Optimizing the neural network training algorithm in predicting kerma in mammography

In regard to the enhanced use of mammography screening tests for screening breast cancer, some concerns on the enhancement of the patient's absorbed dose have increased as well. Therefore, the assessment of the patient's dose before mammography is very important, and being aware of the dose level by its estimation can be helpful before radiation.

To this end, an artificial neural network (ANN) was used in this study to estimate the entrance surface air kerma (ESAK). A phantom with similar characteristics of the breast tissue was also used to collect the required data and the network was trained using some measurable parameters. To conduct the current research, multilayer perceptron (MLP) neural network architecture with training algorithms of LMBP, SCGBP, Rprop, BFGS, and GDBP, as well as radial basis function (RBF) neural network were used.

The results show that the neural network with BFGS training algorithm and 38 hidden layer neurons has the best performance with 7.40% root mean square error (RMSE) and coefficient of determination (R2) was obtained as 0.91.

According to the results of this study, there is a good correlation between the estimated network output and the measured values of the ESAK. The present method will remove the limitations and costs associated with the preparation process of dosimeter instruments.