Artificial neural network model to distinguish pleomorphic adenoma from adenoid cystic carcinoma on fine needle aspiration cytology.

Abstract

To differentiate pleomorphic adenoma (PA) and adenoid cystic carcinoma (ACC) on fine needle aspiration cytology (FNAC) is often challenging to cytologists. In the present study, we tried to make an artificial neural network (ANN) model from the FNAC smears to differentiate PA from ACC. The detailed cytomorphological features were analysed on the FNAC of histopathology proven cases of PA (n = 35) and ACC (n = 33) and enumerated semi-quantitatively by two independent observers. These data were used to make an ANN model to distinguish PA from ACC on FNAC material. We used neuro-intelligence software to build the ANN model. The network architecture was 10-2-1. The heuristic search engine was applied to have this model. We used backpropagation neural network to teach ANN. At least 500 iterations were done to train the model. The efficacy of this ANN model was assessed with the help of the confusion matrix and receiver operating characteristic curve. The data were separated automatically by the software as a training set (n = 48), validation set (n = 10) and test set (n = 10). The ANN model was able to differentiate every case (10/10) of PA and ACC in the final test set. The area under the receiver operating characteristic curve was 1. The currently built ANN model is competent to identify PA and ACC cases on FNAC. Additional parameters and new cases can be incorporated into this open-ended ANN model to make it more robust.