HEDN: multi-oriented hierarchical extraction and dual-frequency decoupling network for 3D medical image segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Existing 3D encoder-decoder networks face challenges in fine-grained feature decomposition and hierarchical fusion.
• Blurred contour boundaries in medical images hinder the extraction of high-frequency contour features.

Purpose of the Study:

• To introduce a novel network, the Multi-oriented Hierarchical Extraction and Dual-frequency Decoupling Network (HEDN), for improved 3D medical image segmentation.
• To enhance feature hierarchy and contour feature representation in medical image segmentation.

Main Methods:

• The proposed HEDN integrates an Encoder-Decoder Module (E-DM), a Multi-oriented Hierarchical Extraction Module (Multi-HEM), and a Dual-frequency Decoupling Module (Dual-DM).
• Multi-HEM enriches feature hierarchy through 3D fusion of spatial and slice-level features.
• Dual-DM employs self-supervision to separate and integrate high-frequency contour features with hierarchical features.

Main Results:

• HEDN demonstrated superior performance on the Synapse dataset, increasing Dice Similarity Score (DSC) by 1.38% and reducing 95% Hausdorff Distance (HD95) by 1.03 mm.
• On the Automatic Cardiac Diagnosis Challenge (ACDC) dataset, HEDN achieved a 0.5% performance improvement across all categories.
• The integration of self-supervised high-frequency features enhanced contour feature representation and hierarchical feature interactions.

Conclusions:

• HEDN effectively addresses limitations in previous 3D segmentation architectures by improving feature decomposition and fusion.
• The network's ability to enhance both contour and hierarchical features leads to more accurate medical image segmentation.
• HEDN represents a significant advancement in deep learning for medical image analysis, offering improved segmentation accuracy and robustness.