Diseases such as drusen or glaucoma are among the leading causes for legal blindness in the U.S. Typically, an ophthalmologist diagnoses the disease by visually inspecting the retinal background using a stereoscopic microscope which provides 3D vision. However, surgical microscopes and slit lamps provide only a shallow depth perception decreasing the diagnosis and treatment accuracy of the ophthalmologist. Therefore physicians generally analyze the retina only qualitatively and only while the patient is still present.
The objective of this project is to investigate the possibilities for an automatic 3D reconstruction of the retina using images from a stereoscopic slit lamp, a surgical microscope or fundus images.
Fundus reconstruction
Stereo fundus images were reconstructed by a multi-label graph-cut approach using mutual information as a similarity metric.
Slit lamp reconstruction
The more challenging stereo slit lamp images were reconstructed using a pyramid based dense disparity map computation.
Slit lamp calibration
An evaluation of the intrinsic calibration has been performed with a micro stage. By moving the calibration pattern at intervals of 500μm we produced a reference grid of known world coordinates.
The thus evaluated algorithm attains an overall mean unsigned error of 64.71 ± 50.9μm for 44 images with 384 control points per image, which is sufficient for the dimensions of the optical disc (~ø2mm).
