Osteoporosis Presence Verification Using MACE Filter Based Statistical Models of Appearance with Application to Cervical X-ray Images. M. Aouache1, A.
Osteoporosis Presence Verification Using MACE Filter Based Statistical Models of Appearance with Application to Cervical X-ray Images M. Aouache1, A. Hussain1, S.A. Samad1, A.H. Hamzaini2, A.K. Ariffin3 1
National University Malaysia , Electrical, Electronic & Systems Engineering Dept, , Bangi 43600, Malaysia 2 National University Malaysia, Radiology Dept, Faculty of Medicine, , Bangi 43600, Malaysia 3 National University Malaysia, Mechanical and Materials Engineering Dept, Bangi 43600, Malaysia
Abstract — vertebral fracture is a very common outcome of osteoporosis, which is one of the major public health concerns in the world. Early detection of vertebral fractures is important because timely pharmacologic intervention can reduce the risk of subsequent additional fractures. Our goal seeks to develop a computerized method for detection of vertebral fractures by measuring the shape and appearance of vertebrae on cervical xray radiographs in order to assist radiologist’s image interpretation and thus allow the early diagnosis of osteoporosis. The statistical models of shape and appearance are powerful tools for interpreting medical images. This work introduces the application of correlation filter classifiers for identification and verification of the osteoporosis presence in cervical vertebrae training/testing set. Correlation filter classifiers have been previously applied to other biometric classification tasks, but not to classification of cervical vertebrae images. We describe how the extraction of an appropriate region of interest in the cervical vertebrae surface can be used to design correlation filters that accomplish 90 % recognition on a database of 50 cervical bone shapes. Keywords — x-ray radiographs, Segmentation, ASM modeling, Correlation filter, vertebral deformity, Osteoporosis.
I. INTRODUCTION Medical imaging allows scientists and physicians to decide about life saving actions with regard to the human physiological physical activities. Medical images playing an important role to detect and functional information of the body part for diagnosis, medical research and education [1]. A key feature of spine disease manifested in these images is the presence of osteophytes which are bony processes that alter the shape of vertebrae [2]. The segmentation of vertebral bodies in such spine X-ray images is of great interest to bone morphometrists and radiologists since they exhibit pathologies such as anterior osteophytes (AO), disc space narrowing (DSN) [3], subluxation and spondylolisthesis that are detectable consistently and reliably by vertebra boundary shape. Other pathologies such as vertebral fractures, ossification of posterior longitudinal ligament(OPLL), spinal stenos is caused by posterior osteophytes, tumors, and osteoporosis may also be detected from this dataset. The feasibility of com-
puter assisted techniques for the segmentation of vertebral bodies in spine x-ray images has been of great interest [1,2] to biomedical researchers, in particular the osteoarthritis research community. Reliable extraction of vertebrae boundaries is a prerequisite for subsequent pathology validation and Content-Based Image Retrieval (CBIR) research. However, fully automated segmentation of spine X-ray images is a very challenging problem. Many published segmentation algorithms with results using collection of spine X-ray images [4, 5, 6] depend on user intervention or some initial guess to achieve satisfactory performance. The general problem of developing algorithms for the automated indexing by structure contents is a significant challenge [7], where the structures of interest are commonly irregular. Our eventual goals of the project are to: 1-Acquire and annotate a large database of imvages of the spine. 2-Develop algorithms which can automatically locate and measure all the vertebrae in the image. 3-Develop new techniques for classifying vertebral fracture using the full shape and texture.4 Produce a tool that would be used regularly by clinicians. 5Develop a fully automatic mode for use in large-scale clinical trials and epidemiological studies II. AVFAS SYSTEM PROTOTYPE The AVFAS prototype system is short for Automatic Vertebral Fracture Assessment System. In this current system, the user manipulates GUI tools to provide one of the three main interface systems .First, the Medical Image Measurement & Decision System (MIMDS) for the task of pre-process x-ray image including the selection and the enhancement of the region of interest and search for the shape in image to fit the model selected. Second, the Medical Image Training & Verification System (MITVS) for the role to build and train the model based on active shape modeling and verify the similarity to the shape in the image selected. Finally, the Medical Image Registration System (MIRS) record the information about the image, database source, coordinate system and origin.
N.A. Abu Osman, F. Ibrahim, W.A.B. Wan Abas, H.S. Abd Rahman, H.N. Ting (Eds.): Biomed 2008, Proceedings 21, pp. 607–610, 2008 www.springerlink.com © Springer-Verlag Berlin Heidelberg 2008
608
M. Aouache, A. Hussain, S.A. Samad, A.H. Hamzaini, A.K. Ariffin
III. FRACTURE
PATTERNS
A. 9-point Distribution Model The model that will be used to describe a shape and its typical appearances is based on the variations of the spatial position of each landmark point within the training set. Each point will thus have a certain distribution in the image space and therefore the shape model is being referred to as a Point Distribution Model (PDM). In order to obtain the PDM, we first need to determine and label the landmarks, to align the shapes, and finally, to summarize the landmark variations in a compact form. These steps are being described in some detail as shown in Figure 1. Points(8,9) indicate the existence of anterior osteophytes for normal vertebrae, points 8 and 9 will coincide with points (3,6)respectively. The semantic relevance of the 9 points in the sagittal view is as follows: Points (1,4) mark the upper and lower posterior concerns of the vertebra respectively. The Points(3,6) marks the upper and lower anterior concerns of vertebra respectively. The Points (2,5) are the medians along the upper and lower vertebra edges respectively. Point 7 is the median along the anterior vertical edge of the vertebra Points (8,9) indicate the presence of the upper and lower anterior osteophytes respectively .These points are typically marked at the osteophytes extremities B. Object Shape Representation An object shape is represented by a set of labeled points or landmarks. The number of landmarks should be large enough to show the overall shape, and the details where it is needed as showing below
Fig. 1 Radiologist market 9 point model
Fig. 2 Cervical
x-ray image segmented vertebra and a localized view showing inferior AO on vertebra with 36 boundary points
_________________________________________
Figure 2 illustrate an example of applying segmentation technique to detect osteophytes. The boundary points of vertebra shape helps to determine the existence of superior and inferior anterior osteophytes (AO), which are pathologies along the anterior superior or inferior edges of the vertebral bodies. Points(2,3,8,7) describe superior anterior osteophytes on a sagittal spinal X-ray and points(7,9,6,5) describe an inferior anterior osteophytes. IV. MODELING SHAPE VARIATION A. Active Shape Model (ASM) An ASM [8, 9] is a statistical model that describes "what an object looks like" in terms of its shape and its imaging appearance. An ASM is created by "training" it with sample images on which the boundaries of the objects of interest have been annotated by an expert. After the ASM is trained, it can be used to locate the objects on a new image by matching the model, which describes the expected shape and appearance, to findings on the new image. An ASM is a type of deformable template model that does not just describe a single fixed object shape and appearance but also describes the ways in which the objects were observed to vary in shape and appearance over the set of training images. ASMs have an important advantage over other methods for locating objects on images because they are specific to the type of object under study and describe only the variation observed in the training examples and do not allow "illegal" variation. ASMs have been applied successfully to a range of medical image interpretation problems in two and three dimensions [8,10]. An ASM contains two separate components that describe the object shape and appearance. Object shape is described by means of a point distribution model (PDM), which is generated by performing statistical analysis of the object shapes observed over the set of training images. The contour around the structure on each training image is described with a set of n landmark points, which are manually annotated by an expert. Each contour can then be described with a vector, x = {x1, y1, x2, y2, . . ., xn, yn}, where (xi, yi) is the position of the ith landmark point on the contour. The training contours are aligned as closely as possible by means of scaling, rotation, and translation. Then, to describe the main independent ways in which the training shapes vary, principal component analysis [11] is performed by using the deviations of each training shape vector from the mean shape vector . The PDM represents shape in terms of a mean shape and a set of linearly independent modes of shape variation that describe the variation over the training set. Use of only the most common modes of shape variation is required to describe members of the training set to a chosen level of accuracy.
IFMBE Proceedings Vol. 21
___________________________________________
Osteoporosis Presence Verification Using MACE Filter Based Statistical Models of Appearance with Application...
Any new shape, x, of the same type as those observed in training can then be generated by adding combinations of the modes of variation contained in a matrix P to the mean shape, , with a vector of weights b controlling the influence of each mode: x = + Pb. A PDM of the cervical vertebra represented with 9 landmark points is shown in Figure 1. This shows the mean shape of the vertebra and the three most common modes of variation, or ways in which its shape varied from the examples with which the model was trained. B. Training Set In order to build a model that is flexible enough to cover the most typical variations of an object, a sufficiently large training set has to be used ,the segmented object is stored with unique object using the active shape mode (ASM) Figure 3 and 4 shows the main task steps modeling
Correlation filters have been applied successfully to automatic target recognition (ATR) [13] problems. The most basic correlation filter is the matched svpatial filter (MSF), whose impulse response (in 2-D, point spread function) is the flipped version of the reference image. While the MSF performs well at detecting a reference image corrupted by additive white noise, it performs poorly when the reference im-
Testing and Verification
Weight Adjustment
Aligned Training set
Unaligned Training set
Labeling Image: 1 - LM: 10(+1)
ASM: unaligined training set
750
ASM: aligined training set 740
740
720
720
700
700
680
680
660
660
640
640
620
620
600
600
ASM: Shape resulting from image search and original image
ASM: test shape with b = [0]
100
700
200
300
650
400
500 Labeling Image: 2 - LM: 10(+1)
600
600
700
550 800
Correlation
as if we started in HI res
1
FFT
0.8
IFFT
0.6
0.4
0.2
0 30 25
30
20
25
15 10
10
5 0
Correlation Filter
15
20
5 0
Training Images as i f we started in HI res
as if we st arted in HI res
as if we started in HI res
Filter Design
Fig. 5 Mace filter block diagram age appears with distortions (e.g., rotations, scale changes). Figure 5 shows schematically how the cross-correlation is implemented efficiently using (FFTs). The correlation output is searched for peaks, and the relative heights of these peaks are used to determine whether the object of interest is present or not.
Mahalanobis et al [13] developed the minimum average correlation energy (MACE) filter. The Resulting MACE filter equation HMace= D-1 X (X + D-1 X)-1 c
(1)
We perform 2-D FFTs on these training images and convert the 2-D FFT arrays into 1-D column vectors by lexicographic ordering. These vectors are the column vectors of the d×N matrix X in Eq. (1). Column vector c with N elements contains the prespecified correlation peak values of the training images and the dddiagonal matrix D contains along its diagonal the average power spectrum of the training images. Note that the synthesized HMACE is a column vector with d elements and the 2-D correlation filter is obtained by reordering the column vector back to a 2-D array.
900
500
450 350
Correlation
Test Images
A. Minimum average correlation energy (MACE) filter
V. MATCHING BASED CORELATION FILTER
Shape Set Collectio
609
400
450
500
550
580 600 420
440
460
480
500
520
540
560
580 420 580
100
440
460
480
500
520
540
560
200
300
400
500
600
700
800
580
B. Osteoporosis Presence Based Shape Verification
Fig. 3 Active shape modeling main task steps
A single MACE filter was synthesized for each of the 50 persons x-ray images using a 7 of training images from that person. In the test stage, for each filter, we performed cross correlations with all the vertebrae images from all the cervical x-ray images . For authentic, the correlation output should be sharply peaked and it should not exhibit such strong peaks for impostors. Fig.6 shows peak to sidelobe ratio (PSR) defined below is used to measure the peak sharpness .
PSR Fig. 4 Training stage, a) initialization parameter, b) Obtain landmark coordinates for each shape in the training set ,c) calculating the weight d) Result: unaligned training set shape coordinates
_________________________________________
peak mean
V
(2)
Figure 7 (a) shows a typical correlation output for an authentic normal vertebrae . Note the sharp correlation peak resulting in a large PSR value of 49.
IFMBE Proceedings Vol. 21
___________________________________________
610
M. Aouache, A. Hussain, S.A. Samad, A.H. Hamzaini, A.K. Ariffin
abnormal vertebrae using a universal threshold. We are currently improving the filter design methods and testing the correlation filters on the much larger (NHANES II) database collected.
ACKNOWLEDGMENT
Fig. 6 shows how the peak to sidelobe Ratio (PSR) is estimated.
The Authors like to thanks The NHANES II for their work on the original version of this document. The authors also we would like to acknowledge the financial support of the National University of Malaysia
REFERENCES Fig. 7 Correlation outputs when using a MACE filter designed for Shape . (a): Input is a shape image belonging to Normal Vertebrae . (b): Input is a Shape image not belonging to Normal Vertebrae
1. 2.
The bottom correlation output in Figure 7(b) shows a typical response to an impostor abnormal vertebrae exhibiting low PSRs (