MLMSeg: A multi-view learning model for ultrasound thyroid nodule segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate thyroid nodule segmentation is crucial for distinguishing benign from malignant nodules.
• Existing deep learning methods struggle with multi-scale nodule segmentation in complex ultrasound environments due to limited view learning.

Purpose of the Study:

• To develop a multi-view learning model (MLMSeg) for enhanced thyroid nodule segmentation in ultrasound images.
• To improve the accuracy and robustness of deep learning models for thyroid nodule analysis.

Main Methods:

• Introduced a multi-view learning model (MLMSeg) incorporating local, global, and structural views.
• Utilized a deep convolutional neural network for local feature encoding.
• Employed a multi-channel transformer for global view correlations and a cross-layer graph convolutional module for inter-layer feature relationships.
• Integrated a channel-aware graph attention block for effective view fusion.

Main Results:

• MLMSeg achieved superior performance compared to 14 baseline methods on two thyroid datasets.
• Achieved high Dice coefficients (92.10%, 83.84%) and Intersection over Union scores (86.60%, 73.52%).
• Demonstrated exceptional segmentation capability for thyroid nodules of varying scales.

Conclusions:

• MLMSeg offers a significant advancement in automated thyroid nodule segmentation from ultrasound images.
• The model's accuracy aids in nodule localization and precise diameter measurements, supporting clinical diagnosis.
• This approach holds substantial clinical relevance for early and accurate thyroid nodule diagnosis.