This paper presents a novel approach for robust and fully automatic segmentation with active shape model search. The proposed method incorporates global geometric constraints during feature point search by using interlandmark conditional probabilities. The A* graph search algorithm is adapted to identify in the image the optimal set of valid feature points. The technique is extended to enable reliable and fast automatic initialization of the ASM search. Validation with 2-D and 3-D MR segmentation of the left ventricular epicardial border demonstrates significant improvement in robustness and overall accuracy, while eliminating the need for manual initialization.