DAC-Net: A light-weight U-shaped network based efficient convolution and attention for thyroid nodule segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Thyroid nodule segmentation algorithms are becoming more complex, increasing parameter counts and computational needs.
• Resource limitations in clinical settings hinder the implementation of sophisticated segmentation models.
• There is a need for efficient and high-performing algorithms for thyroid nodule segmentation.

Purpose of the Study:

• To develop a lightweight U-shaped network, DAC-Net, for efficient and accurate thyroid nodule segmentation.
• To reduce the number of parameters and computational complexity compared to existing state-of-the-art models.
• To achieve competitive segmentation performance suitable for clinical applications.

Main Methods:

• Introduced DAC-Net, a lightweight U-shaped network incorporating Depthwise Separable Convolution and Squeeze-Excitation (DWSE) for enhanced feature extraction.
• Utilized an Attention-based Dual-level Attention (ADA) module with Split Atrous convolution for comprehensive global and local feature capture.
• Implemented Channel and Spatial-Scale Connections (CSSC) for effective fusion of multi-stage features across different scales.
• Evaluated the model on the DDTI and TN3K datasets.

Main Results:

• DAC-Net demonstrated superior segmentation performance compared to state-of-the-art thyroid nodule segmentation architectures.
• The model achieved significant reductions in parameters (73x) and computational expenses (56x).
• DAC-Net outperformed TransUNet in segmentation accuracy.

Conclusions:

• DAC-Net offers a computationally efficient and high-performing solution for thyroid nodule segmentation.
• The proposed architecture effectively integrates multi-scale features while maintaining a lightweight design.
• DAC-Net presents a viable alternative for clinical implementation due to its reduced resource requirements.