Skip to main content

What do Abraham Lincoln, Carl Sagan and artificial intelligence (AI) have to do with one other? More than you would think, according to Prof. Elizabeth Burnside from the University of Wisconsin, who presented the Plenary Session ‘Leading Through Technology: Valuing Artificial and Human Intelligence’ at RSNA 2023 last month.  

Using the American collective imagination, Burnside, a Chicago Cubs fan, tackled the challenges brought by generative AI using the analogy with a baseball ground, while touching on the meeting’s overarching theme of Leading Through Change. 

Prof. Elizabeth Burnside giving the Plenary Lecture at RSNA 2023 last month.

‘I’ll take you to first base and talk about leadership and how it matters when talking about the AI revolution,’ she said. ‘We’ll then move on to second base and talk about AI examples relative to the imaging chain, stakeholders, and continue to revere the importance of human expertise. Third base will take us where we are now. And last, as we reach home plate, we will envision the future.’ 

The challenges that come with generative AI are a wicked problem – one that is not easily solved, she went on. Different approaches to leadership – situational, transformational, functional – might be needed, but everyone can take the lead.

‘I encourage you to see yourselves as leaders,’ she added. ‘Leaders are not born, they’re made. And everyone leads at some point, and everyone follows.’

Discriminative AI and generative AI

It is important to distinguish between discriminative AI and generative AI. ‘Discriminative models, when they’re given an input, generate consistent outputs,’ she said. ‘Generative models come up with a myriad of outputs. It’s not the easiest concept to grapple with and there are many grey areas in between.’

Tasks performed by discriminative AI in radiology include, for example, assessing a chest x-ray to determine if a pneumonia is linked to Covid, or reading a mammogram to determine the presence or absence of cancer. 

Generative AI models train on similar data, but they do things such as use images to generate a report; improve performance on image segmentation; create summaries for patients in lay language; and simulate disease progression. 

‘Discriminative and generative AI models can often do the same task,’ she said. ‘I want to emphasize the nature of the output, which is either discriminative or generative.’

Decades ago, neural networks were mainly used for discriminative tasks, but with the explosion of computational abilities, new large language models (LLMs), transformers and convolutional neural networks (CNNs) are now doing more generative work.

AI software should come up with tools that increase trust. ‘I will always ask my AI colleagues to present me with an uncertainty map,’ she said. ‘It’s great for transparency and knowing how accurate the AI is.’   

CNNs can make intentional modifications to an image. ‘It might be useful to a nefarious agent who tries to mock up our diagnostic ability, or it might be incredibly useful for education and training,’ she said. ‘A lot of work needs to be done to increase trust in these types of image transformations.’ 

One should consider where AI may fit in the imaging chain. ‘In radiology, we are aware how we return value to patients through a pretty complex imaging chain. AI is present in the chain – e.g. detection of findings or post processing – but not everywhere,’ she said.

LLMs are likely to fit at many levels of the chain, she went on. ‘They can develop output that is instructional, they can complete information in radiology reports, or they can transform words to images or images to words.’

Abraham Lincoln and Carl Sagan

Coming to third base, Burnside invoked Lincoln, who approached leadership in the context of a wicked problem: abolishing slavery while making sure the young nation preserved unified constitutional democracy.

Abraham Lincoln, the 16th President of the United States of America

‘When Lincoln was elected in 1860, he assembled his cabinet from those people who competed against him in the election,’ she said. ‘He chose them because they had opinions that were widely divergent from his own. He talked, listened and tried to understand before he made any decisions.’

Lincoln also believed in public opinion. Measuring public opinion about AI may also be useful, according to Burnside, who shared recent surveys on the matter.
In a questionaire sent to all its members in 2020, the ACR found that only 3-4 % of respondents actually used AI, but also asked non-users for their opinion. ’80% saw no benefit and 35% said they couldn’t justify the expense,’ she said.

An Italian survey highlighted attitudes towards AI. Among the benefits brought by the technology, respondents mentioned workload reduction, work optimization, and efficiency. But they also thought AI might damage the reputation of radiologists, and demanded policies.

Patients approved replacement partial scenario and were pretty open to a triage and companion scenario, a survey from the UK showed. 

To conclude, Burnside looked up to Carl Sagan, who spent his life tackling nuclear proliferation. ‘He asked the question of how can we hope to be responsible informed decision makers regarding the inevitable challenges posed by newly acquired technology so few understand?’, she said. ‘I would like to encourage all of the luminary educators out there to really engage and help our community learn about AI.’