Fully automated endocardial contour detection in time sequences of echocardiograms by active appearance motion models

Abstract

A novel fully automated border detection technique for phase-normalized echocardiographic image sequences is developed: Active Appearance-Motion Models (AAMM). AAMM finds shape and appearance eigenvariations of the heart over the full cardiac cycle from a set of examples, capturing typical motion patterns. AAMM segments sequences by adjusting eigenvariation coefficients to minimize model-to-target differences. This results in a time-continuous segmentation. The method was applied on 4-chamber sequences from 129 unselected patients, split randomly into training (TRN, n=65) and test set (TST, n=64). In all sequences, an independent expert manually drew endocardial contours (MAN). On TST, fully automated AAMM succeeded in 97% of cases (AUTO) and performed well (average contour distance 3.3 mm, area regression AUTO = 0.91*MAN + 1.7 cm2, r = 0.87). Results outperformed single-frame AAM segmentation and human interobserver variabilities.