Automatic Retinal Image Analysis to Triage Retinal Pathologies

Abstract

Fundus retinal imaging is a non-invasive way of imaging the retina popular among the ophthalmic community and the targeted population. Over the past 15 years, extensive research and clinical studies using fundus images have been done for automatizing the screening and diagnosing process of three significant conditions affecting vision: macular edema, diabetic retinopathy, and glaucoma. These are the most important causes of preventable blindness around the globe, yet they can be successfully screened using the fundus image of the retina. Such diseases are associated with an observable variation in the structural and functional properties of the retina. Manual triage/diagnosis of these diseases is time-consuming and requires specialized ophthalmologists/optometrists; it is also expensive. Computer-aided medical triage/diagnosis can be applied to fundus retinal image analysis, thereby automatizing the triage. The process involves successfully combining sub-tasks focused at analyzing, locating, and segmenting different landmark structures inside a retina. The preliminary objective of this thesis is to develop automatic retinal image analysis (ARIA) techniques capable of analyzing, locating, and segmenting the key structures from the fundus image and combine them effectively to create a complete automatic screening system. First, the retinal vessel, which is the most important structure, is segmented. Two methods are developed for doing this: the first uses adaptive histogram equalization and anisotropic diffusion filtering, followed by weighted scaling and vessel edge enhancement. Fuzzy-C-mean classification, together with morphological transforms and connected component analysis, is applied to segment the vessel pixels. A second improved method for vessel segmentation is proposed, which is capable of segmenting the tiny peripheral vessel pixels missed by the first method. This method uses curvelet transform-based vessel edge enhancement technique followed by modified line operator-based vessel pixel segmentation. Second, a novel technique to automatically detect and segment important structures such as optic disc, macula, and fovea from a retinal image is developed. These structures, together with the retinal vessels, are considered as the retinal landmarks. The proposed method automatically detects the optic disc using histogram-based template matching combined with the maximum sum of vessel information. The optic disc region is segmented by using the Circular Hough Transform. For detecting fovea, the retinal image is uniformly divided into three horizontal stripes, and the strip including the detected optic disc, is selected. The contrast of the horizontal strip containing the optic disc region is then enhanced using a series of image processing steps. The macula region is first detected in the optic disc strip using various morphological operations and connected component analysis. The fovea is located inside this detected macular region. Next, an algorithm capable of analyzing the retinal image quality and content is developed. Often, methods focusing on ARIA use public retinal image databases for performance evaluation. The quality of images in such databases is often not evaluated as a pre-requisite for ARIA. Therefore, the performance metrics reported by such ARIA methods are inconsistent. Considering these facts, a deep learning-based approach to assess the quality of input retinal images is proposed. The method begins with a deep learning-based classification that identifies the image quality in terms of sharpness, illumination, and homogeneity, followed by an unsupervised second level that evaluates the field definition and content of the image. The proposed method is general and robust, making it more suitable than the alternative methods currently adopted in clinical practice. Finally, an automatic deep learning-based method for clinically significant macular edema triage is proposed. The classified high-quality retinal images are used as inputs. Both full image and ARIA processed image are experimented as the possible inputs. Deep convolutional neural networks are used as feature extractors. The extracted features are over-sampled to balance the highly skewed database samples across the examined classes. Finally, using the reduced feature set obtained through feature selection, a simple k-NN classifier demonstrates significant classification performance, thereby validating the preliminary objective of this study.