Mitral annulus segmentation from 3D ultrasound using graph cuts

Abstract

The shape of the mitral valve annulus is used in diagnostic and modeling applications, yet methods to accurately and reproducibly delineate the annulus are limited. This paper presents a mitral annulus segmentation algorithm designed for closed mitral valves which locates the annulus in three-dimensional ultrasound using only a single user-specified point near the center of the valve. The algorithm first constructs a surface at the location of the thin leaflets, and then locates the annulus by finding where the thin leaflet tissue meets the thicker heart wall. The algorithm iterates until convergence metrics are satisfied, resulting in an operator-independent mitral annulus segmentation. The accuracy of the algorithm was assessed from both a diagnostic and surgical standpoint by comparing the algorithm's results to delineations made by a group of experts on clinical ultrasound images of the mitral valve, and to delineations made by an expert with a surgical view of the mitral annulus on excised porcine hearts using an electromagnetically tracked pointer. In the former study, the algorithm was statistically indistinguishable from the best performing expert (p=0.85) and had an average RMS difference of 1.81+/-0.78 mm to the expert average. In the latter, the average RMS difference between the algorithm's annulus and the electromagnetically tracked points across six hearts was 1.19+/-0.17 mm .

Fig. 1

Flow chart for the mitral annulus segmentation algorithm designed for closed valves which locates the annulus by first finding the thin leaflet tissue and then finding where the thin tissue meets the thick tissue of the surrounding heart wall. Specific details for the respective processes can be found in the indicated sections.

Fig. 2

(Left) Slice from a 3DUS volume of a prolapsed mitral valve showing the location of the mitral valve annulus (MVA), left atrium (LA), left ventricle (LV), and ventricular septum (VS). (Right) Corresponding slice from the thin tissue detector (inverted for clarity) computed for the 3DUS volume.

Fig. 3

Generalized graph structures used in the mitral annulus segmentation algorithm: (a) 2D graph with a min-cut example, (b) 3D rectilinear graph

Fig. 4

Position, orientation, and cylindrical boundary of the 3D rectilinear graph used to find the mitral leaflet surface

Fig. 5

(a) Slice normal to the valve plane from 3DUS of a prolapsed mitral valve (same data as shown in Fig. 2) showing MV_surf (black) and adjacent regions (white striped) used to form P_int, (b) intensity projection P_int, and (c) thin tissue detector projection P_ttd. Projection images are only defined within the dotted circles.

Fig. 6

Contour initialization and evolution scheme for finding the annulus in the projection plane

Fig. 7

Graph used in the max-flow algorithm for contour initialization in the projection space, along with an example of a min-cut

Fig. 8

Ray system on which snake nodes are forced to reside during contour refinement and convergence, shown with an example snake

Fig. 9

Simplified flow chart for the mitral annulus segmentation algorithm; the algorithm iterates until the computed annulus center stops changing.

Fig. 10

Typical comparison between the algorithmic annulus (solid line) and points delineated manually by experts on clinical 3DUS images

Fig. 11

Typical view of a loaded mitral valve in a water tank that was used in the surgical view delineation study.

Fig. 12

Typical comparison between the algorithmic annulus (solid line) and points delineated by an expert on a porcine mitral annulus using an EM tracked device.

Fig. 13

Estimated shape and size of the region of convergence relative to a baseline segmentation and described using the axial and radial directions

Fig. 14

Mitral valve leaflet appearance of the same valve in several scenarios. (a) Closed mitral valve in a long-axis view (LA - left atrium; LV - left ventricle). Leaflet surface is well defined. Dashed red lines are the approximate border of the LV. (b) Open mitral valve in a long-axis view. Leaflets are poorly defined. (c) Closed mitral valve in a short-axis view shown with the border taken from (a). Leaflet surface is poorly defined and chordae insertions are visible.

Fig. 15

Slice from 3DUS of the mitral valve used in the manual segmentation study showing typical distributions of the annulus points selected by the experts. The first standard deviation (solid line) and second standard deviation (dotted line) are shown. The red point is the location of the algorithm-generated annulus in the slice. This suggests that human observers, much like the algorithm, select annulus points along the leaflet structure.