Results and Discussion

The same dataset of cadavers in Sect. 4.4.2 was used for validation. The segmentation accuracy was evaluated by the Jaccard index (JI) between the segmented region and corresponding true liver label.

Figure 4.31 shows the JIs of the graph cut segmentation when using the eight different SSMs presented in the previous subsection. The figure tells that the segmentation using SSMB+r achieved the highest performance (average JI = 0.806). Note that SSMB+r was the only model whose segmentation performance is significantly superior to those of all conventional SSMs (SSMB, SSML, and SSMB+i). Here, a Wilcoxon test with a significance level of 0.05 was employed for statistical test. It is worth mentioning that the findings from the figure and statistical test are consistent with the conclusion of the previous subsection, where SSMB+r was proven to be the best model.

Figure 4.32 shows examples of graph cut segmentation when using SSML and SSMB+r, in which the contours of the regions are shown in yellow. It was found from the figure that the shape extracted by the algorithm with SSMB+r was obviously superior to that of SSML. The failure of SSML was explained by failure in shape estimation which was caused by a limited ability to delineate a postmortem liver shape.

Box plots of the graph cuts segmentation performance of all eight SSMs (Revised version of figure 1 in [182])

Fig. 4.31 Box plots of the graph cuts segmentation performance of all eight SSMs (Revised version of figure 1 in [182])

Examples of segmentation results from the conventional SSMl and from SSM^+r, which showed the best performance [182]

Fig. 4.32 Examples of segmentation results from the conventional SSMl and from SSM^+r, which showed the best performance [182]

 
Source
< Prev   CONTENTS   Source   Next >