Medical Ultrasound Image Segmentation Using a Hybrid Top-Down/Bottom-Up Approach

Abstract

This paper proposes a hybrid framework in the study of ultrasound image segmentation that combines bottom-up (BU) with top-down (TD) strategy. The BU part with high precision constructs a weight matrix combining region-and edge-cues in terms of a novel idea of designing support regions, whereas the TD part incorporates location-driven prior knowledge from the doctor and spatial coherence information from image data. Each of these approaches has a domain of applicability, so to facilitate cue combination we introduce spectral clustering method. For reasons of algorithm optimization, a crucial idea is embedded in it of Nyström approximation and manifold assumption. The results indicate that the new algorithm provides not only higher accuracy and efficiency, but also robustness of medical ultrasound image segmentation compared to the traditional NCut method.