MLU-Net: Efficient Segmentation for Retinal Layers In Optical Coherence Tomography Images

Abstract

The automatic segmentation of retinal images obtained by optical coherence tomography is increasingly important for ophthalmologists to diagnose and monitor many kinds of ophthalmic diseases. U-Net is the most widely used deep learning network in retinal segmentation, but the limited number of data-flow paths made it hard to capture complex features. We proposed here an optimized Mobile Ladder U-Net (MLU-Net), which consists of a Ladder Connection for increasing the network’s data-flow paths and a depthwise separable convolution for reducing the model’s parameters. Experiments on 100 B-scans from 10 human eyes demonstrated that the 9 retinal layer boundaries can be segmented accurately with the MLU-Net. In addition, compared with the original U-Net and LadderNet, the segmentation result of our method is closest to the expert label.