Thyroid nodule classification in ultrasound imaging using deep transfer learning.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Oncology Diagnostics

Background:

• Accurate diagnosis of thyroid nodules is crucial but challenging in clinical practice.
• Developing precise diagnostic methodologies for thyroid nodules is essential.
• This study explores advanced computational techniques for improved thyroid nodule classification.

Purpose of the Study:

• To evaluate the predictive efficacy of traditional machine learning and deep transfer learning models for distinguishing benign from malignant thyroid nodules.
• To develop and validate a fused model integrating multiple AI approaches for enhanced diagnostic performance.
• To advance the diagnostic paradigm for thyroid nodules using artificial intelligence.

Main Methods:

• Retrospective analysis of 1134 thyroid nodule ultrasound images from 630 patients.
• Image preprocessing and feature extraction using ITK-Snap software.
• Feature selection via LASSO regression, followed by model development using Support Vector Machine (SVM) and Inception V3 (transfer learning), and subsequent model fusion.

Main Results:

• The Support Vector Machine (SVM) model achieved an AUC of 0.748, and the Inception V3 transfer learning model achieved an AUC of 0.763.
• The fused model demonstrated a superior AUC of 0.783, showing a statistically significant improvement over traditional methods (p=0.036).
• Decision curve analysis confirmed the superior clinical utility and practical applicability of the fusion model.

Conclusions:

• A fusion model integrating convolutional neural networks (CNNs) with traditional machine learning and deep transfer learning effectively differentiates benign and malignant thyroid nodules.
• This model fusion approach significantly enhances diagnostic performance for thyroid nodule classification.
• The developed intelligent tool offers a robust solution for clinical detection of thyroid diseases.