A fully automated left atrium segmentation approach from late gadolinium enhanced magnetic resonance imaging based on a convolutional neural network

Background: Several studies suggest that the evaluation of left atrial (LA) fibrosis is a relevant information for the assessment of the appropriate strategy in catheter ablation in atrial fibrillation (AF). Late gadolinium enhanced (LGE) cardiac magnetic resonance imaging (MRI) is a non-invasive technique, which might be employed for the non-invasive quantification of LA myocardial fibrotic tissue in patients with AF. Nowadays, the analysis of LGE MRI relies on manual tracing of LA boundaries and this procedure is time-consuming and prone to high inter-observer variability given the different degrees of observers' experience, LA wall thickness and data resolution. Therefore, an automated segmentation approach of the atrial cavity for the quantification of scar tissue would be highly desirable. Methods: This study focuses on the design of a fully automated LGE MRI segmentation pipeline which includes a convolutional neural network (CNN) based on the successful architecture U-Net. The CNN was trained, validated and tested end-to-end with the data available from the Statistical Atlases and Computational Modelling of the Heart 2018 Atrial Segmentation Challenge (100 cardiac data). Two different approaches were tested: using both stacks of 2-D axial slices and using 3-D data (with the appropriate changes in the baseline architecture). In the latter approach, thanks to the 3-D convolution operator, all the information underlying 3-D data can be exploited. Once the training was completed using 80 cardiac data, a post-processing step was applied on 20 predicted segmentations belonging to the test set. Results: By applying the 2-D and 3-D approaches, average Dice coefficient and mean Hausdorff distances were 0.896, 0.914, and 8.98 mm, 8.34 mm, respectively. Volumes of the anatomical LA meshes from the automated analysis were highly correlated with the volumes from ground truth [2-D: r=0.978, y=0.94x+0.07, bias=3.5 ml (5.6%), SD =5.3 mL (8.5%); 3-D: r=0.982, y=0.92x+2.9, bias =2.1 mL (3.5%), SD =5.2 mL (8.4%)]. Conclusions: These results suggest the proposed approach is feasible and provides accurate results. Despite the increase of the number of trainable parameters, the proposed 3-D CNN learns better features leading to higher performance, feasible for a real clinical application.