Sparse Representation and Variational Methods in Retinal Image Processing

Abstract

Relations between different types of cameras used for retinal imaging were studied with the purpose of improving the quantitative precision of the imaging data (used for diagnostics and medical research). Based on the differences in visual quality and quantitative parameters, we designed analytical models of the effects that cameras introduce into the retinal data and described possible ways of digital post-processing. Some processing tasks involve detection and separation of features (such as the retinal microvessels) prior to subsequent analysis of underlying retinal pathology. Mathematical techniques for feature detection and inpainting are variational, implemented via numerically stable gradient descent schemes. Other tasks involve the estimates of translation - invariant sparse image coefficients allowing to separate the background and significant scales of the image from the texture-like auxiliary information. The above techniques are based on the recent work on the wavelet Ginzburg-Landau energy and methods of adaptive thresholding of the stationary wavelet transform coefficients. We consider algorithms with partial specialist supervision and deliberate choice of processing methods for different eye areas as well as separate processing of healthy vs. pathological eye data.