ACPDNLS: Adaptive convexity preserving double nonzero level set for cardiac MR image segmentation

Abstract

Cardiovascular disease has become a major cause of global mortality. Clinically, quantitative assessment of cardiac MR image is usually used to determine the type and severity of cardiovascular disease, in which segmentation of cardiac MR image is a fundamental but important step. However, due to the inhomogeneity and special anatomical structures, accurate segmentation of cardiac MR images is still a challenging task. This paper proposes a double nonzero level set model for the segmentation of cardiac MR images, incorporating an adaptive convexity preserving mechanism and an improved distance regularization term. The double nonzero level set is capable of simultaneously and rapidly segmenting the left ventricle (LV) and right ventricle (RV). The adaptive convexity preserving mechanism guarantees that the segmentation of LV encompasses the cavity, papillary muscles and trabeculae while preserving convexity to meet clinical criteria. In addition, it ensures that RV retains its inherent physiological form, i.e. crescent-like shape. The improved distance regularization term effectively eliminates the need for reinitialization of double level set functions. The proposed model is evaluated on the data of ACDC MICCAI 2017. Experimental results show that in the end-diastolic (ED) and end-systolic (ES) phases, the mean Dice coefficients of LV segmentation are 0.961 (ED) and 0.936 (ES), with mean Hausdorff distances of 4.89 (ED) and 5.79 (ES), while the mean Dice coefficients of RV segmentation are 0.952 (ED) and 0.914 (ES), with mean Hausdorff distances of 8.52 (ED) and 9.60 (ES). The prominent advantage of our model is that, without the requirement of manual annotation and tedious training, it exhibits segmentation accuracy and robustness comparable to deep learning-based cardiac segmentation models. Especially, the segmentation accuracy of RV surpasses that of current state-of-the-art models.