Diabetic retinopathy detection using fusion of textural and optimized convolutional neural network features / Uzair Ishtiaq

Uzair , Ishtiaq (2024) Diabetic retinopathy detection using fusion of textural and optimized convolutional neural network features / Uzair Ishtiaq. PhD thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (231Kb)
	PDF (Thesis PhD) Restricted to Repository staff only until 31 December 2026. Download (1879Kb)

Abstract

One of the most prevalent chronic conditions that can result in permanent vision loss is Diabetic Retinopathy (DR). The diabetic retinopathy can broadly be categorized as Non-Proliferative DR (NPDR) and Proliferative DR (PDR) and it occurs in five stages: no DR, mild, moderate, severe, and proliferative DR. Early detection of DR is essential for the diabetic patients to prevent vision loss. DR can be detected either manually by an Ophthalmologist or using an automated system. Usually, DR can have mild signs which are negligible and very hard for an ophthalmologist to diagnose, making it difficult to be categorized in its particular class. However, an automated system is capable enough to distinguish even mild signs of DR by extracting salient and discriminative features from retinal images. In this study, a method for the detection and classification of DR stages is proposed to determine whether it is in any of the non-proliferative stage or the proliferative stage. The hybrid approach based on image preprocessing and fusion of features is the foundation of the proposed classification method. The preprocessing steps involved in this study include image resizing, data augmentation, applying median filter and image sharpening. A Convolutional Neural Network (CNN) model was created from scratch for this study. Combining Local Binary Patterns (LBP) based texture features and deep learning features resulted in the creation of the fused features vector which was then optimized using Binary Dragonfly Algorithm (BDA) and Sine Cosine Algorithm (SCA). Moreover, this optimized feature vector was fed as input to the machine learning classifiers including SVM (Linear, Quadratic, Fine Gaussian, Medium Gaussian, and Coarse Gaussian) and KNN (Fine, Medium, Coarse, Cosine and Weighted). SVM classifier achieved the highest classification accuracy of 98.85% on a publicly available dataset i.e., Kaggle EyePACS. Rigorous testing and comparisons with state-of-the-art approaches in the literature indicate the effectiveness of the proposed methodology and it can widely be applied to different DR datasets in future.

Actions (For repository staff only : Login required)