April 2009
Volume 50, Issue 13
Free
ARVO Annual Meeting Abstract  |   April 2009
Multimodal Image Registration Using the Fully Automated Harris-Invariant Feature Descriptor
Author Affiliations & Notes
  • N. M. Pumariega
    Dept. of Ophthalmology, Harkness Eye Institute, Columbia Univ., New York, New York
    Dept. Biomedical Engineering, Fu Foundation SEAS, Columbia Univ., New York, New York
  • J. Chen
    Dept. Biomedical Engineering, Fu Foundation SEAS, Columbia Univ., New York, New York
    Institute of Automation, Chinese Academy of Science, Beijing, China
  • N. Lee
    Dept. Biomedical Engineering, Fu Foundation SEAS, Columbia Univ., New York, New York
  • A. Laine
    Dept. Biomedical Engineering, Fu Foundation SEAS, Columbia Univ., New York, New York
  • R. T. Smith
    Dept. of Ophthalmology, Harkness Eye Institute, Columbia Univ., New York, New York
  • Footnotes
    Commercial Relationships  N.M. Pumariega, None; J. Chen, None; N. Lee, None; A. Laine, None; R.T. Smith, None.
  • Footnotes
    Support  New York Community Trust, R01 EY015520-01, and unrestricted funds from Research to Prevent Blindness
Investigative Ophthalmology & Visual Science April 2009, Vol.50, 306. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      N. M. Pumariega, J. Chen, N. Lee, A. Laine, R. T. Smith; Multimodal Image Registration Using the Fully Automated Harris-Invariant Feature Descriptor. Invest. Ophthalmol. Vis. Sci. 2009;50(13):306.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose: : To propose an automated method for precise spatial alignment of two or more retinal images from different modalities and for serial images from the same modality for clinical review of disease progression.

Methods: : Applying the well-known Harris detector we first locate approximately 200 corner points in each image for feature extraction. A main orientation is assigned to each corner point. An Intensity Invariant Feature Descriptor (IIFD) at each corner point is then calculated from the local orientation histogram, achieving invariance to image rotation and intensity. IIFDs are then bilaterally matched across the two images to identify two subsets of control points. Finally, suboptimal matches are locally spatially refined and a suitable transformation is applied. We evaluated 150 pairs of multimodal images (autofluorescence, infrared and red-free photographs). All images contained interfering retinopathies. Image resolution ranged from 230x230 to 2400x2400 pixels. Image pairs could overlap as little as 20%, could be rotated up to 40 degrees and could differ in scale factor by 1.8. Accuracy of registered images was analyzed by flicker comparison of the superimposed images for vascular alignment. Our algorithm was tested against SIFT (D.G.Lowe, "Distinctive Image Features from Scale-Invariant Keypoints", Internatl J Computer Vis, 2004).

Results: : Our registration time was 14.6 seconds per pair, with 144 (96%) registered pairs judged "excellent" (all major corresponding vessels remained stationary) and the remaining 6 pairs deemed "acceptable" (vessel shifts < 150 microns). SIFT was faster (3.5 seconds per pair) but far less accurate, yielding only 27% excellent, 13% acceptable, and the rest unacceptable or failed.

Conclusions: : Experimental results suggest that our image registration method far surpasses existing algorithms in accuracy, allowing multimodal and serial lesion comparisons for measurement, discovery, and hypothesis testing.

Keywords: image processing • imaging/image analysis: clinical • retina 
×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×