wound analysis method

8/19/2019 Wound Analysis Method

http://slidepdf.com/reader/full/wound-analysis-method 1/7

WOUND ANALYSIS METHOD:

Wound Image Analysis System Overvie:

Our quantitative wound assessment system consists of several functional modules includingwound image capture, wound image storage, wound image preprocessing, wound boundary

determination, wound analysis by color segmentation and wound trend analysis based on a time

sequence of wound images for a given patient.All these processing steps are carried out solely bythe computational resources of the smartphone. The functional diagram of our quantitative

wound assessment system is shown as in Fig. 1 and explained later.



ote that the words highlighted in italics in the text correspond to specific bloc!s in figures with

bloc! diagrams. "hile the image capture is the first step in the flowchart, the image capture box

is not one of the image processing steps and is therefore presented later in #ection $%. A exus &smartphone was chosen due to its excellent '()*+() performance and highresolution camera.

Although there are li!ely performance variations across the cameras of modern smartphones,

such a study was considered beyond the scope of this paper. After the wound image is captured,the -(+ file path of this image is added into a wound image database. This compressed image

file, which cannot be processed directly with our main image processing algorithms, therefore

needs to be decompressed into a /&bit bitmap file based on the standard 0+ color model. $nour system, we use the builtin A($s of the Android smartphone platform to accomplish the -(+

compression and decompression tas!. The 2image quality3 parameter was used to control the

-(+ compression rate. #etting 2image quality3 to 45 was shown empirically to provide the best

balance between quality and storage space. For an efficient implementation on the smartphonealone, no method was used to further remove the artifacts introduced by -(+ lossy

compression.

$n the $mage preprocessing step, we first downsample the highresolution bitmap image to

speed up the subsequent image analysis and to eliminate excessive details that may complicatewound image segmentation. $n our case, we downsample the original image 6pixel dimensions

7/8& 9 /&&4: by a factor & in both the hori;ontal and vertical directions to pixel dimensions of418 9 81/, which has proven to provide a good balance between the wound resolution and the

processing efficiency. $n practice, we use the standard A($ for image resi;e on theAndroid

smartphone platform to ensure high efficiency. #econd,we smooth the images to remove noise6assumed mainly to be +aussian noise produced by the image acquisition process by using the

+aussian blur method whose standard deviation < = 5.> was empirically ?udged to be optimal

based on multiple experiments.

To determine the boundary of the wound area, we first determine an outline of the footwithin the image. @ence the initial $mage segmentation operation is to divide the original image

into pixel groups with homogeneous color values. #pecifically, the Foot outline detection is

performed by finding the largest connected component in the segmented image under thecondition that the color of this component is similar enough to a preset standard s!in color.

ased on the standard color chec!ers provided in , both the light and dar! s!in color thresholds

in '$ A space are incorporated into the system, which means that our algorithm is expectedto wor! for most s!in colors. Afterwards, we carry out a "ound boundary determination based

on the foot outline detection result. $f the foot detection result is regarded as a binary image with

the foot area mar!ed as 2white3 and rest part mar!ed as 2blac!,3 it is easy to locate the wound

boundary within the foot region boundary by detecting the largest connected 2blac!3 componentwithin the 2white3part. $f the wound is located at the foot region boundary, then the foot

boundary is not closed, and hence the problem becomes more complicated, i.e., we might need to

first form a closed boundary. "hen the wound boundary has been successfully determined and thewound area calculated,

we next evaluate the healing state of the wound by performing 'olor segmentation, with the goal

of categori;ing each pixel in the wound boundary into certain classes labeled as granulation,slough and necrosis . The classical selforgani;ed clustering method called Bmean with high

computational efficiency is used . After the color segmentation, a feature vector including the

wound area si;e and dimensions for different types of wound tissues is formed to describe the



wound quantitatively. This feature vector, along with both the original and analy;ed images, is

saved in the result database.

The "ound healing trend analysis is performed on a time sequence of images belonging to agiven patient to monitor the wound healing status. The current trend is obtained by comparing

the wound feature vectors between the current wound record and the one that is ?ust one standard

time interval earlier 6typically one or two wee!s:. Alternatively, a longer term healing trend isobtained by comparing the feature vectors between the current wound and the base record which

is the earliest record for this patient.

Mean!S"i#t!$ased Segmentation Algorit"m



"e chose the meanshift algorithm, proposed in , over other segmentation methods, such as level

set and graph cutbased algorithms, for several reasons. First, the meanshift algorithm ta!es intoconsideration the spatial continuity inside the image by expanding the original 7C color range

space to >C space, including two spatial components, since direct classification on the pixels proved to be inefficient . #econd, a number of acceleration algorithms are available . Third, for

both meanshift filtering and region merge methods, the quality of the segmentation is easilycontrolled by the spatial and color range resolution parameters . @ence, the segmentation

algorithm can be ad?usted to accommodate different degrees of s!in color smoothness by

changing the resolution parameters. Finally, the meanshift filtering algorithm is suitable for parallel implementation since the basic processing unit is the pixel. $n this case, the high

computational efficiency of +()s can be exploited. The meanshift algorithm belongs to the density estimation based nonparametric clusteringmethods, in which the feature space can be considered as the empirical probability density

function of the represented parameter. This type of algorithms adequately analy;es the image

feature space 6color space, spatial space or the combination of the two spaces: to cluster and can provide a reliable solution for many vision tas!s . $n general, the meanshift algorithm models

the feature vectors associated with each pixel 6e.g., color and position in the image grid: as

samples from an un!nown probability density function f6x: and then finds clusters in this

distribution. The center for each cluster is called the mode . +iven n data points xi , i = 1, . . . , nin the ddimensional space 0d , the multivariate !ernel density estimator is shown as follows D

fh,B 6x: = c!,d nhdnEi=1!EEEEEx xihEEEE/E 61: where h is a bandwidth parameter satisfyingh

G 5 and c!,d is a normali;ation constant H18I. The function !6x: is the profile of the !erneldefined only for x J 5 and KK L KK represents the vector norm. $n applying the meanshift algorithm,

we use a variant of what is !nown in the optimi;ation literature as multiple restart gradient

descent. #tarting at some guess at a local maximum y! , which can be a random input data pointxi , the mean shift computes the density estimate f6x: at y! and ta!es an uphillstep using the

gradient descent method. The gradient of f6x: is given as followsD where g6r: = !E6r: and n is

the number of neighbors ta!en into account in the five dimension sample domain. $n our case, we

use the panechini!ov !ernel , which ma!es the derivative of this !ernel a unit sphere. asedon , we use the combined !ernel function shown in 6>: where hs and hr are different bandwidth

values for spatial domain and range domain, respectively. $n , the two bandwidth values are

referred to as spatial and range resolutions. The vector m6x: defined in 67: is called the meanshift vector , since it is the difference between the current value x and the weighted mean of the

neighbors xi around x. $n the meanshift procedure, the current estimate of the mode y! at

iteration ! is replaced by its locally weighted mean as shown next in 6&: D y!*1 = y! * m6y! : 6&:

This iterative update of the local maxima estimation will be continued until the convergencecondition is met. $n our case, the convergence condition is specified as the uclidean length of

the meanshift vector that is smaller than a preset threshold. The threshold value for the mean

shift iteration is the same for the tas! of locating the foot in the full image and for locating the

wound within the foot boundary. After the filtering 6also referred to as the mode see!ing: procedure above, the image is usually oversegmented, which means that there are more regions

in the segmentation result than necessary for wound boundary determination . To solve this

problem, we have to merge the oversegmented image into a smaller number of regions which are



more ob?ectrepresentative based on some rules. $n the fusion step, extensive use was made of

0A+s . The initial 0A+ was built from the initial oversegmented image, the modes being the

vertices of the graph and the edges were defined based on &connectivity on the lattice. Thefusion was performed as a transitive closure operation on the graph, under the condition that the

color difference between two ad?acent nodes should not exceed hf , which is regarded as the

region fusion resolution. The meanshift filtering and region fusion results of a sample footwound image Hsee Fig. /6a:I are shown in Fig. /6b: and 6c:, respectively. "e can see that the

oversegmentation problem in 6b: is effectively solved by region fusion procedure. From the

region fusion result in 6c:, the foot boundary is readily determined by a largest connectedcomponent detection algorithm, which will be introduced in the next #ection. A '** based

implementation method of the meanshift algorithm can be found in.

Hardware Requirement Specification

(rocessor D $ntel (entium $$$ or later Main Memory60AM: D />8 M

'ache Memory D >1/ B

Monitor D 1N inch 'olor Monitor

Beyboard D 154 BeysMouse D Optical Mouse

@ard Cis! D 185 +

Software Requirement Specification

Front ndanguage D Matlab

ac! ndCatabase D il

Operating #ystem D "indows P( #ervice (ac! /"indows %ista"indows N"indows 4

wound analysis method

Documents