A video vision transformer model based on dynamic contrast-enhanced ultrasound outperforms conventional convolutional neural networks and radiologists in identifying malignant thyroid nodules and predicting BRAF V600E mutations in papillary thyroid carcinoma, Nature reported on Tuesday.
The diagnostic model, known as ViViT-CEUS, achieves an area under the curve of 0.938 for diagnosing malignancy. Researchers retrospectively collected thyroid nodules with contrast-enhanced ultrasound videos, dividing them into training, validation, and testing sets in a ratio of 4:3:3 to build and validate the system.
Performance of the ViViT-CEUS model was measured against two-dimensional convolutional neural networks, three-dimensional convolutional neural networks, and traditional radiomics models. Human comparators in the study included radiologists with varying levels of clinical experience.
The dataset comprised 428 thyroid nodules, consisting of 69 benign cases and 359 malignant cases. All malignant nodules identified in the cohort were papillary thyroid carcinoma, which included 267 cases with the BRAF V600E mutation and 92 non-mutated cases.
For diagnosing malignancy, the ViViT-CEUS area under the curve of 0.938 surpassed the two-dimensional convolutional neural network score of 0.767 and the three-dimensional convolutional neural network score of 0.790. Human radiologists achieved area under the curve results ranging from 0.617 to 0.815.
| Model or Evaluator | Malignancy Diagnosis (AUC) | BRAF V600E Prediction (AUC) |
|---|---|---|
| ViViT-CEUS | 0.938 | 0.865 |
| 3D-CNN | 0.790 | 0.597 |
| 2D-CNN | 0.767 | 0.704 |
| Traditional Radiomics | N/A | 0.682 |
| Radiologists | 0.617–0.815 | N/A |
When predicting the specific BRAF V600E mutation, the ViViT-CEUS model recorded an area under the curve of 0.865. That figure outperformed the two-dimensional convolutional neural network at 0.704, the three-dimensional convolutional neural network at 0.597, and traditional radiomics models at 0.682.
The results establish a quantitative basis for identifying molecular features directly from dynamic contrast-enhanced ultrasound videos.
