AI-based spectral metal artifact reduction algorithm for improved iodine detection in virtual monoenergetic images: a quantitative assessment for interventional oncology applications

In this work, we developed a deep CNN based spectral metal artifact reduction algorithm for interventional oncology applications (sMARIO) and quantitatively assessed its performance pertaining to iodine contrast detectability. CNN models were trained specifically for each basis images for spectral CT using patient data with the virtual insertion of metallic probes. Two solid rods (2.5 cm in diameter) representing water and 2mgI/ml iodinated solution, respectively, were inserted in a anthropomorphic phantom and scanned on a spectral CT scanner in both presence and absence of metallic probes used in IO. Different probe configurations were used to generate varying artifacts, and both routine and low dose protocols were evaluated. Each set of reconstructed basis images were processed with sMARIO, and monoenergetic data ranging from 40 to 150 keV were generated for both the original and the sMARIO basis images using vendor-provided offline algorithms. Independent ROIs (27 mm3) were semi-automatically placed within the water and iodine rod for each dataset at matching locations, and the average value within each ROI was recorded to access the precision of the generated data to classify iodine vs water. For each distribution of water and iodine ROI values, the receiver operating curve (ROC) and the area under the ROC curve (AUC) were computed. Spectral MARIO was able to effectively correct a substantial contribution of metal artifacts from the image data across all keVs. Specifically, sMARIO recovered ~77% of the AUC loss due to metal artifact at low keV (40-70 keV) and 57% of the AUC loss at high keV (120- 150 keV). The recorded AUC of 0.98 at low keV with sMARIO demonstrates the ability to accurately classify small iodine concentrations even in the vicinity of metallic objects used in IO procedures.