AI – From Soloist To Conductor

Imaging will be the first discipline to integrate Artificial Intelligence and its broader approach “data science” into the clinical setting. There is still a long way to go, but data science is constantly picking up speed.

Survival of Adaptors

Keith Dreyer, Harvard Medical School, Boston/MA, USA, started by correcting a false oral lore attributed to Charles Darwin: Darwin is often wrongly cited to have said “survival of the fittest,” while he in fact said “survival of those who can adapt,” noted Dreyer. He believes the latter principle to also be true for radiologists dealing with artificial intelligence (AI): “Radiologists should say yes to help from AI,” he said, comparing their role to that of a soloist, who needs to change to conducting a full orchestra.

Intense Change

“We had no idea how intense this would be,” said Dreyer about the rise of Artificial Intelligence in radiology. “Data science is the key term,” clarified Dreyer: “AI” is considered as a somewhat outdated expression; “machine learning” is really only about machines, while deep learning uses multilayered artificial neuronal networks.

The data science input at RSNA alone has doubled between 2016 and 2017. Dreyer believes that after this first flash regarding data science, there will be disappointment. If radiologists define the process their way, data science will achieve success. “It is friendly, but it needs stewards,” he said.

At the moment, radiologists teach empty neural networks how to recognize a certain pathology. At the beginning, pattern recognition is like flipping a coin, but throughout the teaching process the computer learns to recognize patterns by strengthening numerous “switches” in series connection. This may result in accuracies as high as 93%. Google and Apple use similar processes for picture recognition in their operating systems (OS).

Data Science Will Become Real

“AI will come in imaging first,” predicted Dreyer. This will require a combination of big data, powerful companies, robust algorithms and massive investments.

Radiologists from the clinical environment should define their needs, outline a concept for a solution to fulfill this need, then generate clinical applications, have them surveyed in an application environment (comparable to the FDA postmarket surveillance) and put that back into the clinical environment. This would make the full market cycle. Radiology cannot carry out these changes alone, believes Dreyer. “It takes massive companies with massive dollars,” said Dreyer, mentioning Apple, Google, Microsoft, Amazon and Facebook as potential leaders with adequate economic capacities. The process is likely to gain momentum: Things like making driving safer, using a camera in the oven to watch a cake bake on one’s phone or suitcases following their owners have already become reality.

As the Director of the Center for Clinical Data Science at Massachusetts General Hospital in Boston/MA, USA, Dreyer has become part of this “fast development”, initiating 44 new projects within the last 12 months. Lung cancer screening has been one of these projects: Dreyer called the population of 20 million US Americans requiring lung cancer screening per year “ideal” for AI assistance. “In AI diagnostics, we have thousands of solutions to approach,” he added.

Define Use Cases

As no standard method for AI model training or testing is established so far, this needs to a next step. So-called “use cases” are currently the major missing piece. A “use case” is a well-known technique in system analysis. It identifies and organizes all system requirements that might be important to the users.

This means radiologists need to identify what they want and which scenarios the want to work on. The algorithms-to-come will include imaging data plus genetic and genomic data, they will qualify findings and provide structured recommendations.

Getting Data Right

Dreyer pointed out that defining specific protocols and using just the right amount of data is crucial to achieve good accuracy.
Data ownership remains another fundamental point: “There is no answer yet about how access, acquire and use data – it will get more challenging when vendors use data.”