The variation between images obtained with different scanners or different imaging protocols presents a major challenge in automatic segmentation of biomedical images. This variation especially hampers the application of otherwise successful supervised-learning techniques which, in order to perform well, often require a large amount of labeled training data that is exactly representative of the target data. We therefore propose to use transfer learning for image segmentation. Transfer-learning techniques can cope with differences in distributions between training and target data, and therefore may improve performance over supervised learning for segmentation across scanners and scan protocols. We present four transfer classifiers that can train a classification scheme with only a small amount of representative training data, in addition to a larger amount of other training data with slightly different characteristics. The performance of the four transfer classifiers was compared to that of standard supervised classification on two magnetic resonance imaging brain-segmentation tasks with multi-site data: white matter, gray matter, and cerebrospinal fluid segmentation; and white-matter-/MS-lesion segmentation. The experiments showed that when there is only a small amount of representative training data available, transfer learning can greatly outperform common supervised-learning approaches, minimizing classification errors by up to 60%.

Many methods have been proposed for tissue segmentation in brain MRI scans. The multitude of methods proposed complicates the choice of one method above others. We have therefore established the MRBrainS online evaluation framework for evaluating (semi)automatic algorithms that segment gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) on 3T brain MRI scans of elderly subjects (65–80 y). Participants apply their algorithms to the provided data, after which their results are evaluated and ranked. Full manual segmentations of GM, WM, and CSF are available for all scans and used as the reference standard. Five datasets are provided for training and fifteen for testing. The evaluated methods are ranked based on their overall performance to segment GM, WM, and CSF and evaluated using three evaluation metrics (Dice, H95, and AVD) and the results are published on the MRBrainS13 website. We present the results of eleven segmentation algorithms that participated in the MRBrainS13 challenge workshop at MICCAI, where the framework was launched, and three commonly used freeware packages: FreeSurfer, FSL, and SPM. The MRBrainS evaluation framework provides an objective and direct comparison of all evaluated algorithms and can aid in selecting the best performing method for the segmentation goal at hand.

Many medical image segmentation methods are based on supervised classification of voxels. Such methods generally perform well when provided with a training set that is representative of the test images to segment. However, problems may arise when training and test data follow different distributions, for example due to differences in scanners, scanning protocols, or patient groups. Under such conditions, weighting training images according to distribution similarity has been shown to greatly improve performance. However, this assumes that part of the training data is representative of the test data; it does not make unrepresentative data more similar. We therefore investigate kernel learning as a way to reduce differences between training and test data and explore the added value of kernel learning for image weighting. We also propose a new image weighting method that minimizes maximum mean discrepancy (MMD) between training and test data, which enables the joint optimization of image weights and kernel. Experiments on brain tissue, white matter lesion, and hippocampus segmentation show that both kernel learning and image weighting, when used separately, greatly improve performance on heterogeneous data. Here, MMD weighting obtains similar performance to previously proposed image weighting methods. Combining image weighting and kernel learning, optimized either individually or jointly, can give a small additional improvement in performance.