In a joint research project, an international team has developed a deep learning model that can read off ethnic origins with a high degree of certainty from medical X-rays. The researchers from the United States, Canada, Australia and Taiwan found that the AI ​​made this determination with over 95 percent certainty, based on recordings of all anatomical regions of the human body. In the preface to their study, they warned that this effect harbors dangers in the automated evaluation of medical recordings and in AI-supported treatment proposals.
In the past it has been shown several times that AI systems can adopt prevailing prejudices. For example, artificially intelligent systems specifically rated dark-skinned and female applicants worse in the preselection for vacancies. Other applications had given African-American offenders a significantly poorer social prognosis than comparable white inmates when examining the suspension of their remaining prison term. A medical application that recognizes the ethnic origin of a patient could make inappropriate therapy suggestions due to racially influenced training data.
In their work, the researchers trained a neural network with hundreds of thousands of X-ray and computed tomography images, including images of the upper body, breast mammograms, extremities and cervical vertebrae. Although their neural network determined the ethnic origin with great certainty, they did not succeed in identifying the anatomical differences on which the AI ​​was guided.
With c’t 20/2021 you are one step ahead of hackers: With our checklists you can secure your smartphone, computer, WiFi router, online accounts and much more. It’s easy and usually only takes a few minutes. Also in the booklet: Windows 11 in the hardware check, concepts for hybrid conference systems and the best broadband tariffs for the fixed network. You will find issue 20/2021 from September 10th in Heise shop and at the well-stocked newspaper kiosk.
(agr)
