Recognizing and addressing bias with AI and radiologists

According to Nina Kottler, MD, MS, Associate Chief Medical Officer of Clinical AI and Vice-President of Clinical Operations at Radiology Partners, combining artificial intelligence (AI) detection algorithms with radiologist assessment can increase detection rates between two and 45 percent improve.

In a recent video interview, Dr. Kottler, however, that acknowledging and adequately educating about the potential biases of AI algorithms and inherent biases of radiologists are key to maximizing the impact of AI on patient outcomes.

dr Kottler noted that factors such as the patient’s medical history, the type of exam ordered, an unusual location for a lesion, and distracting pathologies can contribute to variability in the radiologist’s assessment. She also pointed out that educational biases due to one’s own experience, recent missed diagnoses, and satisfaction with the search can also factor into image interpretation.

dr However, Kottler added that there are also biases or limitations in the development of AI algorithms.

“Most AI systems are trained on a series or a few images within an entire image set,” noted Dr. Kottler, who recently gave a talk on AI at the Radiological Society of North America (RSNA) conference. “If you look at a chest X-ray, many chest algorithms are only trained on the frontal image. If you are looking at a head CT for intracranial hemorrhage, most of these (AI) algorithms are only trained on the non-contrast axial soft tissue thin slice series. It doesn’t look at all the other components.”

(Editor’s note: For more video interviews with RSNA instructors, click here.)

dr Kottler emphasized adequate training beyond the sensitivity rates and area under the curve (AUC) performance of AI systems.

READ :  Possible use cases of AI in the education sector

“We spend a lot of time in our practice trying to understand the bias of AI (the model),” says Dr. kottler “Where does it exaggerate? Where is it underestimated? We pass this information on to our radiologists and inform them. We just don’t give them info like (this AI system) has an AUC of 95 percent, it’s going to be awesome. no We go to them and say (this system) can overwhelm an intracranial hemorrhage when the patient is moving and there is a streak artifact on the exam. It will be missing some of the results in X, Y and Z cases. You take these things to the radiologists and educate them about the biases of the AI. They do it like this (radiology) and AI are stronger together.”

More insights from Dr. Kottler is available in the following video.