PURPOSE: The presence of metallic dental fillings is prevalent in head and neck PET/CT imaging and generates bright and dark streaking artifacts in reconstructed CT images. The resulting artifacts would propagate to the corresponding PET images following CT-based attenuation correction (CTAC). This would cause over- and/or underestimation of tracer uptake in corresponding regions thus leading to inaccurate quantification of tracer uptake. The purpose of this study is to improve our recently proposed metal artifact reduction (MAR) approach and to assess its performance in a clinical setting. METHODS: The proposed MAR algorithm is performed in the virtual sinogram space to overcome the challenges associated with manipulating raw CT data. The corresponding bins of the virtual sinogram affected by metallic objects are obtained by forward projection of segmented metallic objects in the original CT image. These bins are then substituted by weighted values of three estimates: the affected bins in the original sinogram, the bins in the corrected sinogram using spline interpolation, and the sinogram bins in the neighboring column of the sinogram matrix. The optimized weighting factors (alpha, beta, and gamma) were estimated using a genetic algorithm (GA). The optimized combination of weighting coefficients was obtained using the GA applied to 24 clinical CT data sets. The proposed MAR method was then applied to 12 clinical head and neck PET/CT data sets containing dental artifacts. Analysis of the results was performed using Bland and Altman plots and a method allowing analysis in the absence of gold standard called regression without truth (RWT). The proposed method was also compared to an image-based MAR method. RESULTS: Optimization of the weighting coefficients using the GA resulted in an optimum combination of parameters of alpha=0.26, beta=0.67, and gamma=0.07. According to Bland and Altman plots generated for both CT and PET images of the clinical data, the proposed MAR algorithm is efficient for reduction of streak artifacts in CT images and such reduce the over- and/or underestimation o tracer uptake. The RWT method also confirmed the effectiveness of the proposed MAR method. The obtained figures of merit revealed that attenuation corrected PET data corrected using CTAC after applying the MAR algorithm are more similar to the assumed gold standard. Comparison with the knowledge-based method revealed that the proposed method mainly corrects the artifactual regions without modifying the unaffected regions. The knowledge-based method globally modifies the images including those that do not include metallic artifacts. CONCLUSIONS: The proposed MAR algorithm improves the quality and quantitative accuracy of clinical head and neck PET/CT images and could be easily integrated in clinical setting.