Multi-class retinal fluid joint segmentation based on cascaded convolutional neural networks

Abstract

Objective. Retinal fluid mainly includes intra-retinal fluid (IRF), sub-retinal fluid (SRF) and pigment epithelial detachment (PED), whose accurate segmentation in optical coherence tomography (OCT) image is of great importance to the diagnosis and treatment of the relative fundus diseases.Approach. In this paper, a novel two-stage multi-class retinal fluid joint segmentation framework based on cascaded convolutional neural networks is proposed. In the pre-segmentation stage, a U-shape encoder-decoder network is adopted to acquire the retinal mask and generate a retinal relative distance map, which can provide the spatial prior information for the next fluid segmentation. In the fluid segmentation stage, an improved context attention and fusion network based on context shrinkage encode module and multi-scale and multi-category semantic supervision module (named as ICAF-Net) is proposed to jointly segment IRF, SRF and PED.Main results. the proposed segmentation framework was evaluated on the dataset of RETOUCH challenge. The average Dice similarity coefficient, intersection over union and accuracy (Acc) reach 76.39%, 64.03% and 99.32% respectively.Significance. The proposed framework can achieve good performance in the joint segmentation of multi-class fluid in retinal OCT images and outperforms some state-of-the-art segmentation networks.

Keywords: convolutional neural network; medical image segmentation; optical coherence tomography.