At SIIM-CAIMI 2026, the AI conference of the Society for Imaging Informatics in Medicine, the three program chairs agreed on one point: imaging AI is moving from evaluating algorithms to implementing them. What holds that transition back, in their view, is less about accuracy than about regulation, reimbursement, trust and getting tools into hospital systems.
The meeting ran October 26–27 at the University of Pennsylvania’s Smilow Center for Translational Research in Philadelphia. CAIMI stands for Conference on AI in Medical Imaging, the name that replaced the former CMIMI. Speaking to AuntMinnie, Katherine Andriole, PhD, Peter Chang, MD, and Eliot Siegel, MD, summarized what they saw across keynotes, abstracts and builder sessions.
From evaluation to implementation, in the chairs’ words
Siegel put it most plainly. “We’re sort of at a watershed moment right now as we end up making the transition to implementation rather than just looking at evaluation of AI algorithms,” he said. “Now we are looking at multimodal frontier models and multimodal models and also looking at implementation in real world applications.” He expects LLMs and agentic AI to dominate future programs, and LLM deployment to become routine in imaging departments.
Andriole pointed to practical tools, from detecting image artifacts to easing time-consuming reporting. “Many of them are combined models, so multimodal, where they’re using text information, healthcare information, as well as the imaging data to do a better job at detecting difficult findings,” she said. She also noted educational tools built to help residents.
Chang focused on accessibility. Traditional vision models need specialized expertise; with language models, he said, “people just come up with an interesting, practical, accessible idea, and they’re able to turn it into practice almost right away.” That echoes the society’s June annual meeting, where the top trends from SIIM 2026 already gave governance and adoption as much weight as raw performance.
Apollo and the case for models that cross specialties
All three chairs stressed that AI reaches patients only when radiologists work with pathologists, oncologists and data scientists. Their example was the October 26 keynote by Faisal Mahmood, PhD, of Harvard Medical School and Brigham and Women’s Hospital, who comes from a pathology background. He presented Apollo, which combines data from multiple specialties into a kind of search engine for following a patient’s health over time.
The group’s April 2026 arXiv preprint describes Apollo as a multimodal temporal foundation model trained on more than three decades of records from a large U.S. hospital system: 25 billion records from 7.2 million patients, spanning 28 data modalities and 12 specialties. The authors tested 322 prognosis and retrieval tasks on 1.4 million held-out patients, from five-year disease onset risk to treatment response. The work has not yet been peer reviewed.
Andriole, now at the University of California, Los Angeles (UCLA), called AI a team sport and described projects that brought together radiologists, pulmonologists, ICU physicians, transplant surgeons and data scientists. The weak link is pathology: “I think pathology is kind of where radiology was maybe 20 years ago or even 25 years ago because they’re not digital, and I’m not seeing a huge impetus for them to move to digital pathology, which is unfortunate.” Without digitized slides, multimodal models miss the layer that would confirm an imaging finding.
Agentic workflows blur builders and users
Chang, who ran a coding-agent demonstration at the meeting, described CAIMI as “a conference of developers and tweakers, folks that are building algorithms and getting them into the clinic, whereas most of the other conferences focus on taking those tools and deploying them.” With the technology and the agentive workflow, he added, “we’re really seeing a mix of those two audiences.”
An agent is an LLM-based system that does more than answer a prompt: it breaks a task into steps, calls tools such as a code editor, a database or a PACS interface, checks the output and fixes its own mistakes. A researcher can describe a dashboard or website in plain language and let the agent write, run and debug the code, which is how Chang sees people with little coding experience building tools. In radiology, the same pattern shows up in agents that pull prior exams or identify patients who need follow-up.
The chairs did not dwell on the cost of that openness. Code written with an agent still needs validation, a security review and an owner. Persuasive LLMs also make mistakes; a Radiology study found that radiologist expertise is what keeps readers from accepting bad AI advice.
Regulation, payment and the EHR: what blocks deployment
For Andriole, regulation, reimbursement and trust have to be addressed before AI becomes routine. Few AI tools carry reimbursement, she noted, even though “many places” adopt them for efficiency. In the U.S., most algorithms have no dedicated billing code and must pay for themselves in productivity. The Health Tech Investment Act, introduced in the Senate in 2025, proposes a Medicare payment pathway for algorithm-based services; the issue has reached Congress but is not settled. Supply keeps growing regardless: a recent analysis found FDA radiology AI authorizations up 42%.
Siegel named the most concrete obstacle. A recurring theme, he said, was “people presenting really excellent work and projects and development that they’ve done and innovations that they can’t get into the actual practice because it’s difficult to integrate these into the electronic medical record and standard proprietary tech systems.” An algorithm has to receive the right study at the right time and return results the PACS and reporting system understand, such as a DICOM structured object or an HL7 or FHIR message. IHE profiles such as AI Results (AIR) and AI Workflow for Imaging (AIW-I) standardize part of that path, but only if vendors implement them. When integration lags, results arrive late: across 20 Swiss sites, 7.2% of AI outputs landed after the report was signed.
What it means outside the U.S., including Brazil
The same barriers apply in Brazil and much of Latin America, often more sharply. Integration is harder with mixed PACS, RIS and EHR installations and small IT teams. Payment rarely covers algorithm analysis as a separate item, so the cost is folded into the exam price. Software that qualifies as a medical device in Brazil falls under ANVISA’s RDC 657/2022, and patient data sent to LLMs is subject to the LGPD, Brazil’s data protection law, including in in-house tools built with agents.
The practical takeaway matches the case for a radiology AI management plan before purchase: define the problem, the value metric and the integration path before choosing an algorithm.
Caveats and what to watch through 2027
The report gathers the views of three program chairs, not outcome data, and the idea that 2027 will bring more evidence is an expectation. Apollo, the strongest example cited, is still a single-system preprint. Worth watching: external validation of multimodal models, U.S. reimbursement, and whether PACS and EHR vendors open their systems to third-party tools.
Source: AuntMinnie — SIIM-CAIMI program chairs talk medical imaging AI’s evolution; Apollo preprint: arXiv — A multimodal and temporal foundation model for virtual patient representations at healthcare system scale
