Segmentation of 3D intravascular ultrasonic images based on a random field model.

Segmentation of intravascular ultrasound images provides important information about the degree of vessel obstruction as well as about the shape and size of plaques. To address the problems of inter- and intra-observer variances associated with conventional manual tracing, a fully automated segmentation was developed. The algorithm is based on the optimisation of a maximum a posteriori estimator, implementing the Rayleigh distribution of speckle and a priori information about the contours. Within 3D image sets, additional information by the blood flow resulting in a decorrelation of the pixels within the luminal boundary is used to initialise the segmentation. To accelerate the estimation, dynamic programming was used. The segmentation algorithm was realised as a Windows 95 application on a Pentium II/233 MHz and delivered reliable and reproducible results independent of the catheter position and the total image brightness (except overflow). In contrast, contours drawn by two physicians for an evaluation of 29 clinical cases showed large intra- and inter-observer variances. In vivo images were acquired with a 20 MHz transducer array (EndoSonics InVision). Comparison with the contours drawn by the physicians and histology demonstrates the potential of the segmentation algorithm.