An improved LinkNet-based approach for Retinal Image Segmentation

The characteristics of Retinal Blood Vessels helps identify various eye ailments. The proper localization, extraction and segmentation of blood vessels are essential for the treatment of the eye. Manual segmentation of blood vessels may be error-prone and inaccurate, leading to difficulty in further treatment, and causing problems for both operators and ophthalmologists. We present a novel method of semantic segmentation of Retinal Blood Vessels using Linked Networks to account for spatial information that is lost during feature extraction. The implementation of the segmentation technique involves using Residual Networks as a feature extractor and Transpose Convolution and Upsample Blocks for image-to-image translation thereby giving a segmentation mask as an output. The use of Upsample Blocks arises from its ability to give noise-free output while 18 layered Residual Networks using skip connections are used in the feature extraction without the vanishing gradient issues. The main feature of this architecture is the links between the Feature Extractor and the Decoder networks that improve the performance of the network by helping in the recovery of lost spatial information. Training and Validation using the Pytorch framework have been performed on the Digital Retinal Images for Vessel Extraction (DRIVE) Dataset to establish quality results.