Computational Anatomy and Segmentation of Postmortem Liver
As discussed in Sect.18.104.22.168, increasing training labels that describe the postmortem-specific shape is essential to improve the performance of a postmortem SSM. This section presents a method that solves the abovementioned problem by synthesizing postmortem liver labels, which is inspired by synthesis-based learning . Performance comparisons of SSMs trained using different sets of synthesized postmortem liver labels are presented, followed by a proposal for a postmortem liver segmentation algorithm.
Postmortem Liver SSMs Using Synthesized Postmortem Labels
Three transformations are developed to simulate the shape deformation from in vivo livers to postmortem livers. They are categorized into a geometrical transformation Fa and two statistical transformations, FT and Ftr. Details of the methods can be found in .
In this study, the transformations yielded five different sets of synthesized postmortem liver labels, or~DT from Ft ,~DTR from Ftr,~DA from Fa, ~Dat from Fa followed by Ft, and~DATR from Fa followed by Ftr, respectively. Five postmortem liver SSMs were trained using combinations of the five synthesized liver label sets with original postmortem liver labels, D. The relationships between the five SSMs and the five synthesized label sets are summarized as follows. Note that the training label sets are shown in parentheses:
- • SSMD+T model (D andDT )
- • SSMD+TR model (D andDTR)
- • SSMD+A model (D and~DA)
- • SSMD+AT model (D and ~DAT)
- • SSMD+ATR model (D and~DATR)
In addition, three conventional SSMs constructed solely from original labels were prepared for comparison.
- • SSMD model (D only)
- • SSML model (L only)
- • SSMD+L model (D and L)
Figure 4.30 summarizes the performance of the three conventional and five proposed SSMs in terms of the sum of generalization and specificity explained in the previous section. It was found from the figure that most of the proposed SSMs learnt by synthesized postmortem liver labels outperformed conventional SSMs trained without synthesis-based learning. In particular, the D + T model achieved the highest
Fig. 4.30 The sum of generalization and specificity (Fig.7c of )
score which demonstrated the superiority of the proposed statistical transformation for synthesis-based learning.