Investigation of the effect of training set parameters on deep neural network prediction accuracy of fluoroscopic procedure-room scatter dose distributions

Purpose: During fluoroscopic image-guided interventional procedures, medical staff in the procedure room could receive high scattered-radiation dose due to the complexity and length of the procedure. Accurate real-time prediction and visualization of procedure-room scatter distributions could provide awareness of the scatter dose rate as it changes during the procedure, facilitating the avoidance of high-dose regions and minimizing health risk. Methods: We investigated the effect of varying training-set parameters on the deep neural network (DNN) prediction accuracy of fluoroscopic procedure-room scatter dose distributions. The ground-truth dataset used for DNN training was generated via EGSnrc Monte-Carlo (MC) simulation. Network accuracy as a function of dataset size was assessed on the original MC dataset and on the expanded datasets, obtained via linear interpolation of MC data between parameter values. Other evaluated training set parameter changes included variation of distribution spatial resolution, the dropout of data near the isocenter and logarithmic data compression. The DNN input layer was a 1x5 vector whose elements are the factors which alter the shape of the procedure-room scatter distributions including CRA-CAU angle, RAO-LAO angle, lateral patient shift (determined from table position), entrance field area and beam energy. Results: When compared with ground truth, network predictions using an expanded dataset obtained via linear interpolation provided superior accuracy with an average DNN MAPE which is below 10% for scattered distributions across all gantry angles in the test set. Reduced resolution of the scatter distribution obtained by downsampling substantially reduced the network size and prediction time. Conclusions: Our DNN was able to achieve high accuracy for predicting 2D scatter distributions when appropriate training set parameters are used. This DNN method can be incorporated into our scatter display system (SDS), which provides a real-time, color-coded visualization of procedure room scatter distributions and robust dose management for staff during fluoroscopic image-guided procedures.