An artificial neural network for the prediction of the risk of malignancy in category III Bethesda thyroid lesions.

Background: The diagnosis of cases of atypia of undetermined significance/follicular lesion of undetermined significance (AUS/FLUS) by fine needle aspiration cytology (FNAC) is challenging for both cytopathologists and clinicians. It is extremely difficult to predict the risk of malignancy based on cytological features alone. Aims and Objectives: In this study, we attempted to construct an artificial neural network (ANN) model to predict the risk of malignancy in FNAC cases of AUS/FLUS in thyroid lesions based on cytological features. Materials and Methods: We included two groups of AUS/FLUS cases: (1) 29 cases of histopathologically proven malignancy, and (2) 32 cases that had either been histopathologically proven to be benign, or for which no progress of malignancy on follow‐up had been observed in the last 2 years. Cytological characteristics were analysed semi‐quantitatively by two independent observers (TS and PD). Based on these data, we tried to generate an artificial neural network (ANN) model to differentiate between malignant and benign cases. The performance of the ANN was assessed using the confusion matrix and receiving operator curve. Results: There were 29 malignant cases of AUS/FLUS (histopathologically proven) and 32 benign/follow‐up cases in this study. There were 41 cases in the training set, 9 cases in the validation set and 11 cases in the test set. In the test group, the ANN model successfully distinguished between all benign (5/5) and malignant cases (6/6). The area under the receiver operating curve was 1. Conclusion: The present ANN model is well structured and coherent to distinguish malignant from benign outcomes in AUS/FLUS cases on cytology smears with no error. This is an open‐ended ANN model, and additional parameters and more cases could be included to make the model more robust. [ABSTRACT FROM AUTHOR]