Semi-supervised peripapillary atrophy segmentation with shape constraint.

Area of Science:

• Ophthalmology
• Medical Image Analysis
• Computer Vision

Background:

• Peripapillary atrophy (PPA) is a key indicator in eye diseases such as myopia and glaucoma.
• Accurate segmentation of PPA is crucial for monitoring disease progression.
• Challenges include blurry edges and limited labeled data for PPA segmentation.

Purpose of the Study:

• To develop a novel semi-supervised method for accurate PPA segmentation.
• To enhance PPA segmentation by incorporating prior shape knowledge.
• To effectively utilize large unlabeled datasets for improved model training.

Main Methods:

• Proposed a semi-supervised segmentation approach incorporating a novel shape constraint module based on active shape models.
• Implemented a Siamese-like network with exponential moving average for pseudo-label generation from unlabeled data.
• Utilized region connectivity correction to refine pseudo-labels.

Main Results:

• The proposed method demonstrated strong qualitative and quantitative performance in PPA segmentation.
• The shape constraint module effectively guided the network to learn PPA appearance.
• Leveraging unlabeled data significantly improved segmentation accuracy.

Conclusions:

• The novel semi-supervised method enhances PPA segmentation accuracy by integrating prior shape knowledge and unlabeled data.
• This approach offers a promising solution for clinical applications in ophthalmology.
• The method addresses key challenges in PPA segmentation, including blurry boundaries and data scarcity.