Table 5 Weighted mean Dice score across the test set with 10 prompt points for the fine-tuned SAM 2 model after ablating major components (image encoder, mask decoder, and prompt encoder), with fine-tuning performed using 400 training samples per class.
From: A fine-tuned foundational model SurgiSAM2 for surgical video anatomy segmentation and detection
| Â | Finetuned components of the SAM 2 model | |||
|---|---|---|---|---|
Image encoder + mask decoder + prompt encoder | Mask decoder + prompt encoder | Image encoder + prompt encoder | Image encoder + mask decoder | |
Weighted mean dice score | 0.91 | 0.86 (-0.05) | 0.92 (+ 0.01) | 0.92 (+ 0.01) |
Parameters frozen | – | 69.1 million | 4.2 million | 6000 |
