A Semantic-Wise Convolutional Neural Network Approach for 3-D Left Atrium Segmentation from Late Gadolinium Enhanced Magnetic Resonance Imaging

Several studies suggest that the assessment of viable left atrial (LA) tissue is a relevant information to support catheter ablation in atrial fibrillation (AF). Late gadolinium enhanced magnetic resonance imaging (LGE MRI) is a new emerging technique which is employed for the non-invasive quantification of LA fibrotic tissue. The analysis of LGE MRI relies on manual tracing of LA boundaries. This procedure is time-consuming and prone to high inter-observer variability given the dierent degrees of observers' experience, LA wall thickness and data resolution. Therefore, an automatic approach for the LA wall detection would be highly desirable. This work focuses on the design and development of a semantic-wise convolutional neural network based on the successful architecture U-Net (U-SWCNN). Batch normalization, early stopping and parameter initializers consistent with the activation functions chosen were used; a loss function based on the Dice coefficient was employed. The U-SWCNN was fed with the data available from the 2018 Atrial Segmentation Challenge with two different approaches: the model was trained end-to-end using stacks of 2-D axial slices as a preliminary attempt; then, with the appropriate changes in the baseline architecture, with 3-D data. The training was completed using 95 LGE MRI data, and a post-processing step based on the 3-D morphology was then applied. Mean Dice coefficient of unseen data (5) predicted masks were 0.89 and 0.91 for the 2-D and 3-D approach, respectively. These results suggest that, despite the increase of the number of trainable parameters, the 3-D U-SWCNN learns better features leading to a higher value of the Dice coefficient.