Two-stage segmentation network with feature aggregation and multi-level attention mechanism for multi-modality heart...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Disease Research

Background:

• Accurate segmentation of cardiac substructures is crucial for diagnosing and treating cardiovascular diseases.
• Current cardiac image segmentation faces challenges including multi-target interference and imbalanced sample sizes.
• Existing methods struggle to effectively handle the complexity of multi-modality cardiac imaging.

Purpose of the Study:

• To propose a novel two-stage segmentation network, TSFM-Net, to address challenges in cardiac image segmentation.
• To enhance the effectiveness of multi-target feature representation and attention mechanisms.
• To achieve accurate and balanced segmentation of cardiac substructures in multi-modality images.

Main Methods:

• Developed a two-stage segmentation network (TSFM-Net) utilizing an encoder-decoder backbone.
• Incorporated a feature aggregation module (FAM) for multi-level feature representation in Stage 1.
• Designed a multi-level attention mechanism (MLAM) for improved target focus and feature fusion in Stage 2.

Main Results:

• TSFM-Net demonstrated superior segmentation performance on the 2017 MM-WHS multi-modality whole heart image dataset.
• The proposed method achieved better segmentation balance compared to state-of-the-art techniques.
• Feature aggregation and multi-level attention effectively addressed segmentation challenges.

Conclusions:

• TSFM-Net offers a feasible and effective solution for accurate cardiac image segmentation.
• The novel network architecture successfully overcomes limitations of existing segmentation methods.
• The approach shows significant potential for clinical applications in cardiovascular disease management.