Non-Invasive Image Processing and Deep Learning Methods for Retinal Vessels Segmentation for Early Detection of Eye Diseases

Automatic analysis of medical images is essential to diagnose eye disease progress. Fundus retinal images are used for observation of disease progression. There are two modes in which the fundus camera operates to capture retinal images the angiogram mode and colour-filtered fundus
images. Colour-filtered fundus images suffer from various issues, such as varying and low contrast, uneven illumination and noise, which makes it difficult to analyse the retinal vessels. Image-enhancement techniques are required to establish the impact of disease on the segmentation of retinal blood vessels. There is also a need for novel steps based on image processing and machine learning techniques to achieve well-defined retinal blood vessel segmented images for quick analysis of the disease so timely treatment can be recommended.

This thesis presents the implementation of a new image-enhancement model based on morphological operations along with stationary wavelet transform for retinal fundus images and contrast limited adaptive histogram equalisation for vessel enhancement. The main aim of the proposed
image-enhancement technique is to compare the contrast improvement of colour fundus against fundus fluorescein angiography images. The findings experimentally prove that image contrast enhancement techniques perform much better than fundus fluorescein angiography images,
while giving comparable contrast levels.

For more validation, this research implemented a contrast-sensitive steps-based image contrast enhancement technique. The main purpose of this technique is to observe its performance on pathological images because of the difficulty in enhancing vessels from challenging images. The proposed image-enhancement technique achieved a contrast improvement factor of 5.7. It also increased segmentation of retinal blood-vessel performance by accuracy from 84.4% to 95%, and sensitivity from 60.9% to 74.6%.

For further validation based on improvement of the methodology of the proposed technique, a novel image-enhancement technique is implemented and successfully validated. This newly proposed technique played a vital role in handling the issue of varying and low contrast and gave a well-segmented image. It contained independent component analysis architecture II that is used to enhance retinal blood vessels against their background. The performance of retinal vessel segmentation improved when employing the independent component analysis architecture-II based image-enhancement technique, giving a better result compared to existing methods. An image enhancement technique improved the sensitivity of the segmentation of retinal blood vessels, and sensitivity is increased by detecting more retinal vessels, especially tiny ones.

Contrast normalisation step-based methods are implemented to boost the sensitivity of retinal vessels. The impact of these additional steps is assessed on the retinal database. It outperformed existing methods, achieving the required sensitivity of around 75% through image-processing techniques and detecting more tiny vessels. Further boosting sensitivity and overall performance, a deep convolutional neural network and stride convolutional neural network-based methods for segmentation of retinal vessels is implemented. The main purpose of implementing these deep learning methods is to observe the performance improvement of segmentation of retinal blood vessel in comparison with image processing techniques.

A fully convolutional neural network model along with pre- and post-processing gives segmented vessel images with sensitivity up to 75%, leading to proper detection of tiny vessels with an accuracy of around 95%. However, there is room for improving this method because post-processing is adopted only for removing noisy pixels, so a well-trained convolutional neural network model is required that detects fine vessels without post-processing. Finally, this research proposed a stride-based convolutional neural network model without post-processing, achieving good vessel images. The impact of the implemented segmentation method is experimentally validated and achieved higher sensitivity of 87% with accuracy of 96.8%, which outperformed other existing methods.

These methods based on image processing and deep learning techniques for the detection of retinal vessels achieved good performance. The segmentation of retinal blood vessels has been studied for 34 years, and our proposed methods give much better performance in comparison to
those existing methods.