The making of an AI-powered radiology experience
Designing a standalone experience for demonstrating and reviewing VitalView AI
VitalView AI: the Overview dashboard and the study review screen showing findings and key images
VitalView AI is developing an AI-powered system that analyses CT studies, identifies abnormalities, highlights their locations in medical images, and generates draft radiology reports.
I worked on evolving a founder-built MVP into a product that could support both demonstrations and clinical review, bringing user research, clinical workflow understanding and human-AI interaction principles into an engineering-led product.
0→1 ProductHuman-AI InteractionClinical WorkflowsRoleProduct DesignerCollaboratorsFounder · Engineering team · Clinical stakeholdersStatusImplemented and demonstrated · Clinical evaluation ongoingResponsibilitiesExperience strategy · Product definition · UX research · Information architecture · Interaction design · UI designEarly outcomeDemonstrated to 2 prospective clients, both interested in further collaboration. The report comparison feature helped the team measure how closely system-generated reports matched original hospital reports.A functional MVP, a broader product brief
VitalView AI’s founder built a functional MVP that allowed users to upload CT studies, run AI analysis and generate reports without hospital PACS integration.
While effective for demonstrations, the MVP offered limited support for reviewing system-generated findings. As the startup began engaging prospective clients and radiologists, the product needed to support both clear demonstrations and detailed clinical verification.
The founder-built MVP overview dashboard: study counts, recent studies and quick links.
Two immediate use cases
BUSINESS STAKEHOLDERS & POTENTIAL CLIENTSProduct demonstrationsThe founder needed to demonstrate VitalView AI to prospective clients, investors, hospitals and PACS vendors with varying levels of medical and technical expertise.RADIOLOGISTS REVIEWING FINDINGSSystem output verificationRadiologists needed to inspect system-generated findings, identify potential errors and assess their accuracy as part of the product’s clinical evaluation.The design opportunity was to build on the working MVP, introducing a user-centred experience that could support both use cases without adding unnecessary complexity.
The design challengeHow might we design an experience that allows radiologists to verify system-generated findings before clinical deployment, while keeping the system easy to demonstrate to prospective clients?
Understanding what users needed beyond system-generated reports
Research approach
I evaluated the MVP with doctors and radiologists, explored existing clinical workflows and healthcare AI literature, and gathered stakeholder input to understand user needs and product constraints.
MVP evaluationDoctors & radiologistsSecondary researchHealthcare AI literatureClinical workflowsHospital visit and reportsStakeholder insightsClients and engineering teamKey insights and design implications
01 / Clinical verificationSeeing a finding isn’t the same as verifying itBoth radiologist and doctor needed access to the original CT images to verify system-generated findings. A report and selected key images alone were insufficient.Design implicationIntegrate a DICOM viewer so radiologists can navigate slices, adjust windowing and inspect the complete CT study.02 / Evidence TraceabilityEvery finding needs traceable evidenceRadiologists needed to locate reported abnormalities within the original CT study. Without clear slice and plane references, verifying individual findings required additional effort.Design implicationLink findings to their corresponding series, imaging plane and slice so reviewers can navigate directly to supporting images.03 / Human oversightSystem-generated findings need human reviewStakeholder discussions highlighted the need for radiologists to identify errors and retain control over the final report rather than accepting system-generated conclusions without review.Design implicationIntroduce a review workflow where radiologists can accept, reject, edit or add findings before approving the report.04 / Information architectureDifferent users need different levels of detailProspective clients needed a clear overview of the system’s capabilities, while radiologists required detailed imaging evidence and tools to examine the findings.Design implicationUse progressive disclosure: a high-level Study Summary for understanding the output and a detailed Review Workspace for clinical verification.Patient data privacy was an additional consideration throughout, particularly when using studies for demonstrations and evaluation.
Rethinking the interaction model
The original MVP followed a linear flow from CT upload to system-generated report, with limited opportunities for human review.
Based on the research findings, I introduced a human-in-the-loop workflow with checkpoints for verifying study information, reviewing evidence and approving the final report.
Clinical review checkpoints are available for radiologists; demonstration users can explore system-generated results without completing the approval process.
Adding these checkpoints introduced more steps, but gave clinical reviewers greater control over the output. The challenge was to support this deeper verification without making the demonstration experience unnecessarily complex.
Making findings verifiable
To help radiologists independently verify system-generated findings, I focused on three aspects of the review experience: connecting findings to evidence, enabling DICOM inspection and making abnormalities easier to locate.
01. Connecting findings to their source
The original interface presented findings and key images separately, leaving radiologists to identify which images supported each finding.
The new interface connected findings with their supporting images along with plane, series and slice references. This allows reviewers to navigate directly to the relevant evidence.
02. Bringing DICOM inspection into the workflow
I integrated a DICOM viewer into the review workspace, allowing radiologists to examine original CT images alongside system-generated findings.
By bringing imaging tools and findings into one interface, reviewers could inspect supporting evidence without switching to a separate viewing application.
VitalView AI Review findings screen: three CT viewports (axial, sagittal, coronal) with a viewer toolbar on the left and a panel of system-generated findings on the right.
Hover or tap a point to explore.
03. Making abnormalities easier to locate
Previously, key images lacked markings indicating where the system had detected an abnormality.
I introduced system-generated markings on key images and within the DICOM viewer, helping radiologists locate suspected abnormalities while allowing them to toggle annotations off for independent inspection.
Axial CT slice of the head shown without any abnormality marking.
The same axial CT slice with the system-generated abnormality marking outlined in yellow on the left side.
One experience, two levels of interaction
Both audiences interacted with the same CT studies and system-generated results, but needed different levels of detail. Rather than designing separate interfaces, I used progressive disclosure to keep demonstrations straightforward while making detailed clinical review accessible when needed.
01. Understanding the results
The Study Summary brings together system-generated findings, key images and the impression in an accessible overview, allowing users to understand the results without navigating complex clinical tools.
VitalView AI Study Summary for VV-0242: a findings list with plane, series and slice references beside a key image marked by the system.
02. Investigating the evidence
Radiologists can move into a detailed review workspace to inspect original CT images, verify findings and make corrections before approving the report.
VitalView AI Review findings workspace: viewer toolbar and axial, sagittal and coronal CT views beside a Key images panel with accept, reject and edit actions.
By revealing clinical tools only when needed, I could support detailed verification without making the initial experience unnecessarily complex.
The experience in action
The final walkthrough brings together the complete experience, from uploading a CT study and reviewing system-generated findings to inspecting supporting evidence, approving reports and comparing them with original hospital reports.
Early outcomes and what’s next
This project reinforced an important lesson about designing for AI: making an output visible is not the same as making it verifiable.
Working in an engineering-led startup taught me to introduce user-centred design while balancing clinical needs, technical constraints and business priorities. Although further evaluation is needed, the project helped establish a more structured approach to reviewing system-generated findings and supporting human oversight.
