Slice-mask based 3D Cardiac Shape Reconstruction from CT volume

Xiaohan Yuan, Cong Liu, Fu Feng, Yinsu Zhu, Yangang Wang; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 1909-1925

An accurate 3D ventricular model is essential for diagnosing and analyzing cardiovascular disease. It is challenging to obtain accurate patient-specific models on scarce data via widely accepted deep-learning methods. To fully use the characteristics of medical volume-based images, we present a slice-mask representation to better regress the parameters of the 3D model. A data synthesis strategy is proposed to alleviate the lack of training data by sampling in the constructed statistical shape model space and obtaining the corresponding slice-masks. We train the end-to-end structure by combining the segmentation and parametric regression modules. Furthermore, we establish a larger left ventricular CT dataset than before, which fills the gap in relevant data of the healthy population. Our method is evaluated on both synthetic data and real cardiac scans. Experiments demonstrate that our method can achieve advanced results in shape reconstruction and segmentation tasks. Code is publicly available at https://github.com/yuan-xiaohan/Slice-mask-based-3D-Cardiac-Shape-Reconstruction.