ORCID
https://orcid.org/0009-0006-0761-9561
Year
2024
Season
Fall
Paper Type
Master's Thesis
College
College of Computing, Engineering & Construction
Degree Name
Master of Science in Computer and Information Sciences (MS)
Department
Computing
NACO controlled Corporate Body
University of North Florida. School of Computing
First Advisor
Dr. Xudong Liu
Second Advisor
Dr. Indika Kahanda
Third Advisor
Dr. Karthikeyan Umapathy
Department Chair
Dr. Zornitza Prodanoff
College Dean
Dr. William Klostermeyer
Abstract
Media Haze (MH) is a condition that affects an individual’s quality of life by affecting their eyes. Current practice is to detect MH by manually examining retinal fundus (retinal) images. The analysis of images being used as the prevalent technique for identifying the MH condition strongly suggests that automation of this process may be possible. In recent years, machine learning, specifically computer vision, has allowed for the automation of tasks relating to image analysis. This ability to automate has also recently been shown in the medical field for some eye conditions and diseases. This thesis centers around the problem of the detection of MH not being effectively automated. Therefore, this thesis develops a system that can effectively and satisfactorily detect whether MH is present within a retinal image. Furthermore, this thesis accomplishes this by proposing a method of preprocessing retinal images, and examines the effectiveness of 10 deep learning models. LeNet-5, the modified form of LeNet, AlexNet, and Vision Transformer were trained from scratch, and MobileNet, ResNet152v2, DenseNet, NASNetMobile, EfficientNetV2 B0, and VGG-16 were all pretrained and fine-tuned on our dataset. Then, an ensemble of the 6 fine-tuned models was also examined. EfficientNetV2 B0, MobileNet, and ResNet, being the most effective independent models in testing, were then added, along with an ensemble, to the web application created for this thesis, which demonstrates a user-friendly potential real-world application of this work. When combined with the proposed preprocessing, MobileNet can achieve 96.1% accuracy, and EfficientNetV2 B0 can achieve 96.6% accuracy, both of which are better than the current state-of-the-art attempts at automating the detection of MH. The ensemble of fine-tuned models also has a 96.6% accuracy. The precision, recall, F1 score, and AUC results further evidence this improved performance.
Suggested Citation
O'Berry, Jonathan, "Media haze classification in retinal images using deep learning" (2024). UNF Graduate Theses and Dissertations. 1303.
https://digitalcommons.unf.edu/etd/1303
