Abstract
Purpose :
Significant spatial misalignment may exist between discrete spectral slices captured by Multispectral Imaging (MSI) from retina because image acquisition time is often longer than natural timescale of eye’s saccadic movement. Our goal is to develop a technique for aligning automatically MSI images from retina, which discovers spatial correspondence relationships between any spectral slices.
Methods :
A digital image database comprising 26 MSI image sequences was acquired by using an Annidis RHATM instrument each of which bears at least 8 slices from wavelengths of amber, green, infrared, red and yellow. These images are in a format of dicom, a bit depth of 16 and a size of 2048x2018. They are binocular images from 4 patients with hypertensive retinopathy and 8 healthy subjects. We developed a novel technique for a joint alignment of sequential MSI images by searching the lowest matching costs between automatically-detected salient feature points, which performs as solving a low-rank semidefinite matrix via a convex optimization. The proposed technique is unique for the global consistency of the generated spatial mappings between images. A trained rater manually picked 15 salient points for each MSI sequence and marked them in all MSI images, which were mixed with the algorithm-detected feature points and treated as the ground-truth in our experiments. The agreement of point-matches between the proposed computer algorithm and manual marks was assessed by computing the percentage of correct matches.
Results :
The proposed automated MSI alignment technique showed an almost perfect match with human’s manual works. The percentage of correct matches produced by the automated technique reached 99.3%. When the feature matching costs were added a Gaussian noise with zero mean and variances of 0.001, 0.05, 0.1, 0.15 and 0.2, respectively, percentages of correct matches became 98.5%, 97.8%, 96.3% and 95.9%, respectively, showing the robustness of the proposed technique to image noise. An exampling MSI sequence and the point-matches generated by our algorithm are shown in Figure 1.
Conclusions :
The proposed automatic joint alignment technique demonstrated not only a good agreement with manually-specified matches between MSI spectral slices but also a good robustness to image noise. It also holds promise in helping to fuse retinal features measured by MSI in different spectral bands.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.