The Age of Narrow AI

Artificial Intelligence in radiology is gaining momentum. Radiologists should steer this process and start by defining the “use cases” they consider most important. Radiologists may become “data conductors” in the end.

Dreyer assessed the current status of AI, calling it “narrow AI.” It has grown from pure programming to being “like a child learning and making decisions focusing on very specific tasks.“ Medicine is still very far from “general AI”, he said.

Presently systems need a single question, for example whether a TCL (tibial collateral ligament of the knee) tear exists or not. The question can only be answered for a single, standardized modality. For larger solutions, a huge amount of case data from different modalities will have to be fed into the systems.
“Perfect AI would be superhuman, but I doubt we will get there,” said Dreyer.

Getting To Work

Dreyer sees radiologists already doing an “incredible amount of work” regarding AI.
This is in line with the general change in radiology. The American College of Radiology (ACR) has proposed “Imaging 3.0” as a framework to move imaging beyond interpretation to the entire spectrum of patient care with radiologists become involved in the whole healthcare process. This will require a cultural shift for radiologists – they will do less reporting, especially of normal cases, but instead focus on interpreting complex images and giving clinical advice.
Dreyer believes radiologists should be the ones to set the framework for IT to develop tools physicians can use at the point of care—right from the PACS workstation.
Economic incentives should change with the process: Politicians and payers need to be convinced to move from volume-based care to value-based care.

Building Healthcare AI

One of the next steps will be to make sure that AI “is working in thousands of clinics in many countries,” said Dreyer, adding that “a lot of societies will have to take this approach.” Radiologists will need to define “use cases” in specific areas, e.g. thoracic imaging with CT. For this, radiologists need to define data sets and AI models. “We need to create consistency,” demanded Dreyer.
These cases will be embedded in scanners or cloud systems and eventually lead to consistent processes for clinical care.

Dreyer pointed out that the healthcare industry does not have this kind of information, but knows how to set up the IT processes: “They have the algorithms, but not the data.” The final goal would be to have a solution in the cloud at hand that could be placed on images. A large number of findings could be made available to guidelines. Instead of a mere data creator, radiologists would become more of a data orchestrator.

Normal is Difficult

Dreyer’s talk was followed by a vivid discussion exploring different approaches towards AI. An audience member referred to the huge deficit of radiology consultants in the UK (which leads to a backlog of one quarter of a million scans per month) and asked whether ruling out normal plain radiographs with AI might help in saving consultant time.

“I agree with the goal,” said Dreyer, “but the concept of normal is often more difficult than we think.” It would likely need a huge amount of data to separate normal from ab-normal. He proposed focusing on the most dangerous diseases and narrow AI down to them. An AI standard would then suggest which patients need to be looked at quickly and which might not need urgent care.