TC-SemiSAM3-new

Semi-supervised SAM3 model for coronary vessel segmentation.

Model Description

This is the final checkpoint from semi-supervised training using:

• 5 labeled videos
• 31 unlabeled videos
• Mean Teacher framework with temporal consistency

Usage

import torch

# Load checkpoint
checkpoint = torch.load("checkpoint_final.pt", map_location="cpu")

# Get student model weights
state_dict = checkpoint["student_state_dict"]

# Load into SAM3 model
model.load_state_dict(state_dict, strict=False)

Training Details

• Framework: Mean Teacher + Temporal Consistency
• Labeled data: 5 videos (140 frames)
• Unlabeled data: 31 videos
• Confidence-aware regularization enabled

Performance

Dataset	Dice	clDice
36 Training Videos	0.6811	0.7492
36 Videos (Adaptive)	0.7232	0.7775

Related

• GitHub: https://github.com/qimingfan10/TC-SemiSAM.git