Classification of a Small-data-set Thyroid Nodules Based on Supplementary Feature Layer Improved VGG16
Yifei Chen, Xin Zhang, Dandan Li, Jing Jin, Yi Shen · 2020
Thyroid nodule, a common disease of the endocrine system, has attracted international attention. How to apply deep learning in the field of diagnosis of thyroid nodule becomes more crucial in recent years. However, the size of samples obtained in medical field usually cannot meet the requirement size of samples needed in training deep learning model. In order to improve the classification accuracy when meeting small data set, the proposed method based on improved VGG16 utilize supplementary features to add more effective information in training. The supplementary features consist of medical features and features selected by ReliefF algorithm. The features before selection are statistical features and texture features, which are extracted from thyroid nodule ultrasound images. Then, a model based on the improved VGG16 is trained by utilizing thyroid nodule ultrasound images with the assistant of medical features and selected features. The result obtained by this method can improve the classification accuracy by 2.24%, which is from 76.68% to 78.92%. Furthermore, some commonly used computer vision methods, including Local Binary Patterns, Histogram of Oriented Gradients, Haar-like features are compared with the proposed method to demonstrate the improvement of the proposed method more clearly. As can be seen from the result, this method has positive effect on solving the negative effect caused by small data.