StyleGAN2-ADA Synthetic Brain MRIs Fail to Boost Tumour Classification
A study on arXiv (2605.23094) tested whether synthetic brain MRIs generated by StyleGAN2-ADA improve tumour classification. Twelve class-plane generators were trained on BRISC 2025 partitions. Synthetic samples were added to real training data at 1:1 and 1:2 real-to-synthetic ratios. Three classifier families were evaluated: random forest on InceptionV3 features, a compact two-headed CNN, and MobileViTV2. InceptionV3 feature-space filtering was also tested. An independent GPT-5.5 blind test achieved only 57.73% accuracy (95% CI: 54.48–60.92%) in discriminating real from synthetic images, indicating limited utility of the synthetic data for downstream tasks.
Key facts
- Twelve class-plane StyleGAN2-ADA generators trained on BRISC 2025 partitions.
- Synthetic samples added at 1:1 and 1:2 real-to-synthetic ratios.
- Three classifier families: random forest on InceptionV3 features, compact two-headed CNN, MobileViTV2.
- InceptionV3 feature-space filtering tested.
- GPT-5.5 blind test achieved 57.73% accuracy (95% CI: 54.48–60.92%) in real-versus-synthetic discrimination.
- Synthetic augmentation did not reliably improve tumour classification.
- Study published on arXiv with ID 2605.23094.
- Generative augmentation proposed as remedy for small medical-image datasets.
Entities
Institutions
- arXiv
- BRISC