congdanh99/dnbcd-busuclm

LGFVM-UNet — Synapse Multi-organ Segmentation

Implementation of "A Local-Global Fusion Vision Mamba UNet Framework for Medical Image Segmentation" published in Engineering Applications of Artificial Intelligence (2026).

Model Description

LGFVM-UNet combines:

• LGF-VSS block: parallel multi-scale convolutions + Mamba SSM fused via QuadGate dynamic gating
• MCFB: Multi-level Cross-scale Feature Fusion Block with spatial-channel dual attention
• Adaptive Hierarchical Loss: gradient statistics-based dynamic supervision weighting

Performance on Synapse Multi-organ Dataset

Method	DSC (%) ↑	HD95 ↓
UNet	71.98	22.39
TransUNet	82.42	25.54
MSVM-UNet	87.41	8.91
LGFVM-UNet (paper)	88.74	6.65

Training Config

• Optimizer: AdamW (lr=1e-4, weight_decay=1e-4)
• Batch size: 32
• Epochs: 200 (early stopping patience=15)
• Scheduler: CosineAnnealingLR
• Input size: 224×224
• GPU: NVIDIA RTX 4090

Usage

import torch
from models.vmunet.vmunet import LGFVMUNet

model = LGFVMUNet(
    num_classes=9,
    input_channels=3,
    depths=[2, 2, 2, 2],
    depths_decoder=[2, 2, 2, 1],
    drop_path_rate=0,
    use_full_scale_skip=True,
)

checkpoint = torch.load("best_model.pth", map_location="cpu")
model.load_state_dict(checkpoint)
model.eval()

# Input: (B, 3, 224, 224)
# Output: (B, 9, 224, 224) logits
with torch.no_grad():
    logits, _ = model(image)
    pred = torch.argmax(torch.softmax(logits, dim=1), dim=1)

Citation

@article{li2026lgfvmunet,
  title={A Local-Global Fusion Vision Mamba UNet Framework for medical image segmentation},
  author={Li, Yanbo and Mao, Zihan and Qin, Feiwei and Peng, Yong and Zhang, Guodao and Xi, Xugang and Ma, Xiaoqin and Yu, Huanhuan and Zhou, Yu and Zhu, Zhu},
  journal={Engineering Applications of Artificial Intelligence},
  volume={169},
  pages={113987},
  year={2026},
  publisher={Elsevier}
}

Source Code

https://github.com/NicoleDyson/LGFVM-UNet