Estimation of glandular dose in mammography based on artificial neural networks

This work proposes using artificial neural networks (ANNs) for the regression of the dosimetric quantities employed in mammography. The data were generated by Monte Carlo (MC) simulations using a modified and validated version of the PENELOPE (v. 2014) + penEasy (v. 2015) code. A breast model of a homogeneous mixture of adipose and glandular tissue was adopted. The ANNs were constructed using the Keras and scikit-learn libraries for mean glandular dose (MGD) and air kerma (K_air ) regressions, respectively. In total, seven parameters were considered, including the incident photon energies (from 8.25 to 48.75 keV), breast geometry, breast glandularity and K_air acquisition geometry. Two ensembles of five ANNs each were formed to calculate MGD and K_air . The normalized glandular dose coefficients (DgN) were calculated using the ratio of the ensemble outputs for MGD and K_air . Polyenergetic DgN values were calculated by weighting monoenergetic values by the spectrum bin probabilities. The results indicate a very good ANN prediction performance when compared to the validation data, with median errors on the order of the average simulation uncertainties (≈ 0.2%). Moreover, the predicted DgN values are in good agreement compared with previously published works, with mean (maximum) differences up to 2.2% (9.4%). Therefore, it is shown that ANNs could be a complementary or alternative technique to tables, parametric equations and polynomial fits to estimate DgN values obtained via MC simulations.