A 2.5D Coarse-to-Fine Framework for 3D Cardiac CT View Planning

Xiaohan Yuan, Yinsu Zhu

Abstract: Usually, the directly acquired CT images are from the axial views with respect to the major axes of the body, which do not effectively represent the structure of the heart. If CT imaging is first reformatted into the typical cardiac imaging planes, it will lay the foundation for the subsequent analysis. In this paper, we propose an automatic CT view planning method to acquire standard views of the heart from 3D CT volume, obtaining the equation of the plane by detecting landmarks that can determine this view. To face the challenge of memory cost brought by 3D CT input, we convert the 3D problem into a 2.5D problem, taking into account the spatial context information at the same time. We design a coarse-to-fine framework for the automatic detection of anatomical landmarks. The coarse network is used to estimate the probability distribution of the landmark location in each set of orthogonal planes, and the fine network is further used to regress the offset distance of the current result from the ground-truth. We construct the first known dataset of reformatted cardiac CT with landmark annotations, and the proposed method is evaluated on our dataset, validating its accuracy in the tasks of landmark detection and view planning.