RSF-Conv: Rotation-and-Scale Equivariant Fourier Parameterized Convolution for Retinal Vessel Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Ophthalmology

Background:

• Retinal vessel segmentation is crucial for diagnosing eye diseases but challenging due to complex vascular structures.
• Convolutional Neural Networks (CNNs) excel at translation symmetry but struggle with rotation and scale variations in retinal images.

Purpose of the Study:

• To introduce a novel rotation-and-scale equivariant convolution (RSF-Conv) for enhanced retinal vessel segmentation.
• To improve the generalization capabilities of deep learning models in retinal image analysis.

Main Methods:

• Developed a rotation-and-scale equivariant Fourier parameterized convolution (RSF-Conv) module.
• Integrated RSF-Conv into existing networks like U-Net, Iter-Net, DE-DCGCN-EE, and FR-UNet.
• Conducted comprehensive in-domain and out-of-domain evaluations.

Main Results:

• RSF-Conv significantly outperformed existing methods, especially in out-of-domain evaluations, demonstrating superior generalization.
• RSF-Conv-enhanced models showed improved performance in retinal artery/vein classification.
• The RSF-Conv module reduced parameter count while boosting performance.

Conclusions:

• RSF-Conv offers a powerful solution for retinal vessel segmentation, addressing limitations of traditional CNNs.
• The method's strong generalization is vital for clinical applications facing cross-device and cross-hospital variability.
• RSF-Conv shows promising potential for broader clinical use in ophthalmology.