Purpose: To present a novel method to automate weighted logarithmic subtraction (WLS) of dual energy (DE) images using convolutional neural networks (CNN).
Methods: A novel algorithm using CNN was developed to automate WLS for DE imaging. Briefly, the algorithm decomposes the image into basis materials followed by the calculation of weighting factors using the basis materials thicknesses. A CNN architecture was trained to decompose high-low image pairs into basis materials of aluminum (Al) and polymethyl methacrylate (PMMA). To train the model, a custom phantom was built consisting of Al and PMMA step wedges. Predicted equivalent thicknesses along with projections were used to calculate the effective attenuations of Al at high and low energies. The optimal weighting factor is then determined as the ratio of attenuation coefficient for Al at high to low energy. Alternatively, the weighting factor was manually determined by iteratively minimizing the contrast between bone and surrounding soft tissue. The CIRS thorax phantom composed of tissue and lung-equivalent epoxy materials, with five simulated spherical tumors (0.5 â€“ 2.5 cm) embedded in lung-cavity, was utilized. The phantom was imaged using fast-kV DE imaging (120 and 60 kVp) on a commercial linac. The relative contrast of the tumors were compared between the two methods for a complete 360áµ’ rotation.
Results: For all simulated tumors, our analysis demonstrated consistency between the two methods. The mean relative contrast difference between the two methods were 3.54% Â± 0.85%, 0.52% Â± 0.22%, 0.05% Â± 0.15%, -0.12% Â± 0.15% and -0.13% Â± 0.14% for 5, 10, 15, 20 and 25mm targets, respectively. A two sampled t-test demonstrated no-significant difference between two methods (p>0.5).
Conclusion: We present a novel technique to automatically optimize weighting factors for the WLS method. Having the same soft-tissue contrast as manual optimization, this method allows for real-time processing of DE images.
Funding Support, Disclosures, and Conflict of Interest: Research reported in this publication was supported by the National Cancer Institute of the National Institutes of Health under Award Number R01CA207483. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.