Spectral Neural Attenuation Fields for Cone Beam CT

The modelling of Cone Beam Computerized Tomog-raphy reconstructions as Neural Attenuation Fields has shown progress when compared to more traditional explicit model-based reconstruction approaches. We propose an extension to prior work in this field to model the Attenuation field at different narrow-band spectrums for a polychromatic X-ray source using photon-counting detectors in an unsupervised manner. Our results show on real-world, raw CT data a significant increase in the accuracy of the five identified materials from 81.24% to 86.64% in a single scan. We also show a significant increase in sensitivity for three of the five materials and specificity for two of the five materials. Our approach is currently limited by an increase in reconstruction noise and a decrease in reconstruction sharpness.