关键词: CT, 卷积神经网络（CNN）, 迁移学习, 检测

Abstract: The application of artificial intelligence to medical images can reduce the workload of doctors and the repeated examination of patients. Aiming at the problems of the existing thyroid nodule detection methods such as complex processing procedures and difficult feature extraction, a thyroid nodule detection method based on Convolutional Neural Network（CNN） is put forward. For the small data sample size restrictions, the strategy for improving network performance using pre-training and transfer learning is proposed. According to the trait that different structure CNN can extract different levels of features, a method to simultaneously fuse two networks to improve the accuracy is advanced. The proposed method is validated on 3414 images collected from hospital with the accuracy of 91.60%, sensitivity of 90.08%, specificity of 93.24% and AUC of 96.55%.

Key words: computed tomography, Convolutional Neural Network（CNN）, transfer learning, detection