Determine the human performance ceiling for apical tooth-boundary delineation

Determine whether residual segmentation error on lower anterior teeth in panoramic radiographs reflects limitations of the segmentation model or genuine clinical ambiguity in tooth-boundary annotation, thereby establishing the human performance ceiling for the task.

Background

The study uses one practitioner per image, preventing direct measurement of inter-annotator agreement and the level of performance that expert human annotators can achieve on the same radiographs. This is particularly consequential for the lower anterior teeth and the apical region, where radiographic boundaries are poorly defined and the model exhibits its largest residual errors.

The authors explicitly identify the unresolved distinction between model inadequacy and genuine clinical ambiguity. Resolving it would require repeated, independent annotation of the same images or tooth boundaries, especially in the regions where the radiographic edge is least determinate.

References

Each image was annotated by one practitioner, so inter-annotator agreement could not be computed and we cannot establish the human ceiling for this task. This is a material gap: it prevents us from determining whether the residual error on lower anterior teeth reflects model limitation or genuine clinical ambiguity.

— Detection is solved, delineation is not: what governs tooth segmentation on panoramic radiographs  (2609.21628 - Rehan et al., 18 Sep 2026) in Section 7, Limitations