Abstract
Neural network models have been used to analyze thyroid ultrasound (US) images and stratify malignancy risk of the thyroid nodules. We investigated the optimal neural network condition for thyroid US image analysis. We compared scratch and transfer learning models, performed stress tests in 10% increments, and compared the performance of three threshold values. All validation results indicated superiority of the transfer learning model over the scratch model. Stress test indicated that training the algorithm using 3902 images (70%) resulted in a performance which was similar to the full dataset (5575). Threshold 0.3 yielded high sensitivity (1% false negative) and low specificity (72% false positive), while 0.7 gave low sensitivity (22% false negative) and high specificity (23% false positive). Here we showed that transfer learning was more effective than scratch learning in terms of area under curve, sensitivity, specificity and negative/positive predictive value, that about 3900 images were minimally required to demonstrate an acceptable performance, and that algorithm performance can be customized according to the population characteristics by adjusting threshold value.