Preclinical validation of a novel deep learning‐based metal artifact correction algorithm for orthopedic CT imaging

PURPOSE: To validate a novel deep learning‐based metal artifact correction (MAC) algorithm for CT, namely, AI‐MAC, in preclinical setting with comparison to conventional MAC and virtual monochromatic imaging (VMI) technique. MATERIALS AND METHODS: An experimental phantom was designed by consecutively inserting two sets of pedicle screws (size Φ 6.5 × 30‐mm and Φ 7.5 × 40‐mm) into a vertebral specimen to simulate the clinical scenario of metal implantation. The resulting MAC, VMI, and AI‐MAC images were compared with respect to the metal‐free reference image by subjective scoring, as well as by CT attenuation, image noise, signal‐to‐noise ratio (SNR), contrast‐to‐noise ratio (CNR), and correction accuracy via adaptive segmentation of the paraspinal muscle and vertebral body. RESULTS: The AI‐MAC and VMI images showed significantly higher subjective scores than the MAC image (all p < 0.05). The SNRs and CNRs on the AI‐MAC image were comparable to the reference (all p > 0.05), whereas those on the VMI were significantly lower (all p < 0.05). The paraspinal muscle segmented on the AI‐MAC image was 4.6% and 5.1% more complete to the VMI and MAC images for the Φ 6.5 × 30‐mm screws, and 5.0% and 5.1% for the Φ 7.5 × 40‐mm screws, respectively. The vertebral body segmented on the VMI was closest to the reference, with only 3.2% and 7.4% overestimation for Φ 6.5 × 30‐mm and Φ 7.5 × 40‐mm screws, respectively. CONCLUSIONS: Using metal‐free reference as the ground truth for comparison, the AI‐MAC outperforms VMI in characterizing soft tissue, while VMI is useful in skeletal depiction.