Comparisons Between an SSM of an In Vivo Liver and That of a Postmortem Liver
An SSM trained using in vivo liver labels was compared with one trained from postmortem liver labels, in terms of performance in delineating postmortem livers . This study employed a level set distribution model (LSDM) [175, 180] that does not require correspondence between boundaries of shape labels. Materials, performance indices, and SSMs to be compared are given below followed by experimental results:
Materials, Performance Indices, and SSMs
Datasets of 144 in vivo liver labels, L144, as well as 32 postmortem liver labels, D32, were used. The liver labels were manually delineated on unenhanced CT volumes of size 512 x 512 x 191 - 3201 voxels and were reduced to 170 x 170 x 170 voxels for the sake of computational efficacy (voxel size: 2.0 mm isotropic). A spatial standardization of liver labels was carried out before constructing SSMs so that the gravity points of the labels were identical among the training labels.
We computed generalization and specificity of the SSMs as performance indices. Note that generalization is a measure of the ability to describe unknown shapes and specificity a measure of the ability to represent only valid instances of the object. To calculate both indices, in vivo livers and postmortem livers in the database were divided as follows: The L144 and D32 sets were randomly divided into two equally sized subsets, named L72 and D16, respectively. In addition, two L16 subsets were randomly selected from the two L72 subsets. This study constructed three SSMs, namely, SSML72, SSML16, and SSMD16, that were trained from subsets L72, L16, and D16, respectively. Performance indices of the constructed SSMs were calculated using the L16 and D16 subsets that were not used for training.
Fig. 4.29 The sum of generalization and specificity (Fig.4c of )