Ready for testing artificial intelligence in radiology clinical practice

In the near future, I believe AI support systems will play a crucial role in the daily practice of radiologic reading. Radiologists are faced with an ever-increasing number of examinations and it remains vital to identify urgent cases as quickly as possible. An initial certified support system on the market has reported interesting performance data that proved the positive effect of the support system and displayed superior performance compared to 32 radiologists.

Even with over 210,000 CT scans of the brain as training data, some rare diseases were underrepresented and had to be excluded from the final model. Therefore, we as radiologists need to take care to evaluate performance in daily practice and have a close eye on rare, but important diseases, like venous thrombosis, basilar thrombosis, or herpes encephalitis, which are underrepresented in training datasets and therefore not directly reported. Using a support system might decrease the attention to unreported findings. Therefore, we urgently need multicenter validation of such systems with clinically meaningful endpoints, e.g. time to treatment onset.

I can only speculate on a potential outcome, but it might prove that overall patient management is improved by such systems, as the time to treatment onset for relevant and common findings such as cerebral hemorrhage decreases; but rare subgroups might suffer from increased time to treatment onset if radiologists only trust support systems and are unaware of clinical red flags. Therefore, we require careful scientific monitoring, and eventually an ethical discussion in the radiologic community, on how to deal with this risk, but we should not hesitate to implement and use such systems even today.

Article: Ready for testing artificial intelligence in radiology clinical practice: We would do well to be in the front line leveraging their strengths but also highlighting today weaknesses

Authors: Benjamin Bender