Injecting Textual Spatial Context into Vision-Language Models for Surgical Scene Understanding

Accurate anatomical landmark identification is important for safe laparoscopic navigation, yet limited view and strong tissue similarity make multi-label organ classification diffi- cult. Existing vision-language models mainly rely on appearance and overlook the spatial structure of surgical scenes (Zhang et al., 2025). We propose SpatialContext, a multi- modal framework that injects scene geometry into classification through natural language prompts derived from segmentation masks, together with a context-conditional training strategy centered on the primary surgical target. Results on DSAD (Carstens et al., 2023) and Endoscapes (Mascagni et al., 2025) show improved recognition of scene-defining and off-target anatomy, suggesting that explicit spatial semantics can improve surgical scene understanding.