Radiomics have the ability to comprehensively quantify human tissue characteristics in medical imaging studies. However, standard radiomic features are highly unstable due to their sensitivity to scanner and reconstruction settings. We present an evaluation framework for the extraction of 3D deep radiomics features using a pre-trained neural network on real computed tomography (CT) scans for tissue characterization. We compare both the stability and discriminative power of the proposed 3D deep learning radiomic features versus standard hand-crafted radiomic features using 8 image acquisition protocols with a 3D-printed anthropomorphic phantom containing 4 classes of liver lesions and normal tissue. Even when the deep learning model was trained on an external dataset and for a different tissue characterization task, the resulting generic deep radiomics are at least twice more stable on 8 CT parameter variations than any category of hand-crafted features. Moreover, the 3D deep radiomics were also discriminative for the tissue characterization between 4 classes of liver tissue and lesions, with an average discriminative power of 93.5%.