Answers are grounded in this paper and cite their sources.
Thyroid nodule classification by ultrasound presents challenges in that nodule categories with varying risk degrees can have overlapping ultrasound characteristics. In this paper, we assess deep learning-based TI-RADS classification of thyroid nodules based on a case-based evaluation strategy that minimizes information leakage. The Digital Database of Thyroid Ultrasound Images (DDTI) was re-evaluated and resulted in 349 ultrasound images belonging to 298 cases. TI-RADS 2–3 nodule categories are categorized as low suspicion cases, whereas TI-RADS 4a–5 cases are categorized as high suspicion. Three deep learning architectures, DenseNet121, EfficientNetB0, and ViT-Tiny, were evaluated using five-fold case-level cross-validation. We calculated class weights only based on training data. The evaluation metrics used were accuracy, sensitivity, specificity, precision, F1 score, balanced accuracy, ROC-AUC, and PR-AUC with bootstrapped confidence intervals and statistical tests. DenseNet121 had the highest accuracy (0.718) and F1 score (0.813), but EfficientNetB0 had the highest ROC-AUC (0.724) and specificity (0.673). ViT-Tiny had an accuracy of 0.634 and an ROC-AUC of 0.648. DenseNet121 achieved significantly higher accuracy and F1 score than EfficientNetB0, whereas differences in ROC-AUC, PR-AUC, and balanced accuracy were not statistically significant. The TN5000 evaluation demonstrated poor transfer performance of 0.498 balanced accuracy and 0.533 ROC-AUC. The deep learning model showed moderate performance for TI-RADS-based risk stratification. Peritumoral information may help to add additional value, but cross-dataset transfer is still difficult.
No takes yet. Share an insight, caveat, or question.
Azzouz et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: