By: Peter Ngum, Clinical Application Specialist, and Greg Kingston, Chief Marketing Officer (CMO)
The latest stop on our 2022 conference schedule was the European Congress of Radiology (ECR) in Vienna, Austria, at which we were accompanied by Luis Brenes, Combinostics Clinical Application Specialist. The conference highlighted artificial intelligence (AI) in an extensive program, including many sessions on machine learning, deep learning, and big data in its Artificial Intelligence Theatre. Based on the sessions we attended and conversations in our booth, the conference delivered on its theme of “Building Bridges” between research and practice. However, there remains a need to build trust in AI applications in radiology, which is among the top conference takeaways for our team.
AI in radiology: here to stay
After attending three conferences in the United States recently, it was interesting to be talking with our colleagues in Europe again — and interesting to hear where radiologists in Europe are in terms of use of AI in practice. In conversations with new contacts and customers alike, viewpoints ranged from apprehension, but accepting that AI is here to stay, to complete adoption and wanting to expand their AI-supported radiology services. Meaningful discussions in our booth centered around the practical value of AI, how it would fit within existing workflows, and its impact on day-to-day outcomes.
To us, it feels as if the European radiology community has reached the slope of enlightenment, seeing the potential of AI technology to help improve workflows, reduce workload, increase objectivity, improve communication with referring clinicians, and, ultimately, improve patient care.
For example, a question during our research presentation session, Validation of automated MS lesion detection in two independent cohorts, indicated a desire to automatically identify multiple sclerosis (MS) lesion dynamics in real time, to alert the treating physician of any significant changes as soon as possible. While the technology is not capable of this yet, it’s interesting to hear the tools radiologists would like to start incorporating into their practice and exciting to consider what this means for the future of radiological services.
It also brings to mind the multidisciplinary approach required to develop new AI methods, a concept that surfaced at ECR 2022 as well as other recent conferences we attended, including the Annual Meeting of the American Society of Neuroradiology (ASNR). Expertise across radiology, neurology, physics, statistics, and machine learning domains is needed and as early as possible in the development of an AI model. This multifaceted domain knowledge helps answer questions such as:
- What are the pain points being experienced by the end users?
- What would they like to be able to accomplish?
- Which features would make their job more efficient or provide insights they weren’t able to achieve previously?
- What is possible with existing technology?
- How do we ensure it is accurate and valid within this context?
- Which datasets do we need to train and develop new models?
Getting the right people in the room early can help make sure that the technology will be valuable, accepted, and trustworthy — minimizing the need for costly, time-consuming changes after the technology has been deployed.
Moving past the “black box”: gaining an understanding of AI
A particularly strongly recurring theme at ECR 2022 was trust in the AI behind radiology applications. The concept of “Explainable AI” captured the need by radiologists to understand how a model reaches its conclusions. For many, AI remains a “black box” that mysteriously produces an output, prompting the question: Can we trust a machine to accurately predict disease state, especially if we don’t know the methods it’s using?
Interestingly, we heard about a recent study of radiologists’ confidence in AI in one of the sessions. In the study, all diagnoses were performed by human experts based on chest x-rays, but some of the diagnostic advice was labeled as coming from AI. After viewing the recommendations, the radiologists participating in the study were more likely to rate the diagnostic advice they believed was from AI as lower quality, highlighting the potential bias toward AI methods. This same skepticism was presented in another way, by asking clinicians whether clinicians would inform their patients that an AI system was used to determine their diagnosis. If so, how much information did they think should be shared about how the algorithm works? It would be interesting to know how the public views the ability of AI vs their human doctors to reach an accurate diagnosis. Does the general public have the same preconceptions?
Although a number of scientific sessions described how AI systems are being validated, their content focused on the technical aspects and science behind the system validation, attracting an audience of people with a technical background and researchers and reinforcing the idea of AI as a black box in radiologists’ minds. To increase their comfort with using AI, radiologists and clinicians want to know how that validation process translates to the AI-derived recommendations they’re seeing in practice. Can they really believe what the system is predicting? In fact, the conference presentations by radiologists asked for greater transparency in the specific inputs and an overview of how they’re used to calculate a prediction.
However, is an in-depth understanding necessary? Speakers in one session debated this very question, with one speaker comparing AI to driving a car: do most people driving a car understand how it works? Another speaker asked whether oncologists understand the pharmacology behind the treatments they administer to their patients? For both of these examples, there are processes governing their development and safety. The rigor and transparency in the processes instill some confidence that a car will be manufactured correctly and safely and that drugs will work as intended and be safe for their patients. The same level of scrutiny has not historically been applied to AI platforms, although this is changing with the adoption of more stringent medical device regulations in markets such as Europe and the United States.
Some researchers have taken steps to provide some transparency in this area. Characteristics of the different commercial AI systems for radiology (subspeciality, modality, main task, CE marking, FDA clearance, deployment method, pricing model, clinical feasibility, and more) are captured on a website, AI for Radiology, with the methodology behind the information retrieval documented in Imaging Informatics and Artificial Intelligence. We had the pleasure of speaking to one of the authors at ECR about the project. Interestingly, at the time of the study, only 36 of the 100 evaluated CE-marked products the authors listed had peer-reviewed evidence on the efficacy of the AI software, and few companies had addressed the clinical impact of their product.
This lack of transparency has prompted some in the industry to advocate for labeling of AI systems, similar to how approved drugs are labeled. The required information for AI could include data sources and types, the population in the dataset, algorithm validation methods, evaluation of how bias was assessed, and performance metrics. Some systems, such as Combinostics cDSI™, part of our cNeuro® platform, report the contribution and relevance of each biomarker and data point to the system’s decisions, making it easy for end users to review the sources themselves and determine their own level of confidence in the prediction.
Trust in the output: new opportunities for value-added services
It is this type of transparency in our platform and its reporting that has contributed to end user confidence in its predictions. We heard firsthand at the conference that some attendees previously considered our products helpful only for patients with dementia; however, after viewing our clinically focused reports for MS, epilepsy, and traumatic brain injury (and the methodology behind them), they were interested in expanding the breadth of patients they could support using our platform. This feedback extends to our Dementia Differential Analysis report, which we presented at the Artificial Intelligence Theatre. The value of the report continues to be viewed positively, especially since the contribution of each imaging biomarker to the predicted diagnosis is openly available for review. Watch the video below to learn more about the imaging biomarkers and how they’re used to determine the most probable dementia diagnosis.
Up next: improving education and transparency
As we settle back into our day-to-day responsibilities, one thing was clear to us: more work is needed to demystify AI and its value for radiology. This includes general education and open conversations about the development and validation of our systems — and how that supports day-to-day use in practice with real patients. In other words, what is the clinical relevance?
We’re looking forward to returning next year and evaluating whether we’ve been successful in moving the radiology community further up the slope of enlightenment. In the meantime, our next stop is the Alzheimer’s Association International Conference in San Diego, CA, July 31 to August 3. Will we see you there?
