C3S3: Complementary Competition and Contrastive Selection for Semi-Supervised Medical Image Segmentation

Official PyTorch implementation of "C3S3: Complementary Competition and Contrastive Selection for Semi-Supervised Medical Image Segmentation" (ICME 2025)

🎉 News

[2025.03] 🔥 Paper accepted at ICME 2025!

📖 Abstract

For the immanent challenge of insufficiently annotated samples in the medical field, semi-supervised medical image segmentation (SSMIS) offers a promising solution. Despite achieving impressive results in delineating primary target areas, most current methodologies struggle to precisely capture the subtle details of boundaries. This deficiency often leads to significant diagnostic inaccuracies.

To tackle this issue, we introduce C3S3, a novel semi-supervised segmentation model that synergistically integrates complementary competition and contrastive selection. This design significantly sharpens boundary delineation and enhances overall precision.

Overview

🌟 Key Features

🔥 Core Innovations

Outcome-Driven Contrastive Learning (ODCL) 📍
- Dedicated to refining boundary localization
- Spatial Position Binary Masking
- Dual-Space Intersection-Union Loss
Dynamic Complementary Competition (DCC) ⚡
- Leverages two high-performing sub-networks
- Generates high-quality pseudo-labels
- Minimizes reliance between student and teacher networks

ODCL Module Details

ODCL Contrastive Learning

Figure: Detailed illustration of the Outcome-Driven Contrastive Learning (ODCL) module, showing the spatial position binary masking and dual-space intersection-union loss mechanism for boundary refinement.

📊 Performance Highlights

State-of-the-art results on LA and Pancreas-CT datasets
Boundary precision: Significant improvements in 95HD metric (LA: 5.14 vs 5.63 voxels; Pancreas: 6.96 vs 7.34 voxels)
Segmentation accuracy: Best Dice scores on LA dataset (91.24%) and competitive on Pancreas (80.93%)
Efficient training with only 2500 iterations per dataset

🏆 Results

Left Atrium Segmentation (LA Dataset)

| Method | 95HD (voxel)↓ | ASD (voxel)↓ | Dice (%)↑ | Jaccard (%)↑ | | --------------------- | -------------- | -------------- | --------------- | --------------- | | UA-MT | 7.54 | 2.31 | 88.82 | 80.18 | | SASSNet | 8.75 | 3.10 | 89.29 | 80.85 | | DTC | 7.36 | 2.15 | 89.45 | 81.01 | | SS-Net | 6.96 | 1.84 | 90.21 | 82.36 | | MCF | 6.68 | 2.02 | 88.02 | 79.51 | | CAML | 6.11 | 1.68 | 90.78 | 83.19 | | TraCoCo | 5.63 | 1.79 | 90.94 | 83.47 | | C3S3 (Ours) | 5.14 | 1.57 | 91.24 | 84.01 |

Pancreas Segmentation (Pancreas-CT Dataset)

| Method | 95HD (voxel)↓ | ASD (voxel)↓ | Dice (%)↑ | Jaccard (%)↑ | | --------------------- | -------------- | -------------- | --------------- | --------------- | | UA-MT | 17.06 | 5.23 | 74.02 | 60.10 | | SASSNet | 13.89 | 3.56 | 73.58 | 59.73 | | DTC | 13.18 | 3.75 | 73.31 | 59.26 | | SS-Net | 12.64 | 3.59 | 74.27 | 60.56 | | MCF | 11.59 | 3.27 | 75.00 | 61.27 | | CauSSL | 8.11 | 2.34 | 79.97 | 67.32 | | TraCoCo | 7.34 | 1.84 | 80.90 | 69.26 | | C3S3 (Ours) | 6.96 | 1.74 | 80.93 | 68.06 |

Visual Comparisons

Left Atrium Dataset Static Results

Pancreas Dataset Static Results

Animation Results

Animation Analysis: These dynamic visualizations demonstrate the 3D segmentation process on the Pancreas-CT dataset, showing slice-by-slice progression through the volume. Key observations:

C3S3 exhibits the most consistent and accurate boundary delineation across all slices
Boundary precision: C3S3 maintains sharp, well-defined pancreas boundaries compared to other methods
Volume consistency: Our method shows superior volumetric coherence with minimal segmentation gaps or false positives
Temporal stability: C3S3 demonstrates the most stable segmentation quality throughout the entire volume sequence

Technical Advantages Demonstrated

The animation sequences highlight several key technical advantages of C3S3:

🎯 Superior Boundary Localization: Through Outcome-Driven Contrastive Learning (ODCL), C3S3 achieves precise boundary delineation that remains consistent across different anatomical regions and slice positions.
🔄 Dynamic Adaptation: The Dynamic Complementary Competition (DCC) mechanism enables adaptive pseudo-label generation, resulting in more robust segmentation performance across varying image quality and contrast conditions.
📐 Volumetric Coherence: Unlike baseline methods that may produce inconsistent results between adjacent slices, C3S3 maintains excellent 3D consistency, crucial for clinical applications requiring accurate volume measurements.
⚡ Computational Efficiency: Despite superior performance, C3S3 maintains computational efficiency with only 2500 training iterations, making it practical for real-world deployment.

🚀 Quick Start

Environment Setup

# Clone the repository
git clone https://github.com/Y-TARL/C3S3.git
cd C3S3

# Create conda environment
conda create -n c3s3 python=3.7
conda activate c3s3

# Install PyTorch (CUDA 12.1 recommended)
pip install torch==2.2.1 torchvision==0.17.1 torchaudio==2.2.1 --index-url https://download.pytorch.org/whl/cu121

# Install other required packages
pip install numpy==1.21.5 scipy==1.10.1 scikit-learn==1.3.2 pandas==2.0.3
pip install SimpleITK==2.3.1 MedPy==0.4.0 nibabel==5.2.1
pip install scikit-image==0.21.0 opencv-python==4.9.0.80
pip install tensorboardX==2.6.2.2 tqdm==4.66.2 wandb==0.18.7

# Install additional dependencies
pip install -r requirements.txt

Dataset Preparation

We provide the dataset structure for both LA and Pancreas-CT datasets:

dataset/
├── LA/
│   ├── LA_Data/
│   │   ├── 06SR5RBREL16DQ6M8LWS/
│   │   ├── 0RZDK210BSMWAA6467LU/
│   │   ├── 1D7CUD1955YZPGK8XHJX/
│   │   └── ... (100 cases total: 80 train + 20 test)
│   └── Flods/
│       ├── train0.list
│       └── test0.list
└── Pancreas/
    ├── Pancreas-CT/
    │   ├── PANCREAS_0001_norm.h5
    │   ├── PANCREAS_0002_norm.h5
    │   └── ... (82 cases total: 62 train + 20 test)
    └── Flods/
        ├── train0.list
        └── test0.list

Dataset Details:

LA Dataset: 100 total MR volumes (80 for training: 16 labeled + 64 unlabeled for 20% setting; 20 for testing)
Pancreas Dataset: 82 total CT scans (62 for training: 12 labeled + 50 unlabeled for 20% setting; 20 for testing)
Data Format: Preprocessed and stored in H5 format for efficient training

Training

# Train on LA dataset (20% labeled data)
python train_LA.py --exp LA_20percent --max_iterations 2500 --gpu 0 --alpha 0.8

# Train on Pancreas-CT dataset (20% labeled data)
python train_Pancreas.py --exp Pancreas_20percent --max_iterations 2500 --gpu 0 --alpha 0.5

Testing

# Test on LA dataset (modify iters in test.py to match your trained model)
python test.py --root_path ./dataset/LA/ --model LA_20percent --gpu 0

# Test on Pancreas dataset
python test.p

C3S3

Install / Use

README