A COMPARATIVE STUDY ON THE DIFFERENCE OF INCIDENT SURFACE DOSE PREDICTION MODEL BASED ON ARTIFICIAL NEURAL NETWORK

Abstract

Background/Objectives By predicting the quantitative amount of effective dose through machine-learning of the models of the artificial neural network program, the predicted data values between each model are compared and evaluated, and the ESD values derived from actual imaging and the data showing the most similar predicted values are compared. Therefore, practical accuracy and clinical utility were verified and evaluated.

Statistical analysis: Position the phantom upright on the detector face in the chest PA position, and take three shots with each combination of tube voltages 90, 100, 110, 120 kVp tube currents 5,8, 10, 12.5 mAs, SID 180, and 200 cm. The center line was taken with a vertical incidence on the mid-plane at the height of the 6th spine, and the dosimeter position was measured at 4 locations, including the 6th thoracic vertebrae, breast, thyroid gland, and upper abdomen (liver point). After that, after data mining the data values obtained through shooting, it is written in Excel, an artificial intelligence prediction program, and applied to Orange 3.0, an artificial neural network program, to obtain prediction data using data mining.

Findings: A total of 5 algorithms kNN, ,Tree, SVM, Random Forest, and Linear Regression were used. As for the accuracy of the evaluation model, the smaller the mean square error (MSE), the root mean square error (RMSE), and the mean absolute error (MAE) are, the better the model is. Therefore, the model with the best predictive power was analyzed in the order of Tree, Random Forest, Linear Regression, kNN, and SVM.

Improvements/Applications: The SVM was analyzed as not suitable for use as a model for using the ESD prediction rate, but since this is a data value that does not reflect the sensitivity of the variable, if the amount of data set is increased, it is expected that the utilization value will increase sufficiently in the future.