A Multi-View Deep Learning Model for Thyroid Nodules Detection and Characterization in Ultrasound Imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thyroid Ultrasound (US) is the standard for thyroid nodule assessment.
• Deep Learning (DL) shows promise in enhancing thyroid cancer diagnosis.

Purpose of the Study:

• To develop and evaluate a DL pipeline for detecting and classifying thyroid nodules.
• To compare DL performance against expert radiologist evaluation using ACR-TIRADS.

Main Methods:

• A DL pipeline using two You Look Only Once (YOLO) v5 models for nodule detection and classification from transverse and longitudinal US images.
• Ensemble methods with non-max suppression (NMS) and an extreme gradient boosting (XGBoost) model for final malignancy prediction.
• Retrospective analysis of 983 patient cases with an 81-case independent test set.

Main Results:

• The DL pipeline achieved an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.84.
• Sensitivity and specificity were 84% and 63%, respectively, outperforming ACR-TIRADS (76% sensitivity, 34% specificity).
• Ensemble models achieved mAP0.5 scores of 0.797 (transverse) and 0.716 (longitudinal).

Conclusions:

• The proposed DL pipeline demonstrates significant potential for accurate thyroid nodule evaluation.
• This AI-driven approach may improve diagnostic performance in identifying malignant thyroid nodules.
• DL offers a promising tool to augment radiologist capabilities in thyroid cancer screening.