Probabilistic intra-retinal layer segmentation in 3-D OCT images using global shape regularization.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Spectral-domain optical coherence tomography (OCT) offers high-speed retinal imaging.
• Accurate segmentation of 3-D OCT scans is crucial for quantitative analysis.
• Existing segmentation methods may lack speed or robustness.

Purpose of the Study:

• To develop a novel probabilistic approach for fast and accurate 3-D OCT scan segmentation.
• To model both the appearance of retinal layers and global shape variations.
• To enable efficient probabilistic inference for segmentation.

Main Methods:

• A probabilistic model incorporating retinal layer appearance and global shape variations.
• Variational inference for approximate posterior distribution computation.
• Segmentation of 3-D OCT volumes in approximately one minute.

Main Results:

• Accurate segmentation of 3-D OCT volumes with an average unsigned error of 2.46 ± 0.22 μm.
• Accurate segmentation of 2-D scans from normal (2.92 ± 0.5 μm) and glaucomatous (4.09 ± 0.98 μm) eyes.
• The inferred posterior distribution aids in quality assessment, error detection, and disease classification.

Conclusions:

• The novel probabilistic approach provides fast and accurate segmentation of 3-D OCT scans.
• Global shape regularization significantly benefits segmentation accuracy.
• The method demonstrates robustness, requiring no pre/post-processing and consistent parameters across datasets.