Evaluating synthetic neuroimaging data augmentation for automatic brain tumour segmentation with a deep...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Automated glioma segmentation in medical images is crucial for treatment planning.
• Deep learning models require extensive, high-quality training data.
• Existing datasets may be insufficient for robust model training.

Purpose of the Study:

• To develop and evaluate a neuroimaging synthesis technique for augmenting data.
• To train fully-convolutional networks (U-nets) for automatic glioma segmentation using synthetic data.
• To assess the impact of synthetic data augmentation on segmentation performance and computational cost.

Main Methods:

• StyleGAN2-ada was used to generate synthetic FLAIR MRI images and corresponding glioma segmentation masks.
• Synthetic data were added to a real dataset (n=2751) for U-net training.
• U-nets were trained with and without geometric augmentation (translation, zoom, shear) and evaluated using Dice coefficients.

Main Results:

• Synthetic data augmentation yielded marginal improvements in Dice coefficients (validation +0.0409, test +0.0355).
• Geometric augmentation significantly improved model generalization, reducing performance variation across datasets.
• The computational expense of synthetic data generation was substantial.

Conclusions:

• Synthetic data augmentation provides modest performance gains for automatic glioma segmentation.
• Geometric augmentation is more beneficial for improving the generalization of U-nets.
• The current computational cost of synthetic data generation is difficult to justify for this application.