Sparse feature competition and shapes similarity based ultrasound image sequence segmentation method

Abstract

Segmentation of the object region in ultrasound image sequences plays a crucial role in many clinical applications. Active contour models are widely used in medical image segmentation. Yet, the inherent limitations of ultrasound imaging, such as its low signal-to-noise ratio or local fuzzy boundary caused by inhomogeneous distribution of intensity, have prevented the classical active contours from yielding satisfactory results. To cope with these limitations, many methods have been proposed for modeling image features or shapes prior to constraining active contours. However, most features are limited to constructing the object regions image features, and the object shapes priors usually are learnt from a large set of examples. This paper proposes a sparse feature complete and shapes similarity based method for segmentation of ultrasound image sequences. In this method, a novel contour searching strategy, named sparse feature competition, is proposed by exploiting the reconstruction errors based on both the object and the background feature dictionaries, which is used to alleviate the defects in ultrasound images. By proving that the variation in the object shape has the low-rank property in a linear space, the similarity between the object shapes in the image sequence is also exploited as a shapes prior, which can be interpreted as an unsupervised approach to the shapes prior modeling. To validate the performance of our method, sequences of clinical images were used as the training and test set. The proposed method was compared with three well-known methods on the same test set. The results demonstrate that the proposed method can consistently improve the performance of active contour models and increase the robustness against image defects; consequently, it is likely to improve the efficiency of computer-assisted therapy.