Shell‐Net: A robust deep neural network for the joint segmentation of retinal fragments

Abstract

AbstractSegmentation of retinal fragments like blood vessels, optic disc (OD), and optic cup (OC) enables the early detection of different retinal pathologies like diabetic retinopathy (DR), glaucoma, etc. This article proposed a novel deep learning architecture termed as Shell‐Net for the accurate segmentation of the retinal fragments. The main novelty of the architecture relies on the intellectual fusion of two different styled networks for attaining better segmentation results. The lower part of the Shell‐Net (feature condenser) follows the down‐sampling and up‐sampling style, whereas the upper part of the network (feature amplifier) follows the up‐sampling and down‐sampling style of architecture. In addition to this, an additional residual module (feature stabilizer) is integrated with the network to achieve more spatial information from lower levels. The lower part of the network reduces the data through summarization, enabling much scope for precise extraction of heavy details such as thick vessels, OD, and OC. On the contrary, the upper part of the network augments the data using duplication, assisting in the enlargement of minuscule details such as the tiny vessels and boundaries. Experiments were performed on publicly available datasets like Digital Retinal Images for Vessel Extraction (DRIVE), Child Heart and Health Study in England (CHASE_DB1), Structured Analysis of the Retina (STARE), Online Retinal Fundus Image Dataset for Glaucoma Analysis and Research (ORIGA), DRISHTI‐GS1, and Retinal Image database for Optic Nerve Evaluation (RIMONE r1). The network accomplished an average accuracy (ACC) and specificity (SPE) of 0.96 and 0.98 when tested on DRIVE, CHASE_DB1, and STARE datasets respectively for vessel segmentation. Furthermore, it outperformed previously existing models in OD and OC segmentation by achieving an average accuracy of 0.98 with a specificity of 0.99 on the DRISHTI_GS1 dataset.