Novel Human Artificial Intelligence Hybrid Framework Pinpoints Thyroid Nodule Malignancy and Identifies Overlooked...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thyroid nodules require accurate malignancy risk stratification for effective diagnosis.
• Current diagnostic methods, including the Thyroid Imaging Reporting and Data System (TIRADS) and convolutional neural networks (CNNs), have limitations in identifying all malignant nodules.
• There is a need for advanced frameworks that integrate diverse data sources to improve diagnostic performance.

Purpose of the Study:

• To develop and evaluate a Human Artificial Intelligence Hybrid (HAIbrid) integrating framework for enhanced thyroid nodule malignancy stratification and diagnosis.
• To explore novel malignancy-relevant features from color Doppler ultrasound images.
• To identify and leverage second-order interacting features for improved diagnostic accuracy.

Main Methods:

• A Gated Attentional Factorization Machine (GAFM) was proposed to identify second-order interacting features.
• The GAFM was integrated into a HAIbrid framework, reweighting Thyroid Imaging Reporting and Data System (TIRADS) features and CNN-predicted malignancy scores.
• A 10-fold distribution-balanced stratified cross-validation was performed on 3002 retrospectively collected ultrasound-characterized nodules (1270 malignant).

Main Results:

• The GAFM-HAIbrid model achieved a significant improvement in Area Under the Curve (AUC) to approximately 0.92 (p < 10−5), outperforming standalone CNN (~0.87) and senior radiologists (~0.86).
• The model identified a second-order vascularity localization and morphological pattern overlooked by first-order feature analysis.
• Validation on an additional 500 nodules confirmed the framework's advantages over a commercial CNN system.

Conclusions:

• The proposed HAIbrid framework effectively integrates AI and clinical data for superior thyroid nodule malignancy risk stratification.
• The GAFM-HAIbrid model demonstrates potential in discovering novel diagnostic features beyond conventional ultrasonography.
• This framework offers seamless integration into clinical workflows, aiding in more accurate thyroid nodule diagnosis.