A novel aortic valve segmentation from ultrasound image using continuous max-flow approach

Geometric features of aortic valve can be applied in diagnostic, modeling and image-guided cardiac intervention, however methods to accurately and effectively delineate aortic valve from ultrasound (US) image are not well addressed. This paper proposes a novel aortic valve segmentation algorithm for intra-operative 2D short-axis US image using probability estimation and continuous max-flow (CMF) approach. The algorithm first calculates composite probability estimation (CPE) and single probability estimation (SPE) over 5 prior images based on both intensity and distance to the corresponding centroid, then the energy function for the current input image is constructed, followed by a Graphic Processing Unit (GPU) accelerated CMF approach to achieve aortic valve contours in approximately real time. Quantitative evaluations over 270 images acquired from 3 subjects indicated the results of the algorithm had good correlation with the manual segmentation results (ground truth) by an expert. The Average Symmetric Contour Distance (ASCD), Dice Metric (DM), and Reliability were employed to evaluate our algorithm, and the evaluation results of these three metrics were 1.79±0.46 (in pixels), 0.96±0.01 and 0.84 (d=0.95) respectively, where the computational time was 39.23±5.02 ms per frame.