Medical imaging is critical in allowing physicians to identify conditions and diseases; however, these photos are not only sensitive but also protected by privacy laws. Yet, some hackers often try to sneak in and alter the images. To mitigate this issue, researchers from the Missouri University of Science and Technology (MST) are using ACCESS allocations on Jetstream2 at Indiana University (IU). Specifically, Tie “Thomas” Luo is leading a team that develops effective artificial intelligence (AI) methods – machine learning and deep learning – to prevent attacks by enhancing the robustness of medical imaging systems.
The team’s work has recently been published in the proceedings of the 28th Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD), which accepts less than 20 % of submissions. Their study, entitled Adversarial-Robust Transfer Learning for Medical Imaging via Domain Assimilation, received the best paper runner-up award at the conference.
“Transfer learning has been remarkably successful in reducing the demand for data and compute in training deep neural network models, but when applied to medical imaging, a ‘domain discrepancy’ between medical images and natural images has long been overlooked – which makes medical AI models more vulnerable to adversarial attacks,” said Luo, an associate professor in the Department of Electrical and Computer Engineering at the University of Kentucky, formerly an associate professor in the Department of Computer Science at Missouri University of Science and Technology. “To address this issue – bridge that ‘domain gap’ – we proposed and tested a ‘domain assimilation’ strategy which colorizes medical images and adapts their texture to resemble nature images.”
Luo explained that this concept of “domain assimilation” – realized by a colorization and texture adaption method – fixes the vulnerability and adds more robustness to medical AI models.
“Our study used ACCESS allocations on Jetstream2 at IU to train and evaluate our approach against multiple adversarial attacks on medical imaging and have obtained encouraging results,” he said.
Resource Provider Institution(s): Indiana University (Jetstream2)
Affiliations: University of Kentucky, Missouri University of Science and Technology
Funding Agency: NSF
Grant or Allocation Number(s): CIS230363
The science story featured here was enabled by the U.S. National Science Foundation’s ACCESS program, which is supported by National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296.