Multi-view Learning for 3D LGE-MRI Left Atrial Cavity Segmentation

This paper presents a multi-view learning based method for left atrial cavity segmentation in 3D Late Gadolinium Enhanced Magnetic Resonance Imaging (LGE-MRI). Segmenting left atrium is challenging due to the low intensity contrast, motion artifacts, and extremely thin atrial walls. Since the spatial consistency of the atrium could help to alleviate the segmentation ambiguity caused by those problems, the proposed method consists of three deep convolutional streams which construct 3D segmentation likelihood maps from different views, i.e., axial view, coronal view, and sagittal view. Then, those likelihood maps will be fused and contribute to a final 3D segmentation map, where the method further inspects the 3D connectivity of the labeled pixels and discards the disconnected regions that don't belong to the atrium. The proposed method is tested on a publicly available dataset, where 80 scans are for training and 20 scans are for testing. Compared to the other state-of-the-art algorithms, the proposed method demonstrates a considerable improvement, which shows the advantages of using multi-view information.