| William Paul Berriss B.Sc, M.Sc,
AMIEE, wpberriss@iee.org Stephen John Sangwine, Senior Member IEEE S.J.Sangwine@reading.ac.uk Department of Engineering, University of Reading, Reading, England Publishing Details: Submitted: 26th April, 1997 Published: 19th September, 1997 Edition 1.1 Includes hyperlinks to online abstracts where available. Previous edition 1.0 (14 July 1997) also available. |
Abstract, Introduction Colour image processing Image acquisition, digitisation and calibration Conclusion References Acknowledgement, Useful resources. |
There are two main applications of colour image processing in the field of skin imaging. They are the assessment of the healing of skin wounds or ulcers, and the diagnosis of pigmented skin lesions such as melanomas. The analysis of lesions involves more traditional image processing techniques such as edge detection and object identification, followed by an analysis of the size, shape, irregularity and colour of the segmented lesion. However, in wound analysis, although it is necessary to detect the wound border and to calculate its area, analysis of the colours within the wound site is often more important. In short, wounds generally have a non-uniform mixture of yellow slough, red granulation tissue and black necrotic tissue, and the proportions of each are an important determining factor in the healing state of the wound.
In the case of assessing skin lesions in the clinic, clinicians have to decide whether or not a skin lesion should be tested further, and analysis using colour image processing could provide additional information to aid such decisions. A very general review on digital imaging has been written by Perednia et al. [1] . Their review covers the basics of image analysis, transmission and storage on computer. One problem of storage is that image files can be very large. However this can be reduced to some extent by use of data compression techniques without significantly reducing the information content or quality of an image. One of the groups reviewed found that dermatologists were able to diagnose lesions with compressed digital images without significant change from their performance with the original digital image.
At the University of Glamorgan, Wales, Jones and Plassmann [5] [6], of the Department of Computer Studies have developed an instrument, known as MAVIS (Measurement of Area and Volume InStrument), to measure the dimensions of skin wounds. It involves capturing two images of the wound in quick succession whilst the wound is illuminated with colour-coded structured light. This enabled phvolume measurements to be made. A colour CCD video camera with a 250 W tungsten halogen bulb was used for imaging the skin directly. MAVIS is capable of measuring the area phand volume of deep three-dimensional wounds. For each acquisition, a magnesium oxide chip, placed alongside the wound, was used as a white standard. The group have experimented with algorithms that use colour to segment an image into one of three tissue types: healthy skin, wound tissue and epithelialisation tissue. They found that epithelialisation tissue is often a darkened band around the wound, separating skin from wound. In all, they tried six measurement parameters: the R, G, and B intensities; Hue; Saturation; and gray-level intensity. The R, G and B intensities were only examined in isolation and they concluded that, `It is clear from inspection of Red, Green and Blue plane intensity-level histograms for the different tissue types that straightforward thresholding of these planes cannot produce a good segmentation which distinguishes between wound and skin or wound and surrounding connected tissue'. They conclude that in looking at such 1D histograms, segmentation is only partially achievable, but using a 3D RGB histogram space, volume clusters may be more widely separated.
One group has made some progress with such a 3D RGB colour histogram clustering technique. Mekkes et al.[7] at the Department of Dermatology, University of Amsterdam, The Netherlands, have been using colour images to assess the healing of wounds. They recognized that many of the enormous number of wound care materials that have been introduced into the market have not been properly tested in randomized, double-blind clinical trials. They pointed out that such trials are desperately needed to supply clinicians with information to guide them in their choice of wound care products. They compared a debriding product (Intrasite Gel) with an old form of treatment using saline soaked gauzes [8]. They found that for a proper evaluation of the cleansing effect of both treatments, colour aspects were more important than wound size. Their technique measured the shift from black to yellow necrotic tissue to red granulation tissue. Their aim is to create an automatic computerised method which can be used as a reference standard or `gold-standard' for colour wound analysis. In their system, images were acquired directly with an RGB video camera and framegrabber. They used two polarised filters to reduce unwanted reflections. A clinicians knowledge of the colours in secondary healing ulcers was used for calibration of the system. The computer had to be instructed in advance as to which colours can be encountered in the granulation region and which in the necrotic region of a wound. They found that clusters in RGB space for a given tissue type formed an irregularly shaped 3D cloud, and so simple thresholding along the R, G and B axes would not help to segment the image into these three tissue types. For this reason, large classification tables, of the colours present in each tissue type, were created semi-automatically by the computer with the aid of a clinician. One problem discovered was that although digital image analysis could detect the wound margins automatically, the colour differences between granulation tissue, surrounding skin and the thin partly transparent layer of newly formed epithelium were too small to allow automatic detection.
Finally, there are a few other groups that have done some work in colour image processing of wounds. El Gammal et al. [9] at the Dermatological clinic of Ruhr University, Germany, wrote a very short paper on the use of the black-yellow-red classification scheme to evaluate the debridement activity of wounds. Solomon et al. [10] at the University of Otago Medical School, New Zealand developed a simple and rapid technique to measure the size of skin wounds and ulceration using two-dimensional colour video images of ulcers. The images were stored on video cassette, thus rendering low image quality. The work did not involve the development of colour image processing algorithms, but a novel method to correct for limb convexity was presented. Smith et al.[11] at the University of Akron, Ohio, USA, evaluated wound repair in humans and animals using video images. Images were stored on VHS video tape, and only basic colour image processing techniques were applied to the digitised images.
Framegrabbers, the digitiser boards in computers that connect to the video camera, do not have such high resolution. Typically framegrabbers digitise images to only 512 pixels by 512 pixels, and resolution does not meet up to the standards of photography or colour slide scanners. Colour resolution is also inferior for framegrabbers, typically a colour framegrabber has a colour resolution of 24-bits, corresponding to 16 million colours. Their advantage is that digitisation of the images takes place as they are acquired, and consequently no photographic processing time is incurred. However, although video cameras can be as compact as a still camera, and use of a laptop computer allows the system to be portable, such a system tends to be less versatile than using a 35mm still camera. This renders imaging outside the laboratory less suitable. The best solution would be to use a phdigital still camera, but these are still fairly new on the market and rather expensive. Still, costs are gradually falling and so they are becoming a viable option. They are quick, as no photographic processing is needed, digitisation occurs immediately, and they render high resolution images, comparable with slide scanners.
Calibration is a very important step and often overlooked by programmers since they often aim to improve results by writing more complex algorithms rather than aim to improve the quality of the original input image. By considering the nature of non-uniformities in an image acquisition system due to the non-linear response of electronic devices and non-uniform lighting, methods can be devised to measure these non-uniformities to enable corrections to be made at the pre-processing stage. The use of a pure white reference object in each scene, or better still a uniform greyscale, can be of great benefit in correcting for non-linearities between the red, green and blue channels as well as correcting for the non-linear reproduction of intensity by the system. In fact, Hall et al. [12] found that It is not sufficient to simply have a reference white and black in the image for calibration purposes, as this would assume a linear relationship for all shades of grey in between'. Such greyscale non-linearities are inherent in all imaging systems. Calibration can also be taken further, to ensure the correct reproduction of colour as well as intensity. Frey and Palus [13]. Others considered the measurement of a colour in a digital image processing system and explained a method for calibrating such a system. In particular, they state that greyscale linearisation of each of the three channels, R, G and B, is not enough to allow the system to reproduce colours or hues correctly. A further step of linearisation must be performed over the three channels together. This ensures that a pure red object which is twice as bright as another object of the exact same red hue, is represented as being twice as bright in the red channel only, rather than becoming marginally brighter in the red, green and blue channels for example. For this stage, a colour look-up table (LUT) must be created and used for each digitisation.