Right Ventricular Segmentation in Multi-view Cardiac MRI Using a Unified U-net Model

Abstract

Accurate ventricle segmentation is an essential step towards cardiac function quantification from magnetic resonance imaging (MRI). Despite the vast efforts made over the last few years to automate this step, few works have specifically targeted the right ventricle (RV) assessment, and even fewer have attempted to integrate the information provided by both short- and long-axis cine MRI views. In this work, a novel automatic method for multi-view RV segmentation in cardiac MRI is proposed, which contemplates: (1) a preprocessing stage that re-orients the standard views according to their DICOM-provided 3D pose; (2) a novel multi-view augmentation module that augments on the fly the input images while conserving the 3D relationship between them; and (3) a unified cross-view model based on the U-net architecture (named xUnet) that simultaneously processes both views. The proposed method was evaluated on the M&Ms-2 challenge and achieved a combined average Dice score and Hausdorff distance of 0.918 and 9.25 mm, and 0.916 and 9.78 mm, on the validation and test sets, respectively. These results demonstrate the potential of the proposed unified model (and associated training scheme) towards accurate RV segmentation in cardiac MRI.