Automated detection and classification of thyroid nodules in ultrasound images using clinical-knowledge-guided convolutional neural networks.

Accurate diagnosis of thyroid nodules using ultrasonography is a valuable but tough task even for experienced radiologists, considering both benign and malignant nodules have heterogeneous appearances. Computer-aided diagnosis (CAD) methods could potentially provide objective suggestions to assist radiologists. However, the performance of existing learning-based approaches is still limited, for direct application of general learning models often ignores critical domain knowledge related to the specific nodule diagnosis. In this study, we propose a novel deep-learning-based CAD system, guided by task-specific prior knowledge, for automated nodule detection and classification in ultrasound images. Our proposed CAD system consists of two stages. First, a multi-scale region-based detection network is designed to learn pyramidal features for detecting nodules at different feature scales. The region proposals are constrained by the prior knowledge about size and shape distributions of real nodules. Then, a multi-branch classification network is proposed to integrate multi-view diagnosis-oriented features, in which each network branch captures and enhances one specific group of characteristics that were generally used by radiologists. We evaluated and compared our method with the state-of-the-art CAD methods and experienced radiologists on two datasets, i.e. Dataset I and Dataset II. The detection and diagnostic accuracy on Dataset I were 97.5% and 97.1%, respectively. Besides, our CAD system also achieved better performance than experienced radiologists on Dataset II, with improvements of accuracy for 8%. The experimental results demonstrate that our proposed method is effective in the discrimination of thyroid nodules.