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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Manual video frame labeling is labor-intensive and can lead to data loss.
• Developing efficient annotation methods is crucial for advancing AI in medical studies.

Purpose of the Study:

• To introduce a semi-automatic labeling approach using Segment Anything Model 2 (SAM 2).
• To augment training datasets for segmentation models in medical video analysis.
• To assess the impact of SAM 2-augmented data on model performance.

Main Methods:

• Collected video data of emergency endotracheal intubation.
• Manually labeled a subset of frames to create a baseline dataset.
• Utilized SAM 2 to automatically generate labels for the remaining frames.
• Trained segmentation models on both baseline and augmented datasets.

Main Results:

• Models trained on the SAM 2-augmented dataset showed improved Dice Similarity Coefficient (DSC) scores.
• The semi-automatic approach significantly reduced the need for manual labeling.
• Enhanced segmentation model accuracy was observed with the augmented dataset.

Conclusions:

• Segment Anything Model 2 (SAM 2) effectively reduces manual annotation effort in video-based medical studies.
• Augmenting datasets with SAM 2 improves the performance of segmentation models.
• This approach offers a scalable solution for creating large, high-quality annotated datasets.