A Novel Merged Strategy with Deformation Field Reconstruction for Constructing Statistical Shape Models

Abstract

Statistical shape models (SSMs) play important roles in the analysis of medical images. SSMs of particular organs are built by learning statistical shape patterns of the target organs from medical images. As a prerequisite step of SSM construction, the anatomical correspondences between the training subjects have to be obtained, requiring nontrivial and tedious organ segmentation from the training images, as well as difficult registration of the segmented organ shapes. To tackle these problems, we design a novel strategy called merged intensity and landmark registration (MILR) which does not require segmentation of the target organ. The MILR method directly registers an organ shape template to the training images by simultaneously optimizing the template-to-image intensity similarity and the key anatomical landmark alignment. Compared to conventional correspondence calculation methods, the MILR method not only saves the time of organ segmentation but also ensures accurate alignment of important anatomical feature points, improving both the efficiency and accuracy of SSM construction. To verify this method, we constructed the human vertebral column SSM from volumetric CT images. We show results of correspondences by three validation metrics anatomical landmark distance (ALD), Dice similarity coefficient (DSC) and average surface distance (ASD)and the corresponding construction of SSMs by generalization ability and specificity of the human vertebral column from the CT images of 17 different subjects. The novel merged strategy, with merged deformation field reconstruction, shows competitive results against the conventional methods.