Explainable Automated TI-RADS Evaluation of Thyroid Nodules

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Thyroid nodules are common, requiring accurate risk stratification for appropriate management.
• The Thyroid Imaging Reporting and Data System (TI-RADS) provides a standardized framework for classifying thyroid nodules based on ultrasound features.
• Limitations exist in TI-RADS, particularly with scarce training data, necessitating complementary diagnostic tools.

Purpose of the Study:

• To develop and validate an automated system for classifying thyroid nodules using TI-RADS criteria and assessing malignancy risk.
• To investigate the efficacy of machine learning models (ResNet-101, DenseNet-201) in conjunction with Grad-CAM for improved TI-RADS classification.
• To enhance the diagnostic accuracy of thyroid nodule assessment by integrating AI-driven insights with clinical guidelines.

Main Methods:

• Implementation of ResNet-101 and DenseNet-201 deep learning models for thyroid nodule classification.
• Utilizing the Grad-CAM algorithm to analyze model decision-making and identify critical nodule features relevant to TI-RADS scoring.
• Developing a precise heatmap by integrating Grad-CAM results with feature probability calculations for enhanced visualization.

Main Results:

• The automated system achieved improved TI-RADS classification accuracy, with an enhancement of up to 10% when using ResNet-101 and DenseNet-201 models.
• Grad-CAM analysis successfully identified risk areas and nodule features contributing to the TI-RADS score, demonstrating model interpretability.
• The integrated approach provided a visual heatmap, aiding in the identification of specific nodule characteristics relevant to malignancy assessment.

Conclusions:

• Machine learning models, particularly ResNet-101 and DenseNet-201 combined with Grad-CAM, significantly improve thyroid nodule classification accuracy within the TI-RADS framework.
• The developed system offers a promising tool for assisting clinicians in the accurate diagnosis and risk stratification of thyroid nodules.
• Further exploration and clinical integration of this AI-powered approach hold potential for advancing thyroid cancer detection and patient care.