Converting dose-area product to effective dose in dental cone-beam computed tomography using organ-specific deep learning

Aim: To convert dose-area product (DAP) to effective dose for cone-beam computed tomography (CBCT) using deep learning. Material and Methods: 24,384 CBCT exposures of an adult phantom were simulated using PCXMC 2.0 (STUK, Finland), using permutations of tube voltage, filtration, source-isocenter distance, beam width/height and isocenter position. Equivalent doses for each organ used in the ICRP 103 calculation of effective dose as well as DAP values were recorded. Next, a series of dense neural networks (NN) were trained for each organ using Keras/TensorFlow. Two methods were explored: (1) ‘Coordinate’ mode, which uses the XYZ-coordinates of the isocenter; (2) ‘AP/JAW’ mode, which uses categorical codes for the anteroposterior (AP) and craniocaudal (JAW) position. For both methods, the six aforementioned scan parameters were used as inputs as well. Each network was trained, validated and tested using a 3/1/1 data split. Effective dose was calculated from the combination of NN outputs using ICRP 103 tissue weighting factors. The performance of the resulting NN model for estimating the effective dose per DAP was compared with that of a multiple linear regression (LR) model as well as direct conversion coefficients. Results: The mean absolute error (MAE) for equivalent dose/DAP on the test data ranged from 0.18% (bone surface) to 2.90% (oesophagus) in ‘Coordinate’ mode and from 2.74% (red bone-marrow) to 14.13% (brain) in ‘AP/JAW’ mode. The MAE for effective dose was 0.23% and 4.30%, respectively, for the two modes, both of which outperformed the LR model (MAE: 5.70%) and conversion coefficients (MAE: 20.19%-32.67%). Conclusion: NNs allow for a fast and accurate estimation of patient dose based on DAP in CBCT.