PROJECT TITLE :

Deep Retinal Image Segmentation With Regularization Under Geometric Priors

ABSTRACT:

Ophthalmology relies on retinal picture vessel segmentation as a critical diagnostic tool. Low contrast, fluctuating artery size and thickness, and the existence of interfering pathologies such as micro-aneurysms and haemorrhages are some of the issues this topic encounters. Hand-crafted filters and morphological post-processing were used in early attempts to solve this challenge. Deep learning algorithms have recently been used to improve segmentation accuracy dramatically. Representation and residual task networks are the two components of a revolutionary domain-enriched deep network that learns geometric properties specific to retinal pictures. For each training set, the representation and task networks are taught together. The expected prior structure of these filters inspires two new limitations on these filters, which we suggest in order to produce physically meaningful and practically effective representation filters: A data-adaptive noise regularizer that penalises false positives and an orientation restriction that promotes geometric diversity of curvilinear features. Thin vessels can be accurately detected thanks to multi-scale extensions. On three challenging benchmark databases, the proposed prior guided deep network exceeds current best practises as assessed by popular evaluation measures, while being more efficient in terms of network size and time.