Retinal Microvascular Feature Extraction Using Faster Region-based Convolutional Neural Network

Artificial Intelligence (AI) more specifically Deep Learning (DL) incorporating with image processing is being employed widely to solve different refractory problems by academia and industry from the ophthalmology discipline. The microvascular structure of the human retina shows remarkable abnormali...

Full description

Saved in:
Bibliographic Details
Main Author: Mohammed Enamul, Hoque
Format: Thesis
Language:English
Published: 2021
Subjects:
Online Access:http://ir.unimas.my/id/eprint/36563/3/Mohammed%20Enamul%20Hoque%20ft.pdf
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Artificial Intelligence (AI) more specifically Deep Learning (DL) incorporating with image processing is being employed widely to solve different refractory problems by academia and industry from the ophthalmology discipline. The microvascular structure of the human retina shows remarkable abnormalities responding to different kinds of hazardous ophthalmic and cardiovascular diseases. The high dimensionality and complex hierarchical microvascular structure of the human retina, and random retinal image accumulation create enormous size data. This scenario is offering the challenge of understanding and managing retinal image data. The original input data need to be projected into output data which has a smaller number of features whilst as much as possible preserving its native information. This process is known as feature extraction. A recently introduced DL approach, Convolutional Neural Network (CNN), is dedicated to extracting and quantifying the complex hierarchical image features with more abstraction. The supervised CNN methods employ different algorithms that iteratively learn from data for analyzing data and predicting outcomes. The implementation of CNN methods has proved their efficiency in the identification, localization, and quantification of interesting retinal image features such as exudates, microaneurysms. These features are considered remarkable signs for detecting Diabetic Retinopathy (DR), Hypertensive Retinopathy (HR), and stroke. The quantitative features such as vessel widening and deviation in bifurcation angle are also relative to these diseases. The recently reported DL-based retinal image feature extraction methods are not dedicated to extracting retinal vessel segments from multiple locations of the retinal image. Extracting retinal vessel segments from the retinal image is important for vessel diameter and bifurcation angle quantification. Moreover, employing inappropriate image processing techniques at the pre-processing level can lead to poor system performance. This work is dedicated to developing an image processing-based AI method for retinal vessel extraction from retinal images. This thesis includes a brief explanation of the proposed method, Faster Region-based Convolutional Neural Network (Faster RCNN) for retinal image feature extraction. At the initial stage of this proposed method, fundamental image processing was used for retinal image preprocessing. The retinal images were taken from the different public databases to train, test, and validate the performance of this proposed method. This proposed method obtained 91.82% Mean Average Precision (mAP), 92.81% sensitivity, and 63.34% Positive Predictive Value (PPV). According to the performance analysis, it can be expected to integrate this proposed method into the ophthalmic diagnostic tools after further development, evaluation, and validation