Date of Award

2-2-2024

Document Type

Thesis

School

School of Electrical & Electroncis Engineering

First Advisor

Dr.A.Srinivasan

Keywords

Diabetic Retinopathy, Image Processing, Convolutional Neural Network, Biological Sensor Network Model, Electro Retino Gram Signal

Abstract

Diabetes is a disorder that arises when blood sugar level increases. Insufficient secretion of insulin hormone is the ground for the evolution of diabetes, and it affects most of the critical organs in our body. Diabetes causes leakage of blood in the retinal blood vessels, inducing an eye disease, Diabetic Retinopathy (DR). The visual identification of micro features in fundus images makes the clinicians complex and challenging tasks. Therefore, timely diagnosis and proper treatment can prevent eye blindness. This research is proposed to identify and classify diabetic retinopathy using an Image Processing technique, Deep Convolutional Neural Network (DCNN), Biological Sensor Network Model (BSN), and the integrated method. Finally, the performances of all approaches are compared.

In the image processing technique, the workflow pre-processes the fundus images. Pre-processing comprises green channel extraction, wiener filtering, and contrast enhancement. Subsequently, the morphological operation is performed on the pre-processed image to extract retinal blood vessels. Graph cut image segmentation is performed to the resultant image to detect lesions, namely microaneurysms, hemorrhages, and exudates. From the segmented image, various features like statistical (Mean, Variance, and Standard Deviation), texture (GLDM), and Histogram of Oriented Gradients (HOG) features are extracted. Further, Cascaded Rotation Forest (CRF) classifier is trained with the extracted features. During the testing stage, the classifier detects the lesion and the severity level of diabetic retinopathy. Accuracy, sensitivity, and specificity are 98.00%, 96.00%, and 98.66%.

A deep convolutional neural network consists of many layers to facilitate extracting features and classifying fundus images into normal images, early-stage DR images (Non-Proliferative Diabetic Retinopathy), and advanced-stage DR images (Proliferative Diabetic Retinopathy). Furthermore, it classifies NPDR according to microaneurysms, hemorrhages, cotton wool spots, and exudates, and the presence of new blood vessels indicates PDR. The accuracy, sensitivity, and specificity of this approach are 98.88%, 96.66%, and 99.32%, respectively. The biological sensor model uses Electroretinogram signals (ERG) and the signal processing algorithm to detect diabetic retinopathy.

The workflow consists of pre-processing the recorded one-dimensional electroretinogram signals to remove noises. Fast Fourier Transform (FFT) extracts spectral information of signals. Then, Mel Frequency Cepstral Coefficients (MFCC) features are extorted. These features train the Support Vector Machine (SVM) classifier. During the testing stage, extraction of MFCC features is done for signal under testing. The classifier will predict and classify the input ERG signal from the extracted features into normal and diabetic retinopathy. The accuracy, sensitivity, and specificity of this approach are 98.50%, 97.00%, and 100%, respectively.

A novel integrated system using DCNN and BSN model is proposed to identify and classify diabetic retinopathy. A comparison of all the methods showed that the integrated method achieved a high accuracy rate of 99.33%, a sensitivity of about 98.80%, and a specificity of about 100%. Using fundus image and the Electroretinogram signal of a patient, it is possible to increase system effectiveness in detecting and classifying diabetic retinopathy.

Download

Share

COinS