SPARSE: Few-Shot Semi-Supervised Learning via Class-Conditioned Image Translation

This repository contains the official implementation of "SPARSE Data, Rich Results: Few-Shot Semi-Supervised Learning via Class-Conditioned Image Translation".

📋 Overview

SPARSE introduces a novel GAN-based semi-supervised learning framework specifically designed for extremely low labeled-data regimes (5-50 labeled samples per class). Our approach achieves robust classification performance in medical imaging scenarios where annotation costs are prohibitive.

Key Features

• Three-Phase Training Framework: Alternating supervised and unsupervised training phases
• Image-to-Image Translation: Modifies real unlabeled images to preserve authentic anatomical features
• Ensemble-Based Pseudo-Labeling: Combines confidence-weighted predictions with temporal consistency
• Medical Imaging Focus: Evaluated on 11 MedMNIST datasets with strong performance in 5-shot settings

🏗️ Architecture

Our framework integrates three specialized neural networks:

• Generator (G): Attention U-Net for class-conditioned image translation
• Discriminator (D): Dual-purpose network for authenticity assessment and classification
• Classifier (C): EfficientNet-B3 dedicated to classification tasks

The training consists of three main phases:

1. Supervised Training Phase: Joint training of all networks on limited labeled data
2. Self-Supervised Pre-training: Ensemble-based pseudo-labeling + class-conditioned image translation
3. Synthetic Data Enhancement: Training classifier on generated samples with known target classes

🚀 Getting Started

Prerequisites

• Python 3.8+
• CUDA-compatible GPU (recommended)

Installation

1. Clone the repository:

git clone https://github.com/username/SPARSE.git
cd SPARSE

2. Install dependencies:

pip install -r requirements.txt

Dataset Preparation

The codebase includes utilities for processing MedMNIST datasets:

1. Download and organize raw data:

python src/dataset/process_medmnist.py -i /path/to/raw/data -o /path/to/organized/data

2. Split into supervised/unsupervised sets:

python src/dataset/split_datasets.py --path_to_dataset /path/to/organized/data --percentage_of_supervised_dataset 0.05 --saving_path ./dataset

3. Create few-shot folders:

python src/dataset/create_shot_folder.py -i /path/to/organized/data -o ./dataset

🎯 Usage

Training

Run training with default parameters for 5-shot learning:

python src/train.py \
    --data_flag bloodmnist \
    --path_to_dt ./dataset \
    --percentage_supervised 5 \
    --epochs 1000 \
    --batch_size 24 \
    --lr 0.0002 \
    --project_id SPARSE_experiment \
    --exp_id bloodmnist_5shot

Key Training Parameters

Parameter	Description	Default
--data_flag	Dataset name (e.g., bloodmnist, pathmnist)	bloodmnist
--percentage_supervised	Number of shots per class (5, 10, 20, 50)	5
--epochs	Total training epochs	1000
--n_unsupervised	Unsupervised training frequency (every N epochs)	5
--percentile	Confidence threshold for pseudo-labeling	0.75
--batch_size	Training batch size	24
--lr	Learning rate	0.0002

Example Commands

5-shot learning on PathMNIST:

python src/train.py --data_flag pathmnist --percentage_supervised 5 --exp_id path_5shot

20-shot learning on DermaMNIST:

python src/train.py --data_flag dermamnist --percentage_supervised 20 --exp_id derma_20shot

Custom unsupervised training frequency:

python src/train.py --data_flag bloodmnist --n_unsupervised 10 --exp_id blood_custom

📊 Results

Performance Comparison (Accuracy %)

Method	5-shot	10-shot	20-shot	50-shot
SPARSE	63.21	68.50	73.44	77.15
SPARSE_ens	66.22	70.95	75.71	78.28
TripleGAN	64.23	68.79	71.40	76.25
SEC-GAN	58.73	63.79	66.95	73.30
CISSL	45.84	46.80	49.19	51.20
MatchGAN	39.88	48.73	51.90	54.15
EC-GAN	35.09	34.66	34.29	47.41
SGAN	25.80	25.96	24.92	27.28

Results averaged across 11 MedMNIST datasets

Key Findings

• Statistical Significance: SPARSE_ens achieves statistically significant improvements over most competitors in 5-shot settings (p < 0.05)
• Ensemble Advantage: Combining discriminator and classifier predictions consistently improves performance
• Optimal Training Schedule: Unsupervised training every 10 epochs (μ=10) provides best performance across all shot settings

🏥 Supported Datasets

The framework has been evaluated on 11 MedMNIST datasets:

Dataset	Modality	Task	Classes	Samples
BloodMNIST	Blood Cell Microscopy	Multi-class	8	17,092
PathMNIST	Colon Pathology	Multi-class	9	107,180
DermaMNIST	Dermatoscopy	Multi-class	7	10,015
RetinaMNIST	Retinal OCT	Multi-class	5	-
TissueMNIST	Kidney Cortex	Multi-class	8	236,386
OrganAMNIST	Abdominal CT	Multi-class	11	58,830
ChestMNIST	Chest X-Ray	Binary	2	112,120
PneumoniaMNIST	Chest X-Ray	Binary	2	5,856
BreastMNIST	Breast Ultrasound	Binary	2	780
OrganCMNIST	Abdominal CT	Multi-class	11	23,583
OrganSMNIST	Abdominal CT	Multi-class	11	25,211

🔧 Implementation Details

Model Architectures

• Generator: Attention U-Net with encoder-decoder architecture
• Discriminator: PatchGAN with dual outputs (authenticity + classification)
• Classifier: EfficientNet-B3 with custom classification head

Loss Functions

The supervised training combines four specialized losses:

1. Prototype Loss: Creates discriminative class-specific features
2. Mutual Learning Loss: Enables knowledge sharing between models
3. Entropy Minimization: Encourages confident predictions
4. Mixup Loss: Provides regularization through data augmentation

Training Schedule

• Supervised Phase: Every epoch using limited labeled data
• Unsupervised Phase: Every μ epochs (default μ=10) with pseudo-labeling and image translation
• Optimizer: AdamW with learning rate 0.0002
• Total Epochs: 1000

📚 Citation

If you find this work useful, please cite our paper:

@misc{manni2025sparsedatarichresults,
      title={SPARSE Data, Rich Results: Few-Shot Semi-Supervised Learning via Class-Conditioned Image Translation}, 
      author={Guido Manni and Clemente Lauretti and Loredana Zollo and Paolo Soda},
      year={2025},
      eprint={2508.06429},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.06429}, 
}

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Keywords: Semi-supervised learning, Few-shot learning, Medical imaging, Deep learning, GAN-based methods, Class-conditioned image translation