engleski

Machine learning approach in discrimination of pigmented skin lesions

Pigmented lesions are the most common lesions of the skin. There are a lot of types of these kinds of lesions, and they can be benign or malignant. The most often benign pigmented lesion is a common acquired nevus, and it’s most common subtype is the intradermal nevus. Dysplastic nevus is also a subtype of common acquired nevus, but a lot of recent medical studies has shown that it is a very important precursor of malignant melanoma. Malignant melanoma is the most malignant tumour of the skin and mucosa. In this study, 20 intradermal nevi, 20 dysplastic nevi and 20 malignant melanoma from the archive of the Department of Pathology (Zagreb, Croatia) were analysed. The microscopical fields from the central part of each lesion were recorded in a database of the PC-compatible software ISSA (Vamstec, Zagreb,Croatia) using a magnification of 40 x.Circumference and area of nuclei are measured and a computer net of 10x10 fields is applied. The number of nuclei is counted in each net field.Analysis of the obtained data was performed using Poisson’s distribution, machine learning system Weka 3.1.7 and neural net application aiNet 1.25.Measurements have shown that nuclei of malignant melanoma are the biggest and that they have the largest nuclear roundness factor, which means that these nuclei have the most unregular shape.Poisson’s distribution showed that nuclei of examined groups have different way of distribution in space and that dysplastic nevi have the largest number of net fields with high number of nuclei.Machine learning system using SMO classificator gave very high level of accuracy of the classification by using all measured parameters (app.95%). Neural net aiNet 1.25, also, reached high level of accuracy of classification (app.95%). The conclusion is that there is a difference between study groups taking nuclei circumference and area, as well as distribution and density in space. The applied methods can be relatively easy and quickly performed, so they can be useful help in problematic pigmented skin lesions.

dysplastic nevi; machine learning; malignant melanoma; neural net; pigmented lesion; Poisson’s distribution; SMO classificator

nije evidentirano