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

Area of Science:

• Cardiovascular Imaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Accurate left ventricle (LV) segmentation from cine MRI is crucial for assessing cardiac function in cardiovascular diseases.
• Current deep learning methods often process cine MRI frames independently, neglecting temporal motion dynamics.
• Radiologists utilize cardiac motion as a key indicator for LV assessment in cine MRI.

Purpose of the Study:

• To develop a novel U-net-based deep learning model (OF-net) that integrates temporal information for improved LV segmentation from cine MRI.
• To enhance LV segmentation accuracy and temporal continuity by incorporating optical flow and specialized modules.
• To emulate expert analysis by leveraging dynamic cardiac motion in automated segmentation.

Main Methods:

• Proposed a U-net-based convolutional neural network (CNN) named OF-net.
• Integrated an optical flow (OF) module to capture cardiac motion along the temporal axis.
• Introduced LV localization and attention modules to improve detection and segmentation accuracy.

Main Results:

• OF-net achieved superior LV segmentation performance on multicenter cine MRI data compared to a classical U-net model.
• Achieved an average perpendicular distance (APD) of 0.90±0.08 pixels and a Dice index of 0.95±0.03 in middle slices.
• Demonstrated enhanced temporal continuity in segmentation, particularly for challenging apical and basal slices.

Conclusions:

• OF-net effectively integrates temporal information into LV segmentation, outperforming frame-by-frame analysis.
• The proposed method offers a more robust and accurate approach to cardiac function assessment using cine MRI.
• This work highlights the potential of deep learning to learn from expert observational strategies in medical imaging analysis.