Image Translation for Medical Image Generation -- Ischemic Stroke Lesions

Deep learning based disease detection and segmentation algorithms promise to improve many clinical processes. However, such algorithms require vast amounts of annotated training data, which are typically not available in the medical context due to data privacy, legal obstructions, and non-uniform data acquisition protocols. Synthetic databases with annotated pathologies could provide the required amounts of training data. We demonstrate with the example of ischemic stroke that an improvement in lesion segmentation is feasible using deep learning based augmentation. To this end, we train different image-to-image translation models to synthesize magnetic resonance images of brain volumes with and without stroke lesions from semantic segmentation maps. In addition, we train a generative adversarial network to generate synthetic lesion masks. Subsequently, we combine these two components to build a large database of synthetic stroke images. The performance of the various models is evaluated using a U-Net which is trained to segment stroke lesions on a clinical test set. We report a Dice score of $\mathbf{72.8}$% [$\mathbf{70.8\pm1.0}$%] for the model with the best performance, which outperforms the model trained on the clinical images alone $\mathbf{67.3}$% [$\mathbf{63.2\pm1.9}$%], and is close to the human inter-reader Dice score of $\mathbf{76.9}$%. Moreover, we show that for a small database of only 10 or 50 clinical cases, synthetic data augmentation yields significant improvement compared to a setting where no synthetic data is used. To the best of our knowledge, this presents the first comparative analysis of synthetic data augmentation based on image-to-image translation, and first application to ischemic stroke.