JIEB-linknet: an Improved Linknet with Joint Input Encoder Block for Segmentation of Retinal Layers and Fluid Accumulation in OCT Images

Age-related Macular Degeneration, Diabetic Retinopathy, Edema as well as Glaucoma are considered as leading ophthalmology diseases result in permanent blindness unless diagnosed earlier. Vision Impairment may be associated with factors such as hereditary, lifestyle and age. Structural changes and fluid formation in the retina helps to investigate the prognosis. Optical coherence Tomography helps to visualize the retinal microstructures in a non-invasive manner. Identification of disease progression is time demanding when these OCT biomarkers are segmented by human experts. Recent findings show that, Semantic segmentation in deep learning is effective for class imbalance problems. This research work proposes an improved LinkNet architecture to delineate retinal layers and fluids concerning OCT images. The primary intention concerning the proposed segmentation network is to categorize 7 retinal layers and fluid as distinct classes. Two Joint input encoder blocks are included in the architecture, which accept multiple inputs from prior layers in order to preserve high resolution characteristics for exact segmentation. In addition, the hybrid loss functions were utilized for improving the prediction accuracy. This model was assessed over the public Duke dataset and the outputs demonstrate that this model attains a dice coefficient of 0.9