Authors: Harlan Stevens, Arwen Oakley
Mentors: Stephen Piccolo
Insitution: Brigham Young University
The number of scientific journal articles published per year now exceeds one million. To help maximize the impact of these articles, researchers must ensure that images in the articles are accessible to people with color-vision deficiencies (CVD). Up to 8% of males and 0.5% of females experience at least one form of color-vision deficiency, thus making it difficult for these individuals to discern patterns in images that use particular color combinations. We sought to shed light on this problem by estimating how often published images use color combinations that are unfriendly to people with a CVD. Examining 6,000 images published in biology-oriented research articles published in the eLife journal between 2012 and 2022, we identified images with potentially problematic color combinations. Using quantitative metrics and manual review, we estimate that 13% of these articles would be difficult for people with moderate-to-severe deuteranopia to interpret. We used a convolutional neural network to automate the ability to label images as being problematic for people with moderate-to-severe deuteranopia. The machine learning model successfully classified images in a testing dataset with an auROC of 91.3%. Based on these results, we created a web application that allows users to upload images and view estimates about whether the images are CVD-friendly. Such efforts are critical to ensuring that papers published in the biology literature are interpretable to diverse audiences.