Rapid and accurate classification of retinal OCT diseases using reparameterized one-dimensional convolutional networks

Abstract

Efficient and precise classification of retinal optical coherence tomography (OCT) images is crucial for accurate diagnosis of eye diseases. However, existing classification networks often struggle with balancing inference speed and accuracy. To address this, we propose a novel retinal disease classification network that leverages prior knowledge of OCT images. By employing structural reparameterization and transforming the convolutional kernel shape to 1D, our network enhances its ability to focus on the inherent layering information of OCT images. Experimental results demonstrate that our approach significantly improves inference speed while maintaining high classification accuracy, compared to conventional and state-of-the-art networks. This advancement addresses real-time diagnostic needs in clinical settings. Our source code is available at: https://github.com/xunlizhinian1124/1D-OCT-Classification.