consumer Eletronis, 53(): p. 593-600, 007. [4] Ibrahim, H. and N.S.P. Kong, Brightness preserving dnami histogram equalization for image ontrast enhanement, IEEE Transations on Consumer Eletronis, 53(4): p. 175-1758, 007. [5] Polesel, A., G. Ramponi, and V.J. Mathews, Image enhanement via adaptive unsharp masking, IEEE Transations on Image Proessing, 9(3): p. 505-510, 000. [6] Stark, J.L., et al., Gra and olor image ontrast enhanement b the urvelet transform, IEEE Transations on Image Proessing, 1(6): p. 706-717, 003. [7] Perona, P. and J. Malik, Sale-spae and edge detetion using anisotropi diffusion, IEEE Transations on Pattern Analsis and Mahine Intelligene, 1(7): p. 69-639. 1990. [8] Gilboa, G., N. Sohen, et al., "Forward-and-bakward diffusion proesses for adaptive image enhanement and denoising." IEEE Transations on Image Proessing, 11(7): 689-703, 00. [9] Ghita, O., K. Robinson, et al., "MRI diffusion-based filtering: a note on performane haraterisation." Computerized Medial Imaging and Graphis, 9(4): 67-77, 005. [10] Li, B. and M.Q.H. Meng, Wireless apsule endosop images enhanement using ontrast driven forward and bakward anisotropi diffusion, International Conferene on Image Proessing, p. 437-440, 007. [11] Pun, T., A new method for gre-level piture thresholding using the entrop of the histogram, Signal Proessing, (3): p. 3-37, 1980. [1] Yan, C., et al., Image transition region extration and segmentation based on loal omplexit, Journal of Infrared and Millimeter Waves-Chinese Edition, 4(4): p. 31, 005. 77