toolkits at no cost to the institutions to perform the annotation creation, transfer learning, and pipeline integration. In addition, Nuance will provide the last-mile technol- ogy required to integrate AI for the participating radiologist. Once the pilot is complete, the initiative is anticipated to progressively expand to all institutions interested in par- ticipating. Sharing local AI models from image data between institu- tions for fine tuning—while patient information remains securely on site at the originating institution—has not previously been done success- fully in radiology at this scale. This is due, in part, to the variability in how medical images are created, including the equipment, software, and protocols used. The pilot sites will use ACR AI-LAB to evaluate AI developed elsewhere, modify- ing the investigational algorithms to improve performance based on testing and evaluating them on local patient data. Creating the local AI models will not require ACR AI-LAB users to have programming skills. ACR AI-LAB allows users to adjust and change AI models without hav- ing to make line-by-line changes to the underlying code. Once the pilot is complete, the consortium is antici- pated to progressively expand to more institutions and vendors inter- ested in participating.”

One of the participating leaders in the initiative, Keith Dreyer, D.O, is the chief data science officer at the Boston-based Partners HealthCare health system, and is vice-chair of radiology at Massachusetts General Hospital and Brigham & Women’s Hospital, two Partners hospitals located in Boston. Dreyer says that “This is the wave of the future. We’re using this network to share information and insights. The prob- lem in taking so long to build these models—there have been missteps. And then small companies that have no access to data grab general data and build a model. So at Partners, Mass General and Brigham, we have 20 billion images, increasing by 1 billion a year, so we can build these models to improve patient care. That’s what’s new and trans- formational” about this initiative, he said. “These tools are on the verge of being in the hands of radiologists and clinicians. They understand what’s needed and the data, and the environments involved; they just need the models.”

Dreyer also spoke with Hagland

more broadly about the overall tra- jectory of AI in radiology. Below are excerpts from their recent interview.

How do you see the overall tra- jectory of the adoption of AI and machine learning in radiology practice?

I think you’ll see adoption happen- ing on a regular basis. There will be transformational things happening continuously. At some point in the next five years, people will be using it more than in the past five years. And reimbursement needs to hap- pen; no one gets paid to use AI.

When will the average radiologist be using algorithms to support the bulk of their work?

It’s going to be a gradual thing. Here’s how I think of this: you use the Internet today, and you didn’t use it when you were five years old, right? But I used email in 1988 in my hospital. And I was already banking online by that time. It’s the same kind of thing here—we will see gradual, continuously growing adoption. And people are going to look back and say, it’s crazy it was called AI. The classic question I get from reporters is, what is the amaz- ing thing that’s going to happen with AI? And it’s not going to work like that. It will just be adopted more and more, and become a part of clinical practice.

Will it change things in the non- diagnostic area?

Yes, people are starting to realize there’s incredible value in the data, so there will need to be heightened secu- rity and control of data. The classic large companies are trying to find ways to get health records, because that’s how you develop AI. That’s what IBM tried in acquiring Merge, but you need much more structured data. And what it reveals is that we don’t put a lot of that structured data into our interpreta- tions. So one will be data privacy and security, because companies will try to reach into hospitals and steal their data. And two, the quality of the data will improve. We’re doing a project at Mass General and Brigham around interpretations, instead of just saying ‘pulmonary embolism,’ on the image, we’re circling the lesion, and when we circle it, we say what it is to define

characteristics, and then we’re putting that into the database.

I thought that was what had hap- pened previously?

What happened was that they ended up with 24 years worth of data, with disparity of different image modali- ties, devices, and reporting styles, and that it made it harder. It takes a lot more data to get the answers, and how the data is organized makes a huge difference.

What would you like CMIOs, CIOs, imaging informatics lead- ers, and other healthcare IT lead- ers to know about all of this?

First of all, absolutely do not release your data to vendors, companies, modality companies, anyone who in theory can see your data, for purposes of giving you an answer for something. For all the companies in our institution that have data access agreements, we have strong language forbidding them from using our data for any reasons. Informatics people become stewards of patients’ data. It’s always been the case, but now it’s more important than ever. Data now has value, so people will be coming out of the woodwork to grab your data.

Might some people become con- fused about these issues?

Of course they might; this is very con- fusing. It’s like people going around to old folks who have coin collections and saying, hey, why don’t I take those coins off your hands?

Who will convene everyone, in terms of gathering together the practicing radiologists, the clini- cal informaticists, and others, to help move the specialty forward?

The ability to make AI, in the future, will be as ubiquitous as creating PowerPoints. As automation starts to happen, more and more people will be able to use this. That’s why the ACR created the Data Science Institute. At its core, its purpose is to teach radiolo- gists what AI is and what it’s not. And just four months ago, we introduced the AI Lab. In terms of building mod- els, you could actually go onto the ACR’s website and do it. You could take data and annotate it, push a but- ton, and create models. Anyone can do it. And that’s the entire point. HI


Page 1  |  Page 2  |  Page 3  |  Page 4  |  Page 5  |  Page 6  |  Page 7  |  Page 8  |  Page 9  |  Page 10  |  Page 11  |  Page 12  |  Page 13  |  Page 14  |  Page 15  |  Page 16  |  Page 17  |  Page 18  |  Page 19  |  Page 20  |  Page 21  |  Page 22  |  Page 23  |  Page 24  |  Page 25  |  Page 26  |  Page 27  |  Page 28