First Advisor

Dr. Nate George

Second Advisor

Dr. Kevin Pyatt

Third Advisor

Dr. Bob Mason

College

College of Computer and Information Sciences

Degree Name

MS Software Engineering

Document Type

Thesis - Open Access

Number of Pages

62 pages

Abstract

Diagnosis of an unknown skin lesion is crucial to enable proper treatments. While curable with early diagnosis, only highly trained dermatologists are capable of accurately recognize melanoma skin lesions. Expert dermatologist classification for melanoma dermoscopic images is 65-66%. As expertise is in limited supply, systems that can automatically classify skin lesions as either benign or malignant melanoma are very useful as initial screening tools. Towards this goal, this study presents a convolutional neural network model, trained on features extracted from a highway convolutional neural network pretrained on dermoscopic images of skin lesions. This requires no lesion segmentation nor complex preprocessing. Further, it doesn’t cost much computational power to train the model. This proposed approach achieves a favorable training accuracy of 98%, validation accuracy of 64.57% and validation loss 0.07 in the model with 46% sensitivity and 64% classification accuracy in testing data.

Date of Award

Fall 2017

Share

COinS