Ultrasound-Based Transfer Learning Model to Assist Partially Cystic Thyroid Nodule Diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Partially cystic thyroid nodules (PCTN) pose diagnostic challenges.
• Accurate preoperative assessment of PCTN malignancy is crucial for effective treatment planning.
• Non-invasive methods for predicting PCTN malignancy are highly desirable.

Purpose of the Study:

• To establish a transfer learning model utilizing ultrasound to predict the preoperative malignant probability of PCTN.
• To offer clinicians a non-invasive primary screening tool for PCTN.

Main Methods:

• Retrospective analysis of 258 PCTNs from 258 patients (January 2020 - January 2024).
• Dataset split into training (80%) and test (20%) sets.
• Five pre-trained transfer learning models (EfficientNet, Inception_v3, Mobilenet_v3, SqueezeNet, VGG19) were evaluated.
• Performance metrics (AUC, accuracy, sensitivity, specificity) were calculated.
• Grad-Class Activation Map (Grad-CAM) used for visual interpretation.
• Comparison with multivariate logistic regression and radiologist performance.

Main Results:

• The Inception_v3 model demonstrated the highest Area Under the Curve (AUC) of 0.9243 in the training cohort for PCTN malignancy prediction.
• Inception_v3 achieved 85.19% accuracy, 85.26% sensitivity, and 85.00% specificity.
• The Inception_v3 model outperformed multivariate logistic regression (AUC 0.8247).
• Grad-CAM visualized important predictive features in ultrasound images.

Conclusions:

• Ultrasound-based transfer learning models, especially Inception_v3, can effectively predict PCTN malignancy non-invasively.
• This AI-driven approach can assist clinicians in preoperative decision-making for PCTN.
• The model offers a promising tool for improving the diagnostic accuracy of PCTN.