thelets/UltraSAM-3D-Ultrasound-MVSeg

UltraSAM — SAM-Med3D fine-tuned for 3D ultrasound (mitral valve)

A fine-tuned SAM-Med3D checkpoint for promptable segmentation of the mitral valve in 3D echocardiography, trained on the MVSeg 2023 dataset with a boundary-aware objective.

🏆 Part of the UltraSAM project — 4th place at AI for Life 2026 Hackathon.

• Code, training & interactive MONAI Label app: https://github.com/thelet/UltraSAM-3D-Ultrasound-Segmentation
• Base model: SAM-Med3D (vit_b_ori, 128³ input)
• Task: interactive (click-prompted) binary segmentation
• Modality: 3D ultrasound / echocardiography

Results

Predicted mitral-valve segmentation on a held-out test case, shown in 3D Slicer. Pink = model prediction, yellow = ground truth. Top: 3D surface renderings; bottom: axial / sagittal / coronal slices with overlays.

Usage

import medim
model = medim.create_model(
    "SAM-Med3D", pretrained=True,
    checkpoint_path="sam_model_dice_best.pth",
)

See the GitHub repo for the full inference pipeline (run_infer_no_baseline.py) and the interactive 3D Slicer workflow.

Training

Fine-tuned for 200 epochs on MVSeg 2023 with a combined DiceCE + 0.5·Boundary objective. Validation Dice plateaus around ~0.87.

Train Dice	Validation Dice

Attribution & license

This model is a fine-tuned derivative of SAM-Med3D (Apache-2.0) and is released under the same license. The base weights and architecture are the work of the SAM-Med3D authors; this repository modifies them by fine-tuning on 3D echocardiography (MVSeg 2023) with a boundary-aware objective.

• Base model: SAM-Med3D (Apache-2.0)
• SAM-Med3D builds on Meta AI's Segment Anything (Apache-2.0)
• Dataset: MVSeg 2023

Citation

@article{wang2023sammed3d,
  title={SAM-Med3D},
  author={Wang, Haoyu and Guo, Sizheng and Ye, Jin and Deng, Zhongying and
          Cheng, Junlong and Li, Tianbin and Chen, Jianpin and Su, Yanzhou and
          Huang, Ziyan and Shen, Yiqing and Fu, Bin and Zhang, Shaoting and
          He, Junjun and Qiao, Yu},
  journal={arXiv preprint arXiv:2310.15161},
  year={2023}
}