Automatic segmentation of head anatomical structures from sparsely-annotated images

Bionic humanoid systems, which are elaborate human models with sensors, have been developed as a tool for quantitative evaluation of doctors' psychomotor skills and medical device performances. For creation of the elaborate human models, this study presents automated segmentation of head sectioned images using sparsely-annotated data based on deep convolutional neural network. We applied the following fully convolutional networks (FCNs) to the sparse-annotation-based segmentation: a standard FCN and a dilated convolution based FCN. To validate the availability of FCNs for segmentation of head structures from sparse annotation, we performed 8- and 243-label segmentation experiments using different two sets of head sectioned images in the Visible Korean Human project. In the segmentation experiments, only 10% of all images in each data set were used for training data. Both of the FCNs could achieve the mean segmentation accuracy of more than 85% in the 8-label segmentation. In the 243-label segmentation, though the mean segmentation accuracy was about 50%, the results suggested that the FCNs, especially the dilated convolution based FCNs, had potential to achieve accurate segmentation of anatomical structures, except for small-sized and complex-shaped tissues, even from sparse annotation.