NeuroGen: Synthetic Brain MRI Generation Using Wasserstein GAN with Gradient Penalty for Medical Data Augmentation

This paper has presented NeuroGen, a WGAN-GP-based system for synthetic Brain MRI generation. Looking back at the work, the four pieces we would call out as concrete outputs are these. There is a training pipeline that produces stable convergence across all four diagnostic classes at 128×128 resolution, with anatomically plausible samples at the selected checkpoints. There is the @tf.function graph caching defect — silently corrupting any sequential multi-model TensorFlow training pipeline — which we identified, diagnosed in detail, and fixed using a class-encapsulated trainer. There is the manual visual checkpoint selection methodology, which we applied uniformly across 500 epochs per class with explicit visual criteria. And finally there is the deployed web application, which puts on-demand synthetic MRI generation behind a one-click interface for users without GPU access. The samples our models produce show the morphological signatures we would want for each diagnostic class — infiltrative texture for Glioma, peripheral masses for Meningioma, skull-base specificity for Pituitary — without mode collapse at the picked checkpoints. Of these four contributions, our suspicion is that the @tf.function failure mode is the most broadly portable. It is not a WGAN-GP issue specifically, nor a medical imaging issue. The same trap will catch any TensorFlow project that sequentially trains multiple independent models using module-scope decorated step functions, regardless of the application domain.

[1]   Litjens, G., et al. (2017). A survey on deep learning in medical image analysis. Medical Image Analysis, 42, 60–88. [2]   Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional networks for biomedical image segmentation. MICCAI. [3]  Shen, D., Wu, G., & Suk, H. I. (2017). Deep learning in medical image analysis. Annual Review of Biomedical Engineering, 19, 221–248. [4]   Cheng, J. (2015). Brain tumour dataset. Figshare. doi:10.6084/m9.figshare.1512427.v5. [5]   Goodfellow, I., Pouget-Abadie, J., Mirza, M., et al. (2014). Generative adversarial nets. Advances in Neural Information Processing Systems (NeurIPS), 27, 2672–2680. [6]    Arjovsky, M., Chintala, S., & Bottou, L. (2017). Wasserstein GAN. Proceedings of the 34th International Conference on Machine Learning (ICML). [7]   Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., & Courville, A. (2017). Improved training of Wasserstein GANs. Advances in Neural Information Processing Systems (NeurIPS), 30. [8]    Radford, A., Metz, L., & Chintala, S. (2016). Unsupervised representation learning with deep convolutional generative adversarial networks. ICLR. [9]  Frid-Adar, M., Klang, E., Amitai, M., Goldberger, J., & Greenspan, H. (2018). Synthetic data augmentation using GAN for improved liver lesion classification. IEEE Transactions on Medical Imaging. [10]   Nie, D., Gou, M., Mao, Y., & Shen, D. (2018). Medical image synthesis with context aware generative adversarial networks. MICCAI.