Segmentation of vertebrae in lateral lumbar spinal X-ray images Eduardo A. Ribeiro, Marcello H. Nogueira-Barbosa, Rangaraj M. Rangayyan, Paulo M. Azevedo-Marques School of Medicine of Ribeirão Preto, University of São Paulo, Ribeirão Preto, São Paulo, Brazil Department of Electrical & Computer Engineering, University of Calgary, Calgary, Alberta, Canada 1
INTRODUCTION Vertebral fractures are important indicators of osteoporosis. Insufficiency fractures of the vertebrae are usually detected as a partial collapse of the vertebral body National Osteoporosis Foundation Working Group on Vertebral Fractures, Assessing vertebral fractures, Journal of Bone and Mineral Research, 10:518 523, 1995. 2
H. K. Genant, C. Y. Wu, C. van Kuijk, and M. Nevitt, Vertebral fracture assessment using a semi-quantitative technique, Journal of Bone and Mineral Research, 8:1137 1148, 1993. 3
INTRODUCTION Analysis of spinal and vertebral deformities could assist in the diagnostic decision-making process and in guiding therapeutic procedures. H. K. Genant, C. Y. Wu, C. van Kuijk, and M. Nevitt, Vertebral fracture assessment using a semi-quantitative technique, Journal of Bone and Mineral Research, 8:1137 1148, 1993. Both semi-quantitative and quantitative methods have been used to achieve objective and reproducible definition of the associated findings. 4
INTRODUCTION The long-term goal of the present work is to develop methods for the evaluation of vertebral deformity, including the segmentation of vertebrae, in lateral X-ray images of the lumbar spine. Segmentation is based on the detection and characterization of oriented edges using Gabor filters and classification using a neural network. F. J. Ayres and R. M. Rangayyan, Design and performance analysis of oriented feature detectors, Journal of Electronic Imaging, 16(2):023007:1 12, 2007. R. C. Gonzalez and R. E. Woods, Digital Image Processing, Prentice-Hall, Upper Saddle River, NJ, 2nd Edition, 2002. 5
METHODS Lateral lumbar spinal X-ray images were obtained at the Clinical Hospital of the Faculty of Medicine, University of São Paulo, at Ribeirão Preto, SP, Brazil. 41 cases were selected by an experienced radiologist: 19 of the images are from patients with confirmed diagnosis of insufficiency fractures 22 images are from a normal control group. images were digitized using a Vidar DiagnosticPro scanner: 84 µm/pixel; 8 bits/pixel, typically 3000 4000 pixels. 6
METHODS The present study is concentrated on L1 L4. Gold Standard : regions manually delineated by the radiologist. 7
METHODS Each image was filtered with a bank of Gabor filters (bandpass filter). Defined with the width of the oriented pattern to be detected as 4 pixels and an elongation factor of 8. The filter was rotated in steps of 1 degree to obtain a bank of 180 filters (zero to 180 degrees). For each pixel, the magnitude response and angle of the Gabor filter providing the highest output were used to compose a Gabor magnitude image and an orientation field, respectively. 8
METHODS An indication of the presence of a dominant oriented structure across several pixels was obtained by computing a measure of coherence: the normalized sum of the squared Gabor magnitude responses for all angles at a given pixel weighted by the sine and cosine of twice the corresponding angle. original image Gabor magnitude response coherence image 9
METHODS A semi-automated procedure was applied to the original image as follows: Five points were marked near the middle of the inter-vertebral spaces spanning the range of L1 L4 by using a mouse; the five points are labeled as P1 P5. The distances between the points were calculated automatically, and identified as D(1,2), D(2,3), D(3,4), and D(4,5). Using 75% of each distance measure, the corresponding line joining the manually marked points was shifted in either direction along its perpendicular to create a quadrilateral region for each vertebra. 10
METHODS The corresponding regions were obtained from the original image, the Gabor magnitude response, and the coherence image for further analysis using a neural network (logistc sigmoid). A leave-one-out procedure was repeated until to use each image as the test case once. The output of the neural network for each pixel was used to label the pixel as belonging to a horizontal vertebral plateau or not. 11
METHODS A measure of overlap was computed by comparing the output of the neural network with the corresponding manual segmentation. Receiver operating characteristic (ROC) curves were derived by varying the parameters in the neural network and the area under the ROC curve (AUC) was then obtained as a measure of detection performance. 12
RESULTS Gabor Filters Oriented fields 13
RESULTS Gabor Filters Magnitude image 14
RESULTS Gabor Filters Coherence image 15
RESULTS Average overlap: 91.7%, with SD = 1.4%. Average AUC: 0.80 with SD = 0.03. 16
CONCLUSION The results obtained for the 41 images tested indicate that the proposed methods can provide accurate results of segmentation of lumbar vertebrae plateus with minimal manual input from the user. The methods are being evaluated with a larger dataset. Further work is in progress to improve the results of segmentation* and also to derive measures related to various types of deformities from the detected vertebral bodies. * Ribeiro et.al. Detection of Vertebral Plateaus in Lateral Lumbar Spinal X-ray Images with Gabor Filters. Accepted to EMBC10-32nd Annual International Conference of the IEEE Engineering in Medicine and Biology Society. Buenos Aires, August 31 September 4. 17
ACKNOWLEDGMENTS This work was partially supported by The National Council for Scientific and Technological Development (CNPq) and by the Teaching, Research and Healthcare Foundation (FAEPA) of the Clinical Hospital at the University of São Paulo in Ribeirão Preto Brazil. We thank Dr. Fábio José Ayres for providing assistance and computer programs related to Gabor filters and coherence. 18