Poster M10 - What is a “Good” Figure: Scoring of Biomedical Data Visualization

Introduction: Biomedical Data Visualization has emerged as a pivotal tool for comprehending a vast array of data among scholars and practitioners. Despite the abundance of information available for creating visual representations, the quality of figures varies significantly. Consequently, while each visualization may convey a narrative to its audience, discerning the intended narrative from the creator can be challenging. This study seeks to explore the diverse applications of biomedical visualization in various societal contexts, while also proposing an evaluation framework for scoring visualizations using a bespoke algorithm. Furthermore, this research aims to present an exemplar of interactive data visualizations for research-oriented datasets, achieved through both the scoring of research-based figures and the development of a fully interactive web-based visualization platform. These objectives are informed by a comprehensive analysis of various data visualization sources sourced from the internet, including syllabi and software. The selected software applications underwent rigorous scrutiny based on several criteria including ease of use, customizability, required background knowledge, and financial cost, while educational resources were compared for similarities and differences. This process informed the creation of the Metrics for Evaluation and Discretization of Biomedical Visuals using an Iterative Scoring Algorithm (M.E.D.V.I.S). M.E.D.V.I.S is a scoring algorithm that allows users to input a set of their figures and compare them and provide relative scores and evaluations of their figures based on complexity, coloration, unused space, and the number of visualizations.

Materials and

Methods: To quantify the quality of figures in journals, specific criteria were devised and coded accordingly. These criteria encompassed the amount of white space, the number of visualizations, the color density, and the complexity of the image. Subsequently, each figure was scored based on these criteria and underwent dimensional reduction using PCA, UMAP, and t-SNE. This approach facilitated the aggregation and visualization of scores in a more accessible manner. Additionally, the data were subjected to K-Means Clustering to categorize the figures more effectively. It was anticipated that each clustering event would result in distinct outcomes, reflecting the varied dimensionality reduction methods employed, thus paving the way for the next phase of the algorithm.
The dimensional reduction techniques were varied to mitigate bias and offer insights into potential clustering changes. Following K-Means clustering, the cluster labels were stored in a data frame and subjected to consensus clustering. This method deemed optimal for combining different data results, yielded data labels ranging from zero to three, indicating varying degrees of agreement within the clusters. Subsequently, to achieve a binary classification of figures into high- and low-quality categories, the data underwent binarization based on the consensus clustering. This process entailed assigning values of one to zero and values of two to three, effectively segregating figures into positive and negative clusters, respectively labeled as high and low quality.

Results, Conclusions, and Discussions: Following testing of M.E.D.V.I.S, it was intriguing to find that although figures were categorized as high or low quality, they would still have aspects that could be improved upon. The examples provided actually composed of a high quality figure and two low figures respectively which provided interesting interpretations as the high quality figure passed all three of the four categories. It was low in number of visualizations, high complexity, high color use, but high in white space. White the second figure was also high in color score and low in visual count and white pixel percentage, it was still marked as a low quality figure which indicates that although a figure may pass through the same number of criteria, it still may not pass. This indicates that the criteria themselves might not be equivalently weighted. Further testing would need to be conducted in order to determine the proper weights of each of the categories. The tools available for creating visualizations exhibit a significant range of capabilities, adding to the versatility of data visualization in biomedical applications. As the field of biomedicine continues to grow, so too will the role and sophistication of biomedical data visualization evolving to meet the increasingly complex needs of the field. In response to this complexity, algorithms have been devised to accurately assess the quality of figures and have developed a user interface that enables users to interact with their figures. These advancements empower users to iteratively enhance their data and its visualization, ultimately elevating the sophistication of biomedical data visualizations. The binarization of the images provides a great way for researchers to quickly determine which of their figures are superb and which figures still require more work. Providing feedback for those figures that did not meet the standards expedites the ability of researchers and other users to improve those same figures. Although M.E.D.V.I.S requires both the figure caption in conjunction with the image, future directions indicate a possibility of determining the number of visualizations without the need for provided captions.

Acknowledgements (Optional):