Affine Registration And Stereoscopic Extraction From Retinal Fluorescein Angiographic Images

Abstract

 

Purpose:

 

Quantitative analysis and automated classification of retinal image sequences, such as fluorescein angiograms (FA), both require precise image alignment. Affine registration is relatively effective for this but can leave portions of stereoscopic image pairs unregistered. In these cases, precise pixel alignment requires elastic registration and necessitates extraction of stereoscopic information. This study evaluates the use of Y–feature detection for affine registration and mutual information for stereoscopic extraction.

 

Methods:

 

Stereoscopic FA image pairs from patients with age–related macular degeneration were analyzed with custom software that matches extracted Y–features (vessel bifurcations) and estimates stereo disparity. Various methods of Y–feature detection were evaluated in 12 FA sequences (249 image pairs) to identify the method with the lowest average pixel error. The epipolar geometry was estimated using standard 7 & 8–point algorithms as well as a plane–and–parallax algorithm. Dense subpixel–resolution disparity maps were estimated from rectified images using mutual information. Human assessment of stereoscopic image pairs and OCT data was used as the ground truth for comparison.

 

Results:

 

Global affine Y–feature registration yielded a lower average pixel error (3.18) than pairwise registration. The remaining pixel error appeared to be due to stereoscopic information. Standard implementations of 7 & 8–point F–matrix approximation algorithms were found to be inferior to the plane–and–parallax approach for epipolar geometry estimation (Fig.). Extracted retinal surface maps were found to match the retinal contours identified by human graders from images and OCT data.

 

Conclusions:

 

Retinal images in an FA sequence can be precisely aligned and the residual disparity in these images can be used to extract stereoscopic information. Elastic registration may be useful for the precise classification and quantification of angiographic features.  

 

Keywords: image processing • imaging/image analysis: clinical • imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound)