Source of incomplete masks in prompted SAM pipelines

Determine whether the incomplete bruise masks produced by bounding-box localisation followed by SAM-style segmentation arise from bounding-box localisation, prompting, or mask generation.

Background

The evaluation compares several large vision-language-model prompting pipelines paired with SAM2 or MedSAM. These systems generally produce conservative or incomplete bruise masks, particularly because diffuse bruises are difficult to localise and delineate using bounding-box and prompt-based workflows.

Although the results establish that these pipelines underperform the proposed BruNet variants, the evaluation does not disentangle the separate contributions of the initial bounding-box localisation, the prompting process, and the subsequent mask-generation stage. Identifying the source of the limitation remains unresolved.