Fully Convolutional Boundary Regression for Retina OCT Segmentation

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate retinal layer segmentation in optical coherence tomography (OCT) is crucial for disease monitoring.
• Current methods often rely on graph-based approaches, which can be inflexible and time-consuming.
• These methods struggle with maintaining correct topology and smooth surfaces.

Purpose of the Study:

• To develop a novel, automated deep learning framework for retinal layer segmentation.
• To achieve smooth, continuous surfaces with guaranteed topological correctness in a single feed-forward pass.
• To improve the efficiency and flexibility of retinal image analysis for disease progression monitoring.

Main Methods:

• A deep network directly models surface position distributions using a differentiable soft argmax.
• A specialized topology module is integrated for training and testing to ensure surface hierarchy.
• An additional output branch predicts pixel-wise lesions and layers.

Main Results:

• The proposed method achieves state-of-the-art sub-pixel accuracy on public datasets.
• Evaluated on healthy controls, multiple sclerosis, and diabetic macular edema datasets.
• Demonstrates superior performance in generating smooth, topologically correct retinal surfaces.

Conclusions:

• The novel deep learning approach offers a more efficient and flexible solution for retinal layer segmentation.
• This method holds significant potential for advancing automated analysis in ophthalmology.
• Accurate segmentation facilitates better monitoring of retinal diseases and their progression.