Model-Based Segmentation, Reconstruction and Analysis of the Vertebral Body from Spinal CT - Spinal Imaging and Image Analysis - page 423

To assess the proposed method under various challenges, we added a zero mean
Gaussian noise with different signal-to-noise ratios (SNR) to

our CT images. The segmentation accuracy is measured for each method using the
ground truths. It should be noted that the ground truths are validated by a radiol-
ogist. We calculate the percentage segmentation accuracy (Acc) as follows:

— from 0 to 100 dB —

þ

FN
The total number of slice pixels Þ;

FP

Acc % ¼ 100 ð 1 ð 62 Þ

where FP represents the false positive (i.e. the total number of the misclassi ed

pixels of the background), and FN is the false negative (i.e. the total number of the
misclassi

ed pixels of the object).
We used a variety of methods to measure the accuracy of our framework. First,
we used the visual inspection to evaluate the segmentation quality of our approach.
Figure 42 compares the results of different examples for the initial segmentation
step using the scalar level set model [ 46 ] and the graph cut method [ 45 ] which is
used in our proposed framework. As shown in this figure, the scalar level sets

method fails to segment the whole vertebra in many cases. However, the graph cut
approach can segment them well. Additionally, the boundaries detected by scalar
level sets are not smooth, and some obvious boundaries are not detected. The graph
cut method segments the image accurately. Figure 43 shows various segmentation
results of three different methods applied on some clinical datasets. These methods
are: (i) The graph cut segmentation (identical to initial labeling in our algorithm),
(ii) The PCA based segmentation described in [ 47 ], and (iii) Our 2D-PCA based
segmentation. The segmentation accuracies of the 2D-PCA based results shown in
row (iii) are: 96.8, 92.6, 91.2 and 93.6 % respectively.
For PCA based results in row (ii), the segmentation accuracies are: 89.3, 87.4,
85.6, and 84.5 % respectively. It is clear that our method is more accurate than the
method in [ 47 ]. Figure 44 shows the segmentation results of the ESP using (i) graph
cut method and (ii) our segmentation algorithm (graph cut + shape prior) under
different noise level. With our proposed approach, we obtain much better results
compared to the graph cut only. Figure 45 studies the effect of the initialization on
our proposed framework. Results indicate that the performance of our method is
almost constant with different initialization parameters.
To quantitatively demonstrate the accuracy of our approach, we calculate the
average segmentation accuracy of our segmentation method on 500 CT images
under various signal-to-noise ratios and compare the results with the PCA based
segmentation method in [ 47 ]. Again, as mentioned before, our 2D-PCA based
framework outperforms the conventional PCA described in [ 47 ] as shown in
Fig. 46 a. Additionally, Fig. 46 b studies the effect of choosing the number of the
projected training shapes N on the segmentation accuracy. From this gure, we can

conclude that the performance of 2D-PCA is better than the conventional PCA
under the same number of training shapes. In another word, to get the same
accuracy of PCA framework, the 2D-PCA needs fewer training shapes.