Automatic detection of malignant melanoma using boundary gradient information in dermoscopy images

This thesis demonstrates an algorithm which uses local border features, primarily derived from the border gradient, to help determine whether the lesion is a malignant melanoma or a benign lesion. Although this algorithm uses manually-marked lesion borders, it could be included in a fully automatic routine. A Prewitt algorithm with a window size proportional to the area of the lesion is used to calculate the gradient over the entire lesion boundary. This gradient value is used to calculate various boundary features, such as mean, standard deviation, and direction of maximum gradient. The dark and white areas inside the lesion combined with the gradient measurements are used to generate additional features. All these features are used as inputs to a neural network to classify the lesion as malignant or benign. A total of 90 different features were fed to the neural network. The features that gave the most accurate separation were a combination of dark area features and white area features. These features when combined with some other features like the boundary length gave a diagnostic accuracy of 80.5% for the test set of 48 malignant and 198 benign images --Abstract, page iii