Journal: Patterns

Article Title: AI-enabled image fraud in scientific publications

doi: 10.1016/j.patter.2022.100511

Figure Lengend Snippet: Images forged by generative models are hard to distinguish We conducted a human-opinion study. This figure shows the normalized histogram of votes per image type. The image used for evaluation consists of five categories: (1) natural images, (2) scanning micrographs of nano materials (nano-micrograph), (3) cell immunostaining images, (4) immunohistochemistry (IHC) images, and (5) histopathological images. In total, 800 images are involved, and each image is rated by at least ten medical experts. The voting scale was between 1 to 4 corresponding to the following: 1 – definitely fake, 2 – probably fake, 3 – probably real, and 4 – definitely real. Mean scores are shown as red dots.

Article Snippet: The images used for evaluation may be classified into five categories: (1) natural images, such as natural sceneries, architectures, flora, and fauna; (2) scanning micrographs of nano materials collected from Internet; (3) cell immunostaining confocal microscope images from the Human Protein Atlas dataset; (4) immunohistochemistry (IHC) images collected from clinical and the Human Protein Atlas datasets; and (5) histopathological images from the breast cancer histopathological dataset (BreCaHAD).

Techniques: Immunostaining, Immunohistochemistry