AI in radiology – is it a risky business?, ET HealthWorld

It’s been a stormy few weeks in the AI ​​world. First, we heard about an open letter from technology leaders, including Elon Musk, to AI labs asking them to delay the training of AI algorithms given the “serious risks to society and humanity” of AI. And just recently, during a US Congressional hearing on Chat-GPT, Open AI CEO said, “…governmental regulatory intervention is essential to de-risking increasingly powerful models.” He strongly recommended government oversight of AI. All this raises questions. Do similar risks exist for healthcare (especially radiology) AI? We recognize the many benefits of AI in radiology, such as dissolution. But could the radiology AI go out of control and confuse diagnoses, if not more?

continued below

Clearly, this is a negative way of thinking that risks slowing or thwarting the amazing potential of AI. But on the other hand, is all of radiology AI totally hunky dolly?

1. ask the wrong clinical questions

The fundamental value that AI provides to healthcare should be determined by clinical needs, not by the fact that a particular technology exists or by the good marketing strategies of AI companies.

Without addressing the right clinical questions, AI in radiology can actually mislead diagnoses. For example, consider an AI algorithm aimed at detecting infarcts in the brain (or dead brain tissue as a result of stroke) in CT scans. Given that stroke is a serious medical problem, this seems worthwhile at first glance. However, a major priority in stroke diagnosis that impacts patient recovery is the detection/or exclusion of hemorrhage in the brain, which determines whether anticoagulants are required to prevent stroke progression. will be By the time an infarction appears, it is already too late to mitigate the effects of the stroke. Efforts to develop infarct detection algorithms therefore add little clinical value in the acute treatment of stroke. Therefore, creating a vision/strategy and roadmap for AI development requires an ongoing and ongoing dialogue between the development team and physicians.

2. Inaccurate AI

As we’ve seen on ChatGPT, the answers aren’t always right. In healthcare, inaccurate AI can pose risks to doctors and patients.

A possible cause of inaccuracy is

2. a) Suboptimally Selected Training Data The key to AI algorithm development is accurate training using carefully selected data, radiology being a prime example. As the saying goes, “Garbage in, Garbage out.” In other words, if the training data is not well-tuned for a particular application, the resulting AI algorithms may produce inaccurate results accordingly. In highly regulated countries, such processes are subject to intense scrutiny, but in parts of the world where regulatory scrutiny is less rigorous, they can be cut short in a rush to market.

2.b) Failure to Properly Validate The essence of quality assurance in deep learning lies in algorithm validation, and the continuous improvement of algorithm performance by providing feedback based on the notion that radiologist interpretation is the gold standard. improved to This requires a carefully managed verification process/workflow. If the validation process is inadequate, the results can be distorted accordingly. And once the algorithms are deployed in a real clinical environment, radiologists are too busy to validate the algorithms in real time.

3. Integration challenges

Algorithm development is clearly the first and most important step in healthcare AI, but equally important is the seamless integration of algorithms into reporting workflows. One of the potentially most important benefits of AI in radiology is its ability to improve the productivity of radiologists. However, AI can actually slow down a radiologist’s work if the workflow integration is not optimal. To be used effectively and efficiently, the AI ​​output should be displayed within the radiologist’s primary workflow rather than requiring him to launch a second, independent AI browser. For algorithms that provide alerts for significant findings such as bleeding or thrombosis, the alerting process must prioritize positive cases on the radiologist’s reporting worklist to see any real benefit.

Four. Too rapid adoption and over-reliance

An overwhelming shortage of radiologists has left radiologists overworked today, which is why radiology AI is needed in the first place. A good workflow would be for the AI ​​to act as an intelligent assistant highlighting abnormalities to the radiologist, who then reviews the findings and incorporates them into the report. However, the flip side of the radiologist’s overworked scenario is that the radiologist can quickly become dependent on her AI results and end up in a state where the assistant dominates the master, rather than vice versa. That’s it. In the early stages of AI development, false negatives and false positives still exist, which can lead to missed diagnoses and overcalls.

Five. legal liability

The legal issues surrounding AI in diagnostics have yet to be clarified. If an AI algorithm misses a diagnosis, can radiologists still be held accountable? The answer for now is yes. The radiologist remains responsible for the report, even with AI assistance. Believing otherwise would be a risk that radiologists must not yield to.

AI in medicine in general, and in radiology in particular, is like Alexander’s sword that cuts through the Gordian knot of medical challenges, while also addressing staffing shortages and medical errors. However, like most swords, this one is double-edged and has the risk of false positives and false negatives. Considering the patient’s life is at stake and the potential medical risk to the physician, it must be developed, implemented and used with great care to deliver benefits and avoid negative consequences. need to do it. As the Hippocratic Oath states succinctly (and this should certainly apply to medical AI), “do no harm first.”

(Disclaimer: The views expressed are those of the author only and not necessarily endorsed by ETHealthworld.com. ETHealthworld.com will not be held responsible for any damages caused directly or indirectly to any person/organisation. We do not owe.)

most read blog