Cine MRI analysis by deep learning of optical flow: Adding the temporal dimension

Abstract

Accurate segmentation of the left ventricle (LV) from cine magnetic resonance imaging (MRI) is an important step in the reliable assessment of cardiac function in cardiovascular disease patients. Several deep learning convolutional neural network (CNN) models have achieved state-of-the-art performances for LV segmentation from cine MRI. However, most published deep learning methods use individual cine frames as input and process each frame separately. This approach entirely ignores an important visual clue—the dynamic cardiac motion along the temporal axis, which radiologists observe closely when viewing cine MRI. To imitate the approach of experts, we propose a novel U-net-based method (OF-net) that integrates temporal information from cine MRI into LV segmentation. Our proposed network adds the temporal dimension by incorporating an optical flow (OF) field to capture the cardiac motion. In addition, we introduce two additional modules, a LV localization module and an attention module, that provide improved LV detection and segmentation accuracy, respectively. We evaluated OF-net on the public Cardiac Atlas database with multicenter cine MRI data. The results showed that OF-net achieves an average perpendicular distance (APD) of 0.90±0.08 pixels and a Dice index of 0.95±0.03 for LV segmentation in the middle slices, outperforming the classical U-net model (APD 0.92±0.04 pixels, Dice 0.94±0.16, p < 0.05). Specifically, the proposed method enhances the temporal continuity of segmentation at the apical and basal slices, which are typically more difficult to segment than middle slices. Our work exemplifies the ability of CNN to “learn” from expert experience when applied to specific analysis tasks.