A Hybrid Deep Learning and Handcrafted Features based Approach for Thyroid Nodule Classification in Ultrasound Images
Jiahao Xie, Le‐Hang Guo, Chongke Zhao, Xiaolong Li, Ye Luo, Jianwei Lu
Abstract
Abstract With the increasing incidence rate of thyroid cancer, the diagnosis of thyroid nodules has become an important task. In this paper, we designed a deep neural network (DNN) to classify whether a thyroid nodule is benign or malignant, and proposed a structure which combines local binary pattern (LBP) with deep learning. Our method mitigates the effects of overfitting in medical image diagnosis tasks. With well-designed transfer leaning, we achieve an accuracy of 85% on our own ultrasound thyroid dataset. To ensure the reliability of our experiments, all examples are estimated by experts in Shanghai Tenth People’s Hospital using fine needle analysis (FNA), which is a gold standard for thyroid nodules diagnosis. The experimental results show that combinations of the traditional medial image features can help the deep learning network get more semantic information from low-level inputs.