Dual Active-Shape Modeling for Efficient Right Ventricular Segmentation from MRI Images

Hossam El-Rewaidy^{1}, El-Sayed H. Ibrahim^{2}, and Ahmed Fahmy^{1}

The developed technique includes two contributions. The
first contribution involves splitting each RV contour in two segments: septal
and free-wall (Figure 1(a)). In the training phase, this process is performed semi-automatically
by selecting the two RV insertion points into the septal wall. Each contour
segment is modeled separately, yielding a dual-ASM model. The two segments are then
merged together using a third-order B-Spline algorithm to obtain a smooth RV contour.
The second contribution of the developed technique involves using the Bookstein
algorithm,^{3} instead of the conventional Procrustes method^{4}
used in conventional ASM, to align the generated contours. In the Bookstein
alignment method, a linear space of shape variations is used to represent the RV
shape (Figure 1(b)). Given a contour from the training dataset, the two RV insertion
points are manually selected and transformed into points (0,0) and (1,0) in the
Bookstein Coordinates. The remaining contour points are then transformed to
their corresponding points in the Bookstein Coordinates.

The developed technique has been tested on a dataset from 10 patients imaged with cine MRI (total of 546 short-axis images covering the whole cardiac cycle at the basal, mid-ventricular, and apical locations). The dataset was randomly divided into two subsets: a training set of 162 images and a testing set of 384 images. The ground truth was defined by manually delineating the RV boundaries in the dataset. It is worth noting that only one RV model was built from the three cross-sectional slices (basal, mid-cavity and apical) of each patient, rather than building three models, for improved segmentation efficiency. The principal component analysis technique was applied to the aligned shapes to estimate the mean shape for every segment. The first 8 and 15 modes of variation were selected to represent about 98% and 95% of the variance in the training set for the free-wall and septal segments, respectively.

1. Ginneken, B., Alejandro, F., Joes, S., et al. Active Shape Model Segmentation With Optimal Features. IEEE Transactions on Medical Imaging, 2002; 21(8):924-933.

2. ElBaz, M., Fahmy, A. Active shape model with inter-profile modeling paradigm for cardiac right ventricle segmentation. MICCAI, 2012; 15(1):691-698.

3. Bookstein, F. Size and Shape Spaces for Landmark Data in Two Dimensions. Statistical Science Journal, 1986; 1(2):181-242.

4. Ordas, S., Boisrobert, L., et al. Active shape models with invariant optimal features (IOF-ASM) application to cardiac MRI segmentation. Computers in Cardiology, 2003; 633-636.

Figure 1. (a)The RV contour is divided into septal
and free-wall segments. (b)Transformation of the RV shape into Bookstein-Coordinates
in 2 steps: 1)registering the two RV insertion points to points 0 and 1; and 2)normalizing
each point on the original RV shape with-respect-to the distance between the
insertion points.

Figure 2. The RV segmentation results in 2 subjects
using the proposed and conventional ASM models at the initial, 5th, and 20th
iterations. The figure shows 3 cross-sections at the basal, mid-cavity, and apical levels.

Table 1. Mean±SD of the
mean absolute distance (MAD), Hausdorff measure, and Dice Index of the
segmented contours at basal, mid-cavity, and apical levels using the proposed dual-ASM
and conventional ASM models with respect to the ground-truth of manual
segmentation. ^{+}p-value < 0.005; *p-value < 0.05; ^p-value < 0.07.

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)

2633