Home & Garden

Detection of vertebral plateaus in lateral lumbar spinal X-ray images with Gabor filters

Description
A few recent studies have proposed computed-aided methods for the detection and analysis of vertebral bodies in radiographic images. This paper presents a method based on Gabor filters. Forty-one lateral lumbar spinal X-ray images from different
Categories
Published
of 4
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Transcript
  Detection of Vertebral Plateaus in Lateral Lumbar Spinal X-ray Imageswith Gabor Filters Eduardo Alvarez Ribeiro, Marcello Henrique Nogueira-Barbosa,Rangaraj M. Rangayyan, Fellow, IEEE, Paulo M. Azevedo-Marques*, Member, IEEE  Abstract —A few recent studies have proposed computed-aided methods for the detection and analysis of vertebral bodiesin radiographic images. This paper presents a method basedon Gabor filters. Forty-one lateral lumbar spinal X-ray imagesfrom different patients were included in the study. For eachimage, a radiologist manually delineated the vertebral plateausof L1, L2, L3, and L4 using a software tool for image displayand mark-up. Each original image was filtered with a bankof 180 Gabor filters. The angle of the Gabor filter with thehighest response at each pixel was used to derive a measureof the strength of orientation or alignment. In order to limitthe spatial extent of the image data and the derived featuresin further analysis, a semi-automated procedure was appliedto the srcinal image. A neural network utilizing the logisticsigmoid function was trained with pixel intensity from thesrcinal image, the result of manual delineation of the plateaus,the Gabor magnitude response, and the alignment image. Theaverage overlap between the results of detection by imageprocessing and manual delineation of the plateaus of L1-L4in the 41 images tested was 0.917. The results are expected tobe useful in the analysis of vertebral deformities and fractures. I. INTRODUCTIONVertebral fractures are important indicators of osteoporo-sis. Spinal radiography continues to have a substantial rolein the diagnosis and follow-up of vertebral fractures [1, 2].Genant et al. [2] proposed a semi-quantitative method forthe evaluation of vertebral fractures related to osteoporosis,which combines numerical classification with visual inspec-tion. Analysis of spinal and vertebral deformities could assistin the diagnostic decision-making process and in guidingtherapeutic procedures. The primary evaluation of verte-bral deformity in individuals with osteoporosis is typicallyperformed via radiographic studies, including morphometricassessment of vertebral bodies [3-7]. Both semi-quantitativeand quantitative methods have been used to achieve objectiveand reproducible definition of the associated findings [2]. Incurrent clinical practice, most of the procedures for quantita-tive morphometric analysis of vertebral bodies are performedmanually, in a manner that is labor-intensive and subject tosignificant inter-observer and intra-observer variability. Aninitial stage of digital image processing is desired to facilitatethe development of automated or semi-automated proceduresthat could lead to more efficient and accurate analysis of vertebral deformities. A few recent studies have proposedcomputed-aided methods for the detection and analysis of  School of Medicine of Ribeirao Preto, University of Sao Paulo, RibeiraoPreto, SP, Brazil 14048-900  dualvarez@usp.br Departmente of Electrical and Computer Engineering. Schulich Schoolof Engineering of Calgary, Calgary, AB, Canada T2N 1N4 vertebral bodies in radiographic images [8-13]. The objectiveof this work is to develop digital image processing and pat-tern recognition techniques [14] for the detection of vertebralplateaus in lateral X-ray images of the lumbar spine. Thetechniques are based on the detection and characterizationof oriented edges or patterns using Gabor filters [14, 15]and classification using a neural network [16].II. M ATERIALS AND  M ETHODS Lateral lumbar spinal X-ray images were obtained fromcase files at the Clinical Hospital of the Faculty of Medicine,University of Sao Paulo, at Ribeirao Preto, SP, Brazil. Forty-one cases were selected for the study by a radiologist(MHNB) with experience in spinal and musculoskeletal ra-diology, in such a manner as to include the typical variationsobserved in the clinical setting; 19 of the images are frompatients with confirmed diagnosis of insufficiency fracturesand the remaining 22 images are from a normal controlgroup. The radiographic film images were digitized usinga Vidar DiagnosticPro scanner with the spatial resolution of 84 m and the gray scale represented by 8 bits/pixel. Thetypical size of the digitized images is about 3000 x 4000pixels. In the present study, interest is focused on the fourlumbar vertebrae L1-L4 . For each of the 41 images, theradiologist manually delineated the horizontal plateaus of the lumbar vertebrae L1-L4 using a software tool for imagedisplay and mark-up. Figures 1 and 2 show a part of ansrcinal radiographic image and the corresponding manuallydelineated plateaus of L1-L4. The lines corresponding tomanually delineated plateaus were set to have thickness 5pixels to allow minor variations or error around the trueplateaus both in the process of manual delineation anddetection via image processing. Each image was filtered witha bank of 180 Gabor filters. A Gabor filter is obtained bymodulating a Gaussian envelope function with a sinusoidalwave, and acts as a bandpass filter [14, 15]. A bank of such filters rotated to various orientations was used to detectoriented objects, edges, and textural patterns.A basic Gabor filter function was defined with the typicalwidth of the oriented pattern to be detected as 4 pixels and anelongation factor of 8 [14, 15]. The real Gabor filter kerneloriented at the angle θ  = − π/ 2 is given as equation 1 32nd Annual International Conference of the IEEE EMBSBuenos Aires, Argentina, August 31 - September 4, 2010 978-1-4244-4124-2/10/$25.00 ©2010 IEEE  4052  Fig. 1. A lateral lumbar spinal X-ray imageFig. 2. Manually delineated plateaus of L1?L4 for the srcinal image inFigure 1 g ( x,y ) = 12 πσ x σ y exp  − 12   xσ x 2  +  y 2 σ y 2  cos (2 πfx ) (1)The filter was rotated in steps of 1 degree to obtaina bank of 180 filters spanning the angular range of zeroto 180 degrees. For each pixel, the magnitude responseand angle of the Gabor filter providing the highest outputwere used to compose a Gabor magnitude image and anorientation field, respectively. Figures 3 and 4 show theGabor magnitude response image and the orientation field,respectively, for the srcinal image in Figure 1. It is evidentthat the Gabor magnitude image provides high values at theedges of the vertebral bodies and also at the edges of otheroriented structures and details in the image. The orientationfield shows a number of mutually aligned lines along themajor edges and oriented structures present in the srcinalimage. An indication of the presence of a dominant orientedstructure across several pixels was obtained by computing ameasure of alignment, given by the normalized sum of thesquared Gabor magnitude responses for all angles at a givenpixel weighted by the sum of the sine and cosine of twicethe corresponding angle. Figure 5 depicts the alignment datafor the test image in Figure 1; it is seen that the alignmentvalues are high along oriented edges and features. Fig. 3. Gabor magnitude response image for the srcinal image in Figure1Fig. 4. Orientation field for the srcinal image in Figure 1. Needles havebeen drawn for every 10th pixel in the row direction and every 10th pixelin the column direction to indicate the local orientation. In order to limit the spatial extent of the image data andthe derived features in further analysis, a semi-automatedprocedure was applied to the srcinal image as follows. Fivepoints were marked near the middle of the inter-vertebralspaces spanning the range of L1-L4 by using a mouse;the five points are labeled as P(1)-P(5) in Figure 6. Thedistances between the points were calculated automatically,and identified as D(1,2), D(2,3), D(3,4), and D(4,5), asshown in Figure 6. Using 75% of each distance measure, the 4053  Fig. 5. Alignment image for the original image in Figure 1 corresponding line joining the manually marked points wasshifted in either direction along its perpendicular to createa quadrilateral region for each vertebra. The combinationof the four regions obtained in this manner demarcates thefour vertebrae of interest, L1-L4, as shown in Figure 6.The corresponding regions were obtained from the srcinalimage, the Gabor magnitude response, and the alignmentimage for further analysis using a neural network utilizingthe logistic sigmoid function [14]. Fig. 6. Localization of the vertebrae L1?L4 using a semi-automatic method In order to remove isolated pixels or small regions andfill holes within the detected vertebral plateaus, a series of morphological opening and closing filters [16] was applied tothe result of the classification, using disk-shaped structuringelements of size 1 and 2 pixels, respectively.The accuracy of the results of detection was analyzedby comparing the morphologically filtered outputs with thecorresponding manual delineation. A measure of overlapwas computed as the ratio of the intersection of the result TABLE IA CCURACY OF DETECTION OF VERTEBRAL PLATEAUS  L1-L4  IN  41 IMAGES IN TERMS OF OVERLAP BETWEEN THE DETECTED PLATEAUSAND THE MANUALLY DELINEATED PLATEAUS . S TD . = S TANDARDDEVIATION .Images Overlap StdAll 41 images 0.917 0.14Normal control group (22) 0.923 0.012Insufficiency fractures (19) 0.911 0.015 of delineation by the radiologist and that of the imageprocessing procedure to their union.III. R ESULTS Figure 7 shows the final result of detection obtainedfrom the proposed methods. Comparing the final result of detection in Figure 7 to the manual delineation shown inFigure 2, it is seen that the result is accurate, with gooddepiction of the horizontal plateaus of the vertebral bodies.Table 1 shows the values of the overlap measure and theassociated standard deviation for the 41 images processed.The high overlap values of 0.91 - 0.92 indicate successfulperformance of the proposed methods. Fig. 7. Final result of the proposed image processing methods for thedetection of the plateaus of lumbar vertebrae L1?L4 for the srcinal imageshown in Figure 1 The complexity of the algorithm for a single image withdimensions AxB is O (A * B). For training the neuralnetwork are needed N pictures which results in a complexityof O (N * A * B). However, the identification of vertebralplateaus in lateral lumbar spine x-ray requires only a singleimage to be analyzed, so the complexity of the algorithm inthe Bachmann-Landau notation is O (A*B).IV. CONCLUSIONSThe proposed methodology has shown high efficiencyin semi-automatic detection of the plateaus of the lumbarvertebrae L1 to L4 in lateral lumbar spinal X-ray images, 4054  with minimal and simple initial manual input. The resultsare promising for computer-aided diagnosis of deformities invertebral bodies and lordotic and kyphotic spinal deformities.Further work is in progress on the derivation of quantitativemeasures of spinal deformity from the results of detectionas obtained in the present work. We also propose to conductfurther work to verify if the proposed methods could achievehigh accuracy in the thoracic spinal region.V. ACKNOWLEDGMENTSThis work was partially supported by the national councilfor scientific and technological development (CNPq) Brazil.We thank Dr. Fabio Jose Ayres for providing assistance andcomputer programs related to Gabor filters and alignment.REFERENCES [1]. National Osteoporosis Foundation Working Group on Vertebral Fractures,Assessing vertebral fractures, Journal of Bone and Mineral Research, 10:518523, 1995.[2]. H. K. Genant, C. Y. Wu, C. van Kuijk, and M. Nevitt, Vertebral fractureassessment using a semi-quantitative technique, Journal of Bone and MineralResearch, 8:11371148, 1993.[3]. T. W. ONeill, D. Felsenberg, J. Varlow, C. Cooper, J. A. Kanis, and A. J.Silman, The prevalence of vertebral deformity in European men and women: TheEuropean Osteoporosis Study, Journal of Bone and Mineral Research, 11(7):101018,1996.[4]. K. K. Jensen and L. Tougaard, A simple x-ray method for monitoring progressof osteoporosis, Lancet, 2:1920, 1981.[5]. D. Nelson, E. Peterson, B. Tilley, W. OFallon, E. Chao, B. L. Riggs, andM. Kleerekoper, Measurement of vertebral area on spine x-rays in osteoporosis:Reliability of digitizing techniques, Journal of Bone and Mineral Research, 5:707716,1990.[6]. N. E. Spencer, P. Steiger, S. R. Cummings, and H. K. Genant, Placement of points for digitizing spine films, Journal of Bone and Mineral Research, 5:S2;S247,1990.[7]. M. Jergas and R. San Valentin, Techniques for the assessment of vertebraldimensions in quantitative morphometry. In: H. K. Genant, M. Jergas, and C. vanKuijk (eds.), Vertebral Fracture in Osteoporosis, Radiology Research and EducationFoundation, San Francisco, CA, pp. 163188, 1995.[8]. J. Boisvert, F. Cheriet, X. Pennec, H. Labelle, and N. Ayache, Geometricvariability of the scoliotic spine using statistics on articulated shape models, IEEETransactions on Medical Imaging, 27(4):557568, 2008.[9]. S. Kadoury, F. Cheriet, C. Laporte, and H. Labelle, A versatile 3Dreconstruction system of the spine and pelvis for clinical assessment of spinaldeformities, Medical and Biological Engineering and Computing, 45(6):591602, 2007.[10]. M. Benjelloun and S. Mahmoudi, X-ray image segmentation for vertebralmobility analysis, International Journal of Computer Assisted Radiology and Surgery,2:371383, 2008.[11]. M. Benjelloun and S. Mahmoudi, Spine localization in X-ray images usinginterest point detection, Journal of Digital Imaging, 22(3): 309318, 2009.[12]. S. Kasai, F. Li, J. Shiraishi, Q. Li, and K. Doi, Computerized detection of vertebral compression fractures on lateral chest radiographs: Preliminary results witha tool for early detection of osteoporosis, Medical Physics, 33(12):46644674, 2006.[13]. J. Zhang, E. Lou, L. H. Le, D. L. Hill, J. V. Raso, and Y. Wang, AutomaticCobb measurement of scoliosis based on fuzzy Hough transform with vertebral shapeprior, Journal of Digital Imaging, 22(5) 463472, 2009.[14]. R. M. Rangayyan, Biomedical Image Analysis, CRC Press, Boca Raton, FL,2005.[15]. F. J. Ayres and R. M. Rangayyan, Design and performance analysis of oriented feature detectors, Journal of Electronic Imaging, 16(2):023007:112, 2007.[16]. R. C. Gonzalez and R. E. Woods, Digital Image Processing, Prentice-Hall,Upper Saddle River, NJ, 2nd Edition, 2002. 4055
Search
Tags
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks