PSPU-Net for Automatic Short Axis Cine MRI Segmentation of Left and Right Ventricles

Abstract

Characterization of the heart anatomy and function is mostly done with magnetic resonance image cine series. To achieve a correct characterization, the volume of the right and left ventricle need to be segmented, which is a timeconsuming task. We propose a new convolutional neural network architecture that combines U-net with PSP modules (PSPU-net) for the segmentation of left and right ventricle cavities and left ventricle myocardium in the diastolic frame of short-axis cine MRI images and compare its results against a classic 3D U-net architecture. We used a dataset containing 399 cases in total. The results showed higher quality results in both segmentation and final volume estimation for a test set of 99 cases in the case of the PSPU-net, with global dice metrics of 0.910 and median absolute relative errors in volume estimations of 0.026 and 0.039 for the left ventricle cavity and myocardium and 0.051 for the right ventricles cavity.