PM-DUnet: Fusing long-range dependencies and attention in a dual-U architecture for thyroid nodule segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate medical image segmentation, especially for thyroid nodules, requires balancing local details and global context.
• Traditional Convolutional Neural Networks (CNNs) struggle with global dependency modeling due to limited receptive fields.
• Thyroid nodule segmentation presents challenges like variable sizes, ambiguous boundaries, and complex imaging contexts.

Purpose of the Study:

• To develop an advanced deep learning model for precise medical image segmentation of thyroid nodules.
• To enhance the modeling of global context and local features in medical images.
• To improve the accuracy and robustness of thyroid nodule segmentation.

Main Methods:

• Proposed the Parallel Mamba Dual-U Network (PM-DUNet), featuring a cascaded dual U-Net architecture for coarse-to-fine segmentation.
• Incorporated a Multi-Path Parallel Mamba (MPM) module utilizing State Space Models (SSMs) for efficient global context modeling with linear complexity.
• Integrated Squeeze-Excitation Downsampling (SED) and Spatial Attention Upsampling (SAU) modules to adaptively enhance features during encoding and decoding.

Main Results:

• The PM-DUNet demonstrated highly competitive performance in thyroid nodule segmentation tasks.
• The proposed model outperformed state-of-the-art methods on most core performance metrics.
• The effectiveness and robustness of PM-DUNet for complex medical image segmentation were validated.

Conclusions:

• The Parallel Mamba Dual-U Network (PM-DUNet) offers a superior approach to thyroid nodule segmentation.
• The integration of Mamba-based modules and attention mechanisms significantly improves segmentation accuracy.
• The developed method provides a robust solution for challenging medical image segmentation scenarios.