Differentiating skin from hair in cancer analysis
Kevin Thon has been working with software to help doctors find out whether a mole is malignant or not. A specialized digital camera – a dermatoscope – is used to photograph moles that might be malignant. The digital image is then analysed using the new software.
|The images above show how different scaling of the software affects the image analysis. With fine scaling (top), more details in the images are captured. All hair is included in the analysis, but some unnecessary noise is also included in the image (shown as shades of colour in the skin around the mole). If coarser scaling is selected (bottom), irrelevant information is removed from the analysis, but the finest hairs are also excluded from the assessment. Thon's research involved exploring how to fine-tune the software to include only information relevant to the diagnosis. Illustration: Tromsø Telemedicine Laboratory.|
Faster and more precise diagnosis of malignant melanoma
Doctors analysing moles for malignant melanoma must consider a variety of factors. Typically, they look at asymmetry, colour, border with the skin, structure, and other features. All these factors will be weighted and together they will form the basis for a diagnosis. A computer-based diagnosis system combined with the doctor's assessments can provide a faster and safer diagnosis. In his PhD project, Thon developed analyses for investigating suspicious moles. He looked more closely at how the surface of the mole could be assessed automatically. He also looked at how the software should be scaled to differentiate between hair and skin in analysis of moles. The aim of the methods that Thon studied is to detect malignant melanomas at an earlier stage.
Statistics in practice
Researchers at the University of Tromsø's Department of Mathematics and Statistics have been working with multiscale methods for a long time. These methods involve assessing the data over a range of scales rather than looking for one optimal scale. The background for Thon's work was the desire to be able to develop and apply these methods in practical contexts. For example, analyses of temperature variations may lead to very different conclusions depending on whether you look at day-to-day variations or fluctuations over many years. Instead of focusing on one scale, one can perform multiscale analysis over a longer period. Thon's thesis describes three statistical multiscale techniques for analysing image data. Kevin Thon is from Tromsø and started on his PhD project in 2007. His work was conducted in collaboration with Tromsø Telemedicine Laboratory.
Smooth or rough, spot or not
Over the past 10 years, dramatic development has taken place in the use of computers as diagnostic tools in the detection of malignant melanoma. To enable assessment of a medical image, criteria must be defined to describe the normal state and possible deviations from it. The conclusions of a statistical analysis will depend on the parameters, that is, how the analysis is scaled. How rough must the area of the mole be in order to classify it as a raised bump? How light or dark must the area be before we can classify it as a dot or globule? A scale must be specified in advance to make these decisions possible. These parameters were set in cooperation with an experienced dermatologist.
Must distinguish between skin and hair in the analysis
Hair growing in a mole is one of the challenges that a digital tool must solve. Because there is no reason to believe that hair provides any diagnostic information, it is important to develop systems that can eliminate hairs from the analysis so that they do not affect the final diagnosis. A hair will stand out as darker than its surroundings and it will usually be a different colour to the mole. In an image, hair will appear as long dark lines. Thon worked on fine-tuning the software to distinguish lines from other formations, and to decide how long and dark the line should be in order to classify it as a hair.
Thon defended his thesis entitled “Multiscale Methods for Statistics Inference on Regular Lattice Data” on 13 December 2013. His work was supervised by Fred Godtliebsen of the Department of Mathematics and Statistics.