Automated contour detection in cardiac MRI using active appearance models: the effect of the composition of the training set.

OBJECTIVE: Definition of the optimal training set for the automated segmentation of short-axis left ventricular magnetic resonance (MR) imaging studies in clinical practice based on active appearance model (AAM).
MATERIALS AND METHODS: We investigated the segmentation accuracy by varying the size and composition of the training set (ie, the ratio between pathologic and normal ventricle images, and the vendor dependence). The accuracy was assessed using the degree of similarity and the difference in ejection fraction between automatically detected and manually drawn contours.
RESULTS: Including more images in the training set results in a better accuracy of the detected contours, with optimum results achieved when including 180 images in the training set. Using AAM-based contour detection with a mixed model of 80% normal-20% pathologic images does provide good segmentation accuracy in clinical routine. Finally, it is essential to define different AAM models for different vendors of MRI systems.
CONCLUSIONS: A model defined on a sufficient number of images with the correct distribution of image characteristics achieves good matches in clinical routine. It is essential to define different AAM models for different vendors of MRI systems.