Leveraging SAM 2 for Semi-Supervised Learning in Endotracheal Intubation Video Segmentation

Manual labeling in video-based studies is often limited by the high effort required, leading to the sampling of frames leading to potential loss of valuable data. This study introduces a semi-automatic labeling approach using the Segment Anything Model 2 (SAM 2) to augment training datasets for segmentation models. Videos of emergency endotracheal intubation were sampled, and a subset of frames was manually labeled to form a baseline dataset. SAM 2 was then employed to generate labels for the remaining frames. Results show that models trained on the augmented dataset achieved improved Dice Similarity Coefficient (DSC) scores. These findings demonstrate that SAM 2 can substantially reduce manual labeling efforts while enhancing model performance.