Li, Hao; Li, Haoxuan. (2026). . Visual Computer, 42(9), 370.
Interactive 3D tumor segmentation allows users to correct computer-generated tumor outlines using visual prompts such as clicks, boxes, and scribbles. However, many existing methods are developed under simplified prompting conditions that do not reflect real-world use, where users may provide different types of prompts, correct only selected image slices, and revise earlier instructions over multiple rounds. We developed a correction-aware framework that treats prompts as ongoing revisions to an evolving tumor outline. The method combines clicks, a 3D bounding box, and scribbles with feedback from the current segmentation, focuses new prompts on remaining errors, supports scribbles on only selected informative slices to reduce user effort, and keeps track of revised instructions. We evaluated the method on the MSD-Colon and KiTS21 kidney tumor datasets and compared it with automatic and other interactive segmentation approaches. The proposed method achieved higher segmentation accuracy across different prompting conditions, with greater improvements when scribbles were limited to a small number of informative slices. It also produced more consistent results when earlier instructions were changed. These findings suggest that treating interactive 3D tumor segmentation as an iterative and revisable correction process can improve accuracy while reducing the amount of user input required.
Fig. 1
High-level overview of the proposed interactive 3D tumor segmentation framework. Starting from an initial mask that may contain under- or over-segmentation, the user inspects the current prediction and adds heterogeneous prompts, including clicks, scribbles, and a bounding box. The model refines the mask by combining the current prompts, the previous prediction, and a revision-aware prompt memory. The framework is designed to improve accuracy across prompt protocols, support sparse interaction on informative slices, and provide a defined prompt-history representation when prompts are revised across rounds