Results and Discussion
Figure 4.29 shows the sums of generalization and specificity of SSM. Note that the higher value indicates better performance.
From the experimental results, the following was observed:
- 1. The performance when applying an SSM to test labels with the same types was always superior to those applying an SSM to test labels with different types. For example, performance of SSMD16 evaluated by D16 was higher than that of SSML16 evaluated by D16.
- 2. Another important finding is that the performance of SSML72 was improved by increasing the number of training labels from 16 to 72. Consequently, even when evaluating with D16, the performance of SSML72 was higher than that of SSMD16.
The first observation indicates that the performance of an SSM trained by in vivo liver labels is suboptimal for delineating postmortem liver shapes because of the difference in shape. This finding tells the difficulties in training a postmortem SSM using solely in vivo liver labels. In contrast, the second observation suggests that a larger number of in vivo liver labels would improve the performance of the postmortem SSM to some extent. To solve the problems caused by difference in shape between in vivo liver and postmortem one as well as the shortage of postmortem labels, synthesis-based learning will be introduced in next section.