Abstract
Purpose :
Reticular pseudodrusen (RPD) confer a 2-6-fold increased risk of progression and represent a contraindication to subthreshold nanosecond laser therapy. Despite their importance, RPD are difficult to distinguish from sub-RPE drusen. A computational approach has been shown to help identify AMD features. Thus, the aim of this study was to develop a proof-of-concept, semi-automated, pattern-recognition approach using en face optical coherence tomography (OCT) images to discriminate RPD from sub-RPE drusen in eyes with intermediate AMD.
Methods :
110 fundus images acquired using the Nonmyd WX-3D retinal camera (Kowa) and Cirrus HD-OCT (Carl Zeiss Meditec) from intermediate AMD eyes seen at the Centre for Eye Health were analysed. Each image series from each eye comprised a reference photograph, reference OCT fundus image, ellipsoid zone slab, custom mid-retina slab, and two custom choroidal slabs. The latter four en face OCT images were pre-processed and the combined pixel value data analysed using unsupervised K-means cluster analysis (PCI Remote Sensing, Geomatica) resulting in a semi-automatically classified image. Processed images were also assessed qualitatively as a quartet image using the merge color channels function in imageJ. Class separability was tested using transformed divergence (TD). For each eye, the reference images, quartet image and the semi-automatically classified image were compared for concordance.
Results :
Figure 1 depicts the usefulness of pattern recognition in intermediate AMD in distinguishing sub-RPE drusen from RPD using input en face OCT images. Each pseudocolour in the classified image represents a unique spectral signature. The RPD prominent in the superior part of the image (pink) were statistically separable from the sub-RPE drusen elsewhere (yellow). The two drusen classes were confirmed by OCT and best distinguished using a statistically separable TD criterion of >1.5 as significant separability, which translates to 83% correct classification.
Conclusions :
A computational, pattern recognition approach was applied successfully to a series of en face OCT images. These results provide proof-of-concept that pattern recognition applied using a pixel-based classification scheme successfully discriminates RPD from sub-RPE drusen.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.