Ultrasound Radiomics and Dual-Mode Ultrasonic Elastography Based Machine Learning Model for the Classification of...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Thyroid nodules (TNs) require accurate differentiation to guide clinical management.
• Current diagnostic methods for TNs have limitations in specificity and sensitivity.

Purpose of the Study:

• To develop and evaluate a random forest (RF) model integrating ultrasound elastography and radiomics for improved TN differentiation.
• To compare the diagnostic performance of the combined model against individual elastography and radiomics models.

Main Methods:

• Retrospective analysis of 152 TNs from 127 patients using ultrasound elastography (sound touch elastography [STE] and strain elastography [SE]).
• Extraction and selection of radiomic features using the Pyradiomics package.
• Development of four RF models: STE, SE, radiomics, and a combined model integrating elastography and radiomics.
• Performance evaluation using Area Under the Curve (AUC) and Decision Curve Analysis (DCA).

Main Results:

• The combined model achieved the highest AUC of 0.911 (95% CI: 0.806-1.000), outperforming the STE (0.699), SE (0.812), and radiomics (0.851) models.
• The radiomics model also demonstrated strong performance with an AUC of 0.851.
• Both the combined and radiomics models showed outstanding diagnostic performance.

Conclusions:

• The combined model integrating elastography and radiomics offers superior predictive accuracy for thyroid nodule diagnosis.
• This integrated approach presents a promising advancement for the non-invasive diagnosis of TNs.
• Further validation is warranted to establish its clinical utility.