DBF-Net: Dual-Branch Fusion Network for Medical Image Segmentation

This repository provides the official implementation of DBF-Net, a novel dual-branch fusion architecture designed for accurate and efficient medical image segmentation. The network combines global context modeling and local detail preservation through two complementary branches, achieving state-of-the-art performance on multiple medical imaging datasets.

---

🔍 Overview

Medical image segmentation requires capturing global semantic information while maintaining local structural details. Conventional CNN-based methods struggle with long-range dependencies, while Transformer-based models are computationally expensive.

DBF-Net introduces:

• Dual-Branch Encoder
  • Local Branch: Captures fine-grained local features using CNN layers.
  • Global Branch: Extracts long-range dependencies using a lightweight attention mechanism.
• Feature Fusion Module (FFM): Efficiently integrates features from both branches.
• Decoder: Recovers spatial resolution with skip connections and progressive upsampling.

---

🏗 Project Structure

├── builders/        # Model building scripts
├── dataset/         # Data loading and preprocessing
├── libs/            # Core library functions
├── model/           # DBF-Net implementation
├── utils/           # Utilities (metrics, visualization, etc.)
├── train.py         # Training pipeline
├── test.py          # Evaluation and inference
└── README.md        # Project documentation

---

📄 Paper Information

Title: DBF-Net: Dual-Branch Fusion Network for Medical Image Segmentation
Authors: [Guoping Xu; Xiaming Wu; Wentao Liao; Xinglong Wu; Qing Huang; Chang Li]
Conference: [2025 IEEE International Conference on Image Processing (ICIP)]
DOI: [10.1109/ICIP55913.2025.11084704]

---

🛠 Requirements

• Python ≥ 3.8
• PyTorch ≥ 1.8
• numpy==1.26.4
• timm==0.4.12
• ptflops
• imgaug

📂 Dataset

├── train  
├── dataset_train_list.txt  
├── dataset_val_list.txt 
└── dataset_test_list.txt             

🚀 Usage

1. Train the model

• python train.py

2. Test the model

• python test.py

🏆 Results BUSI Dataset

Method	DSC ↑	HD ↓
U-Net	65.19±7.19	9.93±0.17
DeepLabV3+	77.76±8.92	7.66±0.27
LinkNet	72.70±9.77	72.70±9.77
DBBS-Net	9.34±9.43	8.05±0.32
UNeXt	78.17±2.57	8.16±0.39
DBF-Net	81.05±10.44	7.35±0.27

🔍 Citation

If you find this work useful, please cite:

@INPROCEEDINGS{11084704,
  author={Xu, Guoping and Wu, Xiaming and Liao, Wentao and Wu, Xinglong and Huang, Qing and Li, Chang},
  booktitle={2025 IEEE International Conference on Image Processing (ICIP)}, 
  title={DBF-Net: A Dual-Branch Network with Feature Fusion for Ultrasound Image Segmentation}, 
  year={2025},
  volume={},
  number={},
  pages={211-216},
  keywords={Deep learning;Image segmentation;Ultrasonic imaging;Accuracy;Codes;Anatomical structure;Artificial neural networks;Breast cancer;Brachytherapy;Lesions;Image segmentation;ultrasound;feature fusion},
  doi={10.1109/ICIP55913.2025.11084704}}