RRI-Net: Classification of Multi-class Retinal Diseases with Deep Recurrent Residual Inception Network using OCT Scans

Optical coherence tomography (OCT) is a label-free, non-invasive imaging technique that is widely used in the diagnosis of various ophthalmic diseases. The diagnostic information related to these diseases is embodied in the texture and geometric features of the OCT scans, which are used by the retinal experts for interpretation and classification. However, due to the large number of OCT scans obtained every day, doctors and hospital staff are unable to meaningfully examine the potential retinal pathological conditions (RPCs), resulting in unexpected delays in the diagnosis and treatment of RPCs. In this paper, we propose an automated deep recurrent residual inception network, RRI-Net, for the classification of retinal OCT scans into diagnostically relevant classes, including healthy, age-related macular degeneration (AMD), diabetic macular edema (DME) and choroidal neovascularization (CNV). The proposed RRI-Net employs residual connections with cascaded multi-kernel convolutions to provide optimal training and classification results. In addition, we conducted extensive training of RRI-Net using 108,312 OCT scans, and tested the performance of the proposed framework over 1,000 OCT scans. The results show that RRI-Net achieves 98.8% accuracy in multi-class classification problem between healthy, AMD, DME and CNV, with 97.6% true positive rate and 99.2% true negative rate, outperforming other state-of-the-art methods.