Transfer Learning for Medical Image Segmentation

Many medical-image-segmentation techniques are based on supervised learning, which assumes training data to be representative of the test data to segment. In practice however, training and test data are often somewhat different, for example because of differences in scanner hardware, scan-sequence parameters, or differences between patient groups. This problem greatly hampers the applicability of such techniques to many real-life segmentation tasks. In this thesis, we therefore investigate whether transfer-learning techniques can aid supervised neuro image segmentation on images from MRI scans with different characteristics. Transfer learning comprises techniques that can cope with certain differences in feature distributions between training and test data. We study different approaches to transfer learning that aim to compensate for these distribution differences at different stages of the classification framework: in the classifier, in the feature representation, or both. Experiments on a variety of neuro-image-segmentation tasks show that these techniques can greatly improve performance on data from different scanners, scanning parameters, and patient groups. Although these experiments focus on neuro image segmentation, most of the presented methods and the drawn conclusions are also applicable to other medical-image-segmentation tasks.