Artificial Intelligence and Structured Reporting - Key Points of our Sessions
This years DGNR-Congress "neuroRAD" was very interesting and instructive. Great lectures, a big poster session, hands on courses and many colleges to meet and exchange with. We thank the DGNR again for giving us the opportunity to organize a part of the "Young Neuroradiologists"-Programm.
For those who did not make it there, the presentations and lively discussions afterwards can be summarized as follows.
Firstly, it was all about Artificial intelligence with B. Wiesler (Munich), J. Trenkler (Linz) and M. Holtmannspötter (Copenhagen). Key points:
We experience exponential growth in AI-development because of the use of new processors (GPU) enabling multi layer deep learning networks.
At the moment AI can effectively detect abnormality. Clinical evaluation is left to the radiologist. However, better functionality is to be expected with self-learning systems in the near future.
It is important for a radiologist to know how AI works or at least where and how mistakes happen to evaluate and use it effectively and safe.
AI carries in itself the potential to take over routine tasks like sorting and segmenting and thus to facilitate our future work flow.
The bottleneck stays data at the moment (at least prepared data). This ties into our next topic of "structured reporting".
You can watch the session here.
Secondly, we discussed Structured Reporting with M. Forsting (Essen) and O. Jansen (Kiel) after a presentation of D. Pinto dos Santos (Köln). Key points:
Prose-reports tend to be too long, carry too much unimportant information and are to heterogeneous.
Structured reports were shown to increase comprehensiveness, comprehensibility and the overall satisfaction of clinicians. They furthermore offer a possibility to include clinical guidelines and the clinicians input.
We have to differentiate between basic reports like a postoperative spine and special indications like tumor staging. Apart from reporting structured, we should report more targeted and relevant. Thus, structured reports should be contextual (look at the paper of Mamlouk et al.).
It would be an important step to implement speech recognition into the process of structured reporting or to find new ways of input (for example tablets).
If you are further interested, we have collected some ressources for Structured Reporting:
You can read a review about Structured Reporting from D. Pinto dos Santos here: https://nyaspubs.onlinelibrary.wiley.com/doi/abs/10.1111/nyas.13741
This paper of Mamlouk et al. compares anatomically structured reports with contextual reports and comes along with 50 contextual structured neuroradiology reports to download: http://www.ajnr.org/content/39/8/1406.long.
SMART-Radiology as an example of a product: https://www.smart-radiology.com/de/
Initiative of the DRG: https://www.befundung.drg.de/de-DE/2908/strukturierte-befundung/
Templates of the RSNA: http://www.radreport.org/
We also would appreciate to know your opinion on these topics, so do not hesitate to comment!