Melanoma detection Colour, clustering and classification

Abstract

Malignant melanoma is the deadliest form of skin cancer, and successful treatment relies on early detection. Undiagnosed skin lesions can be photographed and the images fed into a computer system that potentially differentiates malignant from benign lesions. To develop the melanoma detection system, various methods from statistics, machine learning and image analysis are applied. An image consists of millions of pixels, so reducing the enormous amount of data is an important part of image analysis. This can be done by probability density estimation and clustering. In hierarchical clustering, the dissimilarity measure has great influence on the final clustering, but there has been little focus on how to choose an adequate dissimilarity measure for density-based techniques. In this thesis, six properties for density- based dissimilarity measures are therefore proposed as a guide for the user, based on assumptions and previous knowledge about the data set. An image cannot be fed directly into a classifier because of the amount of data contained in each image; therefore a set of features is extracted from one. In melanoma detection, the colour of the lesion is of special interest. This thesis presents several approaches to colour feature extraction. (i) By clustering the pixel values and then comparing the cluster centres to pre-defined colour values, melanoma-indicative colours are detected. (ii) By estimating the probability density, and then measuring the goodness-of-fit, the variation in colours is accounted for. (iii) By the use of a dissimilarity measure, an unclassified lesion’s similarity to the melanoma class or the class of benign lesions can be calculated. Different methods for feature evaluation are discussed. A thorough presentation of computer systems for melanoma detection is provided, and some of the key elements are discussed. The challenge of feature selection and classifier selection is given special attention. A computer system for melanoma detection, Nevus Doctor, is presented. It is based on semi-automatic feature selection of both new and previously developed features, and a new hybrid classifier. The performance of the system in terms of sensitivity and specificity scores is presented and compared to that of a commercially available system for the same set of lesions. This methodology has previously been used once only, and then in a smaller study. Obsta- cles associated with small data sets are discussed, including cross-validation and clinical relevance. Nevus Doctor performed better than the commercially available system. The new colour features add value in computer-aided melanoma detection, both by improving the existing system and by introducing a new class of features. The properties for dissimilarity measures offer a new perspective on clustering and other fields where dissimilarity measures are a core element. An image consists of millions of pixels, so reducing the enormous amount of data is an important part of image analysis. This can be done by probability density estimation and clustering. In hierarchical clustering, the dissimilarity measure has great influence on the final clustering, but there has been little focus on how to choose an adequate dissimilarity measure for density-based techniques. In this study, six properties for density- based dissimilarity measures are therefore proposed as a guide for the user, based on assumptions and previous knowledge about the data set. An image cannot be fed directly into a classifier because of the amount of data contained in each image; therefore a set of features is extracted from each image. In melanoma detec- tion, the colour of the lesion is of special interest. This thesis presents several approaches to colour feature extraction. (i) By clustering the pixel values and then comparing the cluster centers to pre-defined colour values, melanoma-indicative colours are detected. (ii) By estimating the probability density, and then measuring the goodness-of-fit, the varia- tion in colours is accounted for. (iii) By the use of a dissimilarity measure, an unclassified lesion’s similarity to the melanoma class or the class of benign lesions can be calculated. Different methods for feature evaluation are discussed. A thorough presentation of computer systems for melanoma detection is provided, and some of the key elements are discussed. The challenge of feature selection and classifier selection is given special attention. A computer system for melanoma detection, Nevus Doctor, is presented. It is based on semi-automatic feature selection of both new and previously developed features, and a new hybrid classifier. The performance of the system in terms of sensitivity and specificity scores is presented and compared to that of a commercially available system for the same set of lesions. This methodology has previously been used once only, and then in a smaller study. Obsta- cles associated with small data sets are discussed, including cross-validation and clinical relevance. Nevus Doctor performed better than the commercially available system. The new colour features add value in computer-aided melanoma detection, both by improving the existing system and by introducing a new class of features. The properties for dissimilarity measures offer a new perspective on clustering and other fields where dissimilarity measures are a core element.