A morphological descriptors-based pattern recognition system for the characterization of hip osteoarthritis severity from X-ray images

A morphological descriptors-based pattern recognition system for the characterization of hip osteoarthritis severity from X-ray images
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
  Nuclear Instruments and Methods in Physics Research A 580 (2007) 1093–1096 A morphological descriptors-based pattern recognition system for thecharacterization of hip osteoarthritis severity from X-ray images I. Boniatis a , L. Costaridou a , D. Cavouras b , E. Panagiotopoulos c , G. Panayiotakis a,  a Department of Medical Physics, University of Patras, School of Medicine, 26500 Patras, Greece b Department of Medical Instrumentation Technology, Technological Educational Institute of Athens, 12210 Athens, Greece c Department of Orthopaedics, University of Patras, School of Medicine, 26500 Patras, Greece Available online 30 June 2007 Abstract A pattern recognition system is proposed for the characterization of hip osteoarthritis (OA) severity. Sixty-four (64) hips,corresponding to 32 unilateral and bilateral OA patients were studied. Employing the Kellgren and Lawrence scale, hips were groupedinto three OA severity categories: ‘‘Normal’’, ‘‘Mild/Moderate’’, and ‘‘Severe’’. Utilizing custom-developed software, 64 ROIs,corresponding to patients’ radiographic Hip Joint Spaces (HJSs), were determined on digitized radiographs. A Probabilistic NeuralNetwork classifier was designed employing morphological descriptors of the HJS-ROIs. The classifier discriminated successfully between(i) normal and OA hips (92.2% accuracy) and (ii) hips of ‘‘Mild/Moderate’’ OA and of ‘‘Severe’’ OA (91.3% accuracy). The proposedsystem could contribute in assessing hip OA severity. r 2007 Elsevier B.V. All rights reserved. PACS:  87.59.Bh; 87.57.Ra; 87.57.Nk Keywords:  Hip; Osteoarthritis; Radiograph; Morphological descriptors; Pattern recognition 1. Introduction Osteoarthritis (OA) is a degenerative joint disease,manifested by alterations of the diarthrodal joint tissues.Plain film radiography is considered as the modality of reference for the assessment of joint destruction [1]. Thecharacteristic radiographic features of hip OA include HipJoint Space (HJS) narrowing, subchondral bone sclerosis,osteophytosis, and cyst formation [2]. The evaluation of hip OA severity is mainly relied on qualitative gradingscales. The Kellgren and Lawrence (KL) grading scale [3],which is considered as the reference standard [4], definesfive OA-severity categories: ‘‘Normal (KL  ¼  0)’’, ‘‘Doubt-ful (KL  ¼  1)’’, ‘‘Mild (KL  ¼  2)’’, ‘‘Moderate (KL  ¼  3)’’,and ‘‘Severe (KL  ¼  4)’’ [3].A characteristic morphological alteration associated tohip OA is the narrowing of HJS, perceived in radiographicimages. The specific radiographic finding reflects, indir-ectly, the progressive and nonuniform articular cartilageloss [5], which differentiates both the shape and the size of radiographic HJS in osteoarthritic hips.In previous studies, HJS-width thresholds have beenintroduced for characterizing a hip as normal or osteoar-thritic [6,7]. In previous studies performed by our group,the radiographic texture of HJS has been utilized for thediscrimination among OA-severity categories, as well as forthe quantification of the severity of the disease [8–11].However, to the best of our knowledge, a computer-basedapproach for the assessment of osteoarthritic alterationsby utilizing the morphological properties of radiographicHJS has not been reported. The proposed methodrenders capable both the investigation and the quantifica-tion of parameters of radiographic HJS that cannot bemanually determined. The computerized character of thesuggested approach is accompanied by the advantagesof increased reliability and accuracy. On the other hand,the utility of the method is related to the digitizationof the radiographic images as well as on the availability of proper software. ARTICLE IN PRESS$-see front matter r 2007 Elsevier B.V. All rights reserved.doi:10.1016/j.nima.2007.06.068  Corresponding author. Tel.: +302610996113; fax: +302610996113. E-mail address: (G. Panayiotakis).  The objective of the present study was to develop apattern recognition system for the characterization of hipOA severity from pelvic radiographs. In this context,computational descriptors of the radiographic HJS mor-phology were utilized in the implementation of a classifica-tion system, designed so as to discriminate among variousgrades of hip OA severity. 2. Materials and methods  2.1. Clinical sample and radiographic imaging Sixty-four hips (18 normal, 46 osteoarthritic), corre-sponding to 32 patients with verified unilateral or bilateralhip OA were included in the study. OA diagnosis was basedon the American College of Rheumatology criteria [12].The OA verification was based on clinical and radiographicfindings. Particularly, a hip was characterized as osteoar-thritic by the presence of pain and the restricted mobilityof the joint as well as by the radiographic evidenceof osteophytes and joint space narrowing [12]. Accord-ingly, 18 patients were characterized as of unilateralOA, while 14 as of bilateral OA. A pelvic radiographwas available for each patient, while all radiographs wereobtained following a specific radiographic protocol. Inparticular, the protocol comprised the following para-meters: use of the same X-ray unit (Siemens, Polydoros 50,Erlangen, Germany), tube voltage 70–80kVp, 100cmfocus to film distance, alignment of the X-ray beam 2cmabove the pubic symphysis. Digitization of radiographswas performed employing a laser digitizer (Lumiscan 75,Lumisys, Sunnyvale, CA, USA) and following a specificdigitization protocol (4096 gray levels and 146ppi(0.17mm pixel size)). For the radiographic assessmentof OA severity, three orthopaedists employed the KL scale[3]. In particular, each orthopaedist graded OA byassigning a KL grade to the hips of the sample.Accordingly, three major OA severity categories wereformed, in which the hips of the sample were allocatedinto ‘‘Normal’’ (KL  ¼  0, 1; 18hips), ‘‘Mild-Moderate’’(KL  ¼  2, 3; 16hips), and ‘‘Severe’’ (KL  ¼  4; 30 hips).  2.2. Determination of radiographic hip joint space An algorithm implementing the adaptive wavelet trans-form [13] was applied on the digitized radiographs toenhance the image contrast and thus to emphasize thearticular margins of the hip joint.On each enhanced pelvic radiograph, two ROIs,corresponding to patient’s both HJSs, were manuallysegmented employing custom-developed software [14].The HJS-ROI was enclosed within an acute angle(see Fig. 1, LOM), which was defined by the patient’sstandard anatomical landmarks [15]. Within this specifiedregion, the articular margins of the hip joint were manuallydelineated by the orthopaedists.  2.3. Generation of morphological descriptors The following morphological descriptors [16,17] weregenerated from the segmented ROIs, in order to quantifyshape and size aspects of radiographic HJS:   Area, the number of pixels comprising the ROI.   Convex Area, the area of the Convex Hull of theHJS-ROI. For an object within a digital image, theConvex Hull is defined as the minimal convex shape thatentirely bounds the object.   Equivalent Diameter, the diameter of a circle with thesame area as the HJS-ROI.   Solidity, the ratio of the area of the HJS-ROI to theConvex Area.   Extent, the ratio of the area of the HJS-ROI to the areaof the Bounding box corresponding to the HJS-ROI.In addition, the following geometrical attributes of theellipse that has the same normalized second-order spatialcentral moments as the HJS-ROI, were used as morpho-logical descriptors:   the length of its major axis (Major Axis Length),   the length of its minor axis (Minor Axis Length),   the ratio of the distance between the foci of the ellipseand its major axis length (Eccentricity), and   the angle (in degrees) between the  x -axis and the majoraxis of the ellipse (Orientation).All the generated features were normalized to zero meanand unit standard deviation [18]. ARTICLE IN PRESS Fig. 1. Determination of the Hip Joint Space (HJS)-ROI. O: center of femoral head, L: lateral rim of the acetabulum, M: highest point of thehomolateral sacral wing. I. Boniatis et al. / Nuclear Instruments and Methods in Physics Research A 580 (2007) 1093–1096  1094   2.4. Computer-based classification of hips A classification scheme based on the Probabilistic NeuralNetwork (PNN) classifier [19] was designed for thediscrimination between: (i) normal and osteoarthritic hipsand (ii) hips of ‘‘Mild/Moderate’’ OA and of ‘‘Severe’’ OA.The PNN classifier is a fast-training nonparametricclassification approach that does not require Gaussianforms for the probability density functions of the patternvectors forming a class. The performance of the classifiercan be adjusted by a parameter, labeled as ‘‘sigma ( s )’’,which takes values ranging between 0 and 1. The selectionof sigma affects the estimation error of the PNN and isexperimentally determined by comparing the classificationaccuracies obtained for different values of the parameter[18,19].In order to determine the feature combination providingthe highest classification accuracy with the minimumnumber of features (‘‘optimum’’ or ‘‘best’’ feature combi-nation) the ‘‘exhaustive search’’ procedure was followed inconjunction with the Leave One Out classification errorestimation method. Classifier performance was expressedin terms of overall accuracy [18].  2.5. Statistical analysis The Student’s  t -test was used for the investigation of theexistence of statistically significant differences (  p o 0.05)between normal and osteoarthritic hips for the morpholo-gical feature values. The normality of distributions for thegenerated features was assessed by means of the Lillieforstest [20]. The Coefficient of Variation (CV) was used inorder to assess the reproducibility of the HJS-ROIdetermination process. In particular, each of the orthopae-dists segmented the HJS-ROIs twice, while the evaluationscores were employed for the calculation of the CV. 3. Results and discussion The degenerative process of OA causes the progressiveand nonuniform loss of articular cartilage [5]. In aradiographic image, this loss is indicated by the narrowingof HJS [2,5], which induces alterations in the morphologyof the specific anatomical region. Thus, the shape and thesize of radiographic HJS in osteoarthritic hips are expectedto differ in comparison to normal ones. This differentiationwas verified by the results of statistical analysis of thepresent study, which revealed the existence of statisticallysignificant differences (  p o 0.001) between normal andosteoarthritic hips for the morphological feature values.The latter were found to follow a Gaussian distribution.The segmentation of radiographic HJS-ROI was found tobe reproducible. Regarding the intra-observer reproduci-bility, the CV was found equal to 3.4%, on average,indicating the reliability of the segmentation process. Inter-observer reproducibility was also high, since the corre-sponding value for the CV was 4.2%, on average.Regarding the discrimination between normal andosteoarthritic hips, the highest classification accuracyachieved was 92.2%. The PNN discriminated successfully59 out of 64 hips (see Table 1), employing the optimumfeature combination comprising the features  Minor AxisLength  and  Extent . After multiple trials, the sigma valuewas determined to be equal to 0.3.The PNN classifier was also employed for the character-ization of osteoarthritic hips as of ‘‘Mild/Moderate’’ OA orof ‘‘Severe’’ OA. The highest overall accuracy achieved forthe specific discrimination task was 91.3%, since 42 out of 46hips were assigned to the correct categories by the PNN(see Table 2). The optimum feature combination comprisedthe features  Orientation ,  Equivalent Diameter ,  Solidity , and Extent , while the sigma value was 0.1.In conclusion, structural alteration of the hip joint,associated to OA, can be reliably assessed by computa-tional descriptors of the radiographic HJS morphology.The proposed pattern recognition system discriminatedefficiently normal from osteoarthritis hips while gradedreliably the severity of OA in osteoarthritic hips. Thesuggested approach could be utilized as a decision supporttool for the assessment of hip OA severity. Acknowledgments I. Boniatis was supported by a grant from the StateScholarship Foundation (SSF), Greece. The authors thankthe staff of the Departments of Orthopaedics and Radi-ology for their contribution to this work. References [1] P. Garnero, P. Delmas, Curr. Opin. Rheumatol. 15 (2003) 641.[2] R.D. Altman, J.F. Fries, D.A. Bloch, et al., Arth. Rheum. 30 (1987)1214.[3] J.H. Kellgren, J.S. Lawrence, Ann. Rheum. Dis. 16 (1957) 494.[4] T.M. Link, L.S. Steinbach, S. Ghosh, et al., Radiology 226 (2003) 373.[5] P.A. Ory, Best. Pract. Clin. Res. 17 (2003) 495. ARTICLE IN PRESS Table 1Truth table for the discrimination between normal and osteoarthritic hipsHip Normal Osteoarthritic Accuracy (%)Normal 17 1 94.4Osteoarthritic 4 42 91.3Overall accuracy 92.2Table 2Truth table for the discrimination between hips of ‘‘Mild/Moderate’’osteoarthritis and of ‘‘Severe’’ osteoarthritisSeverity category Mild/Moderate Severe Accuracy (%)Mild/Moderate 14 2 87.5Severe 2 28 93.3Overall accuracy 91.3 I. Boniatis et al. / Nuclear Instruments and Methods in Physics Research A 580 (2007) 1093–1096   1095  [6] P. Croft, C. Cooper, C. Wickham, D. Coggon, Am. J. Epidemiol. 132(1990) 514.[7] T. Ingvarsson, G. Ha ¨gglund, H. Lindberg H, S.L. Lohmander, Ann.Rheum. Dis. 59 (2000) 650.[8] I. Boniatis, L. Costaridou, D. Cavouras, E. Panagiotopoulos,G. Panayiotakis, Br. J. Radiol. 79 (2006) 232.[9] I. Boniatis, L. Costaridou, D. Cavouras, E. Panagiotopoulos,G. Panayiotakis, Nucl. Instr. and Meth. A 569 (2006) 610.[10] I. Boniatis, L. Costaridou, D. Cavouras, I. Kalatzis, E. Panagioto-poulos, G. Panayiotakis, Med. Bio Eng. Comput. 44 (2006) 793.[11] I. Boniatis, L. Costaridou, D. Cavouras, I. Kalatzis, E. Panagioto-poulos, G. Panayiotakis, Med. Eng. Phys. 29 (2007) 227.[12] R. Altman, G. Alarco ´n, D. Appelrouth, et al., Arth. Rheum. 34(1991) 505.[13] P. Sakellaropoulos, L. Costaridou, G. Panayiotakis, Phys. Med. Biol.48 (2003) 787.[14] P. Sakellaropoulos, L. Costaridou, G. Panayiotakis, Med. Inform.Internet Med. 24 (1999) 53.[15] T. Conrozier, A.M. Tron, J.C. Balblanc, et al., Rev. Rhum. Engl. Ed.60 (1993) 105.[16] L.G. Shapiro, G.C. Stockman, Computer Vision, Prentice-Hall,Upper Saddle River, NJ, 2001.[17] J.C. Russ, The Image Processing Handbook, third ed, CRC PressLLC and Springer, Florida and Heidelberg, 1999.[18] S. Theodorides, K. Koutroumbas, Pattern Recognition, second ed,Elsevier Academic Press, Heidelberg, 2003.[19] D.F. Specht, Neural. Networks 3 (1990) 109.[20] H.W. Lilliefors, J. Am. Stat. Assoc. 62 (1967) 399. ARTICLE IN PRESS I. Boniatis et al. / Nuclear Instruments and Methods in Physics Research A 580 (2007) 1093–1096  1096
Similar documents
View more...
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