TELK OMNIKA , V ol. 17, No . 5, October 2019, pp . 2587 - 2594 ISSN: 1693-6930, accredited First Gr ade b y K emenr istekdikti, Decree No: 21/E/KPT/2018 DOI: 10.12928/TELK OMNIKA.v17i5.11964 2587 Chest radiograph ima g e enhancement with wa velet decomposition and morphological operations Anthon y Y . Aidoo *1 , Matilda Wilson 2 , Gloria A. Botc hwa y 3 1 Depar tment, of Mathematical Science , Easter n Connecticut State Univ ersit y , Willimantic , CT 06226, USA 2 Depar tment of Computer Science , Univ ersity of Ghana, Legon, Ghana 3 Depar tment of Mathematics , Univ ersity of Ghana, Legon, Ghana * Corresponding author , e-mail: aidooa@easter nct.edu 1 , matw aa@ug.edu.gh 2 , gaantwi@ug.edu.gh 3 Abstract Medical image processing algor ithms significantly aff ect the precision of disease diagnostic process . This mak es it cr ucial to impro v e the quality of a medical image with the goal to enhance perceiv ability of the points of interest in ord er to obtain accur ate diagnosis of a patient. Despite the reliance of v ar ious medical diagnostics on X-r a ys , the y are usually plagued b y dar k and lo w contr ast proper ties . Sought-after details in X-r a ys can only be accessed b y means of digital image processing techniques , despite the f act that these t echniques are f ar from being perf ect. In this paper , w e implement a w a v elet decomposition and reconstr uction technique to enhance r adiog r aph proper ties , using a ser ies of mor phological erosion and dilation to impro v e the visual quality of the chest r adiog r aphs f or the detection of cancer nodules . K e yw or ds: chest r adiog r aph, image enhancement, mathematical mor phology , w a v elet decomp osition Cop yright c 2019 Univer sitas Ahmad Dahlan. All rights reser ved. 1. Intr oduction A chest r adiog r aph pro vides a g reat measure of medical inf or mation about a patient’ s condition per taining to such diseases as lung cancer and chest inf ections . Ho w e v er , images produced b y X-Ra ys , are filled with noise due to interf e rences from captur ing de vices and anatomical str uctures [1]. The pr ime inspir ation in the v ast major ity of the computer algor ithms f or helping r adiologists in e xamining chest r adiog r aph imag es , is the clinical significance of chest r adiog r aph [2]. Locating cancer nodules in chest r adiog r aphs helps to detect ear ly signs of lung cancer . Ho w e v er , anatomical str uctures usually constitute unw anted ar tif acts in captured X-r a y images . Due to the siz e and density , nodules are usually difficult to detect in a chest r adiog r aph [3, 4]. Image processing algor ithms are designed to impro v e the accur acy of the diagnostic procedure [5], especially , in applications in Computer Assisted Diagnosis (CAD) systems [6]. Despite this , image processing algor ithms are not perf ect. Some of the most utiliz ed algor ithms relied upon to enhance the quality of chest r adiog r aphs include par ameter iz ed logar ithmic image filter ing method based on Laplacian of Gaussian (LoG) [7], the Hessian-LoG filter [8], and the mean and median filter ing f or noise remo v al [9]. These filters are only appropr iate f or cer tain types of noises and are inadequate f or enhancing medical images such as a chest r adiog r aph. Recently , a total v ar iation approach has been used to enhance the local contr ast in chest r adiog r aphs leading to significant impro v ement. [4]. Another common technique used in f eature-based enhancement classification is the classic unshar p masking method. The Fully Con v olutional Neur al Netw or ks (FCNN) is utiliz ed to impro v e the contr ast of delicate lung str uctures in chest r adiog r aphs [10-14]. These methods impro v e the image contr ast b ut not only f all shor t in detecting lung cancer nodules , b ut in addition, lead to unacceptab le n umber of f alse positiv es and f alse negativ es . W e implement a discrete w a v elet tr ansf or m combined with mor phological tools to enhance chest r adiog r aphs . This leads to an efficient method of denoising and enhancement Receiv ed December 4, 2018; Re vised F ebr uar y 8, 2019; Accepted March 12, 2019 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 2588 of x-r a y images , that outperf or ms w a v elet tr ansf or ms based techniques [15]. Our algor ithm is implemented in Python, leading to consistency in the quality of the results obtained. The str ategy str ikingly enhances points of interest in chest r adiog r aphs while preser ving t he details of delicate chest tissue , so r adiologists ma y ha v e a more e xact clar ification of diagnosis [16]. Due to its m ultiresolution capabilities , the w a v elet tr ansf or m has become a po w erful image processing tool [17, 18]. W a v elets ha v e a localizing proper ty and a char acter istic of denoising in a time-scale d omain and hence making a v ailab le local details of an image with minimal loss of detail. W a v elet based image enhancement techniques such as histog r am equalization and gamma adjustments when applied to chest r adiog r aphs , ho w e v er , lead to loss of vital image details [19]. Existing methods such as k er nel and spline estimators , tend not to resolv e local str uctures w ell, and spatial techniques such as the median and mean filters ha v e the demer it of b lurr ing the edges of an image in an attempt to smoothen the image to remo v e noise [17]. This is catastrophic in applications to medical imaging. In order to eliminate the prob lems listed abo v e , w e introduce a technique to remo v e anatomical noise whiles preser ving details . Firstly , w e use the dy adic w a v elet tr ansf or m that has the capability to locally decompose an image to remo v e the unw anted details and then reconstr uct the image using the der iv ed w a v elet coefficients . W e then apply the mor phological erosion and dilation f or se v er al iter ations to enhance the cancer nodules and realiz e a better appear ance , using a small and ellipsoidal str uctur ing element. 2. Morphological Er osion and Dilation Mathematical mor phology is a technique f or e xtr acting and analysing the par ts of an image that are of interest to the researcher . It is based on set theoretical axioms and is der iv ed from the basic Mink o wski set oper ations of addition and subtr action. In this f or m ulation, images are considered as functions mapped from the euclidean space M into R [ f1 ; 1g . Str uctur ing functions which are kno wn as str uctur ing elements are functions of the same f or m as the images . Giv en tw o image sets A and B , the mink o wski addition is defined b y A B = [ 2 B ( A + ) and the subtr action is defined b y A B = \ 2 B ( A + ) . The set A represents the image data and the set B is the str uctur ing element. The str uctur ing element pla ys a similar role that con v olution k er nels pla y in linear image filter ing. The basic mor phological o per ators , dilation and erosion are based on these tw o oper ations . Giv en an image function f ( x ) and a str uctur ing function s ( x ) , the g r a yscale dilation of f b y s is defined as D ( f ; s ) = ( f s )( x ) = sup y 2 M [( f ( y ) + s ( x y )] and erosion of f b y s is defined b y E ( f ; s ) = ( f s )( x ) = inf y 2 M [( f ( y ) s ( y x )] A flat str uctur ing function is defined as s ( x ) = ( 0 ; x 2 S 1 other w ise S M , is the str uctur ing function suppor t. With this kind of str uctur ing element, only tr ue pix els are count ed in the mor phological computation. This thus simplifies the definitions of dilation and erosion to D ( f ; s ) = ( f s )( x ) = sup z 2 S s ( f ( x + z ) and E ( f ; s ) = ( f s )( x ) = inf z 2 S ( f ( x + z ) Chest r adiog r aph image enhancement with... (Anthon y Aidoo) Evaluation Warning : The document was created with Spire.PDF for Python.
2589 ISSN: 1693-6930 respectiv ely , where S s denotes the symmetr ic str uctur ing function suppor t. Applying dilation oper ation to an object causes it to g ro w in siz e b y the str uctur ing element, whereas erosion causes the object to shr ink [20, 21, 22, 23]. Cancer nodules ha v e high intensity v alues than the adjacent usually br ight anatomical str uctures [24]. As such, w e erode the image first to get r id of all noisy details and then w e dilate the result using an elliptical str uctur ing element. This significantly imp ro v es the visibility of nodules to e v en the nak ed e y e in X-Ra ys . 3. Results W e implemented our technique on a database of a set of 247 chest X-r a y images from a standard Pub lic Database; the J apanese Society of Radiological T echnology . This database is endo w ed with diff erent cases which mak es it the appropr iate choice . These images w ere collected o v er a three y ear per iod from 14 medical institutions and are made up of anter ior and poster ior films of measure 14 14 inches . There are 154 images which ha v e lung nodules , out of which 100 are malignant and 54 are benign. 93 of the images are without lung nodules . Nodules are confir med b y CT and their locations are confir med b y three r adiologists [25]. W e im plemented our technique on 100 images with nodules and 60 images with no nodules . Our result sho w ed visib le presence of nodules in 100 % of the images with nodules , eliminating completely the occurence of f alse positiv es and f alse negativ es . Samples of our results are displa y ed in the Figures 1. Figure 1 is a w a v elet decomposition and reconstr uction compared with the or iginal images . Figure 1. (a-c) are the or iginal chest images with nodules , (d-f) are the w a v elet decomposition of images in (a-c), (g-i) are the reconstr ucted images from the decomposed images (d-f) Figure 2 f or ms par t of the sample used in Figure 1. It sho ws the w a v elet decomposition and reconstr uction compared with the or iginal images . Figure 3 sho ws a w a v elet decomposition and reconstr uction compared with the or iginal images . Figure 4 completes the set of images that compare w a v elet decomposition and reconstr uction with the or iginal images . Mor phological erosion and dilation are successiv ely applied to the sample of chest r adiog r aph images . P ar t of the results are sho wn in Figure 5. Figure 6 is the second sample of images that ha v e been processed with mor phological oper ations . The third set of the results of processing the sample TELK OMNIKA V ol. 17, No . 5, October 2019 : 2587 2594 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 2590 of images with mor phological erosion and dilation is displa y ed in Figure 7. The final set of the results of processing the sample of images with mo r phological erosion and dilation is displa y ed in Figure 8. Figure 2. This is par t of the sample of images sho wn in Figure 1. (a-b) are the or iginal chest images with nodules , (c-d) are the w a v elet decomposition of images in (a-b), (e-f) are the reconstr ucted images from the decomposed images (c-d) Figure 3. (a-c) are the or iginal chest images without nodules , (d-f) are the w a v elet decomposition of images in (a-c), (g-i) are the reconstr ucted images from the decomposed images (d-f) Chest r adiog r aph image enhancement with... (Anthon y Aidoo) Evaluation Warning : The document was created with Spire.PDF for Python.
2591 ISSN: 1693-6930 Figure 4. This is par t of the sample of images in Figure 3 which sho ws (a-b) are the or iginal chest images without nodules , (c-d) are the w a v elet decomposition of images in (a-b), (e-f) are the reconstr ucted images from the decomposed images (c-d) Figure 5. (a-c) are the result of applying mor phological erosion to reconstr ucted chest images with nodules , (d-f) sho w dilation of the eroded images in (a-c) (Opening), (g-i) sho w in v er ted images of the images in (d-f) TELK OMNIKA V ol. 17, No . 5, October 2019 : 2587 2594 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 2592 Figure 6. This is par t of the sample of images sho wn in Figure 5. (a-b) are the result of applying mor phological erosion to reconstr ucted chest images with nodules , (c-d) sho w dilation of the eroded images in (a-b), (e-f) sho w in v er ted images of the images in (c-d) Figure 7. Mor phological erosion applied to reconstr ucted chest images without nodules , (a-c) are the result of applying mor phological erosion to reconstr ucted chest images with nodules , (d-f) sho w dilation of the eroded images in (a-c), (g-i) sho w in v er ted images of the images in (d-f) Chest r adiog r aph image enhancement with... (Anthon y Aidoo) Evaluation Warning : The document was created with Spire.PDF for Python.
2593 ISSN: 1693-6930 Figure 8. This is par t of the sample of images sho wn in Figure 7. Mor phological erosion applied to reconstr ucted chest images without nodules ,(a-b) are the result of applying mor phological erosion to reconstr ucted chest images without nodules , (c-d) sho w dilation of the eroded images in (a-b), (e-f) sho w in v er ted images of the images in (c-d) 4. Discussion and Conc lusion W e de v eloped a technique f or im age denoising and enhancement based on a combination of w a v elets and mor phological erosion and dilation which is presented and applied to a large sample of chest x-r a y images , some of which contained cancer nodules in order to enhance the quality and contr ast of the x-r a y images . Our approach is tested on a n umber of pub licly a v ailab le chest r adiog r aph images . The combined w a v elet based and mathematical mor phology technique retains and elucidate more detail image inf or mation o f interest on both cancer nodules and anatomical str uctures captured in chest r adiog r aphs . The technique not only suppresses unw anted noise , it also preser v es the edges of the nodules to enab le accur ate detection. F rom the results obtained, w e conclude that our technique is efficient and compares f a v or ab ly with nonmor phological based techniques f or chest r adiog r aph image enhancement. Ref erences [1] Rajni A. Image Denoising T echniqu es-An Ov er vie w . Inter national Jour nal of Computer Applications . 2014: 86(16): 0975-8887. [2] V an Ginnek en B ., T er Haar Romen y BM, Vierge v er MA. Computer-aided Diagnosis in Chest Radiog r a- ph y: A Sur v e y . IEEE T r ans Med Imaging . 2001: 20(12):1228-1241. [3] Chen SY , Hou HH, Zeng YJ , Xu XM.: Study of automatic enhancement f or chest r adiog r aphs Jour nal of Digital Imaging . 2006: 4: 371-375. [4] Wilson M, Aidoo A Y , Acquah JB .H, Y irenkyi, P . Chest Radiog r aph Image Enhancement: A T otal V ar iation Approach, Inter national Jour nal of Computer Applications . 2017: 163: 0975 - 8887. [5] Sherr ier RH., Chiles C , Wilkinson WE, Johnson GA, Ra vin CE. Eff ects o f Image Processing on Nodule Detection Rates in Digitiz ed Chest Radiog r aphs: R OC Study of Obser v e r P erf or mance . Radiology . 1988: 166(2): 447-450. [6] Firoz R, Ali MS , Khan MNU , Hossain MK, Islam MK, Shahin uzzaman, M. Medical Image Enhancement Using Mor phological T r ansf or mation. Jour nal of Data Analysis and In f or mation Processing . 2016: 4: 1-12. TELK OMNIKA V ol. 17, No . 5, October 2019 : 2587 2594 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 2594 [7] Bao C , Sheng C , A P ar ameter iz ed Logar ithmic Image Processing Method Based on Laplacian of Gaus- sian Filter ing f or Lung Nodules Enhancement in Chest Radiog r aphs 2nd Inter national Symposium on Instr umentation and Measurement, Sensor Netw or k and A utomation (IMSNA), T oront o . 2013: 649–52. [8] Shi Z, Zhao M, W ang Y , He L, Suzzuki K , Jin C . Hessian-LoG: A No v el Dot Enhancement Filter . PICIC Exp Lett . 2012: 6:1987–92. [9] Sar age GN, J ambhor kar S . Enhancement of Chest X-r a y Images Using Filter ing T echniques , Int’l Jour- nal of Adv anced Research in Computer Science and Softw are Engineer ing . 2012: 2:308-312. [10] Sezn MI, T eklap AM, Schaetzing R. A utomatic Anatom ically Selectiv e Image Enhancement in Digital Chest Radiog r aph y . IEEE T r ans Med Imaging . 1989: 8: 154-62. [11] T ahoces PG, Corre a J , S outo M, Gonzalez C , Gomez L, Vidal JJ . Enhancement of Chest and Breast Radiog r aphs b y A utomatic Spatial Filter ing. IEEE T r ans Med Imaging . 1991: 10:330-335. [12] Souto M, Correa J , T ahoces PG, T uc k er D , Malagar i KS , Vidal JJ , F r aser RG. Enhancement of Chest Images b y A utomatic Adaptiv e Spatial Filter ing. J Digit Imaging . 1992: 15(4): 223-229. [13] Saleh SH, Nordin AJ . Imp ro ving Diagnostic Vie wing of Medical Images Using Enhancement Algor ithms . J Comput Sci . 2011;7:1831-1838. [14] Goz es O , Greenspan H. Lung Str uctures Enhancement in Chest Radiog r aphs via CT based FCNN T r aining, https://arxiv .org/abs/1810.05989, 2018. [15] V ese L. A Study in the Bounded V ar iation Space of Denoising-deb lurr ing V ar iatonal Prob lem. Applied Mathematics and Optimization . 2001: 44(2): 131-161. [16] Ala vijeh FS ., Mahda vi-Nasab H., Multi-scale Mor phological Image Enhancement of Chest Radiog r aphs b y a Hybr id Scheme , J Med Signals Sens . 2015: 5(1): 59–68 [17] Anatum , Rajni: P erf or mance Analysis of Image Denoising with W a v elet Thresholding Methods f or Diff er- ent Le v els of Decomposition. Inter national Jour nal o f Multimedia and its Applications . 2014: 6(3):35-46 [18] Ar iv azha gan, S ., D eiv alakshmi, S ., Kannan, K.: P erf or mance Analysis of Image Denoising S ystem f or Diff erent Le v els of W a v elet Decomposition Inter national Jour nal of Imaging S cience and Engineer ing . 2007: 1(3):104-107. [19] Unaldi N, T emel S , Demirci S . Undecimated W a v elet T r ansf or m Based Contr ast Enhancement. Inter- national Jour nal of Computer , Electr ical, A utomation, Control and In f or mation Engineer ing . 2013: 7: 1215-1218. [20] Najman L, T albot H, (Eds)., Mat hematical Mor phology: F rom Theor y to Applications . Gene v a. Wile y 2010. 520. [21] Ronse C , Najman L, Decenci E. (Eds .), Mathematical Mor phology: 40 Y ears On. Ne w Y or k. Spr inger . 2005. [22] Banon GJ , Barrer a J , Br aga-Neto UM. Mathematical Mor phology and its Applications to Signal and Image Processing . Proceedings of the 8th inter national symposium on mathematical mor phology . Rio de J aneiro . 2007: 2-4. [23] Sreed har K, P anlal B . Enhancement of Images Using Mor phological T r ansf or mations . Inter national Jour nal of Computer Science and Inf or mation T echnology . 2012: 4(1):33-50. [24] Rubin GD . Lung Nodule and Cancer Detection in CT S creening, Jour nal of Thor acic Imaging . 2015: 30(2):130-138. [25] Ik ez oe S , Matsumoto J , K omatsu T , Matsui KI, Fujita M, K oder a H, Doi Y , Shir aishi K, Katsur aga w a, J . De v elopment of a digital image database f or chest r adiog r aphs with and without a lung nodule: Receiv er oper ating char acter istics analysis of r adiologists detection of pulmonar y nodules . Amer ican Jour nal of Roentgenology . 2000: 174: 71-74. Chest r adiog r aph image enhancement with... (Anthon y Aidoo) Evaluation Warning : The document was created with Spire.PDF for Python.