Journals Consider Tests to Find Digital Modification of Photos
By Nicholas Wade
THE NEW YORK TIMES
Among the many temptations of the digital age, photo-manipulation has proved particularly troublesome for science, and scientific journals are beginning to respond.
Some journal editors are considering adopting a test, in use at The Journal of Cell Biology, that could have caught the concocted images of the human embryonic stem cells made by Dr. Hwang Woo Suk.
At The Journal of Cell Biology, the test has revealed extensive manipulation of photos. Since 2002, when the test was put in place, 25 percent of all accepted manuscripts have had one or more illustrations that were manipulated in ways that violate the journal’s guidelines, said Michael Rossner of Rockefeller University, the executive editor. The editor of the journal, Ira Mellman of Yale, said that most cases were resolved when the authors provided originals. “In 1 percent of the cases we find authors have engaged in fraud,” he said.
The two editors recognized the likelihood that images were being improperly manipulated when the journal required all illustrations to be submitted in digital form. While reformatting illustrations submitted in the wrong format, Dr. Rossner realized that some authors had yielded to the temptation of Photoshop’s image-changing tools to misrepresent the original data.
In some instances, he found, authors would remove bands from a gel, a test for showing what proteins are present in an experiment. Sometimes a row of bands would be duplicated and presented as the controls for a second experiment. Sometimes the background would be cleaned up, with Photoshop’s rubber stamp or clone stamp tool, to make it prettier.
Some authors would change the contrast in an image to eliminate traces of a diagnostic stain that showed up in places where there shouldn’t be one. Others would take images of cells from different experiments and assemble them as if all were growing on the same plate.
To prohibit such manipulations, Dr. Rossner and Dr. Mellman published guidelines saying, in effect, that nothing should be done to any part of an illustration that did not affect all other parts equally. In other words, it is all right to adjust the brightness or color balance of the whole photo, but not to obscure, move or introduce an element.
They started checking illustrations in accepted manuscripts by running them through Photoshop and adjusting the controls to see if new features appeared. This is the check that has shown a quarter of accepted manuscripts violate the journal’s guidelines.
In the 1 percent of cases in which the manipulation is deemed fraudulent — a total of 14 papers so far -\— the paper is rejected. Revoking an accepted manuscript requires the agreement of four of the journal’s officials. “In some cases we will even contact the author’s institution and say, ‘You should look into this because it was not kosher,’ “ Dr. Mellman said.
He and Dr. Rossner plan to add software tests being developed by Hani Farid, an applied mathematician at Dartmouth. With a grant from the Federal Bureau of Investigation, which is interested in ways of authenticating digital images presented in court, Dr. Farid is devising algorithms to detect alterations.
His work has attracted interest from many people, he said, including eBay customers concerned about the authenticity of images, people answering personal ads, paranormal researchers studying ghostly emanations and science editors.
For the latter, Dr. Farid is developing a package of algorithms designed to spot specific types of image manipulation. When researchers seek to remove an object from an image, such as a band from a gel, they often hide it with a patch of nearby background. This involves a duplication of material, which may be invisible to the naked eye but can be detected by mathematical analysis.
If an object is enlarged beyond the proper resolution, Photoshop may generate extra pixels. If the object is rotated, another set of pixels is generated in a characteristic pattern.
An object introduced from another photo may have a different angle of illumination. The human eye is largely indifferent to changes in lighting, Dr. Farid said, but conflicting sources of illumination in a single image can be detected by computer analysis and are a sign of manipulation.
“At the end of the day you need math,” Dr. Farid said. He hopes to have a set of tools available soon for beta-testing by Dr. Rossner.
Journals depend heavily on expert reviewers to weed out papers of poor quality. But as the Hwang case showed again, reviewers can do only so much. The defined role of reviewers is not to check for concocted data but to test whether a paper’s conclusions follow from the data presented.
The screening test addresses an issue reviewers cannot easily tackle, that of whether the presented data accurately reflect the real data. Because journal editors now have the ability to perform this sort of quality control, “they should do it,” Dr. Rossner said.
The scientific community has not yet come to grips with the temptations of image manipulation, Dr. Mellman said, and he would like to see other journals adopt the image-screening system, even though it takes 30 minutes a paper. “We are a poor university press,” he said, without the large revenue enjoyed by journals such as Nature, Science and Cell. “If they can’t bear this cost, something must be dreadfully wrong with their business models,” he said.
Science, in fact, has adopted The Journal of Cell Biology’s guidelines and has just started to apply the image-screening test to its own manuscripts. “Something like this is probably inevitable for most journals,” said Katrina Kelner, a deputy editor of Science.
She became interested as a quality control measure, not because of the concocted papers of Dr. Hwang, two of which Science published. Dr. Mellman says the system would have caught at least the second of Dr. Hwang’s fabrications, since it “popped out like a sore thumb” under the image screening test.
But other editors are less enthusiastic. Emilie Marcus, editor of Cell, said that she was considering the system, but that she believed in principle that the ethics of presenting true data should be enforced in a scientist’s training, not by journal editors.
The problem of manipulated images, she said, arises from a generation gap between older scientists who set the ethical standards but don’t understand the possibilities of Photoshop and younger scientists who generate a paper’s data. Because the whole scientific process is based on trust, Dr. Marcus said: “Why say, ‘We trust you, but not in this one domain?’ And I don’t favor saying, ‘We don’t trust you in any.’ “
Rather than having journal editors acting as enforcers, she said, it may be better to thrust responsibility back to scientists, requiring the senior author to sign off that the images conform to the journal’s guidelines.
Those guidelines, in her view, should be framed on behalf of the whole scientific community by a group like the National Academy of Sciences, and not by the fiat of individual editors.