Retinal OCTA Image Segmentation Based on Global Contrastive Learning.

Area of Science:

• Medical Imaging
• Ophthalmology
• Computer Vision

Background:

• Automatic retinal vessel segmentation is crucial for diagnosing retinal diseases.
• Imbalanced datasets in retinal vascular imaging pose a significant challenge for current deep learning methods.
• Existing methods often overlook global dataset information, focusing solely on local image features.

Purpose of the Study:

• To propose a novel medical image segmentation method, Contrastive OCTA Segmentation Net (COSNet), to address data imbalance in OCTA datasets.
• To enhance the learning of both local and global information for improved segmentation accuracy.
• To provide a robust method for retinal vessel segmentation applicable to imbalanced data.

Main Methods:

• COSNet utilizes a feature extraction module mapping input OCTA image features to segment and MLP heads.
• A global contrastive learning module employs a memory bank for pixel embeddings and a mixed sampling strategy to create sample pairs.
• A novel contrastive loss function is introduced to simultaneously learn local and global information, followed by fine-tuning to restore positional information.

Main Results:

• The proposed COSNet method demonstrated significant improvements in segmentation accuracy (ACC) and area under the curve (AUC) compared to existing methods.
• COSNet effectively handles imbalanced data challenges in OCTA image segmentation.
• The method shows potential for extending to other medical image segmentation tasks.

Conclusions:

• COSNet offers a powerful solution for segmenting retinal vessels in imbalanced OCTA datasets by integrating global contrastive learning.
• The method enhances diagnostic capabilities for retinal diseases through improved segmentation accuracy.
• COSNet's approach is adaptable and can be applied to various other image segmentation challenges.