Katherine P. Andriole, PhD, FACR, FSIIM, FSPIE, an Associate Professor of Radiology at Brigham and Women's Hospital, researches the clinical implementation of artificial intelligence (AI) in radiology. Dr. Andriole recently presented at the SRG Pillar meeting for the Center for Intelligent Imaging (ci2) at the University of California, San Francisco. Her talk, "AI in Radiology: Experiences from Model Development to Clinical Implementation," showcased her findings.
Mass General Brigham (MGB) AI constructed a pipeline to deploy AI solutions in clinical workflows. The six-step process includes data collection, identifying cohorts, annotating, building models, validating, and deploying. Dr. Andriole notes that clinical domain expertise is required at each step along the pipeline.
Dr. Andriole and her team at MGB AI have garnered numerous clinical examples of AI deployment along the medical imaging chain, including rheumatoid arthritis versus osteoarthritis on hand x-rays, predicting patient no-shows, prostate MRI and lesion characteristics of breast ultrasounds.
The research using machine learning (ML) to improve exam schedules predicted the likelihood of late or absent patients. Analyzing demographic data and past appointment history, Dr. Andriole and the MGB AI team found that older adults with appointments in the winter and during high-traffic times aren't going to show up for their scheduled exam. To implement their research findings, MGB will contact patients who are likely to miss their appointments to schedule transportation accommodations.
The Smart MRI Scanner, developed by the MGB AI team, can detect image quality issues on MRI sequences. It allows technicians to adjust in real time and offers mitigation suggestions to improve the image quality.
To address the prevalent low-back pain in older adults, Dr. Andriole and her team interpreted over 4,000 lumbar scans using AI to predict stenosis. "The AI-generated inference will show what level [of risk the patient] is at and the level of stenosis across left, right and central," Dr. Andriole explained. The radiologist can confirm or disagree with the AI-generated findings, and then the data will be uploaded to the patient's report. This workflow improvement keeps the radiologist in the loop, so the AI tools are under supervision.
In a retrospective trial evaluating the performance of an FDA-cleared AI decision-support tool compared to breast radiologists, the AI tool was found to be more specific but less sensitive, with higher accuracy than breast radiologists. There's potential for the tool to decrease patient follow-ups; however, physician oversight is still required, Dr. Andriole said.
Dr. Andriole identifies that foundation models are of high interest for radiologists. This type of AI can enhance patient communication, report generation, advance diagnostic support, streamline workflows, manage data and predict disease risk. The Gen AI Foundation Model that the MGB AI team utilizes, Radiology CoPilot, can automatically draft report impressions, analyze prior findings, identify errors and improve efficiency and overall report quality. The Foundation Models that Dr. Andriole and the MGB AI team are using improve all aspects of the pipeline: schedule optimization, exam protocol, billing, follow-up, reporting and findings detection.
To ensure quality and patient safety, the MGB AI team monitors the performance of AI tools implemented in patient care. If the tools perform inconsistently and drop below a certain quality threshold, they are removed from use.
Some companies with SaaMD (Software-as-a-Medical-Device) tools partner with MGB AI to test for accuracy before seeking FDA clearance. FDA approval is part of the deployment intake process, where Dr. Andriole and MGB AI evaluate the logistical process for deploying a new AI tool for clinical use. Many agencies, including DICOM and IHE, regulate the standards of clinical integration.
The MGB AI team uses clinical performance metrics to understand the return on investment, not strictly financial returns. Are the implemented AI tools saving time, increasing accuracy and improving patient outcomes? Dr. Andriole seeks to discover that in her work.
On her lessons learned in AI, Dr. Andriole said, "This requires team science. It requires ... the data scientists' expertise, but also clinical domain expertise and implementation of informatics."
According to Dr. Andriole, the future of AI in healthcare is advanced imaging and diagnostics, remote monitoring and telehealth expansion, administration automation and workflow support, and improved trust and adoption.
To learn more about the upcoming SRG Pillar meetings, visit the ci2 events page.
