Ultrasound imaging technology has the advantage of being convenient, less harmful and widely applied, making ultrasonography one of the most popular methods for disease diagnosis. With the rapid development of Computer- Aided Diagnosis (CAD) technology, the use of neural networks to analyze ultrasound images has become a popular method to improve the diagnostic efficiency of ultrasonography. Since the high cost of labeling medical images makes it difficult to train neural networks based on supervised learning, unsupervised CAD techniques without labeling have become a research trend. Most of the current unsupervised approaches focus on the reconstruction task and to some extent ignore the representational capability of models in the feature space. In this paper, we propose a Feature Discretized-based Deep Clustering (FDDC) for improving the deep clustering algorithm by introducing the theory of representation learning, which focuses on improving the representational capability of the model. There are two important strategies proposed in FDDC: 1) the global-local regular discretization method, which improves the expressiveness of the representation network by constraining the feature values; and 2) the greedy-based label reassignment method which is to reduce the loss fluctuations caused by re-clustering. Finally the experiments show that the new FDDC can achieve satisfactory results on six classification tasks, with tumor classification accuracy of 79.06% and machine classification accuracy of 96.17%, which outperforms existing unsupervised baseline methods. Furthermore, we also verify the representational capability of FDDC in feature space using visualization.
Keywords: Disease diagnosis; Feature extraction; Machine learning; Thyroid ultrasound image.
Copyright © 2022 Elsevier Ltd. All rights reserved.