Joint Loss-Based Multi-decoder Network for OCT Fluid Segmentation

Optical coherence tomography (OCT) is a popular and clinically viable tool for diagnosing ocular lesions in ophthalmology. In clinical practice, however, since Macular Edema (ME) segmentation of ocular OCT images is subjective, labor-intensive, and prone to error, it is essential to adopt computer-aided systems to help ophthalmologists perform ME segmentation. In this paper, we propose a novel Joint Loss-Based Multi-Decoder Network, namely MDNet, for OCT Fluid Segmentation. MDNet mainly consists of an encoder and three decoder modules, which are used as segmentation branch for label images, contour branch for edge label images, and diffusion branch for distance maps, respectively. A new loss function corresponding to such three modules is also devised for training. Experiments with a publicly available dataset are conducted to validate the effectiveness of MDNet, and results compared with the existing state-of-the-art methods demonstrate that MDNet is advantagous in achiving better segmentation results.