Comparative Study of Classification Methods for the Mitigation of Class Imbalance Issues in Medical Imaging Applications

Date of Award

2020

Degree Name

M.S. in Electrical Engineering

Department

Department of Electrical and Computer Engineering

Advisor/Chair

Advisor: Russell Hardie

Abstract

The rapid development and popularization of Machine Learning (ML) has paved theway to state of the art solutions in many domains, namely image-based applications suchas image classification, object detection and tracking to name a few. The medical imaging field is a bountiful source for image data with high potential for impacting the commongood. One glaring issue persists; most medical imaging datasets tend to have class imbal-ance. As a result, many ML computer aided detection (CAD) algorithms have surfaced tomitigate this issue. The focus of this work is to comparatively analyze a portion of themon multiple medical imaging datasets. Traditional Deep Learning (DL) classifiers are usedin one and two stage architectures as well as combined with Support Vector Machines. TheCIFAR10 dataset is utilized for benchmarking and determining the relationship betweenclassifier performance and class imbalance ratio. Performances vary across the datasets andalthough the two-stage architectures did not always have the highest overall accuracy, theyare warranted in specific class imbalance scenarios.

Keywords

Medical Imaging, Electrical Engineering, Computer Engineering, medical imaging, machine learning, deep learning, class imbalance, data augmentation, sampling methods, diabetic retinopathy, leukemia, CIFAR-10, CIFAR10, leukocyte, ensemble networks, neural networks, comparative study

Rights Statement

Copyright © 2020, author

Link to Full Text

Share

COinS