Thyroid nodule segmentation in ultrasound images using transformer models with masked autoencoder pre-training

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Thyroid nodule segmentation in ultrasound (US) images is crucial for thyroid cancer diagnosis.
• Challenges include low contrast, speckle noise, and variable nodule appearance.
• Existing methods struggle with boundary detection due to these factors.

Purpose of the Study:

• To develop a transformer-based model for improved thyroid nodule segmentation.
• To leverage Masked Autoencoder (MAE) pre-training for enhanced image representation.
• To explore cross-attention mechanisms for better information flow.

Main Methods:

• Utilized a transformer architecture for capturing global image context.
• Pre-trained a Masked Autoencoder (MAE) on image patches.
• Fine-tuned the model on thyroid US datasets and explored cross-attention.
• Evaluated performance using Dice Similarity Coefficient (DSC) on public datasets (AIMI, TN3K, DDTI).

Main Results:

• MAE pre-training accelerated model convergence and reduced training time.
• Achieved modest DSC scores (0.63-0.65), indicating limitations with small-sample conditions.
• Cross-attention did not consistently improve segmentation performance.
• Limited data and nodule variability constrained overall accuracy.

Conclusions:

• Transformer models with MAE pre-training offer efficiency gains for thyroid nodule segmentation.
• Data volume and diversity are critical factors for improving segmentation accuracy.
• Future research should focus on increasing data, pre-training cross-attention, and hybrid architectures.