This paper studies the use of text in a multi-modal surgery-based model for automating the recognition of critical safety views (CVS), which is essential for the safety assessment of laparoscopic cholecystectomy. Existing CVS recognition models have limitations in that they only use visual information and rely on expensive and time-consuming spatial annotations. Based on a multi-label classification framework, this study proposes a multi-label adaptation strategy called CVS-AdaptNet, which aligns image embeddings and text descriptions for each CVS criterion using positive and negative prompts. Experiments are conducted with the state-of-the-art surgery-based model PeskaVLP on the Endoscapes-CVS2020 dataset. CVS-AdaptNet achieves 57.6 mAP, which is 6 points higher than the ResNet50 image-only baseline model (51.5 mAP). We also propose a text-specific inference method to help analyze image-text alignments. Although there is still a performance gap with existing spatial annotation-based methods, our results demonstrate the potential of applying general models to specialized surgical tasks.
