Title

Design a Convolutional Neural Network for Melanocytic Skin Lesion Identification

Author

Emran Hossen

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Mathematics

Date of Award

Summer 2021

Abstract

Malignant melanoma is a type of skin cancer which stands with the highest mortality rate out of all cancers. Back in the days, dermatologists would use their naked eye to diagnose a lesion. Further, the dermatologists have improved their ways of diagnosing the cancer more accurately by using dermoscops. With the aid of the Convolutional Neural Network throughout the past years, mathematicians and computer scientists are now able to detect skin cancer more easily. In this study, we designed a new Convolutional Neural Network (CNN) for binary classification of skin lesion images to benign and malignant. Further, for the same kind of classification we designed a hybrid deep learning Convolutional Neural Network (CNN) with Support Vector Machine (SVM). The first proposed networks consist of several layers including convolutional layers, max pooling and fully connected layers. ReLu activation function is applied on each convolutional layer. Finally, in the fully connected layer, we used sigmoid activation function for binary classification. To minimize the loss function we applied two machine learning methods from the Tensorflow-Keras library: SGD and ADAM. To analyze accuracy and model performance, we modified our CNN model by adding convolutional layers and max pooling. We applied l2 regularization and drop out layers in our modified CNN model to decrease overfitting. In order to enhance the classification accuracy of our new proposed Hybrid CNN model with SVM, we applied segmentation process by several new dermascopic techniques such as isodata threshold, clustering based method and watershed based methods. To validate our experiments we used ISIC 2020 skin lesion image dataset and we achieved the highest accuracy, sensitivity, specificity and precision is 90%, 93%, 94% and 93% respectively from our CNN and 88%, 94%, 75% and 86% respectively from our CNN-SVM model.

Advisor

Nikolay Sirakov

Subject Categories

Mathematics | Physical Sciences and Mathematics

COinS