Automatic Classification of Volumetric Optical Coherence Tomography Images via Recurrent Neural Network

Automatic and accurate classification of retinal optical coherence tomography (OCT) images is essential to assist ophthalmologists in the diagnosis and grading of macular diseases. Most existing methods classify 3-D retinal OCT volumes by separately analyzing each single-frame 2-D B-scan, and thus inevitably ignore significant temporal information among B-scans. In this paper, we propose to classify volumetric OCT images via a recurrent neural network (VOCT-RNN) which can fully exploit temporal information among B-scans. Specifically, a deep convolutional neural network is first utilized to automatically extract highly representative features from each individual B-scan of the 3-D retinal OCT images. Then, a long short-term memory network is employed to model the temporal dependencies among B-scans and achieve volumetric OCT classification. The proposed VOCT-RNN can be directly learned from volume-level labels, requiring no detailed annotations at each B-scan. Experimental results on two clinically acquired OCT datasets demonstrate the effectiveness of the proposed VOCT-RNN for volumetric retinal OCT image classification.