Automatic Segmentation of Right Ventricle in Cine Cardiac Magnetic Resonance Image Based on a Dense and Multi-Scale U-net Method

Abstract

Evaluating cardiac function from magnetic resonance images requires segmentation of the right ventricle (RV). The traditional segmentation methods, however, are often ineffective due to the complex structure of RV. Herein, a dense and multi-scale U-net (DMU-net) method is proposed to segment RV. Fifty six datasets were preprocessed, and the steps included normalization, enhancement and region of interest (ROI) extraction. Then a DMU-net was constructed by combining multi-scale aggregation with nested dense connection. The preprocessed datasets were then used to train and verify the DMU-net for predicting the results in 15 datasets with ROI extraction only. The mean of Dice index and Hausdorff distance obtained with the proposed method were 0.862 and 4.44 mm, respectively, which were better than those obtained with the conventional methods. Correlation coefficients of the endo-diastolic volume, endo-systolic volume, ejection fraction and stroke volume estimated were 0.992, 0.960, 0.987 and 0.982, respectively. The results demonstrated that the segmentation method based on the DMU-net has better relevance and consistency compared to manual segmentation, providing a promising tool the diagnosis of cardiac diseases.

Keywords

• cardiac magnetic resonance image
• right ventricle segmentation
• deep learning
• dense and multi-scale u-net