Self-supervised pre-training for joint optic disc and cup segmentation via attention-aware network.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Supervised image segmentation requires extensive pixel-level labeling, which is time-consuming and labor-intensive, especially in medical imaging.
• Challenges in optic disc and cup segmentation include designing efficient networks for global context and training with limited, privacy-sensitive medical data.

Purpose of the Study:

• To develop an efficient deep learning model for optic disc and cup segmentation.
• To address the limitations of data scarcity and privacy concerns in medical image segmentation.
• To achieve state-of-the-art performance in optic disc and cup segmentation.

Main Methods:

• A novel attention-aware segmentation model with a multi-scale attention module in an encoder-decoder network was designed to capture global semantics and long-range dependencies.
• A new loss function was introduced to incorporate prior knowledge of the optic cup's location within the optic disc.
• A self-supervised contrastive learning method was proposed for unsupervised feature representation learning, followed by fine-tuning with the proposed loss function.

Main Results:

• The proposed method achieved new state-of-the-art performance on public benchmarks (DRISHTI-GS and REFUGE datasets).
• Achieved F1 scores of 0.9801 and 0.9087 on the respective datasets.
• Demonstrated superior effectiveness in optic disc and cup segmentation, even with limited training data.

Conclusions:

• The developed attention-aware model and self-supervised contrastive learning approach effectively address challenges in medical image segmentation.
• The method offers a promising solution for accurate and efficient optic disc and cup segmentation, particularly in data-scarce scenarios.
• The publicly available code will facilitate further research and application in the field.