A fast multiresolution approach useful for retinal image segmentation

Retinal diseases such as retinopathy of prematurity (ROP), diabetic and hypertensive retinopathy present several deformities of fundus oculi which can be analyzed both during screening and monitoring such as the increase of tortuosity, lesions of tissues, exudates and hemorrhages. In particular, one of the first morphological changes of vessel structures is the increase of tortuosity. The aim of this work is the enhancement and the detection of the principal characteristics in retinal image by exploiting a non-supervised and automated methodology. With respect to the well-known image analysis through Gabor or Gaussian filters, our approach uses a filter bank that resembles the “à trous” wavelet algorithm. In this contribution we show a particular approach to speed-up the computing time. This methodology rotates the kernels and it is a fast enough to extract information useful to assess vessel tortuosity and to segment (not considered explicitly in this paper) retinal images. Furthermore, we compare on the public databases DRIVE and DIARETDB0 our output images against the SCIRD-TS algorithm, which is considered as one of the most effective supervised methods for the detection of retinal thin structures