Abstract
Purpose :
Many optical coherence tomography (OCT) systems incorporate IR fundus imaging for retinal preview and tracking. The acquisition reliability of OCT data depends on the quality of the IR fundus images’ focus. This technique demonstrates a real time image focus measurement using a fast method which can aid an operator during patient alignment. The real time focus measurement is also essential for an automated acquisition system.
Methods :
Our method first computes a focus value for each IR image by calculating the sum of the gradient magnitude above the 95th percentile of the gradient image as the focus value. Prior to the image gradient calculation, stripe artifacts are removed from IR images using a real time algorithm.
To test the algorithm, prototype software was used to collect a series of IR images (768x624 pixels over 11.52×9.36 mm2 with a pixel size of 15 µm/pixel) from a CLARUSTM 500 (ZEISS, Dublin, CA) at a frame rate of 50 Hz, using normal and small pupil acquisition modes.
Sequences of roughly 700-800 images each were collected from 10 subjects (some of which had multiple sequences). Images with artifacts, such as blinks, were removed from the database, leaving roughly 5000 total images. The images that remained in the database could be in-focus or out of focus. The algorithm then determines a “Focus Value” for each image. The focus values are converted to probabilities by transforming the focus values using a cumulative distribution function (CDF) (Fig 1) determined from 5,005 IR images. This plot shows the probability that an image is in focus for a given focus value as determined by the algorithm.
Results :
To evaluate the algorithm’s performance, approximately 100 images were randomly selected from each of the 10 sequences for evaluation. 1,016 independent images (486 out of focus and 530 in-focus) were evaluated subjectively by an expert grader using large vessel sharpness as a measure of focus using two grades (in-focus vs out of focus). We used the grader’s evaluation and the focus measurement calculated by the algorithm to compute receiver operating curves (ROC) (Fig 2). The area under the curve (AUC = 0.97) showed great performance of the algorithm with the test data.
Conclusions :
We demonstrated a functional real time IR image focus assessment algorithm that can help operators with patient alignment and automated acquisition.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.