Automated segmentation of the left ventricle from MR cine imaging based on deep learning architecture

Abstract

Background: Magnetic resonance cine imaging is the accepted standard for cardiac functional assessment. Left ventricular (LV) segmentation plays a key role in volumetric functional quantification of the heart. Conventional manual analysis is time-consuming and observer-dependent. Automated segmentation approaches are needed to improve the clinical workflow of cardiac functional quantification. Recently, deep-learning networks have shown promise for efficient LV segmentation. PURPOSE: The routinely used V-Net is a convolutional network that segments images by passing features from encoder to decoder. In this study, this method was advanced as DenseV-Net by replacing the convolutional block with a densely connected algorithm and dense calculations to alleviate the vanishing-gradient problem, prevent exploding gradients, and to strengthen feature propagation. Thirty patients were scanned with a 3 Tesla MR imager. ECG-free, free-breathing, real-time cines were acquired with a balanced steady-state free precession technique. Linear regression and the dice similarity coefficient (DSC) were performed to evaluate LV segmentation performance of the classic neural networks FCN, UNet, V-Net, and the proposed DenseV-net methods, using manual analysis as the reference. Slice-based LV function was compared among the four methods. Results: Thirty slices from eleven patients were randomly selected (each slice contained 73 images), and the LVs were segmented using manual analysis, UNet, FCN, V-Net, and the proposed DenseV-Net methods. A strong correlation of the left ventricular areas was observed between the proposed DenseV-Net network and manual segmentation (R = 0.92), with a mean DSC of 0.90 ± 0.12. A weaker correlation was found between the routine V-Net, UNet, FCN, and manual segmentation methods (R = 0.77, 0.74, 0.76, respectively) with a lower mean DSC (0.85 ± 0.13, 0.84 ± 0.16, 0.79 ± 0.17, respectively). Additionally, the proposed DenseV-Net method was strongly correlated with the manual analysis in slice-based LV function quantification compared with the state-of-art neural network methods V-Net, UNet, and FCN. Conclusion: The proposed DenseV-Net method outperforms the classic convolutional networks V-Net, UNet, and FCN in automated LV segmentation, providing a novel way for efficient heart functional quantification and the diagnosis of cardiac diseases using cine MRI.