🎯 [MICCAI'25] 🏆 SAM-MaGuP: Mamba-Guided Boundary Prior for Polyp Segmentation 🚀

🌟 Overview

Detecting colorectal polyps with precision is critical for early cancer detection, yet challenges like blurry boundaries and variable polyp shapes make this task complex. SAM-MaGuP is a groundbreaking framework that supercharges the Segment Anything Model (SAM) with advanced boundary-focused innovations.

🔑 Key Features:

• ✨ Boundary Distillation Component (BDC): Refines weak boundaries for superior accuracy.
• 🔍 1D-2D Mamba Adapter: Bridges spatial and channel dependencies at multiple scales.
• 🏆 State-of-the-Art Performance: Excels across five benchmark datasets.

---

💡 Highlights

• 🧠 Boundary Refinement: Tackles low-contrast regions with precision.
• 🌐 Multi-Scale Context: Captures long-range dependencies for robust segmentation.
• ⚡ Generalization Power: Handles unseen datasets with unmatched accuracy.
• 🎯 Clinical Relevance: Designed for real-world medical applications.

---

🛠️ Architecture

SAM-MaGuP combines:

1. MaGuP Module:
   • Multi-scale Spatial Decomposition (MSD): Extracts coarse-to-fine polyp features.
   • Boundary Distillation Component (BDC): Enhances polyp boundary recognition with cross-attention.
2. SAM Backbone: Enhanced with domain-specific knowledge for medical imaging.

---

📊 Results

🚀 Quantitative Metrics:

SAM-MaGuP sets new benchmarks across diverse datasets:

• mDice: 94.7% on seen datasets.
• mIoU: 89.0% on Kvasir-SEG.
• Boundary Precision: 85.4% on challenging ETIS dataset.

🖼️ Qualitative Improvements:

• Accurate polyp segmentation, even in ambiguous regions.
• Strong performance on unseen datasets.

---

🚀 Getting Started

Prerequisites

• Python: >= 3.8
• PyTorch: >= 2.0
• CUDA: Enabled GPU (e.g., NVIDIA A100)

Installation

1. Clone the repository:

   git clone https://github.com/username/SAM-MaGuP.git  
   cd SAM-MaGuP  
   

2. Install dependencies:

    pip install -r requirements.txt
   

3. Download the datasets: Kvasir-SEG, CVC-ClinicDB, ETIS, CVC-ColonDB, CVC-300. and organize the data structure as follows:

   data/
     ├── Kvasir-SEG/
     │   ├── train/
     │   │   ├── images/
     │   │   ├── masks/
     │   ├── test/
     │       ├── images/
     │       ├── masks/
     ├── CVC-ClinicDB/
     │   ...

4. Start training by running:

     python train.py --config configs/train_config.yaml

5. Run the inference script:

     python test.py --input data/test_images --output results/

Training and Inference Code Coming Soon, Stay Tuned !

Citation

If you use this code or framework, please cite:

@inproceedings{SAM-MaGuP,
  title={Mamba Guided Boundary Prior Matters: A New Perspective for Generalized Polyp Segmentation},
  author={Dutta, Tapas K. and Majhi, Snehashis and Nayak, Deepak Ranjan and Jha, Debesh},
  booktitle={MICCAI 2025},
  year={2025}
}