Multi-atlas Segmentation with Learning-Based Label Fusion

Although multi-atlas segmentation techniques have been producing impressive results for many medical image segmentation problems, most label fusion methods developed so far rely on simple statistical inference models that may not be optimal for inference in high-dimensional feature space. To address this problem, we propose a novel scheme that allows more effective usage of advanced machine learning techniques for patch-based label fusion. Our key novelty is using image registration to guide training sample selection for more effective learning. We demonstrate the power of this new technique in cardiac segmentation using clinical 2D ultrasound images and show superior performance over multi-atlas segmentation and machine learning-based segmentation.