Average glandular dose prediction for breast model with patient-specific fibroglandular distribution in mammography and digital breast tomosynthesis: a machine-learning algorithms comparison.

Authors

Abstract

Objective To investigate Machine Learning (ML) methodologies for predicting glandular dose conversion coefficients for breast models with patient-specific fibroglandular distribution (Γpatient) in digital mammography (DM) and digital breast tomosynthesis (DBT). 
Approach We investigated four ML algorithms for predicting Γpatient, namely Generalized Additive Model (GAM), XGBoost, Support Vector Regression (SVR) and Automatic Relevance Determination Regression (ARDR). These were trained with Γpatient data generated with a Monte Carlo software and by adopting a dataset of 126 digital breast phantoms with patient-specific fibroglandular distribution. The ML input features were the compressed breast thickness, the glandular fraction by volume and the total breast volume. DM was simulated at 28 kV (Anode/Filter: W/Rh) and at 36 kV (Anode/Filter: W/Al); DBT at 28 kV (Anode/Filter: W/Rh) and 50 degrees scanning angle. 
Results The four investigated algorithms predicted the Γpatient coefficients with an average difference from the ground truth between -2% (SVR) and +7% (XGBoost). The best model from the GAM fine tuning required the sole compressed breast thickness as input feature. This algorithm presented the smallest model uncertainty, and the lowest cases of dose underestimate. 
Conclusions The GAM algorithm predicted Γpatient with an average difference from the expected value of 4%, in line with the other investigated algorithms. This algorithm showed the best performance in terms of model uncertainty, with average total estimated uncertainty of 12%, including the model accuracy, for DM at 28 kV. No relevant differences were observed in the case of DBT; bias and uncertainty of the prediction reduced for higher tube voltages.&#xD.

Tags

Topics