Date of Award

6-1-2018

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Bioinformatics

First Advisor

Kamran Iqbal

Abstract

We aim to develop a computational core to detect malignant melanoma. To this end, we evaluate abruptness of pigment pattern at the periphery of a skin lesion, which is one of the most important dermoscopic features for detection of malignancy. To achieve this, we quantitatively analyze the texture feature, particularly homogeneity features by having circular windows span along the region within the lesion boundary. To access this region, first we tested a level set propagation (LSP) method that offsets the lesion border to a defined location which is in \textit{constant proximity} to the original border. As lower homogeneity indicates sharp cutoffs and suggests melanoma, we validated the extracted features by a ground truth which consists of 728 benign and 172 melanoma cases. We obtained 87\% f1-score and 78\% specificity for correctly classifying lesions with the fully-connected multi-hidden layer NN classifier using LSP based border contraction method. Quantified abruptness features were then given in to the SVM classifier. 10xCV model is used for accuracy testing. Obtained sensitivity is 100\% whereas specificity is 92.6\%. We obtained 96\% balanced accuracy ratio for malignancy detection in dermoscopy images using abruptness along the lesion border which is obtained through our novel lesion border segmentation method that yields the Dice score of 0.8866 $\pm$ 0.0944 and Jaccard of 0.8074 $\pm$ 0.1339 over the used dataset.

Download

Share

COinS