Hybrid Regional Feature Cutting Network for Thyroid Ultrasound Images Classification

Abstract

With the increasing incidence of thyroid diseases, the detection and classification of thyroid nodule has attracted more and more attention. Deep learning has achieved promising results in computer-aided diagnosis due to the advantages of obtaining high-dimensional features. In this paper, we proposed hybrid cutting network (HCN) based on regional feature cutting method for feature extraction and classification of thyroid ultrasound images. Firstly, we build a backbone network for feature extraction. Next, we add a regional feature cutting (RFC) layer in the network to perform multiple cutting on the image feature matrix, so as to reduce the impact on classification caused by the similarity of local features between benign and malignant nodules. Finally, based on the local feature matrix of different size, we utilize N-first voting method to construct a hybrid classification voting network with three branches for final classification prediction. Experimental results show that the proposed method has achieved 97% accuracy, 95.74% precision, 100% recall, and 97.82% f1-measure in the open access public dataset, outperforming other current mainstream models.