Multimodal self-supervised retinal vessel segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Automatic retinal vessel segmentation is vital for diagnosing eye diseases.
• Limited annotated medical datasets hinder deep learning model performance.
• Self-supervised learning offers a promising solution for leveraging unlabeled data.

Purpose of the Study:

• To develop a novel self-supervised pretraining framework for retinal vessel segmentation.
• To leverage unlabeled multimodal fundus images for generating supervisory signals.
• To enhance segmentation accuracy by fusing multimodal features.

Main Methods:

• Utilized pairs of unlabeled multimodal fundus images.
• Employed Vision Transformer encoding and correlation filtering for feature fusion.
• Constructed a multimodal feature fusion map containing vessel information.
• Learned instance-level discriminative features using INFOMAX loss.
• Transferred learned knowledge to a supervised segmentation network.

Main Results:

• Achieved state-of-the-art performance among unsupervised methods.
• Demonstrated competitive results compared to supervised baselines.
• Significantly reduced the requirement for manual annotation.
• Validated the effectiveness of multimodal feature fusion for vessel segmentation.

Conclusions:

• The proposed self-supervised framework effectively utilizes unlabeled multimodal data for retinal vessel segmentation.
• This approach substantially lowers annotation costs and requirements.
• It offers a viable alternative to fully supervised methods, especially in low-data regimes.