Fully Automated 3D Cardiac MRI Localisation and Segmentation Using Deep Neural Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Cardiac magnetic resonance (CMR) imaging is vital for cardiovascular disease diagnosis.
• Deep learning, particularly 3D fully convolutional networks (FCNs), has advanced CMR image segmentation.
• Prior methods often require extensive pre-processing and focus on lower-resolution images, necessitating improved localization strategies.

Purpose of the Study:

• To develop and evaluate novel deep learning strategies for localizing and segmenting cardiac structures in CMR images.
• To introduce a multi-stage and an end-to-end approach using 3D convolutional neural networks for improved cardiac segmentation.
• To investigate the impact of incorporating coarse localization features on segmentation accuracy.

Main Methods:

• Proposed two strategies: a multi-stage approach and an end-to-end approach, both utilizing a 3D convolutional neural network (3D DR-UNet).
• The multi-stage method involves a coarse localization map prediction followed by high-resolution segmentation.
• Evaluated on the Automatic Cardiac Segmentation Challenge (ACDC) and Left Atrium Segmentation Challenge (LASC) datasets.

Main Results:

• The multi-stage framework demonstrated consistent superior performance across various segmentation metrics compared to state-of-the-art methods.
• The approach proved robust in segmenting cardiac structures with pathological changes, low contrast, and noise.
• Accurate high-resolution segmentations were achieved, highlighting the effectiveness of the proposed localization strategy.

Conclusions:

• The proposed multi-stage deep learning framework offers a robust and accurate solution for cardiac MRI segmentation.
• Incorporating coarse localization features significantly improves segmentation performance.
• The method holds promise for clinical applications requiring precise cardiac structure analysis from CMR images.