Learning-based shape model matching: Training accurate models with minimal manual input

Abstract

Recent work has shown that statistical model-based methods lead to accurate and robust results when applied to the segmentation of bone shapes from radiographs. To achieve good performance, modelbased matching systems require large numbers of annotations, which can be very time-consuming to obtain. Non-rigid registration can be applied to unlabelled images to obtain correspondences from which models can be built. However, such models are rarely as effective as those built from careful manual annotations, and the accuracy of the registration is hard to measure. In this paper, we show that small numbers of manually annotated points can be used to guide the registration, leading to significant improvements in performance of the resulting model matching system, and achieving results close to those of a model built from dense manual annotations. Placing such sparse points manually is much less time-consuming than a full dense annotation, allowing good models to be built for new bone shapes more quickly than before. We describe detailed experiments on varying the number of sparse points, and demonstrate that manually annotating fewer than 30% of the points is sufficient to create robust and accurate models for segmenting hip and knee bones in radiographs. The proposed method includes a very effective and novel way of estimating registration accuracy in the absence of ground truth.