About

David Larson is a Professor of Radiology specializing in Pediatric Radiology at Stanford University. He is affiliated with the Center for Artificial Intelligence in Medicine & Imaging (AIMI), where his work focuses on the application of artificial intelligence and imaging technologies in healthcare. His research involves advancing medical imaging techniques and integrating AI solutions to improve pediatric radiology diagnostics and patient outcomes. As a faculty member at Stanford, he contributes to the center's mission of innovative research and education in AI-driven medicine.

Research topics

• Computer Science
• Medicine
• Artificial Intelligence
• Engineering
• Machine Learning
• Operations management
• Nursing
• Knowledge management
• Computer Security
• Accounting
• Pathology
• Psychology
• Finance
• Systems engineering
• Business
• Process management
• Management
• Physical therapy
• Medical physics
• Risk analysis (engineering)
• Law
• Algorithm
• Internal medicine

Selected publications

• Predicting the Value of Radiology Artificial Intelligence Applications: Large-Scale Predeployment Evaluation of a Portfolio of Models

  American Journal of Roentgenology · 2026-03-04 · 1 citations

  article1st authorCorresponding

  Background: Real-world performance of radiology artificial intelligence (AI) applications frequently diverges from previously reported results, creating challenges in anticipating a model's clinical value and impact.

  PublisherDOI

• Design for Understanding: Sentence-Level Cross-Referencing Improves Comprehension and Confidence in AI-Generated Patient-Friendly Radiology Reports

  Journal of the American College of Radiology · 2026-03-01

  articleSenior author
  PublisherDOI

• Use of Pediatric Imaging is Increasing Again. Now What?

  Hospital Pediatrics · 2026-04-27

  article1st authorCorresponding

  In their article, “Trends in Pediatric Imaging from 1997–2024 in an Integrated Healthcare Setting,” Mahendra et al describe the growth of the use of magnetic resonance imaging (MRI), computed tomography (CT), and ultrasonography imaging of children in the Kaiser Permanente Northern California region1 as a representation of likely national trends. Their research found that CT use declined during the 2010s, whereas the use of MRI and ultrasonography increased, suggesting that imaging may have been shifting from CT to MRI and ultrasonography during this time. Imaging in all modalities then leveled off by the end of the decade, suggesting that overall use of advanced pediatric imaging had appeared to stabilize. Since 2020, however, imaging experienced renewed exponential growth beyond what would be expected from post-pandemic rebound alone. The authors’ findings have significant implications from a pediatric imaging health services research perspective.In general, innovation—such as the use of imaging technology—tends to spread (or be adopted) across a population according to a sigmoid growth model, with 4 general phases2 (Figure 1): Introduction phaseGrowth phaseDeceleration phaseStationary phaseIn reality, the pattern of diffusion of technology can vary, with temporary plateaus, periods of decline, or cycles of decline and renewal, depending on the underlying drivers. Furthermore, even when the growth follows the sigmoid curve precisely, it is usually difficult to predict the saturation rate of use while the technology is still in the growth phase. Nevertheless, the model is helpful to understand the current state and project into the future.The fact that advanced pediatric imaging volumes seemed to have plateaued by the late 2010s suggested that it may have reached the stationary phase. However, this article suggests that the plateau was, in retrospect, only temporary.This renewed increase raises concern because of associated risks—particularly radiation exposure from CT and anesthesia for MRI—as well as both direct and indirect costs of imaging, including system strain and provider burnout.The findings raise 2 pressing questions: what underlying forces are driving renewed growth, and—assuming increased use does not improve outcomes—what can be done to dampen it?The authors attribute the decrease in CT use in the 2010s to campaigns to raise awareness about radiation risks.3 They then attribute the renewed increase in MRI and CT use to 1) increasing complexity and severity of illness, 2) decreasing time that clinicians have to evaluate their patients, 3) a paucity of evidence-based guidelines, 4) defensive medical practice, and 5) low-value imaging use (which is a euphemism for systematic ordering of imaging that is not indicated).4 These proposed drivers are plausible but necessarily speculative because the study was not designed to establish causation. In the authors’ defense, identifying causal mechanisms in complex use patterns is notoriously difficult, particularly when multiple small forces interact over time.5 Although identifying drivers is difficult, designing effective interventions is harder still.Pediatric imaging use reflects a phenomenon described by Thomas Schelling in Micromotives and Macrobehavior: large-scale patterns emerge in unpredictable ways from many individually rational local decisions.6 Although image volumes are measured in the aggregate, imaging studies are ordered 1 patient at a time under conditions of uncertainty, time pressure, incentives, expectations, and availability. Macrobehavior does not reflect macro-intent; rather, subtle shifts in local decisions, repeated at scale, produce what can be dramatic aggregate effects. It is difficult to predict the aggregate results by examining the local decisions and even harder to identify the factors driving local decisions based on the aggregate results.One hypothesis is that the plateau of the 2010s may have reflected temporary countervailing pressures—such as heightened awareness of radiation risk—rather than a durable shift in underlying incentives. If local microincentives continued to favor imaging through scanner availability, clinical time pressure, or defensive practice, the aggregate pattern would be expected to eventually reassert itself. Regardless of the precise drivers, what is to be done about the findings?At its core, imaging is a tool to reduce clinical uncertainty.7 Reframing rising use as increasing reliance on imaging to reduce clinical uncertainty shifts attention upstream to the factors shaping ordering decisions. Therefore, the answer to inappropriate use may lie less in curbing imaging and more in developing other means of managing clinical uncertainty.Schelling reminds us that addressing the accumulated effect of many small decisions requires altering the conditions under which those decisions are made. Several broad strategies for controlling imaging use follow from this framework.Not every actor in a system is the same. Levels of experience, tolerance for uncertainty, and amenability to decision support differ between a newly trained nurse practitioner in an outpatient clinic, a time-pressured emergency medicine physician, and a seasoned subspecialist. Change strategies should account for these distinct archetypes rather than assuming that all ordering providers use a uniform decision calculus.Almost all behavior is influenced by local leaders and peers, even when this influence is not recognized. Therefore, as much attention should be paid to the social dynamics of local systems surrounding ordering of imaging, including committed leadership, group norms, and peer reinforcement, as is paid to policy or technical infrastructure.Simply telling clinicians to image less does not solve the underlying problem of clinical uncertainty. Investing in other local context-specific mechanisms to reduce clinical certainty, such as decision support, clinical pathways, or rapid access to subspecialty consultation, may decrease the need for imaging to answer clinical questions.A significant part of the actors’ decisions occurs at the point of local interaction with “the system.” Small design choices at the point of ordering can have large impacts on behavior. For example, details such as order set defaults, embedded appropriateness criteria, and automatic display of recent past imaging can meaningfully shift ordering decision thresholds.When a behavior is fast, easy, and immediately rewarding, its use tends to expand; the converse is also true. Carefully designed mechanisms to readily facilitate appropriate ordering and impose pauses for inappropriate ordering can meaningfully influence aggregate outcomes.Over time, feedback is among the most powerful drivers of results in a complex system. Transparent reporting of ordering patterns at the clinician or group level can create self-regulating dynamics, especially when accompanied by group discussion and commitment.Health care is local—and ordering of imaging is hyperlocal, involving single or small groups of individuals. Therefore, durable change likely requires engagement at the level of local clinical teams and service lines. This means that meaningful changes are likely to come more from pediatric providers, administrative leaders, and information technology systems managers than from radiologists (although your local radiologist is almost certainly happy to partner with you to improve the value of imaging).The Schelling framework suggests that as imaging becomes more readily available, clinicians’ time pressure becomes greater, and as subspecialty expertise becomes less broadly distributed, the use of imaging is not likely to curb itself. The stationary phase represents an equilibrium where the drivers of increased use are in balance with the limiting factors. As Paul Batalden observed, every system is perfectly designed to get the results that it gets.8 Structural aspects, such as throughput pressures or easier access, tend to outweigh generalized ethical exhortations. Whether the system’s implicit priorities are appropriate cannot be judged in the aggregate; that determination requires case-level assessment.That work is inherently local and more difficult than assessing aggregate outcomes and issuing broad directives. For that reason, solutions will likely need to operate at a higher structural level, such as establishing institutional mechanisms for imaging stewardship, with ongoing monitoring, feedback, and shared accountability. Modern approaches to antibiotic stewardship can serve as a guide.9The persistence of growth in imaging suggests that the system has not reached a stable equilibrium, indicating that shifts in local incentives continue to spur increasing use of imaging.This reality should motivate action while cautioning against both complacency and alarmism—complacency risks lack of urgency to solve the problem whereas alarmism risks rushing to overly blunt mechanisms to restrict imaging. The more effective path lies in systematic study and deliberate redesign of the decision environment—aligning local incentives with stated priorities and supporting professional self-regulation—accompanied by feedback and engagement with local leaders and provider groups.

  PublisherDOI

• Reply to “Alert Burden and Multivendor Interactions: Missing Dimensions in Radiology Artificial Intelligence Predeployment Evaluation”

  American Journal of Roentgenology · 2026-03-25

  article1st authorCorresponding
  PublisherDOI

• Expert panel on monitoring radiation doses from recurrent medical diagnostic procedures: Sixth Gilbert W. Beebe Webinar

  Journal of Applied Clinical Medical Physics · 2025-02-26 · 4 citations

  reviewOpen access

  Recurrent imaging is an essential tool for patient care but with an attendant dose from radiation exposure. Recurrent imaging has been the subject of increasing scrutiny and debate based largely on the risk from increasing cumulative doses. However, the accountability for and actions with recurrent imaging as a special component in the general construct of radiation protection in medicine is unclear. This is demonstrated by the perspectives provided by the various imaging community experts. Some perspectives may be different, but many share a common ground. Understanding these various perspectives illustrates the wide-ranging optics in considering benefits and costs in the recurrent imaging paradigm and, moreover, the value in pursuing multi-stakeholder-derived harmonization for recurrent imaging and radiation dose. This move towards consensus would be to the benefit of the imaging community, referrers, and other related healthcare professionals, as well as patients, their caregivers, and the public.

  PublisherOA PDFDOI

• Expert-level validation of AI-generated medical text with scalable language models

  Research Square · 2025-07-08 · 1 citations

  preprintOpen access
  PublisherOA PDFDOI

• Automated real-time assessment of intracranial hemorrhage detection AI using an ensembled monitoring model (EMM)

  npj Digital Medicine · 2025-10-16 · 2 citations

  articleOpen accessSenior author

  Artificial intelligence (AI) tools for radiology are commonly unmonitored once deployed. The lack of real-time case-by-case assessments of AI prediction confidence requires users to independently distinguish between trustworthy and unreliable AI predictions, which increases cognitive burden, reduces productivity, and potentially leads to misdiagnoses. To address these challenges, we introduce Ensembled Monitoring Model (EMM), a framework inspired by clinical consensus practices using multiple expert reviews. Designed specifically for black-box commercial AI products, EMM operates independently without requiring access to internal AI components or intermediate outputs, while still providing robust confidence measurements. Using intracranial hemorrhage detection as our test case on a large, diverse dataset of 2919 studies, we demonstrate that EMM can successfully categorize confidence in the AI-generated prediction, suggest appropriate actions, and help physicians recognize low confidence scenarios, ultimately reducing cognitive burden. Importantly, we provide key technical considerations and best practices for successfully translating EMM into clinical settings.

  PublisherOA PDFDOI

• Improving Prostate MRI: Lessons Learned from the American College of Radiology’s Prostate MR Image Quality Improvement Collaborative

  Radiographics · 2025-12-04 · 1 citations

  article

  The American College of Radiology’s Prostate MR Image Quality Improvement Collaborative identified the factors of MRI protocol optimization, patient preparation, personnel training, and performance monitoring and feedback as key drivers of prostate MR image quality and developed interventions to address these areas.

  PublisherDOI

• The ACR Medical Image Quality Assessment System (MIQAS): A Unified Approach to Image Quality Assessment in Radiology

  Journal of the American College of Radiology · 2025-12-11

  article1st authorCorresponding
  PublisherDOI

• The Road Map for ACR Practice Accreditation for Radiology Artificial Intelligence

  Journal of the American College of Radiology · 2025-03-08 · 4 citations

  article1st authorCorresponding
  PublisherDOI

Frequent coauthors

• Harold G. Koenig

  Duke Medical Center

  107 shared

• Michael E. McCullough

  64 shared

• David H. Foos

  Carestream (United States)

  49 shared

• Gold Medalist

  University of Minnesota

  49 shared

• David L. Baker

  Western Sydney University

  49 shared

• Gary Gary's Father

  Texas Health Dallas

  49 shared

• Eugene C. Klatte

  University Radiology

  49 shared

• Alby Richard

  49 shared