Foundation models could reshape radiology, but their strongest near-term role is as assistants that draft and accelerate work rather than as autonomous replacements, according to an analysis published November 6 in Current Problems in Diagnostic Radiology.
These AI models could help with operational efficiency by automating routine tasks, blending information, providing decision support, and allowing radiologists to devote more attention to complex tasks, according to the analysis written by Yee Seng Ng, MD, and Elise Dunning, MD, from the University of Washington in Seattle.
“As model capabilities continue to improve and domain-specific training datasets expand, their impact on clinical practice is expected to grow,” Ng and Dunning wrote.
Foundation models could reshape radiology, but their strongest near-term role is as assistants that draft and accelerate work rather than as autonomous replacements, according to an analysis published November 6 in Current Problems in Diagnostic Radiology.
These AI models could help with operational efficiency by automating routine tasks, blending information, providing decision support, and allowing radiologists to devote more attention to complex tasks, according to the analysis written by Yee Seng Ng, MD, and Elise Dunning, MD, from the University of Washington in Seattle.
“As model capabilities continue to improve and domain-specific training datasets expand, their impact on clinical practice is expected to grow,” Ng and Dunning wrote.
Foundation models are pretrained on large, diverse datasets and then adapted to many downstream tasks through fine-tuning or prompting. These models often use less labeled data than traditional narrow-task AI requires.
The review authors credit advances in transformer architectures, self-supervised learning, and vision-language models, which can process both images and text. Rather than relying on manually labeled datasets, these models learn from data such as paired images and radiology reports. The MIMIC-CXR chest radiograph dataset, for example, includes more than 370,000 image-report pairs.
Chest radiography is the most mature area, thanks to large public datasets and a growing ecosystem of open-source and commercial models, the authors wrote. CT models such as CT-Chat, Merlin, and BrainGPT have seen limited clinical validation. MRI models focus mainly on brain imaging and are largely task-oriented rather than built for report generation.
Ultrasound and mammography lag because large datasets are scarce, with the mammography model Mammo-FM a notable exception whose clinical validation remains limited, the authors wrote.
Report drafting leads the applications
Automated report generation has drawn the most attention, driven by workforce shortages and rising imaging volumes. Other applications include text summarization and clinical decision support, which the authors describe as one of the earliest areas of adoption in radiology.
Ng and Dunning wrote that any institutions are working with vendors under IRB-approved protocols to evaluate report generation prospectively, generating data for future regulatory submissions.
Current performance, however, supports a drafting role. In an AI challenge at the 2025 American College of Radiology Annual Meeting, attending radiologists who were blinded to authorship judged 65.4% of AI-generated chest radiograph reports clinically acceptable, compared with 79.6% of radiologist-written reports.
Evaluation and risk challenges
Evaluating generated reports is difficult because no single correct output exists, the authors wrote. Automated tools such as CheXbert, GREEN, and CheXprompt help, but each has limits, and manual radiologist review remains important. Known risks include hallucinated findings, omitted findings, and incorrect laterality or fabricated prior comparisons, all of which can appear in fluent, plausible language.
The authors outlined other barriers to implementation, including the following:
- Regulation: Pathways may include 510(k), De Novo, or premarket approval, and no single metric fully captures report quality.
- Privacy and security: Cloud-based models raise concerns about protected health information and prompt injection, leading many institutions to adopt enterprise-approved platforms.
- Equity: Smaller hospitals and private practices may lack the resources to deploy such platforms, potentially widening disparities.
- Bias: Training data from a few large academic centers may leave models poorly calibrated for community practice.
- Automation bias: Radiologists may place too much trust in AI output and need to stay vigilant.
The near-term outlook
The authors concluded that foundation models may have their greatest early impact indirectly, by speeding development of specialized AI tools with less labeled data while fitting within existing regulatory pathways for narrow devices.
They added that radiologists should develop familiarity with the technology and take part in its evaluation, governance, and implementation.
“As the field continues to evolve, radiologists who understand the capabilities and limitations of foundation models will be best positioned to guide their responsible adoption and maximize their benefit to patients,” Ng and Dunning wrote.
Read the full analysis here.