Abstract
Purpose: :
Slit lamp composites can serve as a powerful retinal imaging modality. However, the technology has not yet been developed to its full potential. A previously proposed algorithm attempts to create the mosaic by means of a frame to frame alignment, minimizing the squared sum of differences (SSD) between pixel values. As this approach appears to be highly susceptible to noise, we suggest an image enhancement step before image alignment in order make it more suitable for retinal images. Furthermore, small alignment errors lead to a "drifting" of the stitched sequence over time due to the high temporal locality of the frame to frame matching. This is particularly problematic in cases where the slit lamp is moved back and forth. We propose a multi-frame alignment algorithm to overcome this limitation.
Methods: :
Image sequences have been acquired using a LSL 532s slit lamp (Carl Zeiss Meditec, Jena, Germany). In order to automatically extract the actual slit lamp image from the video frame, the left and right edges are detected by means of a Hough transform run directly on the gradient image. The video image is then cropped accordingly. Reflections are eliminated by masking very bright pixels. Subsequently, the extracted images are filtered using a laplacian of gaussian kernel. Because of their specific anatomy, blood vessels are considerably emphasized by an edge detection kernel. The frames are then aligned by minimizing the SSD between the filtered images. Currently, a pure translational registration is established. During the composite construction process, the system keeps a list of all frames and their orientations with respect to the global coordinate system. Instead of aligning the current frame with the last one only, the algorithm selects N previously processed frames and aligns the current frame with the N reference frames simultaneously. The references are chosen based on an overlap criterion. After registration the mosaic is constructed by blending all frames together using an appropriate image fusion algorithm. Currently, this is achieved using laplacian pyramids.
Results: :
Experiments on test sequences demonstrated high accuracy under visual inspection. The drifting problem was successfully eliminated.
Conclusions: :
The suggested multi-frame alignment algorithm substantially advances previous work and thereby increases the potential of automatic slit lamp compositing.
Keywords: retina • image processing