From Stroke Detection to Cancer Screening: AI Tools Advancing Imaging Efficiency
In This Article
- Artificial intelligence (AI) tools have the potential to greatly improve medical imaging, but development and implementation present significant challenges
- Mass General Brigham has developed several novel initiatives to address the influx of AI in medical imaging and other fields
- The Imaging AI Governance Committee vets and validates tools for use within the Mass General Brigham system
- The Healthcare AI Challenge crowdsources input from medical professionals on potential AI tools
- Mass General Brigham's AI Clinical Research Organization partners with industry to provide medical expertise and evidence on AI tools
Mass General Brigham has launched several initiatives to help address the overwhelming influx of artificial intelligence (AI) tools in healthcare, particularly radiology.
Subscribe to the latest updates from Neuroscience Advances in Motion
"There's a big gap between training an AI model for research purposes and developing an AI-enabled medical device that's ultimately going to be approved by the FDA, commercialized, and used by healthcare providers in patient care," says Bernardo Bizzo, MD, PhD, a trained diagnostic radiologist and faculty at Massachusetts General Hospital, and senior director of Mass General Brigham AI. "It's really hard to get it to market and get it used by hospitals."
Efforts involving Mass General Brigham AI, such as the Imaging AI Governance Committee, Digital Clinical Research Organization, and the Healthcare AI Challenge aim to tackle the complex challenges of developing, validating, and implementing AI tools in clinical practice.
Training a Convolutional Neural Network to Improve Detection of Acute Infarct
"The technology most used to develop imaging-based AI models in the industry for the past decade has been based on convolutional neural networks (CNNs)," Dr. Bizzo explains. CNNs are a type of deep learning algorithm that uses data to learn patterns and perform tasks. They are often used for object and image recognition, classification, and detection, which makes them particularly useful in developing tools for radiology. "They are what we now call narrow AI tools, as they are typically focused on one specific task for one specific modality, one specific body part."
One example is a tool he and colleagues recently developed to detect acute infarcts on head computed tomography (CT) scans. CT is the imaging technology most commonly accessible in average hospital and emergency department settings, but the test can be insensitive for this medical emergency, particularly in its early phases. However, the finding can be more easily seen with diffusion-weighted magnetic resonance imaging (MRI), which is not always available in emergency settings.
Dr. Bizzo and colleagues used a unique data set of patients at Mass General Brigham who had acute infarcts and head both CT and MRI exams acquired very close by. Using the data, they trained a CNN to detect acute infarcts including those potentially not yet visible to the human eye using CT. It outperformed expert neuroradiologists in detecting acute infarcts, with a sensitivity of 96% versus 61% to 66%. The groundbreaking findings were published in Scientific Reports in 2023. The researchers' industry partner is currently preparing to submit the tool for FDA review.
Challenges Facing AI in Radiology
That innovative solution is considered to be a success from a research perspective, and it could be a very useful tool in radiology practice, Dr. Bizzo says. However, it is a drop in a large bucket, and there are several hurdles to implementing such a tool into clinical practice.
"Detecting an acute infarct represents one single finding in a radiologist's interpretation process. Each one of those narrow AI tools typically just does one task. But when a radiologist is reading a head CT, for example, there are hundreds of potential findings they must assess at to complete their interpretation," he explains. "There are thousands of findings that would need to be addressed by these narrow AI models to do the work of a radiologist across different imaging modalities and organ systems. Although there are thousands of such devices, many are doing the same task, such as looking for brain bleeds and lung nodules. There are only a few dozen unique use cases that are being covered by the existing AI products in the market for radiology."
Clearly, there's an enormous gap between what radiologists need and what narrow AI tools are offering today. Furthermore, there are many challenges to implementing AI tools into clinical workflow, such as:
- Addressing technical challenges in connecting these tools with clinical systems and their maintenance, which are often standalone solutions with their own interfaces and technical requirements
- Assessing cost and contracting with individual vendors that provide individual solutions
- Ensuring a robust framework for cybersecurity risk assessments
- Conducting internal validation and monitoring to ensure that AI tools perform well with your local data and workflows and continue to do so over time
- Understanding and measuring the impact of these solutions on end users, patients, and your health system as a whole
Imaging AI Governance Committee
Dr. Bizzo and colleagues reviewed some of the above challenges in a recent article in Data Science, describing Mass General Brigham's AI Imaging Governance Committee, established to address challenges and implement solutions. The group is led by the radiology department with representation of other groups across the enterprise, including other clinical departments based on use cases.
"We have well-established processes to intake, assess, and help prioritize the use of imaging AI tools across Mass General Brigham through that AI governance so that these solutions can be vetted for safety and effectiveness before being used clinically. The group assesses requests, addresses challenges, and helps facilitate implementation," Dr. Bizzo says. "And if a tool is approved and implemented, the group continues to monitor its use. We perform independent validations, testing whether it lives up to expectations with local patients and local data."
Healthcare AI Challenge
Another strategy is the Healthcare AI Challenge, a new, multi-institutional, virtual, interactive series of crowdsourcing events designed to allow healthcare professionals to explore and assess the latest AI technologies in real-world healthcare scenarios.
The goal is to help healthcare professionals, health systems, and AI manufacturers determine the effectiveness of AI innovations, as well as assess their safety, effectiveness, and value. Healthcare professionals who participate will access the performance of generative AI tools for specific medical tasks, such as medical image interpretation and clinical scenarios related to electronic health records. They can provide feedback and ratings on the tools' performance and utility. The insights generated will help with institutional decision-making about generative AI solutions and inform the whole community, including regulators, about the clinical utility of these tools. The public and industry can follow the initiative's progress online.
Radiology is the first in the series, given the rapid proliferation of AI tools in the field, and is rapidly expanding to medicine and other areas. The initiative is partnering with the American College of Radiology as well as several other prominent institutions across the country. (See also the "Keeping Pace With the Rise of AI" section in this article.)
Partnering With Industry to Develop AI Innovations
Dr. Bizzo also leads a group that partners with industry to develop and validate AI tools that can help improve patient care. The Mass General Brigham AI, with its Digital Clinical Research Organization, was instrumental in his infarct study. The team works with companies around the world who want to introduce AI tools in the U.S. market.
"Our group generates clinically relevant evidence and medical expertise to inform the key steps — from concept to market adoption — before an AI tool reaches physicians and patients. We can really show the promise these tools hold in and the impact that can have in helping with increased demands for radiologists and other physicians, improving patient care."