PIPE-Net: A pyramidal-input-parallel-encoding network for the segmentation of corneal layer interfaces in OCT images.

ABSTRACT

Segmentation of corneal layer interfaces in optical coherence tomography (OCT) images is necessary to generate thickness maps used for cornea diagnosis. In this paper, we propose PIPE-Net, a fully convolutional neural network with a pyramidal input, parallel encoders, and a densely connected decoder to segment four corneal layer interfaces. The pyramidal input is encoded using parallel encoders, which allows the network to process a larger receptive field. The encoders are connected level-wise to the decoder through residual summations. The decoder is densely connected using residual summations between its levels to enhance the gradient flow. We use a linear growth rate for the number of feature maps to limit the network parameters, which allows the network to be trained using a small dataset. A dataset of 295 OCT images was obtained and manually segmented by experienced and trained operators. We implemented other related networks in the literature for comparison with our proposed network. We performed k-fold cross-validation to evaluate all the networks, and their performance was evaluated using precision-recall curves and average precision. PIPE-Net outperformed the other networks with an average precision of 0.95. The layer interfaces were detected and smoothed using the Savitzky-Golay filter, and they were closer to the expert.