Combining Ultrasound Imaging and Molecular Testing in a Multimodal Deep Learning Model for Risk Stratification of Indeterminate Thyroid Nodules
Authors' conclusion
Our multimodal model enhances MT performance by providing statistically significant improvements in PPV and specificity while maintaining high sensitivity.
Study context
- Population
- Patients with indeterminate thyroid nodules (Bethesda III and IV), UCLA Medical Center, 2016–2022; 259 benign, 74 malignant259 benign, 74 malignant
- Participants
- 333 patients
- Comparison
- Ensemble multimodal deep learning model vs clinical baseline
- Outcome
- Diagnostic performance for malignancy
Methodological detailssingle-centre, multimodal deep learning model, five fold cross validation stratified by surgical outcomes
- Single-centre
- Multimodal deep learning model
- Five fold cross validation stratified by surgical outcomes
Performance
As reported in the abstract
| Test or model | AUC | Sensitivity | Specificity | PPV |
|---|---|---|---|---|
| Ensemble model | 0.831 [0.77, 0.89] | 0.946 [0.88, 1.00] | 0.703 [0.66, 0.75] | 0.477 [0.46, 0.50] |
| Bethesda + molecular testing (MT) | 0.728 [0.68, 0.78] | 0.946 [0.88, 1.00] | 0.664 [0.60, 0.73] | 0.448 [0.41, 0.48] |
AUC
Ensemble model0.831 [0.77, 0.89]
Bethesda + molecular testing (MT)0.728 [0.68, 0.78]
Sensitivity
Ensemble model0.946 [0.88, 1.00]
Bethesda + molecular testing (MT)0.946 [0.88, 1.00]
Specificity
Ensemble model0.703 [0.66, 0.75]
Bethesda + molecular testing (MT)0.664 [0.60, 0.73]
PPV
Ensemble model0.477 [0.46, 0.50]
Bethesda + molecular testing (MT)0.448 [0.41, 0.48]
- Ensemble model vs Bethesda + MT: P = 0.0008 — improvements statistically significant
Original publication
Combining Ultrasound Imaging and Molecular Testing in a Multimodal Deep Learning Model for Risk Stratification of Indeterminate Thyroid Nodules. Thyroid. Published 21 April 2025. doi:10.
Composed by AI from the published abstract and checked against it. How Briefs are made →