Good Training Data Needed
It is still a long way to go until automatic diagnosis prediction from imaging and radiology report analysis becomes possible.
The hype about artificial intelligence in radiology is about diagnosing. Promising study results even made it to the public media. Daniel Pinto dos Santos, University of Cologne, Germany, looked behind the curtain and still saw a lot of things to do.
What is already possible
If you train the system with raw data, the answers you get will be raw as well. “Still, this is enough to answer some useful questions,” Pinto dos Santos said. As example, he named predicting the patient waiting time based on DICOM data. Another example is predicting the type of an examination based on the symptoms and lab values extracted from referral texts.
What is still needed
However, for generating more specific answers, you need specific data. “And though we have a lot of imaging data, these data, unfortunately, are very poorly standardized and not always easy to access,” Pinto dos Santos said.
Moreover, radiology reports are text-intensive and often confusing, making them inappropriate for text mining software. Numerous studies have shown that structured reports have more (complete) content and greater clarity than conventional reports. Already, radiology report templates are available. “We developed an open-source prototype for structured reporting at the University of Cologne,” he said. “Yet we need the vendors to provide a practical solution.” An integration of the reports with, a radiology-specific ontology, such as RadLex® should be possible to avoid extra workload for the radiologist. Lastly, Pinto dos Santos drew the attention to the recent ESR paper on structured reporting, which is available online.
Pinto dos Santos made it clear that well-structured data is the key for the useful implementation of artificial intelligence in radiology. Structured reporting would be a good start for generating meaningful labels to analyze the data.
Presentation Title: Artificial Intelligence, Big data and structured reporting
Speaker: Daniel Pinto dos Santos, University of Cologne, Germany
Session code: SF10