SEADNet: Deep learning driven segmentation and extraction of macular fluids in 3D retinal OCT scans

In ophthalmology, symptomatic exudate-associated derangement (SEAD) lesions play an important role in the timely intervention and treatment of maculopathy. Optical coherence tomography (OCT) imaging, due to its ability to visualize early symptoms linked with chronic retinal conditions, is mainly used for screening maculopathy and related SEAD lesions. However, in OCT scans, the inter- and intra-observer variability of manual estimation of SEAD lesions is high, which may lead to serious inconsistencies in the treatment of macular diseases. In this context, an automated SEAD segmentation algorithm can be regarded as a feasible approach. This paper proposes a novel deep encoder-decoder architecture called SEADNet, that performs the joint segmentation and extraction of three SEAD lesions including intraretinal fluid (IRF), subretinal fluid (SRF), and pigment epithelial detachment (PED). SEADNet comprises of three main modules, namely feature encoder, feature decoder and a newly introduced extractor module that further extracts the multi-scale enriched features of candidate SEAD lesions. The proposed framework is trained using 7064 OCT scans and tested over 4270 OCT scans acquired from three different OCT imaging devices. The simulation results show that the segmentation performance of SEADNet is better than the existing algorithms, with mean dice scores of 0.909, 0.913 and 0.918 for IRF, SRF and PED, respectively.