Deep learning based spectral distortion correction and decomposition for photon counting CT using calibration provided by an energy integrated detector

While the benefits of photon counting (PC) CT are significant, the performance of current PCDs is limited by physical effects distorting the spectral response. In this work, we examine a deep learning (DL) approach for spectral correction and material decomposition of PCCT data using multi-energy CT data sets acquired with an energy integrating detector (EID) for spectral calibration. We use a convolutional neural network (CNN) with a U-net structure and compare image domain and projection domain approaches to spectral correction and material decomposition. We trained our networks using noisy, spectrally distorted PCD projections as input data while the labels were derived from multi-energy EID data. For this study, we have scanned: 1) phantoms containing vials of iodine and calcium in water and 2) mice injected with an iodine-based liposomal contrast agent. Our results show that the image-based approach corrects best for spectral distortions and provides the lowest errors in material maps (RMSE in iodine: 0.34 mg/mL) compared with the projection-based approach (0.74 mg/mL) and image domain decomposition without correction (2.67 mg/mL). Both DL methods are, however, affected by a loss of spatial resolution (from 3 lp/mm in EID labels to ~2.2 lp/mm in corrected reconstructions). In summary, we demonstrate that multi-energy EID acquisitions can provide training data for DL-based spectral distortion correction. This data-driven correction does not require intimate knowledge of the spectral response or spectral distortions of the PCD. While the benefits of photon counting (PC) CT are significant, the performance of current PCDs is limited by physical effects distorting the spectral response. In this work, we examine a deep learning (DL) approach for spectral correction and material decomposition of PCCT data using multi-energy CT data sets acquired with an energy integrating detector (EID) for spectral calibration. We use a convolutional neural network (CNN) with a U-net structure and compare image domain and projection domain approaches to spectral correction and material decomposition. We trained our networks using noisy, spectrally distorted PCD projections as input data while the labels were derived from multi-energy EID data. For this study, we have scanned: 1) phantoms containing vials of iodine and calcium in water and 2) mice injected with an iodine-based liposomal contrast agent. Our results show that the image-based approach corrects best for spectral distortions and provides the lowest errors in material maps (RMSE in iodine: 0.34 mg/mL) compared with the projection-based approach (0.74 mg/mL) and image domain decomposition without correction (2.67 mg/mL). Both DL methods are, however, affected by a loss of spatial resolution (from 3 lp/mm in EID labels to ~2.2 lp/mm in corrected reconstructions). In summary, we demonstrate that multi-energy EID acquisitions can provide training data for DL-based spectral distortion correction. This data-driven correction does not require intimate knowledge of the spectral response or spectral distortions of the PCD.