Generalizable Multimodal Retinal Image Registration via Label-free Vessel Segmentation

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Multimodal retinal imaging is vital for diagnosing and managing diseases like diabetic retinopathy and age-related macular degeneration (AMD).
• Accurate registration of images from different modalities (CF, FAG, FAF, ICG, OCT, IR) is essential for integrating complementary pathological insights.
• Existing registration methods often require labeled datasets, limiting their generalizability and application.

Purpose of the Study:

• To develop a generalizable, label-free retinal image registration algorithm applicable across multiple imaging modalities.
• To improve the accuracy of disease quantification, monitoring, and automated diagnosis through integrated multimodal retinal data.
• To overcome the limitations of existing methods by eliminating the need for large labeled training datasets.

Main Methods:

• A novel, label-free retinal image registration approach was developed.
• Vessel structures were extracted using the DexiNed algorithm to facilitate registration.
• The method was evaluated across various multimodal pairings (CF-IR, CF-FAF, CF-FAG, CF-ICG, FAF-FAG, FAF-ICG, FAG-ICG).

Main Results:

• The proposed method achieved a mean landmark error (MLE) between 1.91±0.44 and 4.9±2.32 pixels across different modality combinations.
• Registration of Color Fundus (CF) and Infrared (IR) images yielded an MLE of 3.08±1.47 pixels.
• Registration of CF and Fundus Autofluorescence (FAF) images resulted in an MLE of 4.9±2.32 pixels, comparable to human annotations.

Conclusions:

• The developed label-free registration method effectively integrates multimodal retinal imaging data.
• This approach enhances diagnostic precision and disease monitoring for various retinal conditions.
• The elimination of the need for labeled training data increases the method's practicality and broad applicability.