Explainability, uncertainty, and the human-in-the-loop principles behind SkeletonPlanner.
Trust is a design requirement
A planning tool that produces a number without context asks the surgeon to accept it on faith. In a clinical setting that is not acceptable, and it is not how adoption happens. Trust has to be engineered into the interface as deliberately as accuracy is engineered into the model.
Three principles
- Show the work — every measurement traces back to the landmarks and axes that produced it
- Surface uncertainty — low-confidence segmentation or landmark predictions are flagged for review rather than silently accepted
- Keep the surgeon in control — every automated result is editable, and edits propagate through the full measurement chain
Explainability in practice
Selecting an angle in the report highlights the exact anatomical structures it was derived from in the 3D view. If a value looks wrong, the surgeon can see immediately whether the cause is a mis-segmented condyle, a displaced landmark, or genuine anatomy.
Validation as a continuous process
Model performance is monitored against expert consensus on new cohorts, and edge cases — severe deformity, prior hardware, pediatric anatomy — are tracked separately rather than averaged away. Published, reproducible validation is the foundation the rest of the design rests on.