Semi-automatic border detection method for left ventricular volume estimation in 4D ultrasound data

We propose a semi-automatic endocardial border detection method for LV volume estimation in 3D time series of cardiac ultrasound data. It is based on pattern matching and dynamic programming techniques and operates on 2D slices of the 4D data requiring minimal user-interaction. We evaluated on data acquired with the Fast Rotating Ultrasound (FRU) transducer: a linear phased array transducer rotated at high speed around its image axis, generating high quality 2D images of the heart. We automatically select a subset of 2D images at typically 10 rotation angles and 16 cardiac phases. From four manually drawn contours a 4D shape model and a 4D edge pattern model is derived. For the selected images, contour shape and edge patterns are estimated using the models. Pattern matching and dynamic programming is applied to detect the contours automatically. The method allows easy corrections in the detected 2D contours, to iteratively achieve more accurate models and improved detections. An evaluation of this method on FRU data against MRI was done for full cycle LV volumes on 10 patients. Good correlations were found against MRI volumes [r=0.94, y=0.72x + 30.3, difference of 9.6 +/- 17.4 ml (Av +/- SD)] and a low interobserver variability for US (r=0.94, y=1.11x - 16.8, difference of 1.4 +/- 14.2 ml). On average only 2.8 corrections per patient were needed (in a total of 160 images). Although the method shows good correlations with MRI without corrections, applying these corrections can make significant improvements.