Abstract
Purpose: :
We propose a new filtering approach for canceling noise, laser speckles and vessels pulsatility in fundus images. The method is based on a Principal Component Analysis (PCA) and generates a single image free of dynamic components from a sequence of raw images.
Methods: :
We collected sequences of 50 near-infrared images with a scanning laser ophthalmoscope (Spectralis, Heidelberg Engineering) in 5 healthy volunteers. The images were pre-processed (registration, micro-saccades rejection and artifacts masking) and PCA-filtered to extract the first Principal Component. Raw and PCA filtered images were then compared quantitatively and by medical experts.
Results: :
We found an improved detectability of small features such as localized and faint hypo/hyper-pigmented areas and micro-vessels in PCA filtered images. In some subjects a clear relationship between both can be detected on PCA filtered images and not on raw images. We also identified specific artifacts caused by PCA filtering. One of them is a narrowing of the apparent diameter of the vessels and is due to the residual eye movements after registration and/or to local vessels pulsatility.
Conclusions: :
Filtering of fundus images by PCA potentially improves diagnostic detection power, especially in retinal diseases which present only small or faint pathological features or which affect micro-vessels. The method can easily be incorporated in established imaging resources.
Keywords: image processing • imaging/image analysis: clinical • retina