Abstract
Purpose :
Adaptive optics (AO) retinal imaging datasets can contain hundreds to thousands of individual images that need to be tiled together to create a larger montage, a process that if done manually, generally takes at least several-fold longer than the time it takes to acquire images on a patient. The goal of this work is to build upon two automated algorithms that have been proposed (BOE 7(12):4899,2016; BOE 9(9):4317,2018).
Methods :
Oriented FAST and Rotated BRIEF (ORB) descriptors, combined with Random sample consensus (RANSAC), were used with multithreading to compute and match features across images. A recovery step was implemented to bridge together discontinuous clusters of images and to fill holes within the montage based on the use of “full-frame” images (i.e. images corrected for eye motion using only a simple rigid registration). Native handling of Adobe Photoshop file formats was implemented to improve usability. Normalized Cross Correlation (NCC) and Normalized Mutual Information (NMI) were used to evaluate the performance of the proposed algorithm (“aomontage”) relative to manually-built montages.
Results :
After successful manual and automated montaging of AO datasets from 10 healthy subjects (5 smaller and 5 larger datasets), the resulting NCC and NMI values from aomontage were on average better than those from manually-constructed montages (Table 1). The recovery step reduced the average number of discontinuous clusters from 5.4 to 2.5 (for 8 out of 10 subjects, this resulted in at least 95% of the imaged area being montaged; 2 subjects had 4 and 7 discontinuous clusters that were manually corrected). The average computation time was 79 seconds for small datasets (<40 tiles) and 109 minutes for large datasets (>125 tiles; each tile contains linked AO images from at least 3 different AO modalities, generated from multiple distinct reference frames).
Conclusions :
We demonstrate a robust strategy for automatic montaging of AO images (“aomontage”) which features improved handling of discontinuous image clusters, native handling of Adobe Photoshop file formats, and a semi-automated refinement process that allows for the positions of distinct clusters to be manually adjusted (e.g. for discontinuities), while maintaining accuracy and speed. The overall approach should help to alleviate the bottleneck of post-processing and accelerate the throughput of AO imaging.
This is a 2020 ARVO Annual Meeting abstract.