Retinal layer segmentation in rodent OCT images: Local intensity profiles & fully convolutional neural networks

Area of Science:

• Biomedical Imaging
• Ophthalmology
• Computational Biology

Background:

• Optical coherence tomography (OCT) is crucial for monitoring retinal layer status in humans and animal models.
• Automated OCT analysis in rats is vital for preclinical drug toxicity assessments.
• Accurate segmentation of retinal layers in rat OCT images is essential for reliable data.

Purpose of the Study:

• To present two novel automated methods for detecting significant retinal layers in rat OCT images.
• To evaluate the performance and accuracy of these segmentation approaches.
• To provide a publicly available dataset for future research and comparisons.

Main Methods:

• Approach 1: Combines local horizontal intensity profiles with a novel watershed transformation variant.
• Approach 2: Utilizes an encoder-decoder convolutional neural network architecture.
• Both methods were validated on a rat OCT image database.

Main Results:

• Approach 1 achieved an averaged absolute distance error of 3.77 ± 2.59 µm.
• Approach 2 achieved an averaged absolute distance error of 1.90 ± 0.91 µm on the initial batch.
• The deep learning method achieved 2.67 ± 1.25 µm on an unseen dataset, demonstrating generalizability.

Conclusions:

• The first approach is competitive, outperforming commercial software and aiding ground truth generation.
• The deep learning approach surpasses conventional methods and state-of-the-art techniques in accuracy.
• The proposed deep learning network demonstrates robust generalization capabilities for new rat OCT images.