Identifying images in the biology literature that are problematic for people with a color-vision deficiency

Reviewer #1 (Public Review):

Summary:
The authors of this study developed a software application, which aims to identify images as either "friendly" or "unfriendly" for readers with deuteranopia, the most common color-vision deficiency. Using previously published algorithms that recolor images to approximate how they would appear to a deuteranope (someone with deuteranopia), authors first manually assessed a set of images from biology-oriented research articles published in eLife between 2012 and 2022. The researchers identified 636 out of 4964 images as difficult to interpret ("unfriendly") for deuteranopes. They claim that there was a decrease in "unfriendly" images over time and that articles from cell-oriented research fields were most likely to contain "unfriendly" images.
The researchers used the manually classified images to develop, train, and validate an automated screening tool. They also created a user-friendly web application of the tool, where users can upload images and be informed about the status of each image as "friendly" or "unfriendly" for deuteranopes.

Strengths:
The authors have identified an important accessibility issue in the scientific literature: the use of color combinations that make figures difficult to interpret for people with color-vision deficiency. The metrics proposed and evaluated in the study are a valuable theoretical contribution. The automated screening tool they provide is well-documented, open source, and relatively easy to install and use. It has the potential to provide a useful service to the scientists who want to make their figures more accessible. The data are open and freely accessible, well documented, and a valuable resource for further research. The manuscript is well written, logically structured, and easy to follow.

Weaknesses:
(1) The authors themselves acknowledge the limitations that arise from the way they defined what constitutes an "unfriendly" image. There is a missed chance here to have engaged deuteranopes as stakeholders earlier in the experimental design. This would have allowed to determine to what extent spatial separation and labelling of problematic color combinations responds to their needs and whether setting the bar at a simulated severity of 80% is inclusive enough. A slightly lowered barrier is still a barrier to accessibility.

(2) The use of images from a single journal strongly limits the generalizability of the empirical findings as well as of the automated screening tool itself. Machine-learning algorithms are highly configurable but also notorious for their lack of transparency and for being easily biased by the training data set. A quick and unsystematic test of the web application shows that the classifier works well for electron microscopy images but fails at recognizing red-green scatter plots and even the classical diagnostic images for color-vision deficiency (Ishihara test images) as "unfriendly". A future iteration of the tool should be trained on a wider variety of images from different journals.

(3) Focusing the statistical analyses on individual images rather than articles (e.g. in figures 1 and 2) leads to pseudoreplication. Multiple images from the same article should not be treated as statistically independent measures, because they are produced by the same authors. A simple alternative is to instead use articles as the unit of analysis and score an article as "unfriendly" when it contains at least one "unfriendly" image. In addition, collapsing the counts of "unfriendly" images to proportions loses important information about the sample size. For example, the current analysis presented in Fig. 1 gives undue weight to the three images from 2012, two of which came from the same article. If we perform a logistic regression on articles coded as "friendly" and "unfriendly" (rather than the reported linear regression on the proportion of "unfriendly" images), there is still evidence for a decrease in the frequency of "unfriendly" eLife articles over time. Another issue concerns the large number of articles (>40%) that are classified as belonging to two subdisciplines, which further compounds the image pseudoreplication. Two alternatives are to either group articles with two subdisciplines into a "multidisciplinary" group or recode them to include both disciplines in the category name.

(4.)The low frequency of "unfriendly" images in the data (under 15%) calls for a different performance measure than the AUROC used by the authors. In such imbalanced classification cases the recommended performance measure is precision-recall area under the curve (PR AUC: https://doi.org/10.1371%2Fjournal.pone.0118432) that gives more weight to the classification of the rare class ("unfriendly" images).

Reviewer #2 (Public Review):

Summary:
An analysis of images in the biology literature that are problematic for people with a color-vision deficiency (CVD) is presented, along with a machine learning-based model to identify such images and a web application that uses the model to flag problematic images. Their analysis reveals that about 13% of the images could be problematic for people with CVD and that the frequency of such images decreased over time. Their model yields 0.89 AUC score. It is proposed that their approach could help making biology literature accessible to diverse audiences.

Strengths:
The manuscript focuses on an important yet mostly overlooked problem, and makes contributions both in expanding our understanding of the extent of the problem and in developing solutions to mitigate the problem. The paper is generally well-written and clearly organized. Their CVD simulation combines five different metrics. The dataset has been assessed by two researchers and is likely to be of high-quality. Machine learning algorithm used (convolutional neural network, CNN) is an appropriate choice for the problem. The evaluation of various hyperparameters for the CNN model is extensive.

Weaknesses:
The focus seems to be on one type of CVD (deuteranopia) and it is unclear whether this would generalize to other types. The dataset consists of images from eLife articles. While this is a reasonable starting point, whether this can generalize to other biology/biomedical articles is not assessed. "Probably problematic" and "probably okay" classes are excluded from the analysis and classification, and the effect of this exclusion is not discussed. Machine learning aspects can be explained better, in a more standard way. The evaluation metrics used for validating the machine learning models seem lacking (e.g., precision, recall, F1 are not reported). The web application is not discussed in any depth.

Reviewer #3 (Public Review):

Summary:
This work focuses on accessibility of scientific images for individuals with color vision deficiencies, particularly deuteranopia. The research involved an analysis of images from eLife published in 2012-2022. The authors manually reviewed nearly 5,000 images, comparing them with simulated versions representing the perspective of individuals with deuteranopia, and also evaluated several methods to automatically detect such images including training a machine-learning algorithm to do so, which performed the best. The authors found that nearly 13% of the images could be challenging for people with deuteranopia to interpret. There was a trend toward a decrease in problematic images over time, which is encouraging.

Strengths:
The manuscript is well organized and written. It addresses inclusivity and accessibility in scientific communication, and reinforces that there is a problem and that in part technological solutions have potential to assist with this problem.

The number of manually assessed images for evaluation and training an algorithm is, to my knowledge, much larger than any existing survey. This is a valuable open source dataset beyond the work herein.

The sequential steps used to classify articles follow best practices for evaluation and training sets.

Weaknesses:
I do not see any major issues with the methods. The authors were transparent with the limitations (the need to rely on simulations instead of what deuteranopes see), only capturing a subset of issues related to color vision deficiency, and the focus on one journal that may not be representative of images in other journals and disciplines.