Fluid Segmentation in OCT with an Improved Convolutional Neural Network

Abstract

Multi-scale pathological fluid segmentation is of great importance for the diagnosis and treatment of various eye diseases such as neovascular age-related macular degeneration (nAMD) and diabetic macular edema (DME). Despite significant progress in recent years, there are still several important issues remain unsolved. First, abnormal fluid lesions in optical coherence tomography (OCT) show large variations in location, size, and shape. Second, fluid lesions are contiguous regions with smooth surfaces and without holes inside. In this study, we introduce an adapted fully convolutional neural network (FCN) architecture to improve the ability of the network to extract multi-scale fluid lesions in OCT. Then we introduced a novel curvature loss term to regularize the shape prior in the loss function. The proposed method was extensively evaluated on RETOUCH dataset with a mean Dice score (DSC) of 0.767 and mean absolute volume difference (AVD) of 0.036 mm3, which improved significantly compared with the state-of-the-art methods.