Registration of Multimodal Retinal Images

Aligning retinal images across modalities — color fundus, fluorescein angiography (FA), and ultra-widefield (UWF) fundus — is fundamental for transferring annotations, fusing diagnostic information, and reconstructing wide-field views from multiple captures. Each modality reveals different anatomical details under drastically different appearance statistics, making registration a challenging cross-modal correspondence problem.

Research directions

  • Score-based and diffusion-driven alignment. Casting cross-modal registration as an iterative refinement process guided by learned scores and random-walk correspondence search, robust to large appearance gaps and field-of-view differences.
  • Autoregressive transformation estimation. Decomposing complex non-rigid deformations into a sequence of predicted transformation parameters, enabling accurate alignment of images with severe geometric variation.
  • Fundus–FA registration for fine-scale annotation transfer. Aligning fundus and FA image pairs so that FA-derived vessel maps can serve as high-resolution supervision for fundus-based vessel analysis.
  • UWF enhancement and photomontage construction. Improving UWF image quality with unpaired generative models, and stitching multiple fundus captures into coherent wide-field photomontages through learned registration and blending.