RIANet: Recurrent interleaved attention network for cardiac MRI segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Accurate segmentation of cardiac magnetic resonance images (MRI) is crucial for analyzing heart function.
• Existing deep convolutional neural networks (ConvNets) struggle with precise, automated segmentation of multiple heart structures in cardiac MRI.

Purpose of the Study:

• To present a novel recurrent interleaved attention network (RIANet) for comprehensive cardiac MRI segmentation.
• To improve the accuracy and efficiency of segmenting multiple heart structures from cardiac MRI.

Main Methods:

• The RIANet utilizes a recurrent feedback structure (Clique Block) for parameter reuse and richer feature encoding.
• A plug-and-play interleaved attention (IA) block fuses multi-level contextual information.
• A deep supervision mechanism with weighted losses enhances discrimination capability.

Main Results:

• RIANet achieved high mean Dice scores: 0.942 (left ventricle), 0.923 (right ventricle), and 0.910 (myocardium) in the ACDC 2017 challenge.
• Intermediate feature visualization confirmed the effectiveness of proposed components.
• The model demonstrated competitive segmentation results with fewer parameters than state-of-the-art approaches.

Conclusions:

• The proposed RIANet is effective and robust for cardiac MRI segmentation.
• RIANet offers a promising solution for precise and automated segmentation of cardiac structures.