Keywords

Segmentation, Classification, Glaucoma

Abstract

Medical imaging and related technologies have been widely used in the medical field. As a key to auxiliary diagnosis and treatment, medical image processing and analysis technology has important clinical and research value. Fundus images, obtained by a fundus camera, are crucial for diagnosing fundus diseases such as diabetes, hypertension, macular degeneration, and fundus arteries.

This dissertation focuses on fundus image acquisition, preprocessing, blood vessel segmentation, classification, and the impact of blood vessels on image classification. Precision in identifying micro and macro blood vessels in the retina is crucial for diagnosing retinal diseases but poses significant challenges. Current autoencoding-based segmentation approaches have limitations due to resolution loss during encoding and the inability to recover information during decoding. These limitations restrict the extraction of retinal microvascular structures.

To address this issue, we introduce Swin-Res2-Net, a specialized module designed to enhance retinal vessel segmentation precision. Swin-Res2-Net uses the Swin transformer, which employs shifted windows to reduce network complexity compared to traditional transformers by limiting self-attention calculations to non-overlapping windows. It also incorporates fusion with the Res2Net architecture, leveraging multi-scale techniques to enlarge the receptive field and extract additional semantic information. This combination enhances the localization and separation of retinal microvessels. Additionally, we added a module to eliminate redundant information between encoding and decoding steps, resulting in superior performance compared to other models.

Furthermore, we propose a novel approach to classify and recognize glaucoma using the Res2S Net architecture. The model extracts features using Swin Transformer and Res2Net coding blocks. By integrating these components, the model excels in complex tasks, particularly those requiring multi-scale features and global contextual information.

These approaches aim to enhance vessel extraction and segmentation for improved image classification, ultimately promoting accurate and timely diagnosis and management of eye diseases in health facilities, leading to better health outcomes for patients.

Completion Date

2024

Semester

Summer

Committee Chair

Shunpu Zhang

Degree

Doctor of Philosophy (Ph.D.)

College

College of Sciences

Department

Statistics and Data Science

Degree Program

Big Data Analytics

Format

application/pdf

Identifier

DP0028896

Language

English

Rights

In copyright

Release Date

2-15-2030

Length of Campus-only Access

5 years

Access Status

Masters Thesis (Campus-only Access)

Campus Location

Orlando (Main) Campus

Accessibility Status

Meets minimum standards for ETDs/HUTs

Download

Restricted to the UCF community until 2-15-2030; it will then be open access.

Share

COinS