Automated Fundus Image Standardization Using a Dynamic Global Foreground Threshold Algorithm

A generic fundus foreground extractor is required for the standardization of fundus datasets in machine-learning applications due to the vast range of retinal fundus images. Some fundus images have a large amount of non-essential background data and others have missing data because of clipping. To standardize these varied images for machine learning applications while preserving the aspect resolution, a generalized threshold algorithm is needed to separate the foreground and background. Existing threshold algorithms fail to segment images with low contrast. There is a need for a generalized algorithm to handle varied image conditions in a dynamic manner. The proposed segmentation algorithm uses shifts in histogram frequency using intensity extrema to find the ideal threshold value. The proposed post-processing algorithm crops, pads, and resizes the image to a standardized size of 512x512 pixels using the segmentation map output. To demonstrate the effectiveness of this proposed standardization approach on downstream tasks, an ablation experiment of popular standardization strategies is evaluated on a newly proposed benchmark dataset, EyePACS-light. The experimental results demonstrate the benefits of using this standardization approach for resizing fundus images.