Page 26 - Diagnostic Radiology - Interpreting the Risks Part Two_Neat

SVMIC Diagnostic Radiology: Interpreting the Risks

process, thereby expanding the role of radiologists in the

development of AI.

AI must be able to recognize labeled data, which in radiology
means images from patients who have been given a definitive

diagnosis. Deep learning of AI has been very successful in
compiling labeled images, but there is no repository of radiology.

Right now, these images are controlled by patients, treating
physicians, hospitals, imaging facilities, and even vendors.

Gathering these images and labelling them appropriately for
clinical uses will be a daunting and time-consuming challenge.

Radiologists will need to have a seat at the table, because it
is unlikely that vendors and their contractors will possess the

medical education needed to develop algorithms and programs
that can be used effectively by radiologists.

There are many issues that must be resolved before AI becomes
a tool to be used in daily practice. AI systems will soon be

implemented by healthcare regulators, insurance companies,
and medical billing companies to improve their functioning as

regards determining reimbursement payments and insurance
policy fees.

One issue prominent in physician concerns is whether or not

AI is safe to use. Adversarial attacks or manipulations could
change the behavior of artificial intelligence systems using tiny

pieces of digital data.

Concerned physicians give the example of a lung scan; they

argue that if a few pixels on the scan are changed, you could
fool an AI system into seeing an illness that is not actually there,

or not seeing an illness that is in fact present. Researchers
say that minor changes made to written descriptions of a

Page 26