SurgGAP: Geometry-Guided Automatic Prompting for SAM2-Based Multiclass Surgical Instrument Video Segmentation

Reliable surgical instrument segmentation is essential for tool tracking and surgical scene understanding. However, laparoscopic video segmentation remains challenging because of complex instrument appearances, while Segment Anything Model 2 (SAM2) still relies on target-specific prompts. To address these issues, we propose SurgGAP, a fully automatic framework for multi-class laparoscopic surgical instrument video segmentation based on the SAM2. To remove the dependence on manually specified prompts, a coarse segmentation network first localizes potential instrument regions, from which connected candidate regions are extracted and filtered. For each retained candidate, centroid, foreground-constrained random, and contour points are combined to construct a candidate-specific geometric prompt set describing location, spatial coverage, and complementary boundary cues. The generated prompts are then provided to a frozen SAM2 model to obtain fine-grained masks while retaining its native video-segmentation mechanism. A non-learnable boundary augmentation operation further incorporates complementary coarse foreground information, and a lightweight category decoder assigns an instrument label to each refined mask. Performance is evaluated using the Dice similarity coefficient (DSC), intersection over union (IoU), and mean class IoU (mcIoU). Experiments on EndoVis2017 and EndoVis2018 demonstrate that SurgGAP achieves the highest overall mcIoU among the compared methods, with ablation studies supporting the effectiveness of its main modules. By integrating automatic geometric prompting, promptable video segmentation, and category decoding within a unified pipeline, SurgGAP provides a practical fully automatic solution for surgical instrument segmentation and facilitates subsequent surgical video understanding tasks.