Originally drafted in 2020 and published in 2021, the RCRs guidance document “Integrating AI with radiology reporting workflows (RIS and PACS)”RCR Guidance - Integrating AI with Radiology Reporting - March 2021.pdf (260.2 KB) is now being reviewed ready for updating by a working party jointly formed of members of the RCR AI and RCR RIC SIGs.
For the PACS Managers perspectives, the SCoR RIAG will feed back modern day integration updates.
If you have any comments or suggestions as to improvements or clarifications in the guidance please send me a message and I will ensure they are collected and passed on.
From a technical perspective, it needs updating to address the progression away from RIS-based to PACS-based reporting. Our AI systems use HL7 integration with our PACS system to provide their findings, which allows PACS to build/triage/prioritise reporting worklists, trigger automated workflows and pre-populate report text. We are also considering extending the HL7 integration with our EPR - we want the findings to populate A&E worklists, to improve our AI-supported ED rapid discharge pathway.
One other thought, standard number 2 seems a touch unnecessary, and which accuracy figures should be included? The manufacturer provided ones cannot be assumed accurate, due to algorithm training bias, and it would be very difficult to constantly update the figures with those of pre-deployment audit or post-market surveillance. So far the AI outputs all say some sort of disclaimer that they are just a decision support tool, and we train our users how to correctly use the output clinically. Some systems will also provide a measure of how certain they are in their findings - that could be useful to include as a standard, whether that be certain vs doubt, or a percentage.
Hi Andy! This is purely out of curiosity. Would you be able to share which HL7 message type (ORU or ORM) and segments your AI system uses to send findings to PACS? Thank you so much!