Unsupervised retinal image registration based on D-STUNet and progressive keypoint screening strategy.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate retinal image registration is vital for diagnosing and monitoring eye diseases.
• Existing methods struggle with identifying intricate retinal vascular features, limiting registration accuracy.
• Limitations in current techniques necessitate advanced approaches for improved retinal image analysis.

Purpose of the Study:

• To introduce D-STUNet, a novel fusion network for enhanced retinal image registration.
• To improve the identification and utilization of retinal vascular features in image registration.
• To provide a more accurate and reliable method for retinal image analysis.

Main Methods:

• Developed D-STUNet, a Swin Transformer and U-Net fusion network incorporating a Differential Multi-scale Convolutional Block Attention Module with Residual Mechanism (DMCR).
• Implemented a Progressive Keypoint Screening (PKS) strategy to gradually accumulate effective keypoint information during training.
• The DMCR module enhances focus on vascular features, while PKS concentrates keypoints in vascular regions for improved matching.

Main Results:

• D-STUNet achieved a 98.50% acceptance rate and a 0% failure rate on the FIRE dataset.
• The model demonstrated strong performance across different difficulty levels, with AUCs of 0.929 (Easy), 0.883 (Mod), and 0.724 (Hard).
• The mean AUC (mAUC) was the highest among compared algorithms, significantly outperforming the second-best method.

Conclusions:

• D-STUNet effectively captures local features, including complex vascular structures in retinal images.
• The proposed network offers a significant advancement in retinal image registration accuracy.
• This method provides a new, powerful tool for improving the diagnosis and treatment of eye conditions.