Most of us have experienced some form of medical imaging, whether it was at an eye appointment or after a broken bone.
These images might contain more information than meets the eye. Things artificial intelligence can detect.
“AI is a very upcoming and novel concept in healthcare,” Dr. Jaya Kumar, Chief Medical Officer at Swedish Medical Center, said.
“We’ve been applying AI to medical imaging analysis now for a long time, almost a decade now,” Dr. Jayashree Kalpathy-Cramer, chief of artificial medical intelligence in ophthalmology at the University of Colorado School of Medicine, said. “What we are able to do with them today is phenomenal and quite different than even five years ago.”
Dr. Kalpathy-Cramer said AI is mainly used in medical imaging for diagnosis, but these algorithms can pick up other unintended signals.
“The algorithms can sometimes seem to have superhuman capabilities, so they can do things that a human is not able to see in the signal. So, for instance, they can look at images at the back of somebody’s eye and tell if they are female or male,” Dr. Kalpathy-Cramer said.
A recent study found that AI deep learning models can be trained to identify race in these same medical images. Something radiologists could only determine with 50 percent accuracy.
“The AI system, most of our models, were in the 95 to 99%,” Dr. Judy Gichoya, an assistant professor at the Emory University School of Medicine, said. Dr. Gichoya was one of the authors of the study. “Unfortunately, after two years, we still cannot tell you why.”
“It’s really unclear as to what it is that the algorithm is picking up on that allows it to make those decisions and make it so well, even on low resolution or parts of the image,” Dr. Jayashree Kalpathy-Cramer said.
This is what many researchers in this space are now working to figure out.
“Since we are so early in this type of work, we really don't know the significance of this,” Dr. Gichoya said.
They also want to make sure that the information is used only when necessary to make a medical decision. For example, if a patient's race plays no part in the medical condition, doctors want to make sure it’s not mistakenly or subconsciously used to create the treatment.
“Like if it says the disease is much more severe in women than men,” Dr. Kalpathy-Cramer said.
“Predicting race or any parameter is very important that it’s accurate since it determines the downstream algorithm and what we predict as an outcome for that patient,” Dr. Kumar said.
But as this technology advances, experts believe they will have more tools to help identify situations where the doctor and patient benefit.
“It can work sort of reliably at scale,” Dr. Kalpathy-Cramer said.
“We can use it to harness it for good and even narrow some of the disparities we observe,” Dr. Gichoya said.