Purchase this article with an account.
Zachary Harvey, Alfredo Dubra; Motion distortion correction in scanning ophthalmoscopy by iterative image registration. Invest. Ophthalmol. Vis. Sci. 2014;55(13):4816.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Correction of distortion due to eye motion in scanning ophthalmoscope images is critical for multiple applications, including: increasing signal-to-noise ratio and/or speckle reduction through averaging, creating perfusion maps using motion contrast techniques and registration for longitudinal studies. Here, we demonstrate the improvement of a registration algorithm by applying it iteratively.
An image registration algorithm (Dubra & Harvey, Lect. Notes Comput. Sc. 6204, pp. 60-71) that estimates motion by comparing image strips against a reference frame using a normalized cross correlation (NCC), was modified to: perform multiple iterations, achieve sub-pixel accuracy through NCC interpolation, and account for line skew. The algorithm was tested on image sequences from a confocal adaptive optics (A O) scanning light ophthalmoscope (SLO). Average NCC values of 30 images and the average strip displacement after registration were used as performance metrics.
Results from subjects with normal fixation (achromatopsia and controls) show that most of the quantitative and qualitative improvements take place in the first three iterations (figures 1 and 2 respectively). Actual values vary substantially across image sequences, but NCC increases in the order of 1.5% for iterative registration alone were measured, with an additional 0.5% when incorporation sub-pixel NCC maximum estimation and less than 0.2% due to skew correction. Sources of image variability such as tear film evaporation, poor and/or slow AO correction and electronics noise prevented achieving NCC values of one.
Skew correction appears negligible in subjects with normal fixation, and it is expected to provide its most benefit in slower scanning modalities such as optical coherence tomography or in subjects with nystagmus. Both iterative registration and sub-pixel motion estimation translate in measurable improvement of the selected metrics, with their combination providing the best performance.
This PDF is available to Subscribers Only