Date of Award

January 2019

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Mechanical Engineering

First Advisor

Jeremiah Neubert

Abstract

Melanoma is identified as the deadliest in the skin cancer category. However, early-stage detection may enhance the treatment result. In this research, a deep learning-based model, named “melNET”, has been developed to detect melanoma in both dermoscopic and digital images. melNET uses the Inception-v3 architecture to handle the deep learning part. To ensure quality optimization, the architectural aspects of Inception-v3 were designed using the Hebbian principle as well as taking the intuition of multi-scale processing. This architecture takes advantage of parallel computing across multiple GPUs to employ RMSprop as the optimizer. While going through the training phase, melNET uses the back-propagation method to retrain this Inception-v3 network by feeding the errors from each iteration, resulting in the fine-tuning of network weights. After the completion of the training step, melNET can be used to predict the diagnosis of a mole by taking the lesion image as an input to the system. With a dermoscopic dataset of 200 images, provided by PH2, melNET outperforms the work with YOLO-v2 network by improving the sensitivity value from 86.35% to 97.50%. Also, the specificity and accuracy values are found to be improved from 85.90% to 87.50%, and, from 86.00% to 89.50% respectively. melNET has also been evaluated on a digital dataset of 170 images, provided by UMCG, showing an accuracy of 84.71%, which outperforms the 81.00% accuracy of the MED-NODE model. In both cases, melNET got treated as a binary classifier and a five-fold cross validation method was applied for the evaluation. In addition, melNET has been found to perform the detections in real-time by leveraging the end-to-end Inception-v3 architecture.

Share

COinS