Multi-modal and inter-device retinal OCT registration via learning-based feature correspondence

Accurate registration of widefield OCT volumes is essential in ophthalmology to compensate for motion artifacts, enhance image quality, and enable precise delineation of retinal structures for both single- and multi-modal imaging. We present a learning-guided, feature-based 3D registration framework that combines advanced keypoint detection, transformer-based feature matching, and non-rigid warping to align multiple OCT volumes across repeated acquisitions, encompassing both healthy and diseased cases. Our approach supports both longitudinal, inter-device, and multi-modal registration, including alignment between OCT and standard color fundus images. At a single time point, registered volumes are averaged to produce high-quality, motion-free representations with enhanced signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and high structural similarity index (SSIM), while the derived transformations can be applied to polarization-resolved signals to extract additional functional information. Quantitative evaluation demonstrates substantial improvements in SNR and CNR, with a mean 3D SSIM exceeding 0.97 for single modality cases, and landmark-based target registration error reduced to 2 pixels for multi-modal cases, indicating precise geometric alignment. Qualitatively, the method reduces speckle noise, improves layer connectivity, and strengthens retinal vascular contrast, providing clinicians with a more reliable baseline for interpretation. This framework is versatile, applicable to multi-modal OCT analysis, and enables downstream tasks such as segmentation, quantitative assessment, and longitudinal studies, offering significant potential for enhanced retinal disease monitoring and clinical decision-making.