BreakNet: discontinuity-resilient multi-scale transformer segmentation of retinal layers.

Area of Science:

• Biomedical Imaging
• Ophthalmology
• Artificial Intelligence in Medicine

Background:

• Visible light optical coherence tomography (vis-OCT) offers high-resolution retinal imaging.
• Hemoglobin absorption in vis-OCT causes shadow artifacts, hindering accurate layer segmentation.
• Existing segmentation models struggle with these artifacts.

Purpose of the Study:

• To introduce BreakNet, a multi-scale Transformer-based model for robust retinal layer segmentation.
• To address challenges posed by shadow artifacts in vis-OCT imaging.
• To improve the accuracy of retinal layer segmentation in vis-OCT.

Main Methods:

• Developed BreakNet, a hierarchical Transformer and convolutional network.
• Employed multi-scale feature extraction for global and local context.
• Utilized decoder blocks for enhanced fine detail and semantic information extraction.
• Evaluated on rodent retinal images from a prototype vis-OCT system.

Main Results:

• BreakNet achieved superior segmentation performance compared to TCCT-BP and U-Net.
• The model effectively handled boundary discontinuities caused by shadow artifacts.
• Performance was validated even with limited-quality ground truth data.

Conclusions:

• BreakNet offers a significant advancement in vis-OCT retinal image analysis.
• The model's ability to mitigate shadow artifacts enhances segmentation accuracy.
• BreakNet holds potential for improved retinal quantification and clinical applications.