Next: 6 Discussion and future Up: Robust Contour Tracking in Previous: 4 Improving feature detection

5 Towards classification

  The ultimate goal of this work is to demonstrate that automated image analysis can be used for to perform temporal-based quantification of regional heart function. As a step towards this goal, in this section we present some results of applying the training and tracking procedures outlined in Section 3 to some further real heart image sequences.

5.1 A synthesised abnormal heart:

Abnormal heart motion was simulated by editing images corresponding to the diastole section of the cardiac cycle for a normal heart image sequence (the example used in Section 3 ) so that the posterior (left) wall appeared sluggish. This was done by shifting an image block containing the posterior wall from each of the diastole images by 10 pixels to the right. The image block was then blended in with the data using an exponential weight function. Finally a PCA was performed as before. Figure 6 summarises the results. The key thing to observe is that the nature of the principal modes remain unchanged although the magnitude is affected (compare with Figure 1 ). In particular the second mode (middle plot) shows that the posterior wall scales outwards to a lesser degree which is consistent with the imposed abnormality.

   
Figure 6: Principal component analysis performed on 4 cardiac cycles of a ultrasonic image sequence. The actual data (light curve) is plotted along with the simulated data (dark curve). The mean shape (left) is plotted. The modes represent the addition of standard deviations to the mean shape; From second left mode 1 (the dominant mode) to mode 4.

5.2 A real abnormal heart:

A PCA was performed on four manually segmented non-consecutive cycles of real data for a patient diagnosed with a disease which manifests as a loss in elasticity of the heart. Figure 7 summarises the results of the PCA. In this case seven modes of variation express of the variability as compared to four modes with the normal heart. The first mode appears to be a scaling of the anterior wall, the second mode a translation mode, the third mode a scaling and the fourth mode a mixture of a translation and scaling. The fifth mode is translation.

 

Mode Eigenvalue Variability percent Cumulative variability
1 669.51 0.321 0.321
2 544.00 0.294 0.615
3 284.00 0.139 0.755
4 169.4 0.115 0.871
5 85.29 0.0546 0.925
6 45.54 0.0185 0.944
7 31.35 0.0139 0.958
Table 3: The results of applying a principal component analysis to 4 manually segmented cardiac cycles from an abnormal heart. It is clear that 7 modes of variation explain over of the variability.

   
Figure 7: First five principal modes for an abnormal heart. The mean shape (filled line) is plotted along with flow lines representing how the start of each span behaves with the addition of standard deviations to the mean. From left, mode 1 (the dominant mode) to mode 5.



Next: 6 Discussion and future Up: Robust Contour Tracking in Previous: 4 Improving feature detection

Gary Jacob
Tue Jul 22 17:45:09 BST 1997