OBJECTIVE: We aimed to propose a highly automatic and objective model named online transfer learning (OTL) for the differential diagnosis of benign and malignant thyroid nodules from ultrasound (US) images. METHODS: The OTL mothed combined the strategy of transfer learning and online learning. Two datasets (1750 thyroid nodules with 1078 benign and 672 malignant nodules, and 3852 thyroid nodules with 3213 benign and 639 malignant nodules) were collected to develop the model. The diagnostic accuracy was also compared with VGG-16 based transfer learning model and different input images based model. Analysis of receiver operating characteristic (ROC) curves were performed to calculate optimal area under it (AUC) for benign and malignant nodules. RESULTS: AUC, sensitivity and specificity of OTL were 0.98 (95% confidence interval [CI]: 0.97-0.99), 98.7% (95% confidence interval [CI]: 97.8%-99.6%) and 98.8% (95% confidence interval [CI]: 97.9%-99.7%) in the final online learning step, which was significantly better than other deep learning models (P < 0.01). CONCLUSION: OTL model shows the best overall performance comparing with other deep learning models. The model holds a good potential for improving the overall diagnostic efficacy in thyroid nodule US examinations. SIGNIFICANCE: The proposed OTL model could be seamlessly integrated into the conventional work-flow of thyroid nodule US examinations.
OBJECTIVE: We aimed to propose a highly automatic and objective model named online transfer learning (OTL) for the differential diagnosis of benign and malignant thyroid nodules from ultrasound (US) images. METHODS: The OTL mothed combined the strategy of transfer learning and online learning. Two datasets (1750 thyroid nodules with 1078 benign and 672 malignant nodules, and 3852 thyroid nodules with 3213 benign and 639 malignant nodules) were collected to develop the model. The diagnostic accuracy was also compared with VGG-16 based transfer learning model and different input images based model. Analysis of receiver operating characteristic (ROC) curves were performed to calculate optimal area under it (AUC) for benign and malignant nodules. RESULTS: AUC, sensitivity and specificity of OTL were 0.98 (95% confidence interval [CI]: 0.97-0.99), 98.7% (95% confidence interval [CI]: 97.8%-99.6%) and 98.8% (95% confidence interval [CI]: 97.9%-99.7%) in the final online learning step, which was significantly better than other deep learning models (P < 0.01). CONCLUSION: OTL model shows the best overall performance comparing with other deep learning models. The model holds a good potential for improving the overall diagnostic efficacy in thyroid nodule US examinations. SIGNIFICANCE: The proposed OTL model could be seamlessly integrated into the conventional work-flow of thyroid nodule US examinations.