Abstract
Purpose :
Retinal amyloid deposits exist in patients with Alzheimer’s disease and can be imaged with polarimetry. To analyze size, structure and polarimetry properties of deposits they must be segmented from the background retina. We wished to develop a fast, automatic segmentation method to reduce manual segmentation time and human bias.
Methods :
Eyes were donated from subjects with a variety of neurological disorders, in compliance with the Declaration of Helsinki. Fixed retinal flat mounts were stained with 0.1% Thioflavin-S and counterstained with DAPI. A microscope fitted with a dual-rotating retarder polarimeter was used to take 16 polarimetry images required to reconstruct a 4x4 Mueller Matrix of the deposit. 957 locations with deposits positive in fluorescence and polarization were collected as test images. The developed segmentation was performed in MATLAB by calculating the Euclidean distance between each vector of 16 spatially-related pixels in the raw polarimetry images and the 16 image modes. The resulting Euclidean distance image was thresholded at three standard deviations away from the mean of a Gaussian fit to the image histogram, and with Otsu’s method. Segmentation success was classified by an observer by comparison to the calculated polarization properties of the deposits.
Results :
Segmentation failed on 1% of the deposits (n=9) due to poor image quality. For the remaining 948 deposits, the Gaussian fit threshold successfully segmented 95% (n=898), underfit 2% of the deposits (n=21) and over-fit 3% of the deposits (n=29). Otsu’s threshold segmented an area that was on average 46% smaller than the Gaussian fit threshold; it successfully thresholded 61% (n=575) of the deposits and underfit 39% (n=373). For 100 deposits, the full calculations including both thresholds took an average of 0.61 seconds for a set of 16 raw images of 1280 x 1024 pixels.
Conclusions :
Euclidean distance segmentation is an accurate automatic method to segment retinal amyloid polarimetry images, with potential use in other polarimetry applications. Thresholding using a Gaussian fit to the background provided more accurate segmentation than Otsu’s method, which tended to under-fit the deposits. Using the raw polarimetry images as input allowed for sub-second segmentation after image collection.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.