The American College of Radiology (ACR) Data Science Institute conducted an annual survey of its members. The objective was to understand how radiologists use artificial intelligence in their daily practices. As part of this study, an articlewas published, explaining all the results.

A survey induced by a more frequent use of AI in the field of medicine

Estimates from 2020 predict that the market for AI in medical imaging will grow 10-fold over the next decade. The Food & Drug Administration (FDA), as of July 2020, has validated 56 AI algorithms for clinical use:

While there has been a lot of media hype around the possibility of deep learning being exploited in radiology, leading to widespread use of AI, the reality seems to be more muted about the purpose. From this feeling, researchers from the Data Science Institute of the ACR have published in their institution’s journal, an article on radiologists and the use of AI. Bibb Allen is the lead author, along with Sheela Agarwal, Laura Coombs, Christoph Wald and Keith Dreyer.

For the research team, understanding how radiologists are using AI in clinical practice and monitoring these trends over time seems a critical approach to facilitating the development of AI algorithms that will help improve medical care.

The study process and results

The ACR Data Science Institute sent a brief electronic survey to all ACR radiologists via email. Participants were asked to provide demographic information about their practice, and also about their current uses of AI in their clinical work. They were also asked to evaluate the performance of AI models in their applications and to assess future needs.

According to the survey, about 34% of the radiologists surveyed said they currently use AI in their clinical practices. The more complex or time-consuming the task, the more AI is used. According to 73% of the participants, AI can be exploited to detect lesions, 71% of them think that it will be useful in performing morpho-anatomical analysis. All the results concerning the potential use of AI in radiology are available on the graph opposite:

About 94% of the users reported that “the performance of AI in their practice was sometimes inconsistent” and only 5.7% of them reported that the technology they were leveraging “worked every time it was used.” Among the inconsistencies reported by AI users were issues with visibility, correct use of scanners, or analysis of data from one or more patients.

Questions were focused around possible improvements to AI models in order to perfect the AI models. 60% of the participants answered that the Data Science Institute should provide methods to evaluate the performance of the algorithms and be able to comment on their effectiveness if necessary. They would also like to test the quality of the model with their own databases to decide whether or not to purchase it.

Translated from L’American College of Radiology publie une enquête sur les usages de l’intelligence artificielle en radiologie