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

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate segmentation of corneal layer interfaces in optical coherence tomography (OCT) images is crucial for generating precise corneal thickness maps.
• Corneal thickness mapping aids in the diagnosis of various eye conditions.

Purpose of the Study:

• To introduce PIPE-Net, a novel fully convolutional neural network designed for segmenting four key corneal layer interfaces.
• To evaluate the performance of PIPE-Net against existing methods using a manually segmented OCT dataset.

Main Methods:

• PIPE-Net utilizes a pyramidal input, parallel encoders for a larger receptive field, and a densely connected decoder with residual summations.
• The network employs a linear growth rate for feature maps to minimize parameters, enabling training on a small dataset (295 OCT images).
• K-fold cross-validation and precision-recall curves with average precision were used for performance evaluation.

Main Results:

• PIPE-Net achieved a superior average precision of 0.95, outperforming other implemented networks.
• The proposed method accurately detected and smoothed corneal layer interfaces, closely matching expert segmentations.
• The network's efficiency in parameter usage allowed effective training on a limited dataset.

Conclusions:

• PIPE-Net demonstrates high accuracy and efficiency for corneal layer segmentation in OCT images.
• The developed network offers a promising tool for objective corneal thickness analysis and diagnosis.
• The findings suggest PIPE-Net's potential to advance quantitative ophthalmological assessments.