KINTalk: Artificial Intelligence (AI) in the Radiological Workflow
The spotlight remained on radiology in our latest KINTalk installment, with Stephan Romeyn and Willem Grootjans from Leiden University Medical Center (LUMC) warmly welcomed to our virtual stage to share their first-hand experiences of developing, introducing and using AI throughout the radiological workflow. What lessons can be learned from their experiences? And how do these translate outside of the radiological world? Here we summarise three guiding recommendations which surfaced from the KINTalk so you can follow in their trailblazing footsteps.
Don’t forget the bigger picture. Currently, applications of AI have a narrow, specialised focus. While this may lead to efficiency gains for specific work tasks, a growing network of different AI solutions are starting to infiltrate the workflows of certain occupations such as radiologists and lawyers. This requires end users to simultaneously become savvy at using multiple AI solutions, whilst getting up to speed with alterations to their existing work tasks. The promise of efficiency gains for specific work tasks may therefore be negated when looking at the overall workflow, given the time required by end users to navigate this network of differing AI solutions. What can be done? Depending on the number of AI solutions being used (and those in the pipeline), think about seamlessly integrating the different solutions in a vendor neutral platform to reduce workflow disruption. A front-end(s) can then be designed to facilitate use and maximise efficiency gains on both a work task and overall workflow level.
Always involve the end users. Whether developing in-house AI solutions or co-creating AI solutions with technology companies, end users should be central to the process. From identifying relevant use cases to providing vital feedback, involving end users throughout the process increases the likelihood of developing AI solutions which are both useful and easy to use. Collaboration between technologists and end users equally offers mutual learning opportunities through the process and negotiation of combining technical know-how with domain expertise. This mutual learning can cause positive ripples of change throughout the implementation process, including how impact is defined — moving away from a focus on economic value to a broader set of measures — with improved efficiency, higher quality care, work satisfaction and confidence in diagnosis some of the measures used to assess impact at LUMC.
Give AI solutions a voice. Where we can ask our human co-workers to explain a decision, human-machine interactions do not currently work in the same way. When end users are solely presented with a decision recommended by an AI solution without an explanation of the underlying logic, unintended consequences can arise — especially when the recommendation differs from the expert opinion of an end user. Trust can therefore be hard to build. Thought should be given as to how to optimally display and provide feedback from AI solutions, listening to the requirements of end users whilst pushing the boundaries of technical feasibility. An automated, real-time dashboard, for example, could be crucial in confirming that AI solutions have reached a desired level of efficiency, as well as providing ongoing monitoring to ensure they continue to perform in a desirable way.
And there we have it. Our thanks once again go to Stephan and Willem for a truly insightful KINTalk with many takeaways for the medical world and far beyond!
Author: Lorna Downie
Stay tuned!
Want to be quick of the block when the next KINTalk sign up drops? Then follow the KIN Center for Digital Innovation on Twitter and LinkedIn to keep up to date with our latest events. We hope to see you next time!
