Mitral Valve Segmentation Using Robust Nonnegative Matrix Factorization

Figure 1. Overview of the proposed method. The white filling of the box shows the input and output of our method. The dark purple fill shows the mandatory steps, while the light green fill indicates the optional step of our algorithm.

Figure 2. S˜1 of Algorithm 3 (a) and S˜11 with the calculated centroids displayed by circles, after 11 iterations ((b), zoomed in).

Figure 3. Three frames showing the extraction of the mitral valve (red) from the mis-segmented structures (light gray) by the refinement step.

Figure 4. Three-dimensional view of the mitral valve (red), detected by the refinement step.

Figure 5. Dataset of 10 echocardiographic videos, with their number of frames and their spacial resolution.

Figure 6. (W^H^) (a) and S^ (b) in case of using Bregman iteration and of standard robust non-negative matrix factorization (RNMF) for different values of the regularization parameter λrnmf.

Figure 7. F1-score for automatic segmentation depending on λrnmf, where the dashed lines show the minimum and maximum f1-score values, and the standard deviation is indicated by the filled range.

Figure 8. Segmentation of the mitral valve (red contour), depending on λ2 and λ3 of Equation (8), with the entries in W^H^ appear in green and the entries in S^ in blue; the ground truth is highlighted with a green contour.

Figure 9. F1-score of segmentation result over ten videos for segmentation with windowing (a), automatic segmentation (b), segmentation by Dukler et al. (c) and segmentation by Corinzia et al. (d).

Figure 10. One of the best (a) and the worst (b) segmentation results by automatic segmentation (red), with the corresponding ground truth (green).

Figure 11. Effect of changing segmentation parameter for automatic segmentation, where the dashed lines show the minimum and maximum f1-score values and the standard deviation is indicated by the filled range.

Figure 12. Recall of mitral valve detection over ten videos for Windowing (a), detection by Dukler et al. (b) and detection by Corinzia et al. (c).