AI safeguards imaging devices from ‘malicious’ cyberattacks

A new artificial intelligence technique promises to ward off cyberattacks targeting medical imaging devices and protect against other system-related errors.

The dual-layer platform identifies potentially malicious instructions sent from a host computer to an imaging machine. And when tested on a computed tomography system, it detected up to 99% of abnormal information.

Researchers from Ben-Gurion University of the Negev in Israel will present their work Aug. 26 at the 2020 International Conference on Artificial Intelligence in Medicine.

Tom Mahler, a PhD candidate at BGU, will lead the talk on Wednesday. He noted that medical devices such as MRI, ultrasound and CT are controlled by instructions sent from a primary computer. Cybercriminals can, however, bypass protections and send altered directions to such modalities, manipulating radiation settings or device components to harm patients.

And in some cases, a technician may make a costly error or a virus-infected host computer could automatically disrupt the imaging process.

With this in mind, Mahler et al. created their dual-layer architecture that focuses on rooting out two types of deceptive instructions. The first is known as “context-free” directions, which are “unlikely values” such as administering 100-times more radiation than is normal. The other is “context-sensitive” instructions, which can include mismatching the intended scan type, patients’ age, or diagnosis.

"For example, a normal instruction intended for an adult might be dangerous [anomalous] if applied to an infant,” Mahler said in a statement. “Such instructions may be misclassified when using only the first, context-free, layer; however, by adding the second, context-sensitive, layer, they can now be detected.”

The team tested its AI on a CT system, using 8,277 recorded instructions and put the context-free layer to the test using 14 unsupervised detection algorithms. As part of their study, they evaluated the context-sensitive layer on four various clinical objectives.

After analysis, the second CS layer boosted the overall detection of abnormal instructions from 71.6% (F1 score) to between 82% and 99%.