Computer-aided diagnosis of colour retinal imaging

Abstract

Color images in retina that captured by digital fundus reveal the systemic diseases and offer a harmless way to detect retinopathy that will affect the retina. For retinal images, automated segmentation of specific regions can help ophthalmologists screen larger population. The identification of certain lesions and main regions is able to provide analysis of diseases. This thesis firstly presents to investigate the vessel segmentation of retinal images which is substantially useful for diagnosing different ophthalmologic diseases. The second part proposes a novel system for segmenting several lesions of fundus based on the limited data. Automated vessel segmentation is challenging since the width of retinal vessels could vary in a range. The local intensity and contrast of vessels can be too weak to be detected due to surrounding pathologies of lesions. Instead of current segmentation methods that utilize a vanilla upsample module to retain features for segmentation, our proposed network promotes the U-shape framework in a hierarchical structure by adding a novel multi-scale upsample attention (MSUA) module. An end-to-end nested U-shape framework with innovative attention mechanism is adopted in order to make use of features from cross-layers for discriminating vessels. To further enhance the information flow in the network, the skip connection is replaced with nested connection for feature reuse. Through concatenating mutual connection among multi-stages, the network is able to store rich details of vessels. The cross-testing and separate-testing both demonstrate a state-of-­the-art performance in comparison with other methods. The retinal lesions for DR diagnosis include red lesions (microaneurysms and haemorrhages) and bright lesions (soft exudates and hard exudates). However, common approaches tend to detect mere one single abnormality. Moreover, it is challenging to adopt a data-driven algorithm to achieve fast and reliable lesions localization with limited annotation. To avoid depending on pixel-wise annotation of pathology, we introduce a novel framework to detect various lesions by training a image-level classification network. Through multi-level features and classification score, the proposed network utilizes attention mechanism that is beneficial of discriminative regions. The classifier is regularized by mixing images with refined labels to promote the sensitivity of implicit objects (e.g. small blobs of microaneurysms). The experiment shows that it obtains promising results in comparison with other weakly-supervised methods.