Computer Aided Diagnosis in Ocular Fundus Imaging

The study of the retinal vessels plays a crucial role in many clinically relevant diseases such as systemic hypertension, arteriosclerosis and diabetes. In particular, diabetic retinopathy is the leading cause of new adult blindness. Thought diabetes can affect the eye in a number of ways, the fine network of blood vessels in the retina is usually involved - hence the term diabetic retinopathy. The blood vessel changes that occur in retinopathy are not visible without special instruments. Modern imaging devices such as the Scanning Laser Ophthalmoscope (SLO), in combination with computer analysis make the observation of the capillary macular network possible. In this work a computational approach for detecting and quantifying diabetic retinopathy is proposed. Particular attention has been paid to the study of Foveal Avascular Zone (FAZ). In fact, retinal capillary occlusion produces a FAZ enlargement. Moreover, the FAZ is characterized by qualitative changes showing an irregular contour with notchings and indentations. On this ground, our aim was the development of an automatic system for the quantitative morphological evaluation of the vascular lesions of the fundus oculi occurring in diabetic subjects. The study is mainly focused on the analysis of the FAZ (the degree of its enlargement and its shape change) and on the extraction of a proper set of features to quantify FAZ alterations in diabetic patients. In order to extract such parameters we propose a global system including four modules: enhancement (i.e. registration and integration), segmentation, description, feature extraction. Enhancement is a preprocessing module that is essential for a correct segmentation of the FAZ outline. Small capillaries surrounding the FAZ are not completely visible in a single fluorangiographic frame. The use of SLO allows the observation of several frames during the vascular diffusion of fluorescent die. Therefore, a more complete map of the capillary network can be obtained by integrating a set of consecutive frames. Such integration process requires two distinct phases: alignment, fusion. Different integration techniques have been investigated leading to the use of a temporal matched filter. An automatic segmentation procedure was then proposed to correctly identify the FAZ boundary. The observation of the particular anatomy of the FAZ prompted us to use a robust global segmentation method combining constraints derived from the image data with a priori knowledge about the position, size, and shape of this structure. The method was derived from the theory of active contours, along with genetic optimization. Snake optimization through genetic algorithms proved particularly useful in order to overcome problems related with initialization, parameter selection and local minima. Compared to current methods for segmenting the FAZ (manual selection or threshold methods), the proposed method offers high quantitative accuracy for the measurement of area and perimeter and we expect it will prove sufficiently robust in the aid to ophthalmological diagnosis. As to FAZ shape alterations, these are usually evaluated in a qualitative way by the physician. To obtain reliable quantitative description we referred back to the theory of moments. We used a set of region and boundary moments to obtain a FAZ description which is complete enough for diagnostic purposes and in order to assess the effectiveness of moment descriptors we performed several classification experiments to discriminate diabetic from non-diabetic subjects. In this context, we decided to use a neural network-based classifier with a particular emphasis on a class of networks we called Sparse Neural Network. Sparse Neural Networks are back-propagation networks whose topology is optimized for the specific problem with an automatic design procedure based on genetic algorithms. Moreover, the topology optimization procedure is able to perform feature selection at the same time. The optimization of neural networks through genetic algorithms made it possible to create networks that are optimized for the problem at hand, and at the same time has led to the automatic selection of the most significant parameters. The application of these networks to the specific problem has shown an interesting behaviour in determining a minimal subset of moments. We observed that the moments that are able to characterize shape asymmetries are the most useful features, and this confirms the qualitative observations that can be found in the medical literature. In addition, although potentially more sensitive to noise, the boundary moments appear to be well suited for describing the irregularities of the FAZ. This is supported by the results of our genetic optimization procedure.