There has been increasing interest in analysing left ventricular function using cardiac imaging technology. The clinical demand is for real-time analysis as most pathologies manifest themselves by abnormalities in heart dynamics. Although the ideal would be real-time analysis of temporal sequences of full volumetric data (3D+T) such as MR sequence analysis [ 14 ], ultrasound tomography [ 13 ] or free-hand probe ultrasonography [ 16 ] this is not likely to be achievable at a reasonable price in the near future. Hence, there is considerable clinical interest in developing methods to perform real-time quantification of regional heart function based on an analysis of 2D image sequences (2D+T) of echocardiograms [ 6 , 12 ].
In this paper we present an experimental evaluation of a robust visual image contour tracker [ 1 , 2 ] on extended echocardiographic image sequences. The potential attraction of this method relate to tracking robustness - the approach can track well in the presence of clutter (which includes distracting structures as well as large amounts of spurious sensor noise and imaging artifacts). It achieves this robustness by restricting the class of allowable motions (shape deformations) to an admissible set that has been learnt from tracking on a training data set. In particular, and unlike previous approaches [ 3 , 6 ], working on extended sequences allows us to directly estimate temporal characteristic parameters such as periodicity and asynchronousy. Further, a robust tracker can accommodate part of a contour going out of the measurement window for a limited time. This is an attractive feature in echocardiographic image sequence analysis as due to twisting of the heart a section of the ventricle boundary wall may rotate out of the plane of the sector scan over part of the cardiac cycle. Methods based on tracking image features detected in single frame echograms [ 11 , 3 ] can not do this.
The ultimate goal of this work is to develop the tracking framework as a basis for regional heart function assessment of ischemia and infarcted heart disease. This paper reports on our first studies in this area. We evaluate the limitations of the visual tracking framework for echogram analysis and then go on to describe how to adapt it to better meet the needs of echogram analysis. We are currently working on the full implementation of some of these ideas. Future work will focus on classification issues and extending the ideas to 4D (3D+T) analysis.
The outline of the paper is as follows. In section 2 we briefly review the key ideas behind the tracking algorithm. Section 2.1 explains how shape deformations are defined and can be estimated from training sequences. Section 2.2 considers the tracking model and how tracking dynamics can be estimated. Tracking experiments comparing different models of tracking dynamics are presented in Section 3 . In an attempt to improve tracking performance by enhancing the measurement process, Section 4 presents results of applying energy-based filtering and temporal-based noise-reduction methods. Further results on other heart image sequences are given in Section 5 . We conclude, in Section 6 , with a discussion of directions of current and future work.
Gary Jacob