Quantitative Susceptibility Mapping (QSM) can estimate tissue susceptibility distributions and reveal pathology in conditions such as Parkinson's disease and multiple sclerosis. QSM reconstruction is an ill-posed inverse problem due to a mathematical singularity of the requisite dipole convolution kernel. State-of-art QSM reconstruction methods either suffer from image artifacts or long computation times. To overcome the limitations of these existing methods, a deep-learning-based approach is proposed and demonstrated in this work. 200 QSM datasets were utilized to compare current QSM reconstruction methods (TKD, closed-form L2, and MEDI) with the proposed deep-learning approach using visual scoring assessment of streaking artifacts and image sharpness. These multi-reader study results showed that the deep learning solution can produce QSM images with improved scores in both streaking artifacts and image sharpness evaluation while providing an almost instantaneous inversion computation through neural network inferencing.