Combining Ultrasound Imaging and Molecular Testing in a Multimodal Deep Learning Model for Risk Stratification of...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Indeterminate thyroid nodules (Bethesda III and IV) constitute 15-30% of biopsied cases, necessitating further diagnostic evaluation.
• Current molecular testing (MT) for fine needle aspiration (FNA) samples offers high sensitivity but limited specificity and positive predictive value (PPV) for malignancy.
• Integrating diverse data sources is crucial for improving diagnostic accuracy in thyroid nodule risk stratification.

Purpose of the Study:

• To develop and evaluate a multimodal deep learning model combining ultrasound (US) imaging and molecular testing (MT) for enhanced risk stratification of indeterminate thyroid nodules.
• To improve the positive predictive value (PPV) and specificity of malignancy assessment while maintaining high sensitivity.
• To provide a more robust framework for managing patients with indeterminate thyroid nodules, potentially reducing unnecessary surgeries.

Main Methods:

• Retrospective analysis of 333 patients with indeterminate thyroid nodules (259 benign, 74 malignant) at UCLA Medical Center (2016-2022).
• Development of a multimodal deep learning model integrating whole frame US images, 256x256 patches, and 128x128 patches with clinical baseline data (Bethesda cytology and MT results).
• Model performance evaluated using five-fold cross-validation stratified by surgical outcomes, comparing the ensemble model against the clinical baseline.

Main Results:

• The clinical baseline (Bethesda + MT) achieved an AUROC of 0.728, sensitivity of 0.946, specificity of 0.664, and PPV of 0.448.
• The proposed multimodal ensemble model significantly improved performance, achieving an AUROC of 0.831, sensitivity of 0.946, specificity of 0.703, and PPV of 0.477 (p=0.0008).
• The multimodal model demonstrated statistically significant improvements in PPV and specificity while preserving high sensitivity.

Conclusions:

• The multimodal deep learning model integrating ultrasound imaging and molecular testing offers a significant advancement in the risk stratification of indeterminate thyroid nodules.
• This approach enhances diagnostic performance, particularly in improving PPV and specificity, which can aid in clinical decision-making.
• The framework holds potential for reducing the rate of benign thyroid resections, though further validation on larger, diverse datasets is warranted.