Knowledge-Powered Thyroid Nodule Classification with Thyroid Ultrasound Reports

Abstract

The thyroid is one of the largest endocrine glands in the human body, which is involved in several body mechanisms. Using thyroid ultrasound (US) reports which have latent hierarchical syntactic structures and have hidden rich semantic information in precise thyroid nodule classification is hence essential. In this paper, we proposed a Knowledge-powered Thyroid Nodule Classification model (KTNC), which takes the semantic information as a kind of prior knowledge and incorporated it into deep neural networks. Specially, our proposed model first uses Hierarchical Long Short Term Memory networks (HLSTM) to encode the syntax-aware representation of US reports in a hierarchical way. In the HLSTM, a hierarchical neural network structure, consisting in two encoder layers of LSTM (Long Short Term Memory networks), mirrors the hierarchical structure of US report, and a hierarchical attention mechanism, consisting of two levels attentions, attends to important elements within US report with word-level attention and sentence-level attention. Secondly, our model obtains the semantic information relevant to the US report from the semantic tree, and encodes the semantic representation of US reports by using Tree Structured Recurrent Neural Network with gated recursive units (Tree-GRUs). Finally, we classify US reports by combining both the syntax and semantic representations. We evaluate our method on the real-world thyroid US reports, and results show that our method achieves higher performance.