Analysis of Patterns in Tremor Diagn Source Biomed Tech Berl SO 2013 Sep 7[PMIDT24042940]

ANALYSIS OF PATTERNS IN TREMOR DIAGNOSIS SPIRAL DRA WINGS FOR AUTOMATED CLASSIFICATION

Andreas Wille1,Mohamed Sangaré1and Susanne Winter1

1Institut für Neuroinformatik,Ruhr-Universit?t Bochum,Germany

andreas.wille@ini.rub.de

Abstract:Tremor analysis is a common task in diagnosis of movement disorders.In this paper,we present a new method for an automated classi?cation of tremor strength based on digitized spiral drawings.We used statistical measurements on relative line orientations,which are calculated on the full image.Therefore,this approach is insensitive to draw-ing errors(like line crossings or missing line parts).On a test set of109clinical spiral drawings the difference within the calculated values for different levels of tremor strength was highly signi?cant and allowed a fast and straightfor-ward classi?cation.

Keywords:Tremor,spiral drawing,classi?cation,orienta-tion histogram

Introduction

Tremor is an unintended,rhythmic,oscillatory muscle movement in one or more parts of the body.A light tremor is physiologically normal for humans.This tremor can be-come visible under certain conditions,e.g.stress,anxiety or excitement.Other reasons for tremor are diseases like the Parkinson disease or affections of the cerebellum.As tremor is the most common movement disorder a differen-tiated classi?cation of tremor and tremor strength is neces-sary.A common practice is the individual classi?cation of patient movements in relation to some standard by an ex-perienced neurologist.Another method is the comparison between spiral drawings and standardized spiral images de-picting effects of ten different levels of tremor[1].As the intra-and interindividual variation in this classi?cation is high,an automated evaluation of the drawings could help to perform the evaluation more objectively and allow a precise monitoring of tremor progress.

Kraus and Hoffmann[2]developed a method to automati-cally analyze the tremor amplitude in spiral drawings.How-ever,in this method problems occur whenever it is not pos-sible to?t an ideal spiral-curve to the drawing due to cross-ings or gaps in the drawn line.

To overcome these drawbacks we already examined some measures of two-dimensional histograms[3]and found them to be quite helpful and suitable for classi?cation of tremor strength.Yet,this method still requires visual in-spection of the two-dimensional histograms,which proved to be dif?cult without some experience.

As a logical next step,we present a method to quantify the differences between tremor strengths in spiral drawings by computing a value which correlates to the variation of

the

Figure1:Spiral image after preprocessing relative orientation of the drawn line with respect to the spi-ral center.

Methods

For image creation we follow the standard procedure ac-cording to Bain and Findley[1],i.e.the patient draws a spiral on a piece of paper by following an ideal image of Archimedes’Spiral(which is by some means projected on the paper)as close as possible.Afterwards the resulting im-age is digitized and the drawing area is extracted.To avoid miscounts the image must not contain any additional infor-mation except the spiral drawing.An example for such an image is shown in Fig.1.By application of a Sobel?lter for both horizontal and vertical direction the gradient and it’s orientationφare computed for every foreground image pixel I f(see Eq.1).

φ(x,y)=arctan

?y I f(x,y)

?x I f(x,y)

(1) In this case Sobel?lters of size7×7were used to achieve an appropriate level of detail in the orientation.Addition-ally,the angleαbetween each pixel and the image center (which is equivalent to the center of the spiral)relative to a horizontal line through the center is calculated.A subtrac-tion of these angles according to Eq.2results in a measure for the relative orientationθof all lines.

θ=φ?α(2) After the relative orientation is computed for all fore-ground pixels,mean and standard deviation are calculated. Alternatively the acquired set is visualized in form of a histogram using one bin for every ten degrees.

Biomed Tech 2013; 58 (Suppl. 1) ? 2013 by Walter de Gruyter · Berlin · Boston. DOI 10.1515/bmt-2013-4278

(a)

slight(b)

medium(c)

strong(d)extreme

Figure2:Histograms of the relative line orientation(in bins of10?width)in tremor spiral images for different levels of tremor strength.The data is normalized with respect to the number of foreground pixels in each drawing.

Evaluation data set

For the evaluation109drawings from different patients with varying levels of tremor were available.They were man-ually divided in four groups(slight,medium,strong and extreme tremor).Information about this reference classi?-cation can be found in Tab.1.The described method was performed for every group and the results were compared statistically.

Results

For all four tremor categories the histograms look very much alike within a class and show easily recognizable dif-ferences between classes.Representing examples of all cat-egories are demonstrated in Fig.2.Statistical values for a more detailed analysis are given in Tab.1.For all but the most extreme cases of tremor the distributions of the rela-tive orientation are centered at approximately0?.This cor-responds to gradients pointing directly away from the image center and is typical for centered circular drawings.

The standard deviation,on the other hand,varies strongly when comparing two classes.These differences are highly signi?cant(p<0.001,Welch-Test).Assuming normal distri-butions we applied the Bayes decision rule to create class boundaries minimizing the number of errors in categoriz-ing the images.The numbers given in Tab.2are optimal for the considered training set and resulted in a total of24 misclassi?cations.

Table1:Statistical results for tremor examination data. Classes correspond to slight(1),medium(2),strong(3)and extreme(4)cases.Mean and standard deviation of relative orientation are given in degree.

Class

1234 number of examples5035168 mean of distribution2±12±13±11±4 standard deviation15±119±228±946±7 Discussion

The mean of the relative orientation is almost unaffected by tremor as it’s effects are known to be dependent of the Table2:Class boundaries on standard deviation(in degree) of relative orientation and classi?cation errors(false posi-tive/false negatives)for different levels of tremor

Class Boundaries Errors

1(slight tremor)<16.96/2

2(medium tremor)16.9-22.69/9

3(strong tremor)22.6-36.94/12

4(extreme tremor)>36.95/1 movement directions and thus only certain parts of the spi-ral drawing will be subject to it.While these effects are mostly canceled out by calculating the mean,they still are visible in the standard deviation,because deviations from the ideal drawing path,both towards the center as well as away from it,will add up.

A simple automated classi?cation was similar to the manual ordering,but still had an error rate of22%.A more elabo-rated classi?cation method might be able to improve this.In addition,a larger study to collect more training data should be very bene?cial.

However,our method already provides a proof of concept suitable for a quick check up.Especially,when consider-ing that no more than a standardized spiral drawing paper, a scanner and a usual PC of low computing power are re-quired to gain an objective result in less than2minutes. Acknowledgement

We would like to thank PD Dr.P.H.Kraus of the St.Josef-Hospital in Bochum for providing the spiral images. Bibliography

[1]P.Bain and L.Findley,Assessing Tremor Severity:A

Clinical Handbook.Smith-Gordon,London,1993. [2]P.Kraus and A.Hoffmann,“Spiralometry:Comput-

erized assessment of tremor amplitude on the basis of spiral drawing,”Movement Disorders,vol.25,no.13, pp.2164–2170,2010.

[3]J.-M.Dolnitzki and S.Winter,“Merkmale aus

zweidimensionalen Orientierungshistogrammen zur Beurteilung von Tremorspiralen,”Proceedings of BVM 2012,pp.322–327,2012.