Generative adversarial networks (GAN)-based data augmentation of rare liver cancers: The SFR 2021 Artificial Intelligence Data Challenge

Sebastien Mulé, Littisha Lawrance, Younes Belkouchi, Valérie Vilgrain, M. Lewin, Hervé Trillaud, Christine Hoeffel, Valérie Laurent, Samy Ammari, Eric Morand, Orphée Faucoz, Arthur Tenenhaus, Anne Cotten, Jean-François Méder, Hugues Talbot, Alain Luciani, Nathalie LassauDiagnostic 2022 · DOI

Introduction

Deep learning has revolutionized medical imaging, but its reliance on massive datasets remains a significant hurdle for rare disease research. Macrotrabecular-massive hepatocellular carcinoma (MTM-HCC) is a rare, aggressive subtype of liver cancer where data scarcity often prevents the development of robust AI diagnostic tools. The SFR 2021 Artificial Intelligence Data Challenge was established to investigate whether Generative Adversarial Networks (GANs) could mitigate this "small data" problem by synthesizing high-quality, multi-sequence MRI data from a limited initial cohort.

The Challenge of Rare Disease Data

In oncology, rare subtypes like MTM-HCC often lack the thousands of imaging examples typically required to train deep neural networks. This study addresses whether GANs—models capable of learning underlying data distributions to generate new, synthetic samples—can effectively augment these small datasets. The primary technical challenge was not just generating realistic images, but ensuring anatomical and pathological consistency across three distinct MRI sequences for each synthetic patient: T1 arterial, T1 portal-venous, and T2-weighted images.

Methodology

The challenge involved six multidisciplinary teams comprising radiologists, engineers, and researchers.

The Generative Task

Teams were provided with a training set of 91 real MTM-HCC cases. Their objective was to develop GAN architectures capable of producing 1,000 synthetic "patient triplets." Each triplet required three synchronized MRI sequences that maintained consistent tumor morphology and liver anatomy while correctly reflecting the different contrast properties of each sequence.

Evaluation Framework

A dual-pronged evaluation strategy was employed to ensure both statistical validity and clinical utility:

  1. Quantitative Metrics:
    • Frechet Inception Distance (FID): Used to measure the statistical similarity between the distributions of real and synthetic images.
  2. 1-Nearest Neighbor (1-NN) Accuracy: Employed to detect "mode collapse" or memorization, ensuring the GANs were generating diverse new cases rather than simply replicating the training data.
  3. Qualitative Clinical Review:
    • Three expert radiologists conducted a blinded review of 10 representative cases per team (selected via K-means clustering to ensure a representative sample).
  4. Cases were evaluated on a 0–4 Likert scale across five dimensions: liver morphology, tumor morphology, contrast patterns, presence of necrosis, and cross-sequence consistency.

Dataset Description

The challenge utilized a multicenter dataset curated from seven French academic institutions. The initial pool consisted of 92 MTM-HCC cases (91 usable for training). Each case included:

  • Late arterial phase T1-weighted images.
  • Portal-venous phase T1-weighted images.
  • Fat-saturated T2-weighted images.
  • Expert-provided manual lesion delineations (masks) for all real cases.

Results and Discussion

The challenge demonstrated that GANs can produce clinically convincing results even with a training set as small as 91 samples, a finding that challenges the traditional "big data" requirement for generative modeling.

Five of the six participating teams successfully generated the required 1,000 cases. The winning team, from Philips, achieved a top combined score of 0.64, demonstrating an optimal balance between statistical fidelity (FID) and clinical realism as judged by the radiologist jury.

A critical insight from the challenge was the difficulty of maintaining "cross-sequence consistency." While individual images might appear realistic, ensuring that the tumor and liver anatomy remained identical across different contrast phases proved to be the most significant technical hurdle. Furthermore, the involvement of radiologists in the development process was noted as a key factor for success, as domain expertise helped tune models to capture subtle but clinically vital features like tumor necrosis.

Conclusion and Future Directions

The SFR 2021 Data Challenge serves as a successful proof of concept for GAN-based data augmentation in rare cancers. While the generated images were realistic enough to satisfy expert observers, future work must determine if training diagnostic AI models on this synthetic data leads to improved performance on real-world clinical tasks. Nonetheless, this initiative highlights GANs as a promising tool for overcoming data limitations in the fight against rare diseases.

Summary updated

BibTeX
@inproceedings{mule2022generative,
  author = {Sebastien Mulé and Littisha Lawrance and Younes Belkouchi and Valérie Vilgrain and M. Lewin and Hervé Trillaud and Christine Hoeffel and Valérie Laurent and Samy Ammari and Eric Morand and Orphée Faucoz and Arthur Tenenhaus and Anne Cotten and Jean-François Méder and Hugues Talbot and Alain Luciani and Nathalie Lassau},
  title = {{Generative adversarial networks (GAN)-based data augmentation of rare liver cancers: The SFR 2021 Artificial Intelligence Data Challenge}},
  booktitle = {Diagnostic},
  year = {2022},
  doi = {10.1016/j.diii.2022.09.005},
}
Figure