ARTIFICIAL INTELLIGENCE

FOR WOUND IMAGE UNDERSTANDING

Augustin Prodan, Mădălina Rusu, Remus Câmpean

Department of Mathematics and Informatics, Iuliu Haţieganu University, Cluj-Napoca, Romania

Rodica Prodan

MedFam Group, Str. Constanţa 1, Cluj-Napoca, Romania

Keywords: Web-based Education, e-Learning Scenario, Java and XML Technologies, Artificial Intelligence, Intelligent

Tutoring Systems, Wound Image Understanding, Wound Healing Simulation.

Abstract: This paper presents an e-learning framework for analyzing, processing and understanding wound images, to

be used in teaching, learning and research activities. We intend to promote e-learning technologies in

medical, pharmaceutical and health care domains. Our approach to e-learning is so called blended learning,

which combines traditional face-to-face and Web-based on-line learning, with focus on principles of active

learning. Using Java and XML technologies, we build models for various categories of wounds, due to

various aetiologies. Based on colour and texture analysis, we identify the main barriers to wound healing,

such as tissue non-viable, infection, inflammation, moisture imbalance, or edge non-advancing. This

framework provides the infrastructure for preparing e-learning scenarios based on practice and real world

experiences. We make experiments for wound healing simulation using various treatments and compare the

results with experimental observations. Our experiments are supported by XML based databases containing

knowledge extracted from previous wound healing experiences and from medical experts knowledge. Also,

we rely on new paradigms of the Artificial Intelligence for creating e-learning scenarios to be used in a

context of active learning, for wound image understanding. To implement the e-learning tools, we use Java

technologies for dynamic processes and XML technologies for dynamic content.

1 INTRODUCTION

Medical images are valuable in didactic activities for

students in medicine and pharmacy. Digital pictures

are in great demand, because digital technologies

provide unlimited resources for medical and

pharmaceutical education. Computerized image

processing contains methods for non-invasive

wound evaluation, allowing an accurate diagnosis in

a large category of patients with damaged and

wounded skin. Traditional non-invasive technologies

are limited frequently to subjective visual

evaluations. Colour and texture information provide

the infrastructure for a structured approach to non-

invasive wound assessment. We use the RGB (Red-

Green-Blue) colour space to define a set of image

features for every category of wounds. To identify a

wound in an image, we implement specific methods

based on some generic criteria, such as normal skin

and wounded skin. For some applications we use as

main colours Red, Yellow and Black to assess the

gravity of a wound. Generally, wounds have a non-

uniform mixture of yellow slough, red granulation

tissue and black necrotic tissue. Relying on a high

quality of image acquisition, we can analyse a

succession in time of more images for the same

wound and assess changes in wound healing, i.e. the

recovery or worse evolution.

We intend to develop e-learning tools for

students and residents in medicine, pharmacy and

health care, to be used in both didactic and research

activities. Our aim is to create and implement in

Java an automatic method which can be used as a

reference standard for colour and texture wound

analysis. The purpose is to create e-learning

scenarios for wound image understanding and

wound healing simulation, by applying this method

to large amounts of wound image data stored in

213

Prodan A., Rusu M., Câmpean R. and Prodan R. (2008).

ARTIFICIAL INTELLIGENCE FOR WOUND IMAGE UNDERSTANDING.

In Proceedings of the Tenth International Conference on Enterprise Information Systems - AIDSS, pages 213-218

DOI: 10.5220/0001689002130218

 SciTePress

XML based knowledge bases (Figure 1). Our

objective is to develop appropriate skills in wound

management for a learner that traverses such an

e-earning scenario. The e-learning scenarios are

practice driven and relevant to professional practice,

being used by students in medicine and pharmacy, at

graduate, postgraduate and residency levels.

Figure 1: The general method applied for wound image

understanding and wound healing simulation.

Wound image understanding is a difficult

knowledge-based process and we have to use the

new paradigms of the Artificial Intelligence (e.g.

Bayesian Inference, Case Based Reasoning and

Intelligent Agents) to manage it. Relying on large

amounts of wound image data collected from

medical and health care environments, we intend to

create XML and CBR (Case Based Reasoning)

knowledge bases, working in a continuous

collaboration with physicians and wound care

experts from our university and from health care and

medical units. We have continuous access to actual

medical records, to monitor the wound evolution and

to verify both the accuracy and the consistency of

our system.

The originality of our work consists in relying on

new paradigms of artificial intelligence for creating

intelligent and practical e-learning tools to be used

in a context of blended and active learning. To

implement these e-learning tools, we use Java

technologies for dynamic processes and XML

technologies for dynamic content. We create XML

based databases containing knowledge extracted

from previous wound healing experiences and from

medical experts knowledge. The methods presented

in this paper should be useful as an adjunct to

traditional teaching and learning resources. In a

context of blended learning, the teachers and

learners may combine the colour and texture based

parameters with traditional parameters, such as

smell, venous and arterial status, patient history, etc.

2 IMAGE PROCESSING

The infrastructure of our system is based on a

collection of Java class libraries, containing methods

for processing images specific to various categories

of wounds (Prodan et al., 2006). We implemented

general methods that create many common special

effects and use them in analysing the wounds.

A digital image consists of a two dimensional

array of pixels P

with m rows and n columns.

Using Java language, this image is represented in

internal memory as a three dimensional array P

mn4

each pixel being described in a specific RGB format

by four unsigned 8-bit integers (Prodan and Prodan,

1997). The first three integers represent the base

colour components (Red, Green and Blue), and the

fourth integer, referred to as α (alpha) represents the

transparency. A specific colour is obtained by

mixing different amounts of basic colours (red,

green and blue) with a specific transparency. The

standard Core Java Technologies provides methods

for processing digital images, such as blur, sharpen,

brighten or tone down an image. We will create the

Java framework by implementing the image

processing algorithms into one of the following two

layers: (1) for low-level implementations, allowing

to operate directly on pixels; and (2) for high-level

implementations, based on standard Java libraries

such as JAI (Java Advanced Imaging) API.

For a given wound, we must find out some

quantitative and qualitative attributes for assessing

the healing state. As quantitative attributes we

measure its surface area and its volume (evaluating

depth). The original image is processed with the

purpose to emphasize the distinction between wound

and non-wound area. We use some general methods

to enhance the image, because we must exaggerate

the distinction between wound and non-wound. For

this purpose, we apply a special type of convolution

kernel (Guinot, 2006), named edge detection filter.

In previous works were applied some interpolation

with pixel information for this purpose, but we think

that applying a convolution kernel is more

efficacious. We implement e-tools that will enable to

assess the current state of the wound and to gain

insight into the wound evolution, by comparing the

series of wound data collected over time. Based on

this knowledge we can design an e-tool for

simulating the process of wound healing. The colour

image processing is the most acceptable automatic

non-invasive method of objectively and reproducibly

analysing skin wounds and lesions.

ICEIS 2008 - International Conference on Enterprise Information Systems

214

3 IDENTIFYING THE WOUND

The first task we face with in our system is to

identify the wound in a digital image. Indeed, before

analysing a wound image, it is necessary to identify

it. For this purpose, we implemented specific

methods based on some generic criteria, such as

normal skin and wounded skin. As a general

approach, to identify the wounds it is necessary to

traverse two phases: a pre-process phase and an

identification phase. In pre-process phase, the

original image is transformed with the purpose to

emphasize the distinction between wound and non-

wound area. We use some general methods to

enhance the image, because we must exaggerate the

distinction between wound and non-wound. As an

example, for individuals with fair skin, we lighten

the images and then view them using shades of

green with the red and blue minimized. This way

more clearly exhibit the borders of the wound than

in the original image. Removal of the red and blue

leaves the wound black and the rest of the image

green. For images of individuals with dark skin, both

the red and green are accentuated while the blue is

minimized. This procedure also leaves the non-

wound area green, but colours the wound red. In

either case, the wound can easily be distinguished

from the non-wound without difficulty.

In the identification phase, the image is divided

into little boxes, then start analyzing each box for

colour profile, determining the percentage of main

colours. It is examined the difference in the colour

profile of each box to the colour profile of a box

covering healthy skin, taken from outside the wound

area. The distribution obtained from a box with

healthy skin can be used as a benchmark. Other

distributions are then compared in statistical terms

with this baseline distribution and decisions are

made on determining the edge. Wound area and

different colour percentages follow from this as

well. The degree of deviations from this benchmark

distribution can then be used to classify wounds.

Assuming normality, the first two moments (the

mean and the standard deviation) estimated from a

sample will determine the colour and texture

distributions. The edge identification has an element

of subjectivity which is left to the medic or wound

specialist to set. Say for example, that wound edge

starts if the colour profile changes 40%, 70% or

90%, depending on how sensitive we want the

detector to be.

We implement the process of wound

identification in user oriented applications, endowed

with friendly GUI (Graphical User Interface), to be

used in didactic and research activities. When an

application is launched, it makes the following

general actions: (a) Reads the digital image in main

memory; (b) Converts pixel data of the digital image

into a three-dimensional array that is better suited

for processing; (c) Make a working copy of the

three-dimensional array, in order to avoid having to

make changes to the original array of pixel data; and

(d) Display on the same frame both the original

image and the modified image that contains the

output results.

Figure 2 shows the output result displayed by

such an application launched to identify a wound.

The left hand side contains the original image, while

the right hand side contains the clone image,

processed and marked with the contour of the

wound.

Figure 2: The output result of identifying a wound.

Our system contains two general strategies for

the process of identifying the wounds: a global

strategy and a wound by wound strategy. The user

may choose one of the two general strategies, or

may combine them using a friendly graphical user

interface.

Figure 3: Applying the global strategy.

When apply the global strategy (see Figure 3),

the whole image is traversed from top-left corner

towards bottom-right corner, applying specific

ARTIFICIAL INTELLIGENCE FOR WOUND IMAGE UNDERSTANDING

215

methods for edge-detection and wound

identification. The output result presents all the

wounds identified inside the current image. When

apply the wound by wound strategy (see Figure 4),

each wound is identified in a separate process, based

on a representative area belonging to it. In this case,

only the selected wound is traversed, starting with

representative area and going towards the four main

points: top-left, top-right, bottom-left and bottom-

right.

Figure 4: Applying the wound by wound strategy.

4 CLASSIFICATION METHODS

Our work is based on a continuous collaboration

with physicians and wound care experts, because it

is necessary to make a rigorous classification for

various categories of wounds. We collected large

amounts of wound image data and we calculate

statistical parameters as mean, median, standard

deviation, confidence interval, skewness and

kurtosis for them. These historical data are included

in XML based databases, to be used as inputs to

classification algorithms. The general purpose is to

make distinction between infected and non-infected,

inflamed and non-inflamed wounds. Based on colour

analysis, we build a statistically significant

differentiation of mild, moderate and severe wounds.

Our system analyses the differences in calibrated

hue between injured and non-injured skin, obtaining

a repeatable differentiation of wound severity for

various time intervals. As an example, burn wounds

are characterized according to their depth as: (a)

superficial – with bright red colour and the presence

of blisters (usually with brown colour); (b) deep –

with red-whitish colour and with dark dots; (c) full

thickness – with creamy or dark brown colour.

The system contains classification methods for

classifying wound images into different groups

based on colour and texture information. We

investigated the suitability of statistical parameters

for providing useful inputs to the classification

algorithms.

Assuming normality, the first two moments

(mean and standard deviation) characterize very well

the colour distribution. The mean represents the

centre point of the distribution, separating the values

into two equally probable subsets. Standard

deviation represents the dynamics of the values, how

wide around the mean the colours of the wound

image are distributed. The first two moments (mean

and standard deviation) are used to modify the

contrast and the brightness of an image. The contrast

is determined by the width of the distribution, while

the brightness is determined by the location of the

grouping colour values (Baldwin, 2005). We

implemented Java programs that use the mean and

standard deviation to modify and control both the

contrast and the brightness of an image, by

modifying the distribution of the colour values.

Figure 5 shows the distribution of the colour values

contained in the original image (left), compared with

the distribution contained in the modified image

(right). In processed image, the contrast (width of

the distribution) was increased by a factor of 2.0 and

the brightness (mean value) was increased by a

factor of 1.7.

Figure 5: The distribution of the colour values before

processing (left) and after processing (right).

Sometimes the first two moments alone are

inadequate to discriminate between wound and non-

wound skin. Therefore further details of the colour

distribution are required. Skewness and kurtosis of

the colour data proved to be more useful for this

purpose. Skewness is a measure of the asymmetry of

the distribution around the centre. Skewness is null

for a normal distribution, positive when the

distribution is skewed right (i.e. when the upper tail

of it is predominant) and negative when the

distribution is skewed left. Kurtosis quantifies the

flatness level of the distribution at the mean.

Kurtosis is equal to 3 for a normal distribution. If

kurtosis is lower than 3, the distribution is said to be

platokurtic (i.e. wide-peaked) and if kurtosis is

higher than 3, the distribution is said to be

leptokurtic (i.e. narrow-peaked). The kurtosis is used

ICEIS 2008 - International Conference on Enterprise Information Systems

216

as a measure of the heaviness of the tails in a

distribution.

We build in Java models for various categories

of wounds, due to aetiologies such as pressure, burn,

chilblain, vascular insufficiencies, diabetic foot

ulcer, venous leg ulcer and other chronic disease

states. Based on colour and texture analysis, we have

to identify the main barriers to wound healing, such

as tissue non-viable, infection, inflammation,

moisture imbalance, or edge non-advancing. Our

aim is to implement algorithms for wound healing

simulations.

5 e-LEARNING ENVIRONMENT

In a previous work we defined and implemented a

Java framework for designing and implementing

intelligent and practical e-learning tools, to be used

by both the students and the teaching staff in a

context of open learning. This framework provides

the infrastructure for preparing e-learning scenarios

based on practice and real world experiences, as

practice is essential in learning activities. Our

e-learning scenarios promote active learning, forcing

the students to take part in real world activities

simulated on computer. We rely on new paradigms

of artificial intelligence (Bayesian Inference, Case

Based Reasoning and Intelligent Agents) for

creating e-learning scenarios to be used in a context

of active learning. An e-learning scenario combines

simulation and interactive visualization and allows

the learners to explore the knowledge bases with

some well-defined learning purposes. For each

application object, our system contains a simulation

class and a visualization class. These classes are then

configured to obtain a particular simulation with a

specific visualization. In an e-learning scenario,

visualization is an active part of the system, serving

as an additional interface for modifying dynamically

some parameters. The simulation and visualization

classes are coded in Java, using XML format to

describe the configurations for both the components

and their relationships.

An e-learning scenario is in fact like a traditional

lesson, and the ideal solution is to simulate a

teaching-learning relation with a virtual teacher able

to interact with the learners and to instruct them

(Prodan et al., 2007). A good traditional teacher

learns all the time from previous didactic

experiences. Based on this historical feedback, the

teacher exploits prior specific successful episodes,

and avoids prior failures. We introduce a similar

feedback mechanism in our technology of

elaborating e-courses (see Figure 6). The feedback

information, collected from learners’ remarks and

from prior results and successes, is stored in case

bases. The relevant cases are retrieved and adapted

to fit new situations from new e-learning scenarios,

or to improve the previous ones. In addition, our

approach in creating an e-learning scenario relies

upon a special sort of goal oriented intelligent agents

(Nwana, 1996), able to incorporate knowledge,

teaching methods and pedagogical characteristics

into e-courses. We intend to implement a simulation

of some intelligence based actions and initiatives,

that are to be incorporated into e-learning scenarios,

with the purpose to map, to plan and to monitor the

pace and the progress of a learning process.

Following the traditional model, the cases of

positive experiences from previous e-learning

scenarios are stored into case bases created with

XML and CBR technologies (Leake, 1996).

Figure 6: The generation of the e-learning scenarios.

Our aim is to create and implement in Java an

automatic method which can be used as a reference

standard for colour and texture wound analysis. In

collaboration with medical experts, we will create

e-learning scenarios by applying the method

illustrated in Figure 6 to large amounts of wound

image data stored in XML knowledge bases. By

estimating the percentages for the main colours of

red, yellow and black, it is possible to assess the

gravity of the wound. The image processing

program allows the user to interactively control the

process. The user can set the tolerance for each

colour, that is the width of the band of acceptable

colours. Based on colour analysis and statistical

methods, it is possible to analyse successive states of

a wound, assessing the wound healing evolution.

The final goal is develop a flexible and adaptable

system for wound image understanding, based on

new paradigms of Artificial Intelligence (such as

Bayesian Inference, Case Based Reasoning and

Intelligent Agents). The students create electronic

portfolios, consisting of wound images and reports

about wound healing, based on wound healing

ARTIFICIAL INTELLIGENCE FOR WOUND IMAGE UNDERSTANDING

217

simulation scenarios, allowing to assess the wound

image understanding (Rusu and Prodan, 2006).

The functionality of our system will aim to

creating new e-learning tools, to be used by the

students in medicine and pharmacy, at graduate,

postgraduate and residency levels, for developing

appropriate skills in wound management. Our effort

is supported by a continuous collaboration with

physicians and wound care experts from our

university and from health care and medical units.

We are endowed with a continuous access to actual

medical records, allowing us to have in view the

wound evolution and to verify the accuracy and the

consistency of our system. The observed and the

estimated values of the colour are all the time

compared with each other. Based on colour and

texture analysis, it is possible to identify the main

barriers to wound healing, such as tissue non-viable,

infection, inflammation, moisture imbalance, or

edge non-advancing. These results are used to

implement algorithms for wound healing simulation.

The advantage of using Java for this purpose is the

integration without any difficulty with other Web

based facilities.

The methods presented in this paper should be

useful as an adjunct to traditional teaching and

learning resources. In a context of blended learning,

the teachers and learners may combine the colour

and texture based parameters with traditional

parameters, such as smell, venous and arterial status,

patient history, etc.

6 CONCLUSIONS

This paper presents a Java framework for analysing

and processing wound images, to be used in

teaching, learning and research activities. The colour

image processing methods have many advantages

over traditional human methods in assessment of

wounds. Computer based methods are objective,

repeatable and with a large potential of processing.

The analysis of a wound from a specific distance

involves procedures devoted to identify its

boundaries, to calculate its area and to estimate

proportions of the main colours red, yellow and

black. Generally, wounds have a non-uniform

mixture of yellow slough, red granulation tissue and

black necrotic tissue. To analyse the actual state of

the wound and the healing evolution, it is necessary

to determine the proportions of these main colours.

We create XML based databases containing

knowledge extracted from previous wound healing

experiences and from medical experts knowledge.

The students create electronic portfolios, consisting

of wound images and reports about wound healing,

based on wound healing simulation scenarios,

allowing to assess the wound image understanding.

Our experience demonstrated that electronic

portfolios may improve the teaching-learning

relation. As a future work, we have to implement

e-learning tools and e-learning scenarios enabling to

perform quantitative measurements of wound

evolution in time and to assess changes in wound

healing, i.e. the recovery or worse evolution. This is

our initial work towards a model of colour and

texture based simulation for the wound healing. We

intend to simulate wound healing based on various

treatments and to compare the results with

experimental observations.

REFERENCES

Baldwin, R. G., 2005. Processing Image Pixels using Java:

Controlling Contrast and Brightness. Available (Last

access at January 17, 2008) at address:

http://www.developer.com/java/other/article.php/3441

391.

Guinot, J., 2006. Image Filtering with GLSL –

Convolution Kernels. Available (Acc. March 4, 2008)

http://www.ozone3d.net/tutorials/image_filtering.php.

Leake, D. P., 1996. Case-Based Reasoning: Experiences,

Lessons, and Future Directions (Menlo Park: AAAI

Press / MIT Press).

Nwana, H. S., 1996. Software Agents: An Overview,

Knowledge Engineering Review, 11(3), 205-244.

Prodan, A., Câmpean, R., Rusu, M., Prodan, R., 2007. An

e-learning framework for wound image understanding.

In C. P. Constantinou et al. (Eds.), Proceedings of the

CBLIS 2007 (Computer Based Learning In Science),

Contemporary Perspectives on New Technologies in

Science and Education, pp. 225-235, ISBN 978-9963-

671-06-9, 30 June - 6 July 2007, Crete Island.

Prodan, A., Rusu, M., Campean, R. and Prodan R., 2006.

A Java framework for analyzing and processing

wound images for medical education. In Wolfgang

Borutzky et al. (Eds.), Proceedings of the ECMS 2006

(European Conference on Modelling and Simulation),

pp. 421-426, ISBN 0-9553018-0-7, 28-31 May 2006,

Bonn, Germany.

Prodan, A. and Prodan, M., 1997. Java Environment for

Internet, Editura Promedia-Plus, Cluj-Napoca,

Romania, ISBN 973-9275-07-09.

Rusu, M., Prodan, A., 2006. Electronic portfolios as

means of improving the teaching-learning relation. In

Ken Fernstrom et al. (Eds.), Proceedings of the

ICICTE 2006 (International Conference on

Information Communications Technologies in

Education), 17-21, ISBN: 1-895802-26-8, ISSN:1109-

2084, 6-8 July 2006, Rhodes Island.

ICEIS 2008 - International Conference on Enterprise Information Systems

218