Room: Davidson Ballroom A
Purpose: To improve the accuracy of proton stopping-power-ratio (SPR) estimation by correcting the beam hardening effect in dual-energy CT (DECT) images with a shallow artificial neural network (ANN).
Methods: The proposed ANN approach uses a pixelâ€™s CT numbers and reference SPR as the training dataset. To reduce the inaccuracy resulting from the beam hardening effect, line integrals of CT numbers crossing the pixel of interest were aggregated over 360 degrees to be used as an additional input to the ANN. This data contains the information on the amount of beam hardening effect the rays have experienced, and can correct discrepancies occurring from different sizes of the phantom and different locations within the phantom. A leave-one-out cross validation (LOOCV) strategy was used to evaluate the accuracy of the proposed ANN using Gammex RMI467 phantom images taken using Siemens SOMATOM Force DECT scanner.
Results: A simple ANN comprising one unit and one hidden layer was able to accurately estimate the SPR. Even without the correction of the beam hardening effect, the ANN-based approach excelled in SPR estimation than the conventional DECT method. ANN with correction outperformed both the conventional DECT method and the ANN without correction, particularly for lung and bone tissues. Two units performed better than one unit, but further increasing the number of units or layers resulted in poor performance, due to overfitting and loss of generalizability.
Conclusion: This work demonstrates that ANN-based SPR estimation using DECT images is an accurate and practical method. Line integrals of CT numbers have contributed as an imaging uncertainty correction factor and significantly improved the accuracy of the ANN-based SPR estimation. Such a shallow ANN can be trained fast, and can adaptively tune its parameters based on the CT scanner setting.