Abstract
Purpose: :
To generate a statistical retina atlas from color fundus images. We retrieve the statistical properties from a large number of retinal images, present an atlas coordinate system, map images to the atlas space, and finally obtain statistical characteristics that can serve as a reference retina formation in potential image-oriented ophthalmologic applications.
Methods: :
The database from diabetic retinopathy screening program in the Netherlands is used for random selection of 600 color funds images from left eyes where no pathological feature or imaging artifact presents. We first perform a pair-wise image registration to obtain a larger field of view enough to contain key features of retina atlas. To avoid the variations in magnification, rotation, and illumination, inter-image coordinate alignment by the similarity transformation has been done followed by statistical intensity standardization. We consider the optic disk center (OD), the macular center (MC), and the main vessel arches (VA) as the key atlas features. Mean locations of OD and MC are mapped to the normalized atlas coordinate system. The spatial variation of VA is uniquely represented by a curved line using principal curve estimation and curve fitting technique, and rescaled to fit the OD and MC layout. Finally, the training images are warped from the registration space to the atlas space by the thin-plate-spline mapping method as shown in the figure.
Results: :
Putting OD at the origin, MC is found at (1.000, -0.123±0.002) in atlas coordinate system which corresponds to -7.092±0.001° seen from the OD. The estimated principal curve of VA is fitted to the 4th order polynomials with 0.00036 mean absolute residual error. Training images warped to the atlas coordinate system are observed to have intensity statistics of 174.5±27.5 in red, 77.9±13.5 in green and 62.2±15.4 in blue color channel respectively.
Conclusions: :
The developed retina atlas presents the statistical locations and shape of key features and the spatial intensity distribution of color fundus images. We are currently developing practical applications that are highly driven by the retina atlas properties.
Keywords: image processing • computational modeling • retina