PaperlayerView exact revision on arXiv

Source-linked AI summary

Multimodal Healthcare AI: Identifying and Designing Clinically Relevant Vision-Language Applications for Radiology

Nur Yildirim, Hannah Richardson, Maria T. Wetscherek, Junaid Bajwa, Joseph Jacob, Mark A. Pinnock, Stephen Harris, Daniel Coelho de Castro, Shruthi Bannur, Stephanie L. Hyland, Pratik Ghosh, Mercy Ranjit, Kenza Bouzid, Anton Schwaighofer, Fernando Pérez-García, Harshita Sharma, Ozan Oktay, Matthew Lungren, Javier Alvarez-Valle, Aditya Nori, Anja Thieme

arXiv:2402.14252v1cs.HC

TL;DR

The clinical utility of VLM applications in radiology is underexplored, despite technical capabilities that support image-based tasks. The authors co-designed four concepts with clinical stakeholders and found that participants valued them while identifying design requirements; they also note that risks and mitigation need further study.

  • Problem

    The clinical utility of VLM applications for radiology workflows is underexplored, and the field needs to understand which applications could be relevant to clinicians.

    Source: §1 · p. 2

  • Method

    A multidisciplinary team iteratively developed four VLM concepts and gathered feedback from 13 radiologists and clinicians.

    Source: §1 · p. 2, §9 · p. 16

  • Results

    Participants viewed the concepts as valuable while articulating design requirements for usability and acceptance.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings offer initial insights into VLM design considerations for radiology and healthcare AI.

    Source: §1 · p. 2, §9 · p. 16

    The work focused on potential benefits and clinical uses, with little insight into concrete VLM risks and their mitigation.

    Source: §4.4 · p. 6

Abstract

from arXiv · show

Recent advances in AI combine large language models (LLMs) with vision encoders that bring forward unprecedented technical capabilities to leverage for a wide range of healthcare applications. Focusing on the domain of radiology, vision-language models (VLMs) achieve good performance results for tasks such as generating radiology findings based on a patient's medical image, or answering visual questions (e.g., 'Where are the nodules in this chest X-ray?'). However, the clinical utility of potential applications of these capabilities is currently underexplored. We engaged in an iterative, multidisciplinary design process to envision clinically relevant VLM interactions, and co-designed four VLM use concepts: Draft Report Generation, Augmented Report Review, Visual Search and Querying, and Patient Imaging History Highlights. We studied these concepts with 13 radiologists and clinicians who assessed the VLM concepts as valuable, yet articulated many design considerations. Reflecting on our findings, we discuss implications for integrating VLM capabilities in radiology, and for healthcare AI more generally.

1 INTRODUCTION

The paper asks which VLM applications could be clinically relevant in radiology and how they might support radiologists and clinicians. Through multidisciplinary design and feedback sessions, it develops four concepts that participants valued while identifying requirements for usability and acceptance.

1 INTRODUCTION

  • Clinical use and value of VLM applications in radiology remain uncertain, while technical advances have often developed apart from well-defined user needs and contexts.The authors argue that human-centered, participatory work should begin during ideation and problem formulation.Source: §1 · p. 2
  • The study explores which VLM use cases could enhance radiology workflows and whether, how, and when they may provide value to radiologists and clinicians.Questions include whether clinicians want AI-generated draft reports, image-annotated findings, or the ability to query scans.Source: §1 · p. 2
  • A multidisciplinary team used an iterative, three-phase design process to develop four concepts and gather feedback from 13 radiologists and clinicians.The concepts were Draft Report Generation, Augmented Report Review, Visual Search and Querying, and Patient Imaging History Highlights.Source: §1 · p. 2
  • Participants considered the VLM concepts valuable but identified requirements for AI performance, workflow integration, and context-specificity.They preferred well-defined, tool-based interactions over open-ended queries.Source: §1 · p. 2
  • The paper contributes designed VLM use cases and initial assessments, alongside a reflective account of early-phase AI design with clinical stakeholders.It discusses design implications and future research directions for integrating VLM capabilities into radiology and healthcare more broadly.Source: §1 · p. 2

2 RELATED WORK

Prior work describes the capabilities and challenges of multimodal AI, while leaving open how VLMs should be configured for radiology workflows. This paper brings human-centered design and stakeholder participation to that question.

2.1 VLMs: Multimodal Foundation Models

  • A VLM combines an LLM with an image encoder, supporting radiology report generation, text-image retrieval, visual question-answering, and report error detection.Examples include retrieving cases of left lower lobe pneumonia or checking images for findings absent from reports.Source: §2.1 · p. 3
  • Despite proposed roles such as “AI Mentor” [137] and “Autopilot for Radiologists”, useful VLM configurations for human-AI radiology workflows warrant further study.The authors also note concerns involving model opacity, input quality, bias, prompt sensitivity, hallucinations, and overconfident errors.Source: §2.1 · p. 3

2.2 Human-Centered Medical AI

  • Human-centered medical AI work spans patient autonomy and privacy, workflow integration, clinician-AI relationships, acceptance, trust, and adoption [60].The literature also discusses transparency mechanisms that help clinicians contest or learn about AI outputs,.Source: §2.2 · p. 3
  • For AI used in screening, triage, or treatment recommendations, related challenges also include inequality, evaluation frameworks, organizational barriers, and regulatory approval [29], [76], [102] [106].Source: §2.2 · p. 3

2.3 Designing AI with Domain Stakeholders

  • Prior work calls for domain stakeholders to participate in early AI ideation and problem formulation, not only interface design [34] [40] [70] [78] [151].The authors use VLM capabilities as design material to explore radiology use cases with clinicians.Source: §2.3 · p. 3

3 OVERVIEW OF RADIOLOGY WORKFLOWS

Radiology work moves across clinical roles, from imaging requests and scans to reporting and clinician review. High scan volume can delay reports, while relatively little prior work has investigated current workflows or asked radiologists where support is needed.

3 OVERVIEW OF RADIOLOGY WORKFLOWS

  • Radiology workflows span multiple roles, with referring clinicians requesting imaging, radiologists interpreting images and reporting, and clinicians reviewing reports to inform care.Images may also be discussed in multidisciplinary team meetings.Source: §3 · p. 4, Figure 1 · p. 4
  • Radiology reports combine patient and clinical-question background with findings and an impression, while X-rays may take minutes to review and CTs or MRIs 10–20 minutes.Reports may be narrative or structured, and are often written using dictation, templates, or trainee drafts.Source: §3 · p. 4
  • High scan volumes create backlogs, and report waits may last a few days to a week; some clinicians review images before reports are available.Source: §3 · p. 4
  • Prior radiology AI research has focused more on explaining outputs or diagnostic performance than on studying workflows and support needs; this paper investigates those needs and desired futures for VLM-assisted workflows.A rare earlier needfinding and design study explored AI-assisted radiology for X-rays.Source: §3 · p. 4

4 METHOD

The authors used a three-phase, multidisciplinary design process: explore workflows and use cases, sketch four concepts, then collect feedback from clinicians and radiologists. The prototypes were probes of perceived usefulness and acceptance, not model-performance tests.

4 METHOD

  • The study combined formative workflow exploration and brainstorming, concept sketching, and feedback from 13 radiologists and clinicians outside the research team.The feedback sessions examined whether and how the concepts might be useful in clinical practice.Source: §4 · p. 4

4.1 Phase 1: Brainstorming VLM Use Cases

  • Phase one used in-depth discussions and brainstorming to understand radiology workflows and ideate clinically relevant VLM use cases.Source: §4.1 · p. 4

4.1.1 In-depth Discussions.

  • Seven one-to-one, 30-minute semi-structured discussions with a cardiothoracic radiologist and a general practitioner probed image interpretation, reporting, and clinician-radiologist interactions.Where possible, participants walked through their workflows and discussed VLM capabilities from recent literature [17] [118].Source: §4.1.1 · p. 4

4.1.2 Brainstorming Sessions.

  • Two one-hour brainstorming sessions brought clinical, AI, and HCI expertise together to explore use cases using sketches, scenarios, and wireframes.The team discussed clinical relevance, feasibility, data needs, and desired system behavior before ranking ideas for development.Source: §4.1.2 · pp. 4–5

4.1.3 Data Collection and Analysis.

  • The researchers used affinity diagramming, interpretation sessions, and service blueprinting to derive design implications and trace image workflows across clinical roles.They chose these methods to identify opportunities for future technology rather than build a detailed theory of current practice.Source: §4.1.3 · p. 5

4.2 Phase 2: Sketching VLM Concepts

  • Phase two translated four selected use cases into click-through prototypes populated with MIMIC-CXR images and reports, then validated their plausibility with a radiologist.The high-fidelity prototypes probed context-specific questions rather than proposing finished designs; model-generated content was excluded because performance evaluation was not the study's focus.Source: §4.2 · p. 5

4.3 Phase 3: User Feedback Sessions

The feedback sessions involved 13 clinical stakeholders, and the study reports limitations in its sample, prototype instantiations, and focus on clinician acceptance and desirability.

4.3.1 Participants.

  • 13 clinical stakeholders from eight UK and US hospitals participated, recruited through collaborating hospitals, professional contacts, and snowball sampling.Participants included five radiologists and eight clinicians across several specialties.Source: §4.3.1 · p. 5

4.3.2 Procedure and Data Analysis.

  • One-hour remote feedback sessions probed perceived usefulness of four concepts, and transcripts were coded through affinity diagramming into use-case themes.Sessions were tailored to participants’ roles and included questions about concerns and desired support.Source: §4.3.2 · p. 5

4.4 Study Limitations

  • The convenience sample was UK- or US-based, skewed toward senior and AI-familiar clinicians, included one woman, and was mostly dual clinical-academic.The authors note that expectations and experiences may differ from those of clinicians in private practice.Source: §4.4 · p. 5
  • Feedback on non-functional prototypes remains speculative, so further interaction design and in-situ workflow integration are needed to test the insights and assumptions.Source: §4.4 · p. 6
  • The work examines potential VLM benefits and clinician acceptance in radiology but offers little insight into concrete risks and limitations or their mitigation.Source: §4.4 · p. 6

5 PHASE 1: BRAINSTORMING VLM USE CASES

Phase 1 brainstorming connected radiology workflow challenges with possible VLM uses, surfacing both potential value and practical requirements across draft reporting, visual search, and longitudinal imaging.

5 PHASE 1: BRAINSTORMING VLM USE CASES

  • Workflow brainstorming surfaced opportunities from scan ordering and image quality through prioritization, reporting, and assessment, including redundant-order detection and urgency-based triage.The team organized these challenges and ideas in a radiology workflow customer journey map.Source: §5 · p. 6

5.1 Use Cases for Draft Report Generation

  • Radiologists likened AI drafts to trainee reports, but only senior trainees could make preliminary reports available to clinicians before senior radiologists amended them.Source: §5.1 · p. 6
  • Review approaches varied: one radiologist checked the indication and image before the impression, while another reviewed the indication and impression first; editing ranged from redictation to signing.Junior trainees' reports could require more work than senior trainees' reports.Source: §5.1 · p. 6
  • A useful draft had to be accurate, complete, and error-free while saving time; bullet findings alone would not help if radiologists still had to dictate the report.Radiologists also did not want to spend extra time correcting AI errors.Source: §5.1 · p. 6
  • Responsibility and acceptance concerns arose around showing drafts to clinicians, while both groups proposed using AI findings for triage and early critical-finding flags with limited detail.Source: §5.1 · p. 6

5.2 Use Cases for Visual Search and Querying

  • Radiologists described searching trusted websites and similar cases when unsure of an image pattern or its interpretation, using examples to support diagnostic assessment.Source: §5.2 · p. 7
  • Searches could combine an abnormality and imaging modality, but preferences differed between text queries and selecting or snipping an image region.Clinicians also asked for clarification about finding locations, next steps, and additional imaging; participants discussed whether annotations and image search could reduce back-and-forth.Source: §5.2 · p. 7

5.3 Use Cases for Longitudinal Imaging

  • Longitudinal imaging highlights could summarize prior images and reports, including chronic events, operations, and abnormality trajectories such as nodule growth over time.Source: §5.3 · p. 7

6 PHASE 2: SKETCHING VLM CONCEPTS

Phase 2 co-designed four VLM use concepts for radiology, selecting tasks that span differing expertise and performance requirements. The concepts explored draft reporting, report review, visual search, and patient imaging history.

6.1 Concept Prioritization

  • The team selected image-text VLM tasks across a range of expertise and risk, including uses where lower but “good enough” performance might still be useful.They contrasted lower-risk tasks such as visual search or summarizing prior reports with higher-risk proposals such as report generation.Source: §6.1 · p. 7

6.2.1 Draft Report Generation.

  • Draft Report Generation let radiologists review and edit short-form AI findings alongside a narrative report, with the goal of balancing verification effort against time savings.The short form was intended to help spot mistakes and link findings to source material, such as prior scans or localized image findings.Source: §6.2.1 · p. 7, Figure 2 · p. 8

6.2.2 Augmented Report Review.

  • Augmented Report Review paired an overview of report abnormalities, visually locatable in the image, with an AI assistant offering clinician-oriented prompts.The prototype raised latency as a design concern, including the possibility of pre-running contextual prompts.Source: Figure 3 · p. 8, §6.2.2 · p. 10

6.2.3 Visual Search and Querying.

  • Visual Search and Querying offered image selection and image or text queries, returning groups of similar images rather than an interpretative answer.The concept explored what clinicians would query and what data, such as hospital images and reports, might support similar-image retrieval.Source: Figure 4 · p. 9, §6.2.3 · p. 10

6.2.4 Patient Imaging History Highlights.

  • Patient Imaging History Highlights combined a new scan and prior images with an AI summary of earlier images or reports.The intended highlights included changes in abnormalities, chronic conditions, and operations.Source: Figure 5 · p. 9, §6.2.4 · p. 10

7 PHASE 3: ELICITING USER FEEDBACK

In Phase 3, clinicians and radiologists assessed the four VLM concepts, describing potential benefits alongside requirements and concerns for their use. Their feedback addressed draft reports, report review, visual search, and patient imaging histories.

7.1 Draft Report Generation

  • Most radiologists would correct only 1 out of 10 or 5 to 10 out of 100 AI draft reports, indicating an expectation of near-perfect performance.Participants also distinguished findings from the more interpretative impression, which they felt clinicians may want to retain responsibility for crafting.Source: §7.1 · pp. 10–11

7.2 Augmented Report Review

  • Clinicians valued image-linked findings and prioritized report overviews, while favoring functional, task-specific queries tailored to patient context.They wanted useful context from EHR data and cautioned against chat-based, complex, or high-risk queries, including prognosis and definite negation of abnormalities.Source: §7.2 · pp. 11–12

7.3 Visual Search and Querying

  • Clinicians and radiologists valued searching local patient images for comparable normal, rare, or complex cases, and preferred evidence examples over interpretative answers.They also wanted filters such as condition, age, and sex; some saw retrieved cases as reassurance while still reserving judgment for themselves.Source: §7.3 · pp. 12–13

7.4 Patient Imaging History Highlights

  • Participants valued prior-image summaries for saving time and surfacing key historical events, conditions, and changes in abnormalities.Linked reports and visual highlights could support verification, but even text summaries or pointers to important reports were seen as useful.Source: §7.4 · p. 13

8 DISCUSSION

The discussion frames the work as early-phase design and user research addressing the gap between multimodal AI advances and clinical practice. It identifies co-created VLM use cases in radiology as the study’s focus.

8 DISCUSSION

  • The work aims to narrow the gap between multimodal frontier AI advances and their successful translation into clinical practice.It does so through early-phase design and user research focused on radiology.Source: §8 · p. 13
  • The study engaged users in identifying and co-creating clinically relevant VLM use cases for radiology.The discussion says it will consider user feedback, broader human-AI interaction design implications, and the design process for VLM concepts.Source: §8 · p. 13
  • The paper connects its use-case findings to design considerations for radiology and healthcare more generally.It also reflects on how to ideate, prioritize, and sketch VLM concepts.Source: §8 · p. 13

8.1 Implications for VLMs in Radiology

The discussion favors well-defined, tool-based VLM interactions over conversational clinical decision support, and highlights comparative image retrieval and structured draft findings as promising design directions. It also identifies trainee learning and appropriate AI reliance as open challenges.

8.1.1 Chat and Question-Answering.

  • Clinicians preferred well-defined tools over chat for time-pressured radiology tasks, while the authors note conversational acceptance may evolve and patient use merits study.Participants cited limited time and trust in generated responses, especially for clinical decision support.Source: §8.1.1 · pp. 13–14

8.1.2 Image-Text Search.

  • Comparative image search in local hospital data may offer evidence beyond generic web information, which participants said could be retrieved reliably through web lookups.The authors call for exploring search, filter, and retrieval experiences that go beyond web search.Source: §8.1.2 · p. 14

8.1.3 Report Generation.

  • Participants favored structured, short-form findings over prose, while treating the interpretative impression as harder to automate and important to care-team communication.Open questions include whether AI should draft the entire report and how AI- and human-generated content should be juxtaposed; trainee learning and appropriate reliance also remain challenges.Source: §8.1.3 · p. 14

8.2 Implications for Designing AI in Healthcare

The authors argue that useful healthcare AI should target clinically relevant tasks, support rather than replace expert interpretation, and fit established workflows. Integration also depends on clearly scoped tools, manageable verification and practice-change burdens, and benefits that justify disruption.

8.2.1 Focusing on clinically useful AI applications.

  • Moderate-performance AI may be valuable for lower-expertise tasks such as summarizing prior reports, while high-expertise tasks such as draft report generation prompt expectations of near-perfect performance.The authors suggest starting with simpler, more standardized use cases as a potentially lower-risk and more responsible approach.Source: §8.2.1 · p. 14
  • Participants favored information retrieval and summarization over higher-risk decision-agent uses, valuing reduced time and cognitive effort while retaining radiologists’ role in complex interpretations.Draft reports could save radiologist time if correct; participants also saw AI as a possible prompt for clinicians to seek radiologist assistance.Source: §8.2.1 · pp. 14–15
  • AI can assist and empower clinicians by handling tedious functional tasks or surfacing relevant evidence, while supporting rather than replacing human interpretation.Examples include retrieving patient data, highlighting observations, finding similar images, prioritizing findings, and summarizing imaging history.Source: §8.2.1 · p. 15

8.2.2 Integrating AI seamlessly into clinical workflows.

  • Context-specific tools with clear purposes may help clinicians understand and use AI better than open-ended assistants whose expertise and answer scope are unclear.Verification and correction also need to be quick and useful; extra clinician burden can reduce utility, and changes to established practices require sufficient benefits.Source: §8.2.2 · p. 15

8.3 Sketching VLM Experiences

The authors describe a capability-based, multidisciplinary design approach for sketching VLM experiences with clinicians. Their reflections point to interaction patterns beyond chat, attention to cost and simpler alternatives, and further testing of the resulting concepts.

8.3 Sketching VLM Experiences

  • A capability-based approach matching user needs with AI capabilities supported multidisciplinary brainstorming to identify clinically relevant use cases.Sketches framed as capability instantiations, rather than concrete proposals, appeared to work well for probing clinicians and radiologists.Source: §8.3 · p. 15
  • Separating a VLM capability from its interface form was difficult, so the authors framed capabilities as queries that could appear as questions, prompts, alerts, or annotations.Participants described interactions resembling background data retrieval presented unobtrusively, pointing to design patterns beyond chat or conversational queries.Source: §8.3 · p. 15
  • VLM design must balance seamless interaction against latency and financial costs, and assess whether simpler text-only approaches can provide value.For Patient Imaging History Highlights, text-only models may summarize previous report text or point to important reports without image-text extraction and summarization.Source: §8.3 · pp. 15–16
  • The click-through sketches probed clinician expectations and acceptance, but further work is needed to test the insights and refine interactions and workflow integration.The authors identify task-completion time, error types, implications, and mitigation as areas for future investigation.Source: §8.3 · p. 16

9 CONCLUSION

This study offers an initial investigation of VLM utility and design requirements in radiology, based on brainstorming and four use-case concepts assessed by 13 clinicians and radiologists. Participants favored practical, assistive uses over ambitious predictive or counterfactual concepts, while the authors emphasize trade-offs in integrating AI into clinical work.

9 CONCLUSION

  • The study presents an initial investigation of VLM utility and design requirements in radiology, with feedback from 13 clinicians and radiologists on four use cases.The feedback provided initial insights into acceptance and desirability as well as design considerations.Source: §9 · p. 16
  • Participants showed less interest in predictive diagnosis and counterfactual reasoning, positioning AI instead as a resource for mundane information extraction and processing that supports human interpretation.The conclusion gives patient-specific access to clinical evidence as an example of this assistive role.Source: §9 · p. 16
  • The authors highlight the value of designing around specific practical tasks and contexts to support workflow integration and useful mental models of AI.This contrasts with emphasis on general-purpose models that can be flexibly adapted across healthcare contexts.Source: §9 · p. 16
  • Radiology AI design must balance utility against AI risks, clinician effort to check or correct outputs, work-practice changes, latency, and compute requirements.Source: §9 · p. 16

10 APPENDIX A: EXAMPLE RADIOLOGY REPORT AND PROTOTYPE FLOWS

The appendix provides an example MIMIC-CXR image-report pair and click-through prototypes for three VLM concepts. The supplied captions identify the report and views, and name the concepts illustrated by the prototype flows.

10 APPENDIX A: EXAMPLE RADIOLOGY REPORT AND PROTOTYPE FLOWS

Related papers

Loading 2402.14252v1…