Correction-aware interactive 3D tumor segmentation with sparse and revisable prompts

Abstract Interactive 3D tumor segmentation addresses a practical annotation setting: when an automatic or promptable prediction is incomplete, over-segmented, or anatomically implausible, the user can correct it directly through visual prompts. Most promptable medical segmentation methods, however, are still developed and evaluated around clean-prompt protocols. In realistic volumetric correction, users may combine clicks, boxes, and scribbles, inspect the updated mask, edit only selected informative slices, and revise earlier instructions across rounds. Existing interactive models therefore face three practical limitations: Accuracy can depend strongly on prompt type and density; dense-scribble protocols impose high interaction burden; and cross-round corrections can create ambiguous prompt histories that are not explicitly represented. We propose a correction-aware multi-prompt framework for interactive 3D tumor segmentation. Instead of treating prompts as a static clean condition, the framework models them as revisions to an evolving mask. It integrates clicks, a 3D bounding box, and scribbles with mask-feedback refinement; samples prompts from residual errors; supports sparse scribbles on informative slices; and uses a revision-aware prompt memory to separate the latest positive and negative instructions from locations whose prompt label has changed. Experiments on MSD-Colon and KiTS21 kidney tumor benchmarks compare the proposed method with automatic, promptable, and interactive baselines under standard and sparse-prompt protocols, together with a synthetic analysis of cross-round prompt revision. The proposed method achieves higher segmentation accuracy across prompt settings, with larger differences when scribbles are limited to sparse informative slices and more consistent refinement when prompt labels are revised. These findings support modeling interactive 3D tumor segmentation as a sparse, iterative, and revisable visual correction process rather than as a clean-prompt segmentation task. Code is publicly available at https: //github. com/HaoLi12345/interactiveₛegmentation.