Real-time 3D segmentation of the left ventricle using deformable subdivision surfaces

In this paper, we extend a computationally efficient framework for real-time 3D tracking and segmentation to support deformable subdivision surfaces. Segmentation is performed in a sequential state-estimation fashion, using an extended Kalman filter to estimate shape and pose parameters for the subdivision surface. As an example, we have integrated Doo-Sabin subdivision surfaces into the framework. Furthermore, we provide a method for evaluating basis functions for Doo-Sabin surfaces at arbitrary parameter values. These basis functions are precomputed during initialization, and later used during segmentation to quickly evaluate surface points used for edge detection. Fully automatic tracking and segmentation of the left ventricle is demonstrated in a dataset of 21 3D echocardiography recordings. Successful segmentation was achieved in all cases, with limits of agreement (mean plusmn1.96SD) for point to surface distance of 2.2 plusmn 0.8 mm compared to manually verified segmentations. Real-time segmentation at a rate of 25 frames per second consumed a CPU load of 8%.

1. Introduction

The emergence of volumetric image acquisition within the field of medical imaging has attracted a lot of scientific interest over the last years. In a recent survey, Noble et al. [1] presented a review of the most significant attempts for 3D segmentation within the field of ultrasonics. How-ever, all of these attempts are limited to being used as post processing tools, due to extensive processing requirements, even though volumetric acquisition may be performed in real-time with the latest generation of 3D ultrasound technology. Availability of technology for real-time tracking and segmentation in volumetric data sets would open up possibilities for instant feedback and diagnosis during data acquisition.