Loading...
1 results
Search Results
Now showing 1 - 1 of 1
- Computational Analysis of Fundus Images: Rule-Based and Scale-Space ModelsPublication . Soares, Ivo Miguel da Fonseca Gravito; Pinheiro, António Manuel Gonçalves; Sousa, Miguel Castelo Branco Craveiro deFundus images are one of the most important imaging examinations in modern ophthalmology because they are simple, inexpensive and, above all, noninvasive. Nowadays, the acquisition and storage of highresolution fundus images is relatively easy and fast. Therefore, fundus imaging has become a fundamental investigation in retinal lesion detection, ocular health monitoring and screening programmes. Given the large volume and clinical complexity associated with these images, their analysis and interpretation by trained clinicians becomes a timeconsuming task and is prone to human error. Therefore, there is a growing interest in developing automated approaches that are affordable and have high sensitivity and specificity. These automated approaches need to be robust if they are to be used in the general population to diagnose and track retinal diseases. To be effective, the automated systems must be able to recognize normal structures and distinguish them from pathological clinical manifestations. The main objective of the research leading to this thesis was to develop automated systems capable of recognizing and segmenting retinal anatomical structures and retinal pathological clinical manifestations associated with the most common retinal diseases. In particular, these automated algorithms were developed on the premise of robustness and efficiency to deal with the difficulties and complexity inherent in these images. Four objectives were considered in the analysis of fundus images. Segmentation of exudates, localization of the optic disc, detection of the midline of blood vessels, segmentation of the vascular network and detection of microaneurysms. In addition, we also evaluated the detection of diabetic retinopathy on fundus images using the microaneurysm detection method. An overview of the state of the art is presented to compare the performance of the developed approaches with the main methods described in the literature for each of the previously described objectives. To facilitate the comparison of methods, the state of the art has been divided into rulebased methods and machine learningbased methods. In the research reported in this paper, rulebased methods based on image processing methods were preferred over machine learningbased methods. In particular, scalespace methods proved to be effective in achieving the set goals. Two different approaches to exudate segmentation were developed. The first approach is based on scalespace curvature in combination with the local maximum of a scalespace blob detector and dynamic thresholds. The second approach is based on the analysis of the distribution function of the maximum values of the noise map in combination with morphological operators and adaptive thresholds. Both approaches perform a correct segmentation of the exudates and cope well with the uneven illumination and contrast variations in the fundus images. Optic disc localization was achieved using a new technique called cumulative sum fields, which was combined with a vascular enhancement method. The algorithm proved to be reliable and efficient, especially for pathological images. The robustness of the method was tested on 8 datasets. The detection of the midline of the blood vessels was achieved using a modified corner detector in combination with binary philtres and dynamic thresholding. Segmentation of the vascular network was achieved using a new scalespace blood vessels enhancement method. The developed methods have proven effective in detecting the midline of blood vessels and segmenting vascular networks. The microaneurysm detection method relies on a scalespace microaneurysm detection and labelling system. A new approach based on the neighbourhood of the microaneurysms was used for labelling. Microaneurysm detection enabled the assessment of diabetic retinopathy detection. The microaneurysm detection method proved to be competitive with other methods, especially with highresolution images. Diabetic retinopathy detection with the developed microaneurysm detection method showed similar performance to other methods and human experts. The results of this work show that it is possible to develop reliable and robust scalespace methods that can detect various anatomical structures and pathological features of the retina. Furthermore, the results obtained in this work show that although recent research has focused on machine learning methods, scalespace methods can achieve very competitive results and typically have greater independence from image acquisition. The methods developed in this work may also be relevant for the future definition of new descriptors and features that can significantly improve the results of automated methods.