Abstract
Metal induced artefacts in computed tomography (CT) images are primarily caused by beam hardening, scatter effects, and photon starvation. These artefacts impede the characterization of fine anatomical structures and compromise the diagnostic value of the CT images. We aim to develop an innovative machine learning-based technique called residual dense U-Net (RDU-Net), specifically for spectral photon-counting CT (SPCCT), to mitigate metal artefacts across all energy bins. The proposed model was quantitatively evaluated, with and without the metal artefact reduction (MAR) algorithm, using line profiles, histogram analysis, signal-to-noise ratio (SNR), root mean squared error (RMSE), and structural similarity index measure (SSIM). The results show significant improvements with the average SNR increasing from 3.37 to 17.40 across the five energy bins after the application of the MAR algorithm. The average RMSE decreased from 0.016 to 0.001, and the average SSIM increased by 34.9%. The study also evaluated material density images of hydroxyapatite (HA) and iodine, with and without the MAR algorithm, using the receiver operating characteristic (ROC) paradigm. The results show improved accuracy in the material identification for HA (86% to 91%) and iodine (84% to 93%) after MAR. Overall, the evaluation of the model show promising results and the potential to significantly decrease the metal artefacts in all the parameters used in the energy analysis at p < 0.0001, while preserving the attenuation profile of SPCCT images.