Automatic Detection of Diabetic Eye Disease Through Deep Learning Using Fundus Images

[img]
Preview
SARKI_Rubina-thesis_nosignature.pdf - Submitted Version (6MB) | Preview

Sarki, Rubina ORCID: 0000-0001-5018-9567 (2021) Automatic Detection of Diabetic Eye Disease Through Deep Learning Using Fundus Images. PhD thesis, Victoria University.

Abstract

Diabetes is a life-threatening disease that affects various human body organs, including eye retina. Advanced Diabetic Eye disease (DED) leads to permanent vision loss; thus, early detection of DED symptoms is essential to prevent disease escalation and timely treatment. Studies have shown that 90% of DED cases can be avoided with early diagnosis and treatment. Ophthalmologists use fundus images for DED screening to identify the relevant DED lesions. Due to the growing number of diabetic patients, it is becoming unaffordable for the volume of fundus images to be manually examined. Moreover, changes in the eye anatomy during its early stage are frequently untraceable by human eye due to subtle nature of the features, and a large volume of fundus images puts a significant strain on limited specialist resources, rendering manual analysis practically infeasible. Therefore, considering the popularity of deep learning in real-world applications, this research scrutinized deep learning-based methods to facilitate early DED detection and address the issues currently faced. Despite promising results on the binary classification of healthy and severe DED, highly accurate detection of early anatomical changes in the eye using Deep Learning remains a challenge in wide-scale practical application. Similarly, all previous fundus retinal image classification studies assigned a multi-class classification problems are still a challenge in Deep Learning. While studies conducted in the past have released high classification performance outputs managed by hyper- parameters settings, applying the binary classification model to the actual clinical environment in which visiting patients suffer from different DED diseases is technically tricky. Nevertheless, mild and multi-class DED classification aimed studies have been very minimal. Furthermore, it is observed that previous researches lack in addressing the development of automated detection of early DED, jointly in one system. Detection of DED in one system is considered to be essential for treatment in terms of specific lesions. Identification of the abnormalities in that specific retinal region can provide specific treatment to the target region of the eye, which is mostly affected. In this thesis, we explore different novel Deep Learning methods for automated detection of early (healthy and one mild) and multi-class (three or more) DED employing retinal fundus images. For this purpose, we explore transfer learning based models and build a new convolutional neural network method in automatic feature extraction and classification, based on deep neural networks. To develop an enhanced system certain number of original deep learning approach has been combined with various other advanced techniques such as: (i) image pre-processing, (ii) data augmentation, (iii) DED feature extraction and segmentation (iv) model fine-tune, and (v) model optimization selection. Therefore, the results of the analysis of several retinal image features demonstrate that deep learning can attend a state-of-the-art accuracy for early DED diagnosis.

Item type Thesis (PhD thesis)
URI https://vuir.vu.edu.au/id/eprint/42641
Subjects Current > FOR (2020) Classification > 4602 Artificial intelligence
Current > Division/Research > College of Science and Engineering
Keywords diabetes; eye; diabetic eye disease; DED; fundus images; deep learning
Download/View statistics View download statistics for this item

Search Google Scholar

Repository staff login