Case · 04
Making model outputs adjustable without making them arbitrary
Designing clinical controls that support inspection without eroding provenance.
Radiologists needed to adjust generated networks and tracts to inspect them under different conditions.
The design challenge was to provide control without making the output feel subjective.
The tension
A completely fixed result can be difficult to trust because the clinician cannot interrogate it.
A completely editable result creates the opposite problem: the user may no longer know where the model output ends and their own modification begins.
The interaction needed to support inspection while preserving provenance.
The controls
The viewer included adjustments such as network thresholding, tract thresholding and refinement, opacity, visibility, subregion selection, and left-versus-right comparison.
These were designed as part of the clinical review process rather than as generic display settings.
Key decisions
Keep the generated result as a reference
Adjustments needed to remain connected to the original output. The clinician should understand that they are changing how the result is being inspected, not generating a new scientific result without traceability.
Make changes reversible
Thresholds and visibility states needed to be easy to reset. This reduced the cost of exploration and helped prevent accidental persistence of an unintended configuration.
Show the effect immediately
A control is useful only if the clinician can understand what it changes. Adjustments were designed to update the relevant overlay directly so the relationship between action and visual effect remained clear.
Avoid false precision
The interface could expose scientific parameters, but displaying more numerical precision did not automatically make the system more credible. The level of control and terminology needed to match what the clinician could reasonably interpret and defend.
Why this work matters
Expert control is often treated as the opposite of usability.
In clinical AI, the right controls are part of how usability and trust are created.