|
Advancements in deep learning and GPU computing have exponentially driven the application of neural networks to classic medical imaging problems: denoising, segmentation, artifact removal, etc. Deep learning solutions are particularly attractive for processing multi-channel, volumetric image data, where processing and reconstruction methods are often computationally expensive. Convolutional neural networks (CNNs) are commonly applied to multi-channel image data by matching the number of network input channels to the number of data channels, learning explicit relationships between channels. This provides a high degree of specificity to a particular problem, but may fail to generalize to a broader class of closely related problems. We propose a solution to this generalization problem in the context of spectral x-ray CT, where the scanning kVps (energy bins) and contrast are often variable. Specifically, we propose a novel CNN architecture which handles variable numbers of input channels, variable noise levels between channels, and variable modes of spectral contrast. We demonstrate our architecture in the application of preclinical, photon-counting, micro-CT, effectively replacing 1-2 hours of iterative reconstruction, with <10 minutes of analytical reconstruction and CNN regularization. Experimental validation shows the effectiveness of our approach when applied to both in vivo photon-counting validation data (4 energy thresholds) and to simulated, dual-energy CT data virtually acquired with an energy integrating detector. In both cases, the results output by the CNN provide greater spectral accuracy than analytical reconstruction alone, but suffer from some degradation of spatial resolution. We conclude by proposing several extensions of our work to better preserve spatial resolution.
|
