A Novel Spatio-Temporal Self-Supervised Framework to Improve the Generalization Ability for Left Ventricle Volume Quantification Based on CMR Data
Gongning Luo, Kuanquan Wang, Naren Wulan, Shaodong Cao, Qince Li, Yongfeng Yuan, Henggui Zhang · Computing in Cardiology Conference · 2019
The automated quantification of the left ventricular (LV) volume on MRI is a crucial step for cardiac disease diagnosis. In recent years, deep learning (DL) technology has been widely used in the field of ventricle quantification and achieves relatively higher quantification accuracy. However, the LV volume quantification is still a challenging task, mainly because of limited labelled data. Hence, this study aims to propose an innovative approach to achieve accurate LV volume estimation based on limited labelled data.. The proposed method is three-fold: (1) For the first time, we proposed a self-supervised framework to model the significant spatio-temporal correlation information of adjacent slices from CMR images. (2) We designed a deep learning network based on spatio-temporal ranking loss to achieve self-supervised training utilizing large-scale unlabelled CMR data. (3) An iterative optimization strategy was developed to achieve efficient model optimization. The deep learning network was trained and validated on cardiac MRI datasets from MICCAI 2012 LV segmentation challenge including 100 patients (50 training patients and 50 test patients).