Abstract
Scientists reuse figure elements sometimes appropriately, e.g. when comparing methods, and sometimes inappropriately, e.g. when presenting an old experiment as a new control. To understand such reuse, automatically detecting it would be important. Here we present an analysis of figure element reuse on a large dataset comprising 760 thousand open access articles and 2 million figures. Our algorithm detects figure region reuse, while being robust to rotation, cropping, resizing, and contrast changes, and estimates which of the reuses have biological meaning. Then a three-person panel analyzes how problematic these biological reuses are using contextual information such as captions and full texts. Based on the panel reviews, we estimate that 9% of the biological reuses would be unanimously perceived as at least suspicious. We further estimate that 0.6% of all articles would be unanimously perceived as fraudulent, with inappropriate reuses occurring 43% across articles, 28% within article, and 29% within a figure. Our tool rapidly detects image reuse at scale, promising to be useful to a broad range of people that campaign for scientific integrity. We suggest that a great deal of scientific fraud will be, sooner or later, detectable by automatic methods.
