Image Translation for Medical Image Generation -- Ischemic Stroke Lesions

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