Adopting generalist radiology AI (GRAI) rather than mix-and-match narrow AI approaches will produce viable radiology reports and improve radiologist efficiency with positive downstream effects, according to commentary published September 9 in Radiology.
“Clinical limitations of [task-specific approaches] narrow tools necessitate that radiologists ultimately review and edit imaging examinations for each patient,” wrote lead author Siddhant Dogra, MD, of New York University Langone Health and colleagues from Harvard Medical School and the Long School of Medicine at University of Texas Health in Houston.
Instead, the group has proposed GRAI, suggesting that it will be a “pivotal evolution in medical imaging,” make better financial sense for hospitals and health systems, and serve as a more “holistic approach” to radiology AI. With this work, the group aims to guide generalist radiology AI development going forward.
“GRAI should support radiologists throughout the entire imaging workflow, including detection, differential diagnosis, measurements, and comparisons,” wrote Dogra and colleagues, who offered five key features of the approach:
- Reporting with multifinding detection and characterization
- Producing reports tailored to indications for normal studies
- Producing longitudinal image comparison
- Incorporating patient characteristics
- Offering uncertainty-informed and interactive recommendations
Foundation models have paved the way for generalist radiology AI, according to the team.
“Like generalist medical AI, GRAI will have three distinguishing capabilities: dynamic task specification, where it adapts to new tasks described in plain language without retraining; the ability to accept inputs and produce outputs using multimodality data (e.g., any or all of images, laboratory values, and operative notes); and the capacity to reason through unfamiliar tasks and logically explain outputs,” the authors explained.
GRAI should be capable of providing fully characterized diagnoses with the reference standard based on the input image and patient clinical history data, they noted. In their paper, they highlighted six primary imaging diagnoses: pulmonary embolism, kidney stone, appendicitis, intracranial hemorrhage, fracture, and diverticulitis.
Diagram shows examples of six primary imaging diagnoses (indicated by yellow arrows) with pertinent characteristics that generalist radiology AI (GRAI) should describe if present in (B) the corresponding radiology report.RSNA
Training models for GRAI would incorporate common data elements (CDEs) as proposed by the RSNA and the American College of Radiology and seen on RadElement.org, the group noted.
Performance review and data drift monitoring would still be required, but generalist models could be expected to better adapt to data drift than narrow models, given their inherent adaptability using in-context and few-shot learning, whereby a model generalizes to a task after being shown a handful of examples, Dogra and colleagues said.
The paper also highlighted financial, operational, and clinical limitations of the current approach for radiology AI. Using GRAI, radiologist efficiency would be expected to truly improve with downstream effects of decreased cognitive burden, helping address burnout, worsening imaging backlogs, and delayed reporting, they added.
However, the group pointed out that reimbursement pathways will require a major overhaul to support integration of GRAI into clinical practice. Current reimbursement structures are designed for specific diagnostic and procedural tasks, making them ill-suited for a system that augments radiologists across multiple domains of imaging interpretation.
“To address this, policymakers and payers must develop new reimbursement models that account for the broad, adaptable nature of GRAI, potentially introducing performance-based payment structures that reward AI’s contribution to efficiency, diagnostic accuracy, and patient outcomes,” the team wrote.
Addressing current U.S. Food and Drug Administration (FDA) regulatory limitations, the group recommended adapting regulatory pathways to accommodate broader AI systems such as GRAI and expanding on Predetermined Change Control Plan (PCCP) principles.
Read the rest of the commentary here.