MFS-Unet: A Multi-Path Vision Mamba Network for Precise Thyroid Nodule Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Automated segmentation of thyroid nodules is crucial for clinical diagnosis and treatment.
• Challenges in thyroid nodule segmentation include blurred boundaries, variable scales, noise, and inaccurate annotations.

Purpose of the Study:

• To propose a novel medical image segmentation network, MFS-Unet, for precise thyroid nodule segmentation.
• To enhance segmentation performance by addressing issues of varying nodule size, background noise, and label noise.

Main Methods:

• Developed MFS-Unet incorporating three novel modules: Multi-path Vision Mamba (MPV) for global context and multi-scale features, Feature Gating (FG) for enhancing boundary information, and Supervised Label Rectification (SLR) for handling label noise.
• MPV module utilizes state space models (SSMs) for efficient global context capture with linear complexity.
• FG module employs an attention mechanism to refine features in skip connections, suppressing noise and reinforcing nodule boundaries.
• SLR module dynamically adjusts loss weights to improve robustness against noisy training labels.

Main Results:

• MFS-Unet demonstrated superior performance across all evaluation metrics on three public thyroid ultrasound datasets (DDTI, TG3K, TN3K).
• The proposed network outperformed various state-of-the-art segmentation methods in precision and robustness.
• Experimental results validate the effectiveness of MPV, FG, and SLR modules in improving segmentation accuracy.

Conclusions:

• MFS-Unet offers a significant advancement in automated thyroid nodule segmentation from ultrasound images.
• The network shows substantial potential for precise segmentation in complex clinical ultrasound environments.
• The innovative modules effectively tackle key segmentation challenges, paving the way for improved diagnostic tools.