Robust Retinal Image Matching: A Modality-Resistant Descriptor Using Directional Anisotropic Texton-Like Features and...

Area of Science:

• Ophthalmology
• Computer Vision
• Medical Imaging

Background:

• Retinal image matching is vital for medical diagnostics and disease tracking.
• Multimodal retinal image alignment faces challenges from structural and radiometric differences.
• Current deep learning methods are data-dependent, while traditional methods struggle with complex variations.

Purpose of the Study:

• To develop a robust and adaptable descriptor for multimodal and monomodal retinal image matching.
• To overcome limitations of existing deep learning and handcrafted feature methods.
• To improve accuracy in aligning diverse retinal image datasets.

Main Methods:

• Proposed a novel descriptor using first- and second-order directional anisotropic Leung-Malik (LM) derivatives for feature space construction.
• Fused feature maps using orientation index of maximum directional response.
• Incorporated scale-invariant and rotation-invariant modules with weighted adaptive binning.

Main Results:

• Achieved an average recall of 37.48% without failures.
• Demonstrated superior performance over state-of-the-art methods, including deep learning approaches.
• Successfully adapted to multiple datasets: color fundus (CF), fluorescein angiography (FA), ultra-widefield FA, and SLO images.

Conclusions:

• The proposed descriptor offers a straightforward yet robust solution for retinal image matching.
• It effectively handles geometric and radiometric differences and structural displacements.
• The method's adaptability and high performance make it suitable for various clinical applications.