Objective To aid in the detection of thyroid cancer by using deep learning to differentiate the unique bioimpedance parameter patterns of different thyroid tissues.Methods An electrical impedance system was designed to measure 331 ex-vivo thyroid specimens from 321 patients during surgery.The impedance data was then analyzed with one dimensional convolution neural(1D-CNN)combining with long short-term memory(LSTM)network models of deep learning.In the process of analysis,we assigned 80％of the data to training set(1072/1340)and the remaining 20％data to the test set(268/1340).The performance of final model was assessed using receiver operating characteristic(ROC)curves.In addition,sensitivity,specificity,positive predictive value,negative predictive value,Youden index were applied to compare impedance model with ultrasound results.Re-sults The ROC curve of the two-classification(malignant/non-malignant tissue)model showed a good perfor-mance(area-under-the-curve AUC=0.94),with an overall accuracy of 91.4％.To better fit clinical practice,we fur-ther performed a three-classification(malignant/benign/normal tissue)model,of which the areas under ROC curve were 0.91,0.85,0.92 for normal,benign,and malignant group,respectively.The results indicated that the ar-ea under micro-average ROC curve and the macro-average ROC curve were 0.91 and 0.90,respectively.Moreover,compared with ultrasound,the impedance model exhibited higher specificity.Conclusions A deep learning mod-el(CNN-LSTM)trained by thyroid electrical impedance spectroscopy(EIS)parameters shows an excellent perfor-mance in distinguishing among different in-vitro thyroid tissues,which is promising for applications.In future clini-cal utility,our study does not replace existing tests,but rather complements others,thus contributing to therapeutic decision-making and management of thyroid disease.
NETL
NSTL
万方数据
