ARTIFICIAL INTELLIGENCE MOBILE DERMOSCOPY: A METHOD FOR CLASS IMBALANCE MANAGEMENT

Abstract

Artificial intelligence can help analyze medical images to diagnose skin-related diseases such as melanoma. This thesis presents an end-to-end framework for detecting melanoma in real-time on mole images acquired through mobile devices equipped with magnification lenses. We trained our models with public domain ISIC-2019 and ISIC-2020 datasets using EfficientNet convolutional neural networks. Our aim in this work is to reduce the problem of class imbalance. As a result, we integrated the standard training model with data balance schemes that use oversampling, undersampling, and RCL loss function. We introduce a blurring technique that emulates aberrations caused by magnifying lenses to apply the undersampling method. In addition, we used a novel loss function that incorporates the cost difference between false positive (melanoma misses) and false negative (benign misses) predictions. Our results show significant progress in the AUC scale with 98.64% and an accuracy of 96.91%.

Date of Award	2022
Original language	American English
Awarding Institution	HBKU College of Science and Engineering