An image augmentation method using convolutional network for thyroid nodule classification by transfer learning

Abstract

Thyroid nodules are severely jeopardizing our health. In the diagnosis of thyroid nodules, the ultrasound images serve as an essential tool to discriminate the malignant nodules from the benign ones. In this paper, the method of transfer learning is applied to classify the malignant and benign thyroid nodules based on their ultrasound images. The principal steps are preprocessing, data augmentation and classification by transfer learning. The preprocessing concentrates in extracting the region of interest (ROI). Two techniques of data augmentation are realized in our experiment, the traditional ways of augmenting images and a small convolutional network proposed by our own. After the augmentation of dataset, a pre-trained residual network is adopted to do transfer learning, and the parameters of this pre-trained net are fine-tuned with three different datasets that we have attained, including the original dataset, the augmented dataset via traditional methods and the augmented dataset via our convolutional network. The performances are then evaluated by three indexes, and the final results have proved the effectiveness of our convolutional augmentation network as well as the application of transfer learning. The best accuracy on the augmented dataset via convolutional network attains 93.75%, which exceeds the results of other two datasets and in the meanwhile outperforms other relevant methods.