Designing for Speed and Trust : Visualization Lessons from building AI in Radiology

Abstract

Radiology is an inherently visual domain where the primary input is an image (Xray/CT/MRI) and the output is a diagnosis, in the form of a text-based report. In this high-stakes environment, both efficiency and accuracy are paramount. However, the integration of AI has introduced a dangerous trade-off between Speed and Trust. While AI aims to be an assistant, "black box" computer vision models often increase cognitive load, forcing radiologists to choose between blindly trusting a fast algorithm or slowing down to verify opaque predictions.

How do we design interfaces that optimize for both speed and trust simultaneously? How do we maintain a balance between human agency and AI automation in visualization tools built for explanation, exploration, and decision-making?

This talk discusses a framework to resolve this conflict using Spatial Anchoring, transforming “black box” AI into intuitive, decision-supporting narratives.

The speaker shares his journey to designing representations that build trust with the doctors without compromising reporting speed. The speaker shows how to synthesize insights from multi-dimensional inputs - 2D/3D slices of the scan, 3D volumes and the temporal nature of the scan and the patient’s history - into layered visual encodings that amplify the reporting experience for the Radiologist. The speaker discusses the journey from visual metaphors that failed in non-intuitive ways to building progressive disclosure interfaces that eventually worked.

The talk will demonstrate specific visualization techniques that bridge the gap between raw data(scans), AI output (from computer vision and LLMs) and clinical insights (from doctors).

• Anatomical Heatmaps: Contextualizing complex pathologies by mapping data directly onto 3D patient avatar, using spatial anchoring to ground abstract predictions in familiar anatomical landmarks. • Non-Destructive Uncertainty Layers: Explore the use of confidence intervals and transparent overlays that augment the radiologist's judgment without occluding the source image. Attendees will leave with a transferrable toolkit for designing cognitive-aware interfaces that reduce mental load and turn AI into a transparent, trusted collaborator.

No medical background required. The domain is the case study; the lessons are broadly applicable in all settings that have both visual and temporal components.

Tentative flow of the talk

• Introduction Four universal visualization problems : trust, uncertainty, cognitive load, and AI-human collaboration

• Introduce radiology as an extreme-stakes laboratory for these problems: a brief, visual tour of where the data is the human body and every design decision carries life-or-death consequences. Designing for Trust in High-Stakes Contexts

• Examine how probability heatmap of AI predictions on a medical scan can trigger either blind trust (automation bias) or reflexive dismissal, and share before/after redesign using transparency controls, confidence indicators, and toggle-able layers that restore viewer agency. This will include design decisions that failed either to build trust or improve speed.

• Representing Uncertainty and Managing Cognitive Load Demonstrate different visual encodings of the same uncertain data: binary overlays vs. graduated confidence visualizations using color gradients, contour lines, and discrete buckets.

• Overlaying AI on Human Judgment AI findings are shown as a hard boundary box (implying false certainty) against a redesigned soft highlight with adjacent confidence scores with a one-click access to underlying reasoning as a part of the redesigned vocabulary for this setup. This will also show how users are given visual controls to interrogate, dismiss or accept AI suggestions on their own terms.

• Conclusion Synthesize the four transferable principles into a compact framework

Target Audience

This talk is for anyone who designs visualizations that inform decisions, regardless of domain. Information designers, visual journalists, data visualization practitioners, UX designers working with data-heavy interfaces, newsroom graphics teams, dashboard and analytics designers, and anyone curious about how high-stakes domains push visualization design forward. No medical or technical background required.