An unsupervised automatic texture classification method for ultrasound images of thyroid nodules

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ultrasound Technology

Background:

• Ultrasound is the primary imaging modality for thyroid nodule evaluation, relying on textural analysis for diagnosis.
• A significant challenge exists in creating pixel-level labeled datasets for thyroid nodule textures due to extensive manual effort required from experienced physicians.
• This scarcity hinders the development of automated diagnostic tools for thyroid nodules.

Purpose of the Study:

• To develop an unsupervised method for automated texture classification of thyroid nodules from ultrasound images.
• To reduce the reliance on manually labeled datasets, thereby saving significant manpower.
• To provide a tool that aids physicians in interpreting ultrasound images of thyroid nodules.

Main Methods:

• Development of a spatial mapping network to transform 1D pixel value space into a high-dimensional feature space.
• Implementation of feature selection principles tailored for clustering applications.
• Proposal of a pixel-level clustering algorithm incorporating a region growth pattern and a novel distance evaluation method for texture sets.

Main Results:

• Achieved high pixel-level classification accuracies: 0.931 (cystic/solid), 0.870 (hypoechoic), 0.959 (isoechoic), and 0.961 (hyperechoic).
• Demonstrated the algorithm's efficacy and concordance with human expert observations.
• Visualized the distribution of different textures within benign and malignant thyroid nodules.

Conclusions:

• The developed unsupervised method effectively classifies thyroid nodule textures at the pixel level.
• This approach facilitates the automatic generation of pixel-level labels, aiding in the creation of comprehensive texture datasets.
• The method offers valuable image analysis insights for medical professionals, supporting thyroid nodule diagnosis.