Proceedings of the 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference Shanghai, China, September 1-4, 2005
An Upper-Limb-Movement Classification System of Cerebral Palsy Children Based on Arm Motion Detection Jiann-Der Lee, Kai-Wei Wang, Li-Chang Liu, and Ching-Yi Wu
Abstract—The researches about upper limb palsy patients are the minority areas among the researches about cerebral palsy (CP) patients. This paper presents an upper-limb-movement classification system of cerebral palsy children based on their arm motion information to judge their impairment degree. The system contains three parts: image capture, image segmentation, and information classification processing. Momentum analysis parameters and coordination neural network are used to conduct the data classification. The experimental results are shown that the proposed system has the higher accurate rate of tracking compared with Normalized Cross Correlation and Otsu methods, and the patients are divided into the slight impairment grade or the serious impairment grade.
I. INTRODUCTION
T
PPV: Percentage of Peak Velocity. The better value is above 50%. PPV Frame index at a maximum speed Frame total number in film
MU: Movement Unit. A MU is based on the acceleration and deceleration; the smaller value indicates the better performance. In the proposed system, the actual evaluation parameter only uses NTD, PPV, and MU because MT information is not the major factor. The details of the proposed system are introduced in the second chapter and the experimental classification results of the impairment degree are shown in the third chapter.
HE intact momentum to cerebral palsy patients is one way to analyze neural actives from the movements controlled by the central neural system. In clinical practice, the slight changes of patient’s movements infer the progressive situations. One of the conventional analysis software of human movements is VICON system which utilizes a light-reflection ball attached to the testee’s body to collect movement information. The testee is required to complete some assigned movements so that the testee’s critical position is captured via the parameter analysis [3]. Even thought VICON system requires huge data storage space, a novel movement analysis system is proposed with the aid of VICON system. In the momentum analysis [6][7], the four momentum analysis parameter, including MT (movement time), NTD (normalized total of displacement), PPV (percent of movement where peak velocity occurs), and MU (the number of units) as defined below, are suggested.
II.
SYSTEM FRAMEWORK
The flowchart of the proposed system is shown in Fig. 1. This proposed system consists of three parts: image capture processing, image segmentation processing and data classification/judgment processing.
MT: Movement Time is the time from the beginning of the movement to the end. NTD: Normalized Movement Path. The ideal value is 1. NTD
Actual movement path standard path calculated by computer
This work was partially supported by National Science Council, R.O.C. and Ministry of Economic Affairs, R.O.C. under Grant NSC93-2212-E182-001 and 94-EC-17-A-19-S1-035, respectively. Jiann-Der Lee is with the Department of Electrical Engineering, Chang-Gung University, Tao-Yuan, Taiwan 333, R.O.C. (corresponding author to provide e-mail:
[email protected]). Kai-Wei Wang and Li-Chang Liu are with Chang-Gung University, Tao-Yuan, Taiwan 333, R.O.C. Ching-Yi Wu is with the department of Occupational Therapy, Chang Gung University, Tao-Yuan, Taiwan. (e-mail:
[email protected]).
0-7803-8740-6/05/$20.00 ©2005 IEEE.
Fig. 1. The flowchart of the proposed system.
A. Image Capture The man-machine interface is based on C++ codes. An AVI file with 30 frames per second is imported and each frame is interpolated and converted into a small BMP format with the resolution 180x120 to reduce the computation time
6878
instead of the original resolution 720x480. The bilinear interpolation function using in (2), with the aid of the triangle line function W(XC) defined in (1), is close to a SINC function [9]. °1 XC W XC # ® °¯0
I X P ,YP
0 d XC d1
(1)
According to the types of data, the supervision sheet with self-feedback coordination neural network is used in the data classification as shown in Fig. 2 where di is a relative expectation value, Ș is the learning rate, X1, X2, and X3 are the input parameters, f= W1*X1+W2*X2+W3*X3, and W(t+1)=W(t)- ӯ*W(t)*di.
1 XC
X1 X2 X3
i 1 j 1
¦¦WY m, n I m, n W X m, n C
C
m in j
1 dy 1 dy I i, j dx 1 dy I i, j 1 !!!! (2)! 1 dx dy I i 1, j dx dy I i 1, j 1
B. Image Segmentation The color RGB image is sensitive to light changes with an inaccurate setting valve value. To avoid this difficulty, the linear YIQ color space is utilized to get rid of the influence over brightness on the tracked object [4] instead of a RGB domain. ªY º «I » « » ¬«Q ¼»
ª 0.299 0.587 0.114º ª R º « 0.596 - 0.275 - 0.321» «G » »« » « ¬« 0.212 - 0.523 0.313 ¼» ¬« B ¼»
(3)
In (3), Y indicates brightness, and only I and Q in an image are used in the binary-handling processing. After the binary-handling processing, there are some miscellaneous spots in an image. To remove such undesired spots [4], an opening morphology operator with the functions of erosion, dilation, and shrink is performed. C. Data Classification To seek a central point of a partitioned object, the triangle gravity method is used. First, an image is partitioned and the object is given 1 as the effective points. After the central point of the object is found, the momentum analysis parameter values are calculated. To have the parameter value in a certain scope, all initial parameter values are set in the range between 0 and 1. Before the data are sorted out, the features and correlations of parameters are computed. In the parameter calculation part, MT is computed, PPV is indirectly demonstrative of MT in physics content, and NTD and MU have no relation with MT. Consequently, MT is not considered in the data classification processing. The correlations of parameters are shown as in Table I where P indicates conspicuousness. When P is greater than 0.05, the defined relation doesn’t exist. TABLE I PARAMETER CORRELATION NTD PPV MU MU P=0.447 P=0.015 P=0.087
a( f )
¦
yi
' wi ˘̅̅̂̅
di
K Fig. 2. Supervision sheet with self-feedback coordination neural network.
III.
EXPERIMENTAL RESULTS
A. System Test The successive 50 frames of cerebral palsy patients are tested. In the comparison of the execution speed and loss rate with the Normalized Cross Correlation (NCC) [2] and the Otsu [1], the results are shown in Table II. This system has the best performance in speed and the loss rate. This system, therefore, is the best choice to track the patients’ mevements. TABLE II THE SYSTEM COMPARISON Execution speed Loss rate NCC in [2] 23.217S 20% Otsu in [1] 4.042S 8% Ours 3.756S 0%
B. Data Analysis Cerebral palsy is classified as 8 grades of the exercising ability and is grouped as the muscle impairment group and the nerve impairment group. Muscle impairment contains five types, and nerve impairment contains three types. Grade 1 indicates the most serious degree, and Grade 8 indicates the slightest degree [5][8]. The input images of the experiments are captured from those patients whose intelligence quotient is more than 60 and muscle are repaired in the convulsion type to be capable of hand grasping. Ten cerebral palsy children aged from five to twelve were tracked in this experiment. Five of them were limbs paralysis, and the other five are hemiplegia. In the intact experiments, patients are stipulated to use his non-advantage hand, which is defined as the most seriously symptomatic hand from the same patient, to finish assigned movements as shown in Fig. 3. The man-machine interface of the proposed system is shown in Fig. 4. The analysis parameters of ten tested objects are shown in Table III. With consistency, PPV values are taken as a reciprocal. The cerebral palsy is categorized as muscle impairment and nerve impairment, or is based on the
6879
movement ability. Therefore, when the patients are diagnosed as quadriplegia (Q), it is not necessary to be worse than others in the grade of the movement ability. On the contrary, hemiplegia (H) is not necessarily better in terms of movement grade. In Table III, CP1 patient is in the state of quadriplegia; however, the experimental data are shown that he’s better than CP5 patient who is in the state of hemiplegia. According to the experimental results, quadriplegia and hemiplegia are not necessary to be favorably correlated in terms of experimental data. Consequently, we defined that the cluster occurrence is based on the flowing grade (fluidity) of measurement of tested object’s non-advantage hand. Meanwhile, patients can be grouped into two grades according to the experimental results where CP5, CP8, and CP4 patients are classified as Grade B and others are Grade A. Grade A and Grade are defined below.
Fig. 4. The system man-machine interface of upper limb movements. TABLE III THE EXPERIMENTAL RESULTS OF 10 PATIENTS
Grade A: represents patient’s non-advantage hand with the better fluidity. Grade B: represents patient’s non-advantage hand with the worse fluidity. In the grade analysis of experimental results, NTD and MU have more consist contributions to the grade results rather than PPV such as that the patients whose NTD values are below 1.8 and MU values below 4 belong to Grade A.
NO
NTD
PPV
MU
TYPE
GRADE
CP1
1.193
1.74999
3
Q
A
CP2
1.528
2
3
H
A
CP3
1.406
1.0277
1
Q
A
CP6
1.705
1.55554
3
Q
A
CP7
1
1.18132
4
H
A
CP9
1.35
1.37036
3
H
A
CP10
1.304
1.6098
3
Q
A
CP5
1.896
3.86383
7
H
B
CP8
2.233
1.26087
4
Q
B
CP4
2.433
4.19041
4
H
B
IV.
Fig. 3. The examples of intact experiments.
!
!
CONCLUSION
The proposed system is capable to track the movements of a cerebral palsy children’s non-advantage hand and judge their impairment degree by the momentum analysis parameters. This system adopts the off-line method to gather information, and afterward, the system goes on-line with two cameras to increase the accuracy in calculating parameters. In the current experimental results, MU affects the classification results more than NTD and PPV because each parameter is given the same weighted value in the classification process and MU has much larger values rather than NTD and PPV values. This observation implies that a non-linear weighted classification method should be considered in the future to analyze impairment degrees of cerebral palsy children and normal children. In the future study, more momentum analysis parameters and test samples will improve the system preciseness and more progress grades.
6880
REFERENCES [1] [2]
[3] [4] [5] [6] [7] [8] [9]
N. Otsu , “A Thresholding Selection Method from Gray-scale Histogram” IEEE Transaction on system Man Cybemet, vol. SMC-8, pp. 62-66, 1978. Friemel, B.H.; Boho, U.N.; “Relative performance of two-dimensional speckle tracking techniquesʳ : normalized correlation, non-normalized correlation and sum-absolute-difference” IEEE Confs. vol.2., pp 1481-1484, 1995. Y.L.Lay, R.S.Chang, P.W.Chen, D.C.chern, “CCD Grating Generated Moire Pattern for Close-Range Measurement” *Photonics and Optoelectronics 3(4) (1995)131-138. R. Gonzalez and R. Wood , Digital Image Processing, section edition, 2002. A. Utley, B. Steenbergen, D.A. Sugden. The influence of object size on discrete bimanual co-ordination in children with hemiplegic, [Disabluty and Rehabil-Ation] vol. 26, pp. 603-613, 2004. Chia-Huao Lu, The upper limb functional effect on weight treatment on different positions of arms and legs of parents with intentional trembling: momentum studies, Chang Gung University, 2004. Shu-hsin Ou, Cerebral palsy children’s movement effect on botulinus injection treatment and activity situation: momentum analysis, Chang Gung University, 2003. Yi-Pei Wang, in a mixed situation, studies on cerebral palsy children’s physical education teaching achievement, Kaohsiung Normal University, 2004. Huang Chin-Tsun, Shih Tien-Yuan, fragmentation-based Interpolation and bilinear, comparison of cubic circling interpolation, Central University.
6881