Hannah Rivers

Date of Award

2012

Document Type

Thesis

Degree Name

Bachelors

Department

Natural Sciences

First Advisor

McDonald, Patrick

Keywords

Melanoma, Artificial Neural Network, Computer vision

Area of Concentration

Applied Mathematics

Abstract

Melanoma is an increasingly common and very serious form of cancer, claiming almost fifty thousand lives annually. The single most important factor in successful treatment of melanoma is early detection. In this thesis, an artificial neural network designed to distinguish melanoma from dysplastic nevi was created and trained. Heuristics employed by dermatologists to diagnose melanoma were synthesized into a preprocessing method that extracts relevant features, transforming an image into a feature vector. These vectors were then passed to the artificial neural network so it could learn patterns in the data. The artificial neural network was trained on an amalgamation of images from two publicly available data sets consisting of images of 101 melanoma and 129 benign melanocytic lesions. Because the performance of artificial neural networks is stochastic, results were averaged over a variety of parameter and starting values. On average, the network is capable of distinguishing between the lesions with 64% accuracy, with 53% sensitivity and 73% specificity. At best, its accuracy is 78%, with 86% sensitivity and 68% specificity.

Rights

This bibliographic record is available under the Creative Commons CC0 public domain dedication. The New College of Florida, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.

Download

Share

COinS