Quantitative evaluation of vessel tracking techniques on coronary angiograms

Accurate, automated determination of vessel center lines is essential for two- and three-dimensional analysis of the coronary vascular tree. Therefore, we have been developing techniques for vessel tracking and for evaluating their accuracy and precision in clinical images. After points in vessels are manually indicated, the vessels are tracked automatically by means of a modified sector-search approach. The perimeters of sectors centered on previous tracking points are searched for the pixels with the maximum contrast. The sector size and radius are automatically adjusted based on local vessel tortuosity. The performance of the tracking technique in regions of high-intensity background is improved by application of a nonlinear adaptive filtering technique in which the vessel signal is effectively removed prior to background estimation. The tracking results were evaluated visually and by calculation of distances between the tracked and user-indicated centerlines, which were used as the "truth." Two hundred and fifty-six coronary vessels were tracked in 32 angiograms. Vessels as small as 0.6 mm in diameter were tracked accurately. This technique correctly tracked 255/256 (>99%) vessels based on an average of 2-3 indicated points per vessel. The one incorrect tracking result was due to a low signal-to-noise ratio (SNR<2). The distance between the tracked and the "true" centerlines ranged from 0.4 to 1.8 pixels, with an average of 0.8 pixels. These results indicate that this technique can provide a reliable basis for 2D and 3D vascular analysis.