Alexis is director of content at AIMed, with responsibility for the research, development and delivery of products across events, digital and publishing. A highly experienced events executive with a career focus on the intersection between healthcare and technology, he is also a school governor leading on teaching, learning, and quality of education.
Dr. Christoph Wald, Chairman, Department of Radiology at Lahey Hospital & Medical Center, explains the important role of a radiology AI platform in AI deployment…
“Radiologists often spend quite a bit of time arranging images and scrolling through them to find the same corresponding location for comparisons,” says Dr. Christoph Wald, chairman, Department of Radiology at Lahey Hospital & Medical Center in Burlington, Massachusetts (LHMC). “So, when we learned of a technology that can be plugged directly into the picture archiving and communication system (PACS) to automate the process, we adopted it right away.” That original technology was the hospital’s first touchpoint with Blackford, an Edinburgh-based medical imaging company.
“This registration tool is useful to a radiologist who has to interpret images from the same patient, taken at different times using different modalities,” Dr. Wald adds. “That’s why we were interested in deploying Blackford’s initial product.”
That initial product was ‘Registration’, but utilizing their experience, Blackford identified that their ‘Platform’ system offered a better way to adopt and deploy imaging technologies and AI. Having ‘Registration’ already in place, LHMC were simply able to leverage the Platform to deploy multiple smart imaging and AI applications within existing systems and workflow whilst reducing the cost of implementation and ongoing management via a single vendor.
While Blackford’s AI solution highlights that new applications can be added and deployed quickly and easily on an existing platform, within the industry there still remains hurdles preventing the widespread adoption of AI in radiology. “The challenges we face are general and not specific to any particular AI tool,” explains Dr. Wald. “For instance, many AI tools were built to solve one specific task. Yet, radiologists are required to accomplish many tasks when interpreting even a single imaging study. An algorithm may be deployed to detect blood in the brain from a CT scan, whereas radiologists have to also look beyond the blood, for instance for fractures, atrophy, and everything else in the anatomy included in the scan.
“I can envision a future state where I study a brain CT scan, and use five or more different algorithms, and perhaps these are provided by five different vendors. Each of these vendors would typically require a separate install of their software and will generate individual specific outputs. I would have to review them all and integrate the results with what I see in PACS before I can produce the final report. That whole process would be just cumbersome.
“Another concern is that AI only works well if the datasets used to train it are similar to the patients on which I am using the AI. Because all radiology departments have slightly different combinations of machines and patient populations, I am not able to tell upfront whether a particular AI tool will work well in my practice and on my patients.”
Dr. Wald also believes that the results offered by AI today are not necessarily presented in a way that fits radiologists’ workflow. “There are a couple of common approaches that AI companies have taken,” he explains. “One is that they create little widgets displaying the AI results and radiologists can either agree or disagree with them. As mentioned, radiologists perform many tasks in a day. A future with many different widgets showing results of each AI cluttering our screens is not feasible. Some vendors and practices choose to send the AI results into PACS right away. We don’t permit this at all at the moment in our practice, because AI is not always right. If radiologists ever notice the AI results are incorrect after the fact, it’s difficult to remove it from PACS.
“PACS is designed in a way that makes it hard to remove information as it is part of the official medical record of a patient, and systems are designed to prevent inadvertent or unauthorized removal of information from that record. As soon as anything is stored in PACS, other members of the care team can view the information, and make clinical decisions, which could result in adverse consequences if they happen to read the flawed AI results or AI results that have not been verified by a human radiologist. Therefore, in our department we do not currently release AI results into PACS.”
Dr. Wald also highlights the absence of a reliable and comprehensive reimbursement mechanism for AI. “We are not reimbursed for using AI tools, which is challenging but hopefully also changing over time. This means the healthcare facilities need to budget and pay for the licensing costs out of their existing margin. This limits the number of algorithms we can deploy, making us extremely selective.
“AI in imaging is fairly new and standards and best practices are still emerging. While the sector may be agile, there are no clear ways on how things should be done.” Dr. Wald hopes for a multi-purpose viewer to display AI results at the time when radiologists are interpreting the images, to overcome the multiple widget issue. “I believe this is a logical evolution of a radiology AI platform,” he offers.
Dr. Wald is interested in newer AI which can quantify disease in patients, to help with diagnoses not easily made by radiologists, enabling them to comment on disease progression and prognosis. “Neurodegenerative diseases are conditions where the brain function decays over time. The patterns of shrinkage and the speed at which the brain shrinks over time tells you something about the prognosis of the disease. By providing tools to radiologists to describe these patterns more accurately and quantitatively they can better support their patients and referring neurologists.
“It’s simply not easy and sometimes impossible for a human reader to manually quantify these patterns, particularly in complex anatomy like the human brain cortex, which is folded in an extremely complicated manner. In these situations, AI complements the radiologist, rather than competing for the same task, guiding them in their understanding of the progression of the condition. If we know a patient’s brain is shrinking by 18% per year and 35% per year for another patient, we can conclude the latter has a worse prognosis.”
Meanwhile, platforms such as Blackford’s are proving adept at addressing today’s challenges faced by radiology – supporting the evaluation of AI on data, reducing implementation time, costs and long term maintenance. While continuing to evolve to address the further challenges that lie ahead.
“I believe radiology AI technology will continue to evolve in a meaningful way,” adds Dr. Wald. “Particularly as radiologists and technology companies think critically and discover new ways on how radiology AI can best be implemented as we all work toward our shared goal: providing the best care we can for our patients.”