Boundary-enhanced semi-supervised retinal layer segmentation in optical coherence tomography images using fewer...

Area of Science:

• Medical imaging analysis
• Computer vision
• Ophthalmology

Background:

• Accurate segmentation of retinal layers in optical coherence tomography (OCT) images is vital for diagnosing and managing eye diseases.
• Deep learning models face challenges in retinal layer segmentation due to blurry boundaries and insufficient pixel-wise annotations.

Purpose of the Study:

• To develop an effective semi-supervised learning approach for retinal layer segmentation in OCT images.
• To improve segmentation performance using scarce labeled data and abundant unlabeled data.

Main Methods:

• Proposed a Boundary-Enhanced Semi-supervised Network (BE-SemiNet) incorporating an auxiliary distance map regression task.
• Introduced a novel Unilaterally Truncated Distance Map (UTDM) to address class imbalance and enhance boundary learning.
• Implemented task-level and data-level consistency regularization with pseudo supervision on unlabeled data.

Main Results:

• BE-SemiNet significantly improved supervised baseline performance with only 5 annotations.
• The proposed method outperformed existing state-of-the-art techniques on two public OCT datasets.
• Demonstrated the effectiveness of UTDM and consistency regularization for semi-supervised segmentation.

Conclusions:

• BE-SemiNet offers a promising solution for accurate retinal layer segmentation in OCT images, especially when labeled data is limited.
• The method has significant potential for clinical applications in automated eye disease analysis.
• Reduced reliance on extensive manual annotation facilitates practical deployment in healthcare settings.