Burnout, Bias, and the Black Box

First of all, as Radiologists, our fundamental purpose is to diagnose the underlying condition and guide the referring clinician on the next steps to take in patient management. So “sharing blame if a mistake happens” is not a consideration. If you consider that every single patient must have at least 1 imaging study done in order to to arrive at a diagnosis, you can see how the number of cases that Radiologists have to report on a daily basis can be overwhelming, particularly in the public health sector, where the patient numbers far surpass the number of qualified Radiologists. So in a normal work day, it is understandable how a Radiologist may be prone to human errors as their attention span, visual acuity and focus wanes. It is for this reason, I believe that AI is an asset to Radiologists due to its speed, efficiency and consistency. Instead of viewing it as “”better at finding things” I would re-phrase that and say it acts as an extra pair of eyes for the Radiologist to ensure diagnostic accuracy. Most radiologists already utilise AI in their practice, because it is viewed as an important tool in their arsenal to ensure that the ultimate diagnoses are precise and dependable.

This is a great question that underscores the thought processes that a Radiologist goes through when reviewing a study. When we are scrolling through thousands of images and correlate these with the patient’s clinical history, a number a differentials come to the Radiologists mind. Automatically many of these are disregarded and the Radiologist comes up with 1 or possibly 2 differentials that would “fit” that scenario. This happens in minutes. For every differential that is disregarded -there must be a data- driven reason. So if an algorithm provides a diagnosis without an explanation, then I think it should be ignored as ultimately we, as Radiologists, must base all our diagnoses with clear and convincing data-driven evidence.

Even in the pre-AI era, this happens to Radiologists from time to time. At first glance, a study looks generic and normal and then something gets your attention and causes you to pause. In these situations, we would call the referring physician for additional information to help with the diagnosis and if this is not helpful, additional studies may be suggested to aid in arriving at the diagnosis. There are instances in which a case is not clear-cut and additional investigations are required for further evaluation. The important thing is to be a safe Radiologist and not miss a pathology. So in this hypothetical scenario, the Radiologist has the responsibility to describe the area of interest and request for additional information and /or advise on further investigations for example, requesting a MRCP in a patient with a possible pancreatic lesion detected on CT. Calling something Normal because AI claims it is, even though your gut, which is your years of knowledge and experience, tells you otherwise, would be irresponsible. It is aways better practice to seek more information before committing to a final diagnosis. This protocol I believe will continue to serve Radiologists well into the future as we navigate a new world in AI.