Seq2Seq NETWORK FOR CERVICAL LYMPH NODE METASTASIS PREDICTION OF PAPILLARY THYROID CANCER

Due to the poor sensitivity of ultrasound and enhanced CT, the unreliability of clinical lymph node-negative (cN0) has become a consensus. It is still controversial whether to perform prophylactic central neck dissection (pCND) for patients with cN0 papillary thyroid cancer (PTC). Therefore, accurate noninvasive lymph node prediction is essential for the individualized treatment of patients with PTC. Several studies have investigated deep convolutional neural networks (DCNN) for lymph node metastasis (LNM) prediction and have achieved good performance. To further enhance the algorithms’ performance and reduce overfitting, a sequence-to-sequence (Seq2Seq) model, employing a gated recurrent unit (GRU) as Encoder and Decoder, is proposed, considering the sequential and multifaceted nature of clinical data. A total number of 4,464 sonograms from 1,488 cases are used in this study. The experimental results show that the proposed prediction model has higher accuracy (91.30%) and recall (90.28%) than six traditional prediction techniques, i.e., KNN, SVM, DT, RF, Xgboost, and Stacking.