Image translation for medical image generation: Ischemic stroke lesion segmentation

Deep learning based disease detection and segmentation algorithms promise to improve many clinical processes. However, such require vast amounts of annotated training data, which are typically not available in the medical context due data privacy, legal obstructions, non-uniform acquisition protocols. Synthetic databases with pathologies could provide required data. We demonstrate example ischemic stroke that an improvement lesion is feasible using deep augmentation. To this end, we train different image-to-image translation models synthesize magnetic resonance images brain volumes without lesions from semantic maps. In addition, a generative adversarial network generate synthetic masks. Subsequently, combine these two components build large database images. The performance various evaluated U-Net trained segment on test set. report Dice score 72.8% [70.8±1.0%] for model best performance, outperforms alone 67.3% [63.2±1.9%], close human inter-reader 76.9%. Moreover, show small only 10 or 50 cases, augmentation yields significant compared setting where no used. our knowledge, presents first comparative analysis translation, application stroke.