Comparative study of the multi-atlas segmentation algorithm based on ANTs registration

In order to improve the accuracy and time efficiency of hippocampus multi-atlas segmentation, an algorithm based on Advanced Normalization Tools (ANTs) registration was proposed. In order to reduce the data size, a box with hippocampus as the center was extracted in the preprocessing stage. In the registration stage, ANTs were used to replace the resampling link, and the smoothness, topological retention and continuity of the differential Diffeomorphic Demons algorithm were used to perform accurate registration. In the tag fusion stage, four fusion algorithms including weighted average (Majority Voting, MV) algorithm, GraphCut tag fusion (Generative Model, GM) algorithm based on generated model constraints, metric learning (Metric Learning, ML) algorithm and semi-supervised tag propagation random forest (Integrating Semi-Supervised Label Propagation and Random Forests, RF-SSLP) algorithm were compared. The experimental results show that after replacing resampling with ANTs, the accuracy of four fusion algorithms including MV, GM, ML and RF-SSLP can be improved, respectively. Meanwhile, through the comparison of the above four fusion algorithms, it is found that the semi-supervised random forest algorithm based on ANTs registration has the highest segmentation accuracy, which is improved by 3%similar to 5% compared with MV, GM and ML fusion algorithms.

Automated summary

A segmentation algorithm based on ANTs registration was proposed to improve the accuracy and efficiency of hippocampus multi-atlas segmentation. Experimental results showed that replacing resampling with ANTs improved the accuracy of fusion algorithms, with the semi-supervised random forest algorithm based on ANTs registration achieving the highest segmentation accuracy among the compared fusion algorithms.