Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Complex Style Image Transformations for Domain Generalization in Medical Images

Nikolaos Spanos, Anastasios Arsenos, Paraskevi-Antonia Theofilou, Paraskevi Tzouveli, Athanasios Voulodimos, Stefanos Kollias

TL;DR

The paper addresses domain shift in medical imaging with limited annotated data by introducing CompStyle, a framework that combines input-level high-complexity augmentation with a style-transfer network trained adversarially to expand the domain space from a single source. It integrates fractal-based image mixing and a MaxStyle-inspired style augmentation within encoder-decoder FCN architectures, validating on prostate MRI segmentation and cardiac MRI corruption robustness. Across two FCN backbones, CompStyle achieves superior out-of-domain performance while maintaining competitive intra-domain accuracy, demonstrating the value of explicitly enlarging the training domain through diverse, hard styles. The work highlights practical gains for real-world deployment where multi-site data are scarce and suggests future work on broader architectures and tasks, preserving training efficiency.

Abstract

The absence of well-structured large datasets in medical computer vision results in decreased performance of automated systems and, especially, of deep learning models. Domain generalization techniques aim to approach unknown domains from a single data source. In this paper we introduce a novel framework, named CompStyle, which leverages style transfer and adversarial training, along with high-level input complexity augmentation to effectively expand the domain space and address unknown distributions. State-of-the-art style transfer methods depend on the existence of subdomains within the source dataset. However, this can lead to an inherent dataset bias in the image creation. Input-level augmentation can provide a solution to this problem by widening the domain space in the source dataset and boost performance on out-of-domain distributions. We provide results from experiments on semantic segmentation on prostate data and corruption robustness on cardiac data which demonstrate the effectiveness of our approach. Our method increases performance in both tasks, without added cost to training time or resources.

Complex Style Image Transformations for Domain Generalization in Medical Images

TL;DR

The paper addresses domain shift in medical imaging with limited annotated data by introducing CompStyle, a framework that combines input-level high-complexity augmentation with a style-transfer network trained adversarially to expand the domain space from a single source. It integrates fractal-based image mixing and a MaxStyle-inspired style augmentation within encoder-decoder FCN architectures, validating on prostate MRI segmentation and cardiac MRI corruption robustness. Across two FCN backbones, CompStyle achieves superior out-of-domain performance while maintaining competitive intra-domain accuracy, demonstrating the value of explicitly enlarging the training domain through diverse, hard styles. The work highlights practical gains for real-world deployment where multi-site data are scarce and suggests future work on broader architectures and tasks, preserving training efficiency.

Abstract

The absence of well-structured large datasets in medical computer vision results in decreased performance of automated systems and, especially, of deep learning models. Domain generalization techniques aim to approach unknown domains from a single data source. In this paper we introduce a novel framework, named CompStyle, which leverages style transfer and adversarial training, along with high-level input complexity augmentation to effectively expand the domain space and address unknown distributions. State-of-the-art style transfer methods depend on the existence of subdomains within the source dataset. However, this can lead to an inherent dataset bias in the image creation. Input-level augmentation can provide a solution to this problem by widening the domain space in the source dataset and boost performance on out-of-domain distributions. We provide results from experiments on semantic segmentation on prostate data and corruption robustness on cardiac data which demonstrate the effectiveness of our approach. Our method increases performance in both tasks, without added cost to training time or resources.
Paper Structure (14 sections, 2 equations, 3 figures, 3 tables)

This paper contains 14 sections, 2 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Image Mixing
  5. Style Transfer Network
  6. Network Architecture
  7. Experiments and Results
  8. Datasets
  9. Prostate Data
  10. Cardiac Data
  11. Implementation and experiment set-up
  12. Results
  13. Conclusion
  14. Acknowledgements

Figures (3)

  • Figure 1: Examples of original and augmented images produced by the pipeline on cardiac data.
  • Figure 2: The complete network architecture. Images are augmented before they enter the encoder portion of the network and the latent representations are used to create the adversarial images. Since the entire batch is mixed, the fractal images are mixed together through style mixing in the added decoder to create more varied styles during generation. Adversarial images are fed back to the training loop, while closing the auxiliary decoder, to enhance training.
  • Figure 3: Artifacted cardiac images. From left to right: Bias Artifact (showing as difference in brightness on different images areas), Ghosting Artifact (showing as idols due to movement of the organ or patient), Motion Artifact (showing as displacement), Spike Artifact (showing as stripes due to electromagnetic interference)