Predictions for central lymph node metastasis of papillary thyroid carcinoma via CNN-based fusion modeling of ultrasound images

Abstract

The diagnosis of central lymph node metastasis (CLNM) is very important for the treatment of papillary thyroid carcinoma (PTC), which remains highly subjective and depends on clinical experience. Traditional method based on radiomics tumor feature (RTF) extraction and classifications has its shortages to predict the CLNM and increase the possibility of over-diagnosis and over-treatment leading for PTC. In this paper, a convolutional neural network (CNN) based fusion modeling method is proposed for predictions of CLNM in ultrasound-negative patients with PTC. A CNN and a RTF extraction based random forest (RF) classifier are trained on the context image patches and tumor image patches, and the probability outputs from these two models are combined for predicting the CLNM. It is validated that the proposed method has better diagnostic performance than the conventional method on the test set. The area under the curve (AUC), accuracy, sensitivity, and specificity of the method in predicting CLNM are 0.9228, 83.09%, 86.17%, and 81.46%, respectively. It has the prospect to apply to diagnose ultrasound (US) images with the machine-learning diagnostic system.