Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Data-Efficient Medical Imaging: A Generative and Semi-Supervised Framework

Mosong Ma, Tania Stathaki, Michalis Lazarou

TL;DR

This work tackles data scarcity and class imbalance in medical imaging by introducing SSGNet, a unified framework that integrates class-specific StyleGAN3-based synthesis with iterative semi-supervised pseudo-labeling to enhance both classification and segmentation. By generating class-labeled synthetic images and iteratively refining segmentation masks via pseudo-labels, SSGNet improves robustness across diverse datasets while validating synthetic data quality with Fréchet Inception Distance. The approach yields consistent gains across six public medical imaging benchmarks, demonstrating data efficiency and practical robustness in low-label regimes. These results suggest a viable, scalable strategy to mitigate annotation bottlenecks and improve performance in real-world medical image analysis scenarios.

Abstract

Deep learning in medical imaging is often limited by scarce and imbalanced annotated data. We present SSGNet, a unified framework that combines class specific generative modeling with iterative semisupervised pseudo labeling to enhance both classification and segmentation. Rather than functioning as a standalone model, SSGNet augments existing baselines by expanding training data with StyleGAN3 generated images and refining labels through iterative pseudo labeling. Experiments across multiple medical imaging benchmarks demonstrate consistent gains in classification and segmentation performance, while Frechet Inception Distance analysis confirms the high quality of generated samples. These results highlight SSGNet as a practical strategy to mitigate annotation bottlenecks and improve robustness in medical image analysis.

Towards Data-Efficient Medical Imaging: A Generative and Semi-Supervised Framework

TL;DR

This work tackles data scarcity and class imbalance in medical imaging by introducing SSGNet, a unified framework that integrates class-specific StyleGAN3-based synthesis with iterative semi-supervised pseudo-labeling to enhance both classification and segmentation. By generating class-labeled synthetic images and iteratively refining segmentation masks via pseudo-labels, SSGNet improves robustness across diverse datasets while validating synthetic data quality with Fréchet Inception Distance. The approach yields consistent gains across six public medical imaging benchmarks, demonstrating data efficiency and practical robustness in low-label regimes. These results suggest a viable, scalable strategy to mitigate annotation bottlenecks and improve performance in real-world medical image analysis scenarios.

Abstract

Deep learning in medical imaging is often limited by scarce and imbalanced annotated data. We present SSGNet, a unified framework that combines class specific generative modeling with iterative semisupervised pseudo labeling to enhance both classification and segmentation. Rather than functioning as a standalone model, SSGNet augments existing baselines by expanding training data with StyleGAN3 generated images and refining labels through iterative pseudo labeling. Experiments across multiple medical imaging benchmarks demonstrate consistent gains in classification and segmentation performance, while Frechet Inception Distance analysis confirms the high quality of generated samples. These results highlight SSGNet as a practical strategy to mitigate annotation bottlenecks and improve robustness in medical image analysis.

Paper Structure

This paper contains 28 sections, 6 equations, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Contributions
  4. Related Work
  5. Medical Image Analysis
  6. Datasets in Medical Image Analysis
  7. Generative Models for Data Augmentation
  8. Backbones for Classification and Segmentation
  9. Semi-Supervised Learning with Synthetic Data
  10. Methods
  11. Problem Definition
  12. Class-Specific Synthetic Image Generation with StyleGAN3
  13. Baseline Model Training
  14. Classification.
  15. Segmentation with VM-UNet.
  16. ...and 13 more sections

Figures (4)

  • Figure 1: Overview of the proposed SSGNet pipeline. Class-specific StyleGAN3 models are trained separately for each class to generate high-quality synthetic samples, which are then used to balance or enlarge the original dataset. For classification, synthetic data are directly integrated at varying proportions. For segmentation, fixed-size 10k synthetic images are paired with pseudo-labels produced by baseline models trained on real data. These pseudo-labels are iteratively refined during training through semi-supervised learning, enabling effective use of unlabeled synthetic data.
  • Figure : (a) Comparison with VM-UNet
  • Figure : (a) Comparison with VM-UNet
  • Figure : (b) Comparison with Adaptive t-vMF Dice Loss