Purpose: To produce virtual monoenergetic (VM) cone beam computed tomography (CBCT) images using convolutional neural networks for basis material decomposition of dual-energy (DE) images.
Methods: A custom phantom consisting of two step wedges of aluminum (Al) and polymethyl methacrylate (PMMA) was built to mimic bone and soft tissue, respectively. Low- and high-energy (80/130 kV) planar projections of this wedge and full CBCT images of a pediatric head phantom (CIRS), were acquired using the on-board imager (OBI) of a commercial linac. A convolutional neural network (CNN), using the Keras API (v2.2.4) and TensorFlow (v1.11) engine on the backend, was trained with the wedge phantom images to differentiate relative low and high density materials. After initial training, the model was applied to the pediatric head phantom to decompose the projections into basis material thicknesses and then produce VM-CBCT images of the phantom. The errors measured for CNN were compared to conventional mathematical calibration obtained through error minimization.
Results: Our model decomposed the calibration phantom into Al and PMMA and presented mean squared error of 0.01 mm for Al and 0.15 mm for PMMA. The mathematical approach presented mean squared error of 0.11 mm for Al and 0.18 mm for PMMA. The pediatric head phantom projections were successfully decomposed into relative low and high density materials allowing VM-CBCT production.
Conclusion: We present a CNN method to decompose DE images into equivalent basis material thicknesses, a required step to produce VM-CBCT images, thus avoiding laborious manual calibration for material decomposition.
Funding Support, Disclosures, and Conflict of Interest: This work was funded by a research grant from Varian Medical System.